source: .forge/design/environments-and-identity.md · generated — edit the markdown, not this file

Environments & Identity — concrete configuration reference

Purpose: the concrete, environment-specific values the CRM needs to run — DNS names per environment, the Microsoft Entra ID app-registration facts, and who owns the cloud/identity setup. This is the operational companion to the design in architecture.md (deployment topology) and the auth decision in AD #28 (.claude/CLAUDE.md). Where this doc and architecture.md disagree on intent, architecture.md wins; this doc is the record of the actual values. As of: 2026-10-09 (§4’s “F38-a — the development account, checked”: 2026-10-15, and its AC-5 subsection: 2026-10-21; §4’s “F38-b — the dev estate, checked”: 2026-10-21, and its AC-1, AC-2, AC-3 and AC-7 entries of F38-b-WI-4.1: 2026-10-26, AC-1’s no-change run 06:10 UTC; §4’s “F38-d — the dev deploy pipelines, checked”: 2026-10-19, and its record of F38-d wave 2’s runs: 2026-10-23; §4’s description of forge-et-iaac and the dev estate: 2026-10-21, after forge-et-iaac#3 was merged and applied on 20 Oct). Sources: F38-b-WI-4.1’s read-only reads of the dev account as svc-harness, of GitHub, of forge-et-iaac origin/dev at c728e06b and forge-et-be origin/main at 30fd601c, of the public hosts with curl, dig, nc and openssl, and of https://dev.crm.empire.lk/ in a browser, 26 Oct 2026 05:34–05:40 UTC (§4, “F38-b — the dev estate, checked”, AC-1, AC-2, AC-3, AC-7; the base/secrets bullet’s deploy sentence); F38-d-WI-4.2’s read-only reads of the build runs in forge-et-be and forge-et-fe since F38-d’s wave 2 (forge-et-be origin/main at 06062de8, forge-et-fe at 8e59d076) and of the dev account as svc-harness, 22 Oct 2026 19:16–19:25 UTC (§4, “F38-d — the dev deploy pipelines, checked”, AC-1 – AC-4 and AC-12); F38-a-WI-3.1’s read-only reads of SES in the dev account as svc-harness, of the public DNS and of forge-et-iaac origin/dev at 0eee623 with its dev.yaml run 39553084687, 21 Oct 2026 08:30–08:33 UTC (§4, “F38-a — the development account, checked”, AC-5); forge-et-iaac origin/dev at d230b133, its dev.yaml run 39475850266, and read-only reads of the dev account as svc-harness on 21 Oct 2026 (§4); F38-b-WI-2.1’s read-only reads of the dev account as svc-harness, of GitHub, of forge-et-iaac origin/dev at 0eee623 and forge-et-be origin/main at 9097a68b, and of the public CRM host, 21 Oct 2026 21:11–21:17 UTC (§4, “F38-b — the dev estate, checked”); F38-d-WI-1.1’s read-only review of the dev deploy files — forge-et-be origin/main at 5c457091, forge-et-fe at 32c0dea5, forge-et-pwa at 96d48ba and forge-et-iaac dev.yaml at forge-et-iaac#3’s head 0ff63c5 — with its reads of the dev account as svc-harness and of GitHub, 18 Oct 2026 UTC (§4, “F38-d — the dev deploy pipelines, checked”); F38-a-WI-1.1’s read-only pass over the dev account as svc-harness and over forge-et-iaac at 9eac094, 15 Oct 2026 (§4, “F38-a — the development account, checked”); Shyamin Ayesh’s 30 Sep 2026 message with the crm.empire.lk name servers, forge-et-iaac as read at commit b7e923a on 9 Oct 2026 and forge-et-be main at 03404e8e the same day (§3, §4), the DevOps channel messages of 9 Oct 2026 from Shyamin Ayesh and Dinil Gamage (the dev deploy, the svc-harness access, the Bedrock subscriptions — §3, §4), F5-i Wave 1 (the laptop mode built — §2.5), the Empire Teas Azure portal app registration (screen-shared 10 Sep 2026; Kavindu Anthony’s screenshots of 24 Sep 2026 showing both certificates and all eight reply URLs — §2.2, §2.3, §5), the delivery team’s environment/DNS plan, Tharinda’s notes 13 Sep 2026, the team’s 23 Sep 2026 notes on the mailbox reply URLs, laptop sign-in and certificates (§2.2, §2.3, §2.5, §5), Shyamin Ayesh’s 24 Sep 2026 proposal of two certificate pairs, prod and non-prod (§2.3), and the lead’s access-model decisions of 24 Sep 2026 (AD #34 — the seeded IT Admin and its per-environment email; the info@ decision — §2.4, §5).

⚠️ Secrets boundary — read before adding anything here

This file holds non-secret configuration only. The values below (DNS names, the Entra Client ID and Tenant ID, reply URLs, the Graph permission list) are identifiers, not credentials — they are safe to commit.

Never committed to any repo:

Secret Where it lives instead
M365 user account passwords (the CRM01–05@empire.lk set) A password manager / handed to the account holders directly. Not in the harness, not in the app repos.
The Entra app certificate private keys — one for prod and one non-prod key shared by dev, UAT and the laptops (the login credential — see §2.3) AWS Secrets Manager, referenced by the backend task definition (AD #22) — stays with Altrium, never shared and never in the repo. Until each environment’s Secrets Manager exists, a restricted vault in the team’s password manager (§2.3). Only the public certs are shared with Empire IT to upload.
The non-prod private key on a developer laptop (the laptop sign-in, §2.5) Read from the team’s password manager. On a laptop, only as a shell-session environment variable or in the gitignored application-local.yml — never in a committed .env, a chat or an email. It is the same key dev and UAT use, so a laptop copy that leaks means replacing the non-prod pair everywhere (§2.3).
Any Entra client secret N/A — not used. Empire’s Azure tenant has a tenant-wide policy that blocks adding client secrets, so login uses the certificate approach instead (see §2.3).
Database / RDS credentials, AWS keys Secrets Manager + the task IAM role (AD #22) — no static keys anywhere.
The production administrator’s email — the value of the F5-k Flyway placeholder that seeds the IT Admin user in production (AD #34) AWS Secrets Manager in the production account only, passed to the backend by a task-definition secrets reference (AD #22). Never in git, never in a committed example file. Not an access credential, but it is a real person’s address, and AD #34 keeps real people’s emails out of the repositories. (2026-10-28, AD #40: a plain value in production’s task definition, forge-et-be .aws/td.prod.json, so it is in git once confirmed; it is not secret, and a secret would cost about USD 0.40 a month for a value that need not be one. It stays out of this file and every example file.) Owner: Tharinda, with Kavindu Anthony confirming the address (§5). Outside production the value is the crm01@empire.lk test mailbox, which is not a person and is safe to write down (§5, the test-accounts row).

If a credential ends up needed at build/run time, it goes to Secrets Manager and is referenced, never inlined.


1. DNS / domains per environment

Three apps per environment (CRM web app, backend API, trade-show capture app — the capture app has its own URL per AD #20). All under empire.lk.

Environment CRM web app Backend API Capture app Notes
Production crm.empire.lk api.crm.empire.lk capture.crm.empire.lk Permanent.
Development dev.crm.empire.lk api.dev.crm.empire.lk capture.dev.crm.empire.lk Permanent, build-only — never doubles as UAT (AD #23).
UAT uat.crm.empire.lk api.uat.crm.empire.lk capture.uat.crm.empire.lk Temporary — one instance, applied on 13 Nov 2026, rehearsed on and used for Phase 1’s acceptance round (no rehearsal destroy, F38-h spec Rev 13), and kept to the end of Phase 2’s, then destroyed (AD #23, revised 18 Nov 2026), by hand when the lead decides it is no longer needed, never on a schedule (.claude/rules/tracker.md § Release & UAT Lifecycle).

Frontends are static (S3 + CloudFront) and the backend API is the only Fargate service behind the ALB (AD #22), so the api.* host is where TLS terminates for API traffic and where the OAuth2 redirect lands (see §2). The apps call the backend cross-site (AD #33, 2026-09-24): each frontend’s VITE_API_BASE_URL is its environment’s api.* address, the backend’s CORS list names the CRM and capture-app origins, and the XSRF-TOKEN cookie is shared with the app hosts through a Domain attribute, under a cookie name that differs per environment (register §D).

1.1 Delegation of crm.empire.lk (30 Sep 2026)

Shyamin Ayesh has created a Route 53 public hosted zone for crm.empire.lk in Empire Teas’ AWS account. Its name servers:

ns-344.awsdns-43.com
ns-1313.awsdns-36.org
ns-879.awsdns-45.net
ns-1983.awsdns-55.co.uk
  • What Empire IT does: add four NS records named crm in the empire.lk zone, one per name server. The name servers of empire.lk itself do not change — that would move Empire’s website and email DNS as well. Done 1 Oct 2026: Dinil Gamage sent the request to Kavindu Anthony by email, with Shyamin copied, and Kavindu added the records in Cloudflare that afternoon. Check with dig NS crm.empire.lk +short, which returns the four names; on 1 and 2 Oct 2026 it did so through Cloudflare (1.1.1.1) and Google (8.8.8.8), Route 53 answered the zone’s SOA, and empire.lk’s website, MX (Microsoft 365) and name servers were unchanged.
  • Where Empire’s DNS is hosted (observed 1 Oct 2026 with dig NS and whois): empire.lk is registered at NIC.lk, but its DNS is served by Cloudflare (fatima.ns.cloudflare.com, kip.ns.cloudflare.com), so every empire.lk record is changed in Cloudflare, never at NIC.lk. Kavindu Anthony has access to that Cloudflare account. empireteas.com is registered at Hostinger and also served by the same Cloudflare name servers. Kavindu first tried to create the subdomain at the LK registry and asked whether empireteas.com could be used instead; Shyamin Ayesh pointed out that the record belongs in Cloudflare, and that NIC.lk moving empire.lk back to its own name servers would drop the Cloudflare records Empire’s website and email depend on. The CRM stays on crm.empire.lk: moving to empireteas.com would also have meant re-registering every Entra reply URL (§2.2).
  • The zone is not in Terraform yet. It was created outside forge-et-iaac (no Route 53 resource there at daf4f1c); AD #23 needs it imported before ACM validation records and host records depend on it. (15 Oct: the import moves to F38-i, before production’s records are written — F38 Rev 13; dev writes its records through dns-dev meanwhile.)
  • Decided 30 Sep 2026 (lead): the whole subdomain is delegated. This replaces the earlier design, which delegated only uat.crm.empire.lk and sent Empire IT the dev and production records in two requests per environment. Every CRM record — dev, UAT and production: ACM validation, a CAA record allowing amazon.com, the host records (alias records at the crm.empire.lk apex) and the SES records on a sending domain under crm.empire.lk — is written by Terraform into this zone. The one exception is a sending domain on empire.lk itself, which would need one further request to Empire IT (F38 OQ-E). Recorded as F38 Rev 2, in F38-a FR-10, and in register “Start here” step 11.

2. Microsoft Entra ID — app registration (CRM login)

Auth is Microsoft (Entra ID) OAuth2, sole login path (AD #28), with server-side sessions (Spring Session JDBC, not JWT — AD #5). The backend performs the OAuth2 authorization-code exchange, which is why the reply URL is on the api.* host and the platform is Web, not SPA.

Who can sign in (AD #34). A Microsoft sign-in succeeds only when the verified email matches exactly one active user that an administrator created in the CRM beforehand — there is no auto-provisioning from the tenant and no invitation email. The first user in each environment is the IT Admin seeded by the F5-k migration (§5). Offboarding is done in the CRM: deactivating a user ends their sessions at once and refuses further sign-ins. There is no Entra user sync in V1; Empire disabling the person’s Entra account is a backstop, because Microsoft then refuses the next sign-in before it reaches the CRM — but it does not end a CRM session already open, which runs until it expires; deactivating in the CRM ends it at once.

2.1 Identifiers (non-secret)

Field Value
Display name CRM
Application (client) ID 64f71638-82ea-4172-aab4-4cd84c91a7ad
Object ID 47333335-949a-4c38-b4dc-42d36606b98d
Directory (tenant) ID 26c6eef0-a2ae-4a9e-8bee-27ce39cdc889
Supported account types My organization only (single-tenant)
Tenant Empire Teas (empire.lk)
State Activated

2.2 Reply URLs (redirect URIs) — register as platform = Web (NOT SPA)

The backend handles the code exchange, so the redirect targets the API host’s Spring Security callback path:

Environment Reply URL Purpose Registered?
Production https://api.crm.empire.lk/login/oauth2/code/microsoft Sign-in Registered
UAT https://api.uat.crm.empire.lk/login/oauth2/code/microsoft Sign-in Registered
Development https://api.dev.crm.empire.lk/login/oauth2/code/microsoft Sign-in Registered
Production https://api.crm.empire.lk/authorized/microsoft-mailbox Mailbox consent (F20-a) Registered (24 Sep 2026)
UAT https://api.uat.crm.empire.lk/authorized/microsoft-mailbox Mailbox consent (F20-a) Registered (24 Sep 2026)
Development https://api.dev.crm.empire.lk/authorized/microsoft-mailbox Mailbox consent (F20-a) Registered (24 Sep 2026)
Laptop (local) http://localhost:8080/login/oauth2/code/microsoft Sign-in from a developer laptop (§2.5) — remove before go-live Registered (24 Sep 2026)
Laptop (local) http://localhost:8080/authorized/microsoft-mailbox Mailbox consent from a developer laptop (§2.5) — remove before go-live Registered (24 Sep 2026)

The backend keeps its own api.* address (AD #33, 2026-09-24), so every reply URL is on api.* — the three sign-in URLs registered on 10 Sep keep working, and the capture app reuses them.

The Spring registration id for sign-in is microsoft (hence .../oauth2/code/microsoft). Today the local/test profiles use the profile-scoped stub login (AD #28) and need no Entra reply URL; the two laptop rows exist only for the opt-in real-Entra laptop mode described in §2.5.

Why the laptop rows use port 8080 and plain http. The frontend starts sign-in and mailbox consent by a full-page navigation to VITE_API_BASE_URL + /oauth2/authorization/… (forge-et-fe/src/features/auth/startMicrosoftSignIn.ts, src/features/email/startMailboxConnect.ts), and Spring builds the reply URL from the backend’s own address ({baseUrl}/…) — so a laptop’s reply URL is on the backend’s localhost:8080, not on the Vite port 5173. Caution — the local setup does not reach 8080 today: the working local setup leaves VITE_API_BASE_URL empty and relies on the Vite proxy, which forwards only /api (forge-et-fe/vite.config.ts; .env.example shows http://localhost:8080 but its own comment says empty is the norm). So sign-in and mailbox consent from a laptop need one of two changes, which F5-i’s spec chooses: proxy /oauth2, /login/oauth2 and /authorized to the backend as well (keeping the proxy’s host rewrite, so the reply URL still comes out on localhost:8080), or set VITE_API_BASE_URL=http://localhost:8080 and add http://localhost:5173 to the local CORS list in laptop mode. Entra accepts http reply URLs only for localhost; every other host must be https. Register the exact URLs above, port included. They are platform Web, like the rest of the table.

Calendar and contacts need no new reply URLs. F20-a’s spec (Out of Scope, and OQ-A) had F21-b and F22 add the Calendars.ReadWrite and Contacts.ReadWrite scopes to the existing microsoft-mailbox registration and ask users to reconnect; since the lead’s decision of 2026-10-17 (architecture.md § Mailbox integration, Rev 63) the connection asks for both from the first connect, which happens automatically at sign-in, so no reconnect is needed. They use the same /authorized/microsoft-mailbox callback. Admin consent for both scopes was granted on 10 Sep (§2.4). This holds while the scopes stay on the microsoft-mailbox registration; a spec that chooses a separate registration (for example microsoft-calendar) owes a new reply URL per environment and a new ask to Empire IT.

The capture app (F36-a) needs no new reply URL — decided by the lead 2026-09-23. It signs in and out the same way as the CRM (sign-out added 2026-10-12): Microsoft sign-in through the backend on this one app registration, then the check that the Microsoft account matches an active user in the CRM’s users table. So it reuses /login/oauth2/code/microsoft on the api.* host. What it does need is backend-side, not Entra-side: its own origin on the post-login redirect, its origins in the CORS list, and the shared XSRF-TOKEN cookie, whose Domain (the environment’s CRM host) already covers the capture host under AD #33 (the F36-a spec’s Constraints list these).

The mailbox rows are a second, separate client registration (microsoft-mailbox), not incremental consent on sign-in — F20-a decision D-1: login and mailbox consent are different grants with different scopes and token lifetimes, and a mailbox callback arriving on the login filter’s path would be processed as a sign-in and would replace the session principal. It reuses the same app registration, tenant and certificate (§2.1, §2.3), so the only new ask for the mailbox was the three mailbox reply URLs above. Kavindu Anthony (Empire IT) registered them on 24 Sep 2026. F20-a’s cloud validation (.forge/cloud-gated-validation.md §B F20-a) still waits on a deployed environment; the build runs on the profile-scoped stub.

2.3 Client authentication — certificate, not client secret (one app, two certificates: prod and non-prod)

Empire’s Azure tenant has a tenant-wide restriction that blocks adding client secrets, so login uses a certificate credential: the backend authenticates to Entra with a signed client assertion (private_key_jwt), not a shared secret.

One app registration serves all environments (the single “CRM” app — §2.1; confirmed), and it carries two certificates, each with its own private key: one for prod and one non-prod certificate shared by dev, UAT and the developer laptops (“different certificate” and “different private key” are the same choice — a certificate is a public key + metadata whose private half you keep):

  • An Entra app registration holds multiple certificates at once. Each backend signs its client assertion with the private key it was given and puts the certificate thumbprint in the JWT header; Entra matches that thumbprint to the corresponding registered public cert. So both public certs are uploaded to the one app; prod holds only the prod private key, and dev, UAT and the laptops hold only the non-prod one.
  • Why prod is separate: a leaked non-prod key is revoked — delete its public cert — without touching prod; the prod certificate rotates on its own; nothing outside production can authenticate indistinguishably from production; Entra sign-in logs tell prod apart from everything else by thumbprint.
  • What sharing the non-prod key costs (accepted 2026-09-24): dev, UAT and the laptops cannot be revoked or rotated separately — a leak from any one of them, the laptops being the most exposed, means replacing the non-prod pair for all three — and Entra’s sign-in logs cannot tell them apart. The earlier plan of four pairs (dev, UAT, prod, laptop) was replaced for this, because one registration holds any number of certificates and the extra pairs bought only that separation between non-production environments.
  • Operational split: Altrium (Shyamin Ayesh) generates each keypair, sends the public cert to Kavindu to upload to the CRM app (only he has Azure access), and puts the private key in AWS Secrets Manager (referenced by the task definition; never in the repo). Both pairs are generated before AWS exists and uploaded in one meeting; until Secrets Manager exists the private keys wait in a restricted vault in the team’s password manager, and F38 moves the non-prod key into dev’s Secrets Manager (and UAT’s secret — the table below) and the prod key into prod’s as each environment is built, then deletes the vault copies — except the non-prod one, which stays in the vault while laptop sign-in is in use (§2.5).
  • local/test need no certificate — those profiles use the stub login (AD #28); only the deployed dev/uat/prod envs do real Entra login. The one exception is the opt-in laptop mode (§2.5), which uses the non-prod certificate.
  • Backend: Spring Security OAuth2 client with private_key_jwt client authentication pointed at the env’s key, not client-secret-*.
  • Trade-off of one app for all envs: admin consent and the redirect-URI list are shared across environments (acceptable for an internal tool). The stricter alternative — a separate app registration per environment, or at least a separate app for prod — was considered and not taken.

Status: both certificates are uploaded (24 Sep 2026 — Shyamin Ayesh generated both pairs that day and Kavindu Anthony uploaded the public halves). As shown in the portal:

Description in Azure SHA-1 thumbprint (as Azure shows it) Valid
CRM non-production 5B556BD40E2E45EC8807BC6902E65731616AB8A3 24 Sep 2026 – 23 Sep 2028
CRM production 7066640AFC9E0E1EB44F626DDB5A1FAF0D536206 24 Sep 2026 – 23 Sep 2028

A thumbprint identifies a certificate; it is not a secret. The backend sends the SHA-256 form (x5t#S256), so its value looks different from these but names the same certificate. Moving the private keys into Secrets Manager is F38’s (laptops use the non-prod key through F5-i).

The certificate plan. Two certificates — prod and non-prod — proposed by Shyamin Ayesh (Altrium DevOps) on 2026-09-24 and adopted the same day, replacing the lead’s earlier plan of four (dev, UAT, prod, laptop). The non-prod certificate serves dev, UAT and the developer laptops (Empire IT registered the laptop reply URLs on 24 Sep 2026 — §2.5), so laptop sign-in adds no certificate. Both were generated and uploaded on 24 Sep 2026, before AWS exists, with the private keys in a restricted vault until each environment’s Secrets Manager exists. Still to settle with Shyamin Ayesh (the team’s recommendation is given for each): where UAT reads the non-prod key from, and the rotation routine — the other rows are decided.

Question Recommendation Why
How many certificates? Decided — two (2026-09-24): prod, and one non-prod shared by dev, UAT and the developer laptops (§2.5). Both public certificates go on the one CRM app registration; the prod private key lives only in production Production can be revoked alone, and the laptop copy — the most exposed — is never the key production uses. Dev, UAT and the laptops are revoked and rotated together (the cost bullet above)
Where does the non-prod private key live for UAT, given UAT is destroyed after each round (AD #23)? Decided by the lead, 2026-10-23 (F38-h spec decision 2; F38 Rev 24), for Shyamin Ayesh to confirm: UAT’s task definition reads the development environment’s own Entra secrets (et-dev-entra-tenant-id, -client-id, -certificate, -private-key) — the same key, in the same account — so a uat destroy cannot remove them and nothing is copied per round. First recommended: store the non-prod key in a secret that the uat terraform destroy does not remove (outside the destroyable uat module, or in the dev account), and read it into the task definition at each apply The certificate is long-lived by construction — it is dev’s too — so no UAT round needs a new upload by Empire IT
Who generates the key pairs, and when? Decided — Shyamin Ayesh, both before AWS exists, so Kavindu uploads them in one meeting. Generate on a trusted machine, put each private key straight into a restricted vault in the password manager, and send Kavindu only the public .cer/.pem Kavindu is the only one with Azure access, so one upload round beats two. The private key never travels by email or chat — if one arrived that way, move it into the vault and delete the message
Where do the private keys wait until AWS exists? A restricted vault in the team’s password manager, readable only by whoever builds the environments. The non-prod key is also the one developers read for laptop sign-in (§2.5). F38 moves the non-prod key into dev’s Secrets Manager (which UAT also reads — F38-h decision 2, F38 Rev 24) and the prod key into prod’s as each environment is built, then deletes the vault copies — the non-prod one stays while laptops use it The production key is the one to be strictest about — it is the only credential that signs in as the production CRM. It was generated before AWS exists (24 Sep 2026), so until production’s Secrets Manager exists the vault must be readable only by whoever builds production
Validity and rotation Validity decided — two years for both (24 Sep 2026 – 23 Sep 2028, read from the certificates), chosen because the clock starts on the day a certificate is created, before go-live, not when the environment goes live. Both expiry dates are recorded in §5 (the certificate-expiry row); add a calendar reminder about a month before them (register §B FS-6, the expiry row). Rotate with no downtime: upload the new public certificate next to the old one, switch the secret, then ask Empire IT to remove the old one Entra holds several certificates at once and matches on thumbprint, so an overlap is safe. An expired certificate stops every sign-in on that environment
Key format the backend reads A PEM certificate and a PEM private key, as two values: et.security.entra.certificate and et.security.entra.private-key (env ET_SECURITY_ENTRA_CERTIFICATE, ET_SECURITY_ENTRA_PRIVATE_KEY) — forge-et-be auth/config/EntraLoginConfig.java The Secrets Manager secret, and any laptop copy, must hold exactly these two strings

2.4 Microsoft Graph — delegated permissions

Delegated permissions (NOT application permissions), single-tenant, with one-time org-wide admin consent for the full list. Backs mailbox/calendar/contacts (Microsoft Graph is the only mailbox provider in V1 — AD #16).

Permission Purpose
Mail.ReadWrite Read mail, create/edit AI drafts (covers Mail.Read)
Mail.Send Send, reply, reply-all, forward as the rep
Calendars.ReadWrite View + add/edit/cancel events, respond to invites
Contacts.ReadWrite View + add/edit/delete contacts
User.Read Read the signed-in user’s own profile (login)
User.ReadBasic.All Read basic profiles of other users in the org — not confirmed in the granted set (see below)
offline_access Refresh tokens for background sync — required, easy to miss
openid, profile, email OAuth sign-in

Two entries in this area are unconfirmed, and no feature may assume them (F20-a signal W-4):

  • User.ReadBasic.All appears in the table above but not in the granted list recorded by the 2026-09-10 admin-consent note. Treat it as un-granted until someone re-reads the app registration.
  • Mail.ReadWrite.Shared / Mail.Send.Shared (shared-mailbox access) appear only in discovery/research/m365-mailbox-integration.md §4 and are in neither the table above nor the granted list. No V1 feature needs them any more: the shared inquiries inbox (F20-e) left V1 on 2026-09-24 — info@ reaches the CRM by forwarding to a rep’s own mailbox — so there is no shared-mailbox connection to consent to. If Empire adds info@ as an ordinary CRM user instead, it signs in and connects its own mailbox like anyone else and needs no .Shared scope.

F20-a needs none of the three, so this blocks nothing today. F22 (contact sync; tracked in Phase 1 from 2026-11-06 to 2026-11-11, and again from 2026-11-20, started early) syncs each user’s own Outlook contacts, which needs only Contacts.ReadWrite — granted on 10 Sep and asked for at the first connect since F20-i. It would need User.ReadBasic.All only if its spec also reads other people in the organisation; in that case, confirm against the live app registration before the spec relies on it.

2.5 Real Microsoft sign-in from a developer laptop — Entra side ready, built by F5-i Wave 1 — forge-et-be/docs/auth/laptop-sign-in.md

The problem it solved (September 2026). Before dev existed, nobody could test the real Microsoft sign-in or the real mailbox connection end to end: every test ran against the stub login and the stub mailbox, and waiting for AWS would have left the riskiest integration unproved until late. Dev now signs in through Microsoft (since 13 Oct 2026, §4, the CRM frontend’s dev deploy), and laptop mode stays useful for testing against the real tenant without a deploy.

The proposal. Let a developer run the backend on a laptop against the real Empire tenant and real test accounts (CRM01–05@empire.lk). Three things are needed, and all three are required:

  1. Two laptop reply URLs on the CRM app registration — the two “Laptop (local)” rows in §2.2. Done — Empire IT registered them on 24 Sep 2026, on the shared CRM registration.
  2. The non-prod certificate (§2.3) — laptops share it with dev and UAT, so laptop sign-in needs no certificate of its own and no extra upload. Done — uploaded 24 Sep 2026. The private key is read from the team’s password manager, never from a repo, a chat or an email. On a laptop it reaches the backend either as the ET_SECURITY_ENTRA_PRIVATE_KEY environment variable (exported in the shell session only, not written to a committed .env) or in the gitignored application-local.yml. The certificate goes the same way as ET_SECURITY_ENTRA_CERTIFICATE. The tenant and client IDs are the non-secret values in §2.1.
  3. An opt-in real-Entra laptop mode — built by F5-i Wave 1. On the backend it is the entra-laptop profile, used together with local: SPRING_PROFILES_ACTIVE=local,entra-laptop ./gradlew bootRun. It turns on the real Entra sign-in and mailbox consent and the real Graph mailbox provider, while everything else stays local (Docker Postgres, LocalStack, the local token key, the local file scanner); the stub login and the stub mailbox are off while it is on. The backend refuses to start entra-laptop without local or together with test, so no deployed environment can reach it (AD #28), and in laptop mode it checks the certificate and private key at startup. On the frontend it is the VITE_ENTRA_LAPTOP=true flag in the gitignored .env.local, honoured only by pnpm dev: the login page shows only “Sign in with Microsoft”, every Connect or Reconnect button starts the real consent, and the dev server forwards /oauth2, /login/oauth2 and /authorized to the backend on localhost:8080, so the reply URLs match the two laptop rows in §2.2. With neither switch set, local development is unchanged. The how-to is forge-et-be/docs/auth/laptop-sign-in.md; tracked as F5-i (Phase 1, owner Tharinda). The live run as CRM01 (F5-i-WI-2.1) was recorded in forge-et-harness#322, merged 2026-09-30.

What a laptop still cannot prove. Microsoft Graph sends change notifications to a public https URL, and it cannot reach localhost. Leave et.mailbox.public-base-url unset on a laptop — the backend then makes no Graph subscription (F20-a FR-9) — or use a temporary public tunnel if F20-b’s notifications must be seen. KMS, SES, cross-host cookies and CORS also stay unproved on a laptop. Those rows stay open in .forge/cloud-gated-validation.md until each is proved on dev.

Security trade-off, and the go-live step it creates. The laptop reply URLs sit on the same app registration that production uses, and the laptops hold the non-prod private key. That is acceptable for pre-launch testing because a laptop reply URL only ever sends the sign-in code back to the user’s own machine, and the code is useless without one of the registration’s private keys — all of which Altrium holds. Before production go-live, remove the two laptop reply URLs from the CRM registration, or move laptop testing to a separate app registration. The non-prod certificate stays, because dev and UAT use it; since its private key has been on developer laptops, consider replacing the non-prod pair at the same time (a rotation — §2.3). The stricter option — a separate app registration for laptops and dev — was offered to Empire IT; Kavindu registered the laptop URLs on the shared CRM registration instead.

3. AWS environment ownership & build-out

Empire Teas owns the AWS account (confirmed 29 September 2026). Kavindu Anthony (Empire Teas IT) holds the main account, Empire Teas’ own card pays for it, and AWS bills Empire Teas directly. Empire Teas created an administrator IAM identity and gave its credentials to Shyamin Ayesh, who builds the infrastructure with Terraform, sets up the environments and creates the developers’ access. This returns to the earlier working assumption (AD-O-10) that Empire would open the account and give Altrium access. It replaces the SOW ET-002 model recorded here from 13 September, under which Altrium would have owned the organisation and accounts and billed a fixed USD 650/month; Empire Teas declined ET-002 (aws-sow.md, marked superseded).

  • Region us-east-1 (AD-O-9); dev + prod permanent, UAT temporary: one instance from Phase 1’s acceptance round to the end of Phase 2’s (AD #23). As of 30 Sep 2026 the dev Terraform applies to us-west-2, and its state bucket is in us-west-2 too (§4). Decided on 1 Oct 2026 with Shyamin Ayesh: the estate moves to us-east-1, because AgentCore Web Search is not offered in us-west-2 (meeting record §1). The region switch was pushed to forge-et-iaac’s dev branch and applied on 5 Oct 2026 (commits dff66ce, which destroys the us-west-2 resources, and 5808711); every resource since is in us-east-1. The et-ops state bucket stays in us-west-2 (§4). PR #1, which proposed the switch, was still open on 9 Oct 2026 although its change is on dev.
  • The account stays with Empire Teas on exit; there is nothing to transfer.
  • Accounts seen in the code, as of 9 Oct 2026. The dev estate is in account 570417736903 (the ARNs in forge-et-be .aws/td.dev.json). The crm.empire.lk zone is in a different account, 331342166625, which the Terraform comment calls ets-org; the dev Terraform writes records into it by assuming the role dns-dev, which is scoped to the dev.crm.empire.lk names (forge-et-iaac overlays/dev/providers.tf, env.auto.tfvars). So dev has its own account, with the zone outside it, as F38 FR-6 allows. Read in the account on 13 Oct 2026: dev is a member of the AWS Organization o-imaviqd8nu, whose management account is 331342166625; whether a production account exists is not readable from dev (next bullet).
  • The delivery team’s AWS access and Bedrock, as of 9 Oct 2026 (DevOps channel). Dinil Gamage asked Shyamin Ayesh for access for an identity called svc-harness in dev, used by Claude and the team’s scripts: CLI access (an access key, or an Identity Center user with a permission set used through aws configure sso), bedrock:InvokeModel and InvokeModelWithResponseStream on both inference profiles and on arn:aws:bedrock:*::foundation-model/anthropic.claude-*, bedrock-agentcore:InvokeGateway on the AgentCore Gateway, and the three aws-marketplace subscription permissions. Shyamin said the CLI access, the Bedrock access and the Marketplace permissions were granted; Dinil’s aws login as svc-harness then failed (not resolved in the thread). Bedrock: Claude Haiku 4.5 works and is subscribed. Claude Sonnet 5.5 and Claude Haiku 5.5 both fail with “is not available for this account … contact AWS Sales”, so the synthesis model AD #19 pins cannot be called from this account yet (register §C-5). (Update 14 Oct 2026: Shyamin Ayesh accepted the model agreements for Sonnet 5.5 and Haiku 5.5, and get-foundation-model-availability now reads AVAILABLE and AUTHORIZED for both. A one-line call to each through its us. and global. profile still returns “not available for this account”, because Service Quotas gives both models 0 tokens a minute on every cross-Region and Global quota. Haiku 4.5 answers, at 10 requests a minute. The remaining step is a quota increase for the two models, raised through Service Quotas or an AWS support case by someone with the account’s administrator access — §4 item 4.) (Update 22 Oct 2026: AWS has refused every quota request so far — six on 15 Oct and three on 21 Oct — and no quota has changed; a Support case is next — §4 item 4, the 22 Oct update.) Settled 13–14 Oct 2026: svc-harness is an IAM user with an access key (the aws login route never worked), used through the CLI profile et-dev; the lead accepted it, with its broad access, as a dev-only exception to F38 NFR-1, on the condition that Claude only reads with it and asks before any call that is not a read, test invocations included (.claude/CLAUDE.md § ASK FIRST) (F38 Rev 9; §4 item 7). How the identities fit together is explained in .forge/guides/aws-setup-explained.md.
  • Open until Shyamin Ayesh finishes the setup (answered from his work, not decided in advance): whether dev and prod are separate member accounts under an AWS Organization with Empire’s account as the management account, as AD #4 and F38-a FR-1 expect (half answered 13 Oct: dev is a member account under the Organization; production is not confirmed); how people sign in (F38-a asks for IAM Identity Center with MFA and no long-lived personal keys, which would retire the shared administrator credentials once each person has their own login — answered 2 Oct 2026 by Shyamin Ayesh: each person gets an SSO user with MFA and uses aws configure sso for the CLI; GitHub Actions uses an IAM service user per account with its key in GitHub secrets (meeting record § Confirmed with Shyamin after the call); for dev this was replaced on 14 Oct: the team uses svc-harness’s access key, §4 item 8; production keeps personal logins); and which account the AWS Budgets alerts reach, now that Empire Teas receives the bill (no budget exists yet, §4 item 13; not the delivery team’s to arrange or track — the lead, 18 Oct, F38 Rev 18).

4. Terraform

The Terraform lives in its own repository, altrium-io/forge-et-iaac — the repository for all DevOps work on this project (decided 2026-09-29; AD #23’s location revised from forge-et-be/terraform/, which was never created). Shyamin Ayesh writes the Terraform there for the estate in Empire Teas’ account, and Senuthmi Thimansa, a DevOps engineer working with him, also changes it (GitHub senuthmi07, an admin on the repository). The repository was created empty on 2026-09-29. Tharinda Rajapaksha and Dinil Gamage also have write access. How changes are reviewed (lead, 2026-10-02): the trunk is dev; every change is a PR to dev, with review requested from the others who have write access — Shyamin, Senuthmi, Tharinda, Dinil and, since 2026-11-19, Nanndish (.claude/rules/forge-et-iaac.md, which holds the current list). Still on hold until the setup is finished: which of the other delivery-repository conventions (CODEOWNERS, branch protection, the PR template, the terraform plan-on-PR workflow in architecture.md § CI/CD Pipelines) it takes on.

What it holds, as read on 14 Oct 2026 at commit 9eac094; re-read on 16 Oct at 76fce06, the only commit since, which “Checked against F38” row 1 describes (all on the default branch dev, pushed there directly by the DevOps team; the RDS and ECR commits landed on 5 Oct, the backend’s path on 9 Oct, and the AgentCore Gateway, the CRM frontend’s hosting, the Entra secrets and the cost-allocation tags on 13 Oct; the runs for 87bac05, 162d5de and ac1efdc failed and the next commit fixed each; re-check with git -C forge-et-iaac log --oneline origin/dev). Since 20 Oct 2026 (18:33 UTC; 00:03 on 21 Oct in Sri Lanka) dev was at d230b133 (and since 21 Oct, 06:03 UTC, at 0eee623, the DevOps team’s direct push adding base/ses, applied by dev.yaml run 39553084687 — the base/ses bullet below; re-read 22 Oct) (and since 22 Oct, 07:22 UTC, at 5e078c0d, the merge of the delivery team’s forge-et-iaac#5, F38-b wave 5, applied by dev.yaml run 39740047356 — the capture-app bullet below) (and since 23 Oct, 10:41 UTC, at 1590009, after three more merges of delivery-team PRs, each approved by Shyamin Ayesh and applied by its own dev.yaml run: #8 (d86d2ca9, F11-c, run 39936834361), #6 (acd1cf7, F38-d wave 3, run 39940853448: “0 added, 2 changed, 0 destroyed”) and #7 (1590009, F38-e wave 1, run 39940862972: “8 added, 0 changed, 0 destroyed”); re-read in the account on 25 Oct) (and since 26 Oct, 04:52 UTC, at c728e06b, the merge of the delivery team’s forge-et-iaac#9, F38-b wave 6, applied by dev.yaml run 40271292253 — the base/secrets bullet below) (and since 28 Oct, 16:18 UTC, at 876d634, after three more merges of delivery-team PRs, each applied by its own dev.yaml run: #10 (5401a88, 26 Oct, documentation only, run 40282639877), #13 (a77370d, 28 Oct, the task role’s Haiku 5.5 grant, run 40635870998: “0 added, 1 changed, 0 destroyed” — the base/ecs bullet below) and #14 (876d634, 28 Oct, the frontends’ Content-Security-Policy, run 40724734100: “0 added, 2 changed, 0 destroyed” — the two hosting bullets below); re-read 29 Oct), the merge of the delivery team’s forge-et-iaac#3 (F38-b wave 1, with forge-et-iaac#4 merged into #3’s branch first), and its dev.yaml run 39475850266 applied it (“Apply complete! Resources: 28 added, 5 changed, 1 destroyed” — the one destroyed is the ECR lifecycle policy, replaced); each bullet below says what #3 changed, as read in the account on 21 Oct:

  • base/networking — a VPC (10.0.0.0/16), two public subnets in two Availability Zones with public IPs on launch, an internet gateway and one public route table. No private subnets. Since 4 Nov (#15, F38-i WI-1.1): an option private_subnet_cidrs (default []) adds private subnets in the public subnets’ zones, with a route table of their own declared with no route, and the output private_subnet_ids; dev uses the default, so it still has no private subnets.
  • base/security-groups — the load balancer’s (80 and 443 from anywhere), the backend’s (8080 from the load balancer’s group only) and the database’s (5432 from the whole VPC range). Since 4 Nov (#15, F38-i WI-1.1): an option database_ingress_from_backend_only (default false) makes the database admit 5432 only from the backend’s group; dev uses the default, so its database still admits the VPC range, with the same descriptions.
  • base/rds — RDS PostgreSQL 18, db.t3.micro, 20 GiB, single-AZ, database et, master user app, publicly_accessible = false, placed in the two public subnets. RDS manages the master password in its own Secrets Manager secret (manage_master_user_password), so the password is never in the Terraform state. No storage_encrypted, no KMS key, no backup_retention_period, no max_allocated_storage and no parameter group are set; skip_final_snapshot = true, deletion protection off. Since 20 Oct (#3): the RDS-managed secret’s automatic rotation is off (base/rds/rotation.tf; describe-secret shows RotationEnabled: false), and 76fce06’s rejected master_user_secret_rotation_days line is gone. Since 4 Nov (#15, F38-i WI-1.1): the storage type (gp2), the autoscaling ceiling (max_allocated_storage, none), backup_retention_period (0), deletion_protection (false), skip_final_snapshot (true; when false, a final snapshot <project>-<environment>-final), storage_encrypted (false; true uses the AWS-managed aws/rds key) and availability_zone (none) are module options, each defaulting to dev’s setting; dev uses the defaults, so nothing on its database changed (the merge’s plan read No changes.). Production sets them in F38-i WI-4.1’s overlays/prod.
  • base/ecr — the repository crm/backend, scan on push, tags mutable; untagged images expire after 7 days, and the last 10 images tagged v* or latest are kept (the dev images are tagged build-<sha>, which no rule matched, so they were never expired). Since 20 Oct (#3): a third rule keeps the last 100 build-* images, and the v* rule no longer names latest. Since 23 Oct (#6, F38-d WI-3.1): tags immutable (aws ecr describe-repositories, 25 Oct: imageTagMutability IMMUTABLE; 63 images held).
  • base/ecs — the cluster et-dev (FARGATE and FARGATE_SPOT, Container Insights off); the log group /ecs/dev/backend, 30-day retention; the execution role et-dev-backend-execution (the AWS-managed task-execution policy, plus secretsmanager:GetSecretValue on the RDS secret only); the task role et-dev-backend-task (on 14 Oct: SQS consume on the scheduler queue only — no S3, KMS, Bedrock or SES; for since 20 Oct see the end of this bullet); the service backend, one task, FARGATE, public IP, behind the target group, health-check grace 180 s, deployment circuit breaker with rollback, and ignore_changes = [task_definition] because forge-et-be’s workflow registers every revision. Since 20 Oct (#3): the task role also holds s3, kms, bedrock and sqs-send policies (the attachments bucket, the storage and mailbox token keys, Bedrock, and sending to the ingest and lead-triage queues), and its sqs policy also consumes from those two queues; still no SES (update 2026-10-21, F38-a WI-3.1: since forge-et-iaac 0eee623, applied on 21 Oct, the task role also holds an inline ses policy allowing ses:SendEmail and ses:SendRawEmail on the identity dev.crm.empire.lk and the configuration set et-dev-mail only — base/ecs/task_role.tf; “F38-a — the development account, checked”, AC-5); the service runs in one subnet, in us-east-1b, the database’s zone. (Decided 2026-11-12, the lead: dev’s backend stays on ordinary FARGATE, not Fargate Spot. The cluster also registers FARGATE_SPOT (base/ecs/capacity_providers.tf), which costs less, but AWS can reclaim a Spot task with two minutes’ notice and dev runs one task, so dev would be down for a minute or two each time; base/ecs/service.tf keeps launch_type = "FARGATE" and no change is asked of the DevOps team.) (Update 2026-10-29: since forge-et-iaac#13, applied on 28 Oct, the bedrock policy also allows Haiku 5.5’s Global profile global.anthropic.claude-haiku-5-5 and arn:aws:bedrock:*::foundation-model/anthropic.claude-haiku-5-5; read with aws iam get-role-policy --role-name et-dev-backend-task --policy-name bedrock on 29 Oct, it lists Haiku 4.5’s us. profile and its models in us-east-1, us-east-2 and us-west-2, and the Global profiles and Region-less model ARNs of Sonnet 5.5 and Haiku 5.5, nothing else.) (Update 2026-11-04: since forge-et-iaac#16, applied on 4 Nov (run 41882320196), the bedrock policy allows all six inference profiles — us. and global. of Haiku 4.5, Sonnet 5.5 and Haiku 5.5 — and arn:aws:bedrock:*::foundation-model/<id> for each of the three models; Haiku 4.5’s wildcard replaced its three regional ARNs. Read the same day with aws iam get-role-policy --role-name et-dev-backend-task --policy-name bedrock: those nine ARNs, bedrock:InvokeModel only.)
  • base/alb — the load balancer et-dev-backend: HTTPS on 443 with ELBSecurityPolicy-TLS13-1-2-2021-06 (TLS 1.2 and 1.3), HTTP on 80 redirecting to HTTPS; the target group on 8080 checks /actuator/health/liveness (deregistration delay left at the 300 s default). Since 20 Oct (#3, a359b61): every 10 s with a 5 s timeout, two healthy and three unhealthy checks. Since 23 Oct (#6, F38-d WI-3.1): deregistration delay 30 s (deregistration_delay.timeout_seconds 30, read 25 Oct, the one target healthy).
  • base/domain — the ACM certificate for api.dev.crm.empire.lk, DNS-validated, and an alias A record to the load balancer, both written into the crm.empire.lk zone in the ets-org account through the dns-dev role (§3).
  • base/sqs — the scheduler queue et-dev-scheduler (visibility timeout 30 s, SQS-managed encryption) and its dead-letter queue et-dev-scheduler-dlq, redrive at maxReceiveCount 3, matching forge-et-be/docs/messaging/scheduler-cloud-contract.md. Since 20 Oct (#3 and #4): also the mailbox ingest queue et-dev-mailbox-ingest and the lead-triage queue et-dev-lead-triage, each with its dead-letter queue, and every dead-letter queue keeps messages 14 days. With them, base/kms (the keys alias/et-dev-storage and alias/et-dev-mailbox-token), base/attachments (the bucket et-dev-attachments: versioned, all public access blocked, encrypted with the storage key, SSE-C refused) and base/scheduler (the schedule group et-dev, its schedules all enabled — mailbox-subscription-renewal, mailbox-delta-poll, email-upload-cleanup, price-expiry-sweep and lead-triage-sweep — and the role et-dev-scheduler, allowed to send to the scheduler queue). Since 23 Oct (#8): the lead-research queue and a sixth schedule, lead-research-sweep (the #8 entry in the PR list below). The schedules run on dev: on 24 Oct (UTC) et-dev-scheduler received 173 messages and the backend consumed all 173, with nothing in its dead-letter queue — one day of every schedule (96 delta polls, 4 subscription renewals, 24 runs each of the upload cleanup and the two lead sweeps, and the daily price-expiry sweep, which the backend logged from PriceExpirySweepService at 18:35 UTC) (CloudWatch NumberOfMessagesSent/NumberOfMessagesDeleted and AWS/Scheduler InvocationAttemptCount, read 25 Oct).
  • base/ses (21 Oct, 0eee623, pushed to dev by Shyamin Ayesh) — the SES identity dev.crm.empire.lk with Easy DKIM pinned to 2048-bit keys, the MAIL FROM domain bounce.dev.crm.empire.lk, the configuration set et-dev-mail, and the Route 53 records (three DKIM CNAMEs, the MAIL FROM MX and SPF TXT, and _dmarc with p=none and no report address) written through dns-dev; with it, the task role’s inline ses policy (the base/ecs bullet above). Production access was granted the same day. Checked in “F38-a — the development account, checked”, AC-5.
  • base/alerts (23 Oct, the delivery team’s forge-et-iaac#7, F38-e WI-1.1) — the SNS topic et-dev-alerts with an email subscription for each of the four developers, and the EventBridge rule et-dev-backend-deployment-failed (ENABLED; source aws.ecs, detail type ECS Deployment State Change, event SERVICE_DEPLOYMENT_FAILED, resource the et-dev/backend service) sending a short summary with the service, deployment, reason and time to the topic. Read 25 Oct: all four subscriptions confirmed, none pending. The email itself is proved by the first failed deploy (register §B F38-e). No CloudWatch alarm exists yet. Read 2026-10-25 (F38-e WI-2.1): the topic is not encrypted, its policy allows only cloudwatch.amazonaws.com and events.amazonaws.com to publish when aws:SourceAccount is this account, and the rule’s one target is the topic, through the input transformer; register §B F38-e “F38-e evidence — AC-3”.
  • base/agentcore (13 Oct) — the AgentCore Gateway et-dev-gateway (MCP, inbound auth AWS_IAM) with one target web-search on the managed web-search connector, and its service role et-dev-agentcore-gateway, allowed web search (InvokeWebSearch) and bedrock-agentcore:InvokeGateway on this account’s gateways, nothing else. Used for SP-ARCH-004 stage 2 (.forge/design/ai/lead-research-source-options.md §11). 2026-11-19 (Dinil): Exa was chosen, so nothing uses the gateway; it and svc-harness’s inline policy spike-invoke-gateway are left in place, because neither has a standing charge (Cost Explorer, AgentCore: October $0.61 from the spike’s searches, November to the 19th $0.0002) and the gateway is in overlays/dev only — not uat, and not the production overlay (#17). Removing either is optional tidying: the module is Shyamin Ayesh’s to remove, and the policy needs the lead’s OK (an AWS change). Later on 2026-11-19 the lead asked for the gateway to be removed with the audit fixes: forge-et-iaac#28 removes base/agentcore and dev’s module "agentcore", and its merge destroys the gateway, its target, its role and the role’s policy. The svc-harness policy is removed by hand after that apply (.forge/owed.md).
  • base/secrets (26 Oct, #9; until then base/entra-secrets, 13 Oct, and base/search-secrets, 23 Oct, #8) — the five Secrets Manager secrets, from one list with for_each and prevent_destroy: et-dev-entra-tenant-id, -client-id, -certificate and -private-key, and et-dev-exa-api-key, each tagged Module = secrets. Terraform creates the secrets only; their values were set by hand, so no value passes through a plan, an apply log or the state. A new secret is a new key in the module’s list, never a module of its own (.claude/rules/forge-et-iaac.md). The execution role may read all five (the ecs module’s secret_arns). Read 26 Oct, after #9 applied, with list-secrets and list-secret-version-ids (no value read): the five keep the ARNs they had on 25 Oct, each with its AWSCURRENT version, and the execution role’s secrets policy names the same five and the RDS-managed secret. The next backend deploy, a manual build run on main the same morning (40274424861, backend:67), started and stayed healthy, so the role reads all five (F38-b AC-13, ticked). Since 6 Nov 2026 (F38-i wave 3, forge-et-be#374, deployed as backend:111) the backend reads two of the five — et-dev-entra-private-key and et-dev-exa-api-key; the tenant ID, client ID and certificate are plain task-definition values (AD #40), and their three secrets stay until F38-i wave 8 removes them.
  • base/acm, base/s3, base/cloudfront and a record in overlays/dev (13 Oct) — the CRM frontend’s hosting: the ACM certificate for dev.crm.empire.lk (DNS-validated through dns-dev); the bucket et-dev-spa (versioned, all public access blocked, S3-managed encryption); the CloudFront distribution E2RVC3893O10TJ (origin access control, so only this distribution reads the bucket; HTTPS only with TLSv1.2_2021; PriceClass_100; 403 and 404 answered with /index.html so the SPA’s routes load); and an alias A record dev.crm.empire.lk to it. No WAF and no access logs. Since 20 Oct (#3, e35554d, F38 Rev 20): the response-headers policy et-dev-spa-security-headers sends HSTS (one year, includeSubDomains), X-Content-Type-Options: nosniff, X-Frame-Options: SAMEORIGIN and Referrer-Policy: strict-origin-when-cross-origin, and no CSP (curl -sI https://dev.crm.empire.lk/, 21 Oct). (Update 2026-10-29: since forge-et-iaac#14, applied on 28 Oct, the policy also sends the CRM’s Content-Security-Policy, as given in the #14 bullet under “PRs” below.) The capture app’s hosting is the next bullet.
  • The same three modules and a record in overlays/dev, with name = "capture" (22 Oct, the delivery team’s forge-et-iaac#5, F38-b wave 5, merged as 5e078c0d after Shyamin Ayesh’s approval; dev.yaml run 39740047356 planned “11 to add, 0 to change, 0 to destroy” and applied them) — the capture app’s hosting. As read as svc-harness on 22 Oct: the certificate for capture.dev.crm.empire.lk, ISSUED, in use by the distribution, expiring 7 May 2027 UTC; the bucket et-dev-capture (versioned, all public access blocked, its policy letting only this distribution’s s3:GetObject through, and empty — no build has been uploaded); the CloudFront distribution EZ06T8FTKF7UZ (d2eo2sv80t77es.cloudfront.net, Deployed, alias capture.dev.crm.empire.lk, its own origin access control, HTTPS only with TLSv1.2_2021, PriceClass_100, 403 and 404 answered with /index.html); its own response-headers policy et-dev-capture-security-headers with the CRM’s four headers and no CSP; and the alias A record, which resolves to CloudFront. curl -sSI https://capture.dev.crm.empire.lk/ answers 403 from CloudFront (x-cache: Error from cloudfront) with the four headers and a valid certificate — what an empty bucket gives — until F38-d’s capture deploy uploads a build (its WI-5.1 workflow, first run in WI-7.1). (Update 2026-10-26, F38-d WI-7.1: the bucket holds the capture app’s build since 25 Oct (run 40224118047, then every non-workflow merge), and the host answers 200; “F38-d — the dev deploy pipelines, checked”, AC-5.) The outputs capture_bucket_id, capture_cloudfront_distribution_id and capture_cloudfront_domain_name give the deploy its two targets. (Update 2026-10-29: since forge-et-iaac#14, applied on 28 Oct, et-dev-capture-security-headers also sends the capture app’s own Content-Security-Policy — the #14 bullet under “PRs” below.)
  • The options for a destroyable instance (13 Nov 2026, the delivery team’s forge-et-iaac#11, F38-h WI-1.1, as spec Rev 7 changed it) — base/s3’s force_destroy (default false), base/kms’s deletion_window_in_days (default 30) and base/ecs’s create_service (default true; a moved block keeps dev’s service as aws_ecs_service.backend[0]). There is no bootstrap revision: an apply with create_service = false creates everything but the ECS service, for a first apply before the task definition is registered. Dev uses the defaults; its plan read No changes. before the review (run 43021956495, a move only) and after the merge (run 43376614685).
  • overlays/dev — calls every module; AWS provider >= 6.62, < 7.0 (raised from ~> 5.45 to ~> 6.0 on 13 Oct, b6bee68, for the AgentCore resources, which forge-et-iaac#28 removes, and to a 6.62 floor on 18 Oct, 7dbfb0f, for rotation_enabled = false on the RDS-managed secret; the lock file pins 6.69.1 as of 26 Oct) and random ~> 3.6; region us-east-1; state in the S3 bucket et-ops (key dev/terraform.tfstate, us-west-2, S3 lock files); default tags Project, Environment and Owner = shyamin@altrium.io, and since 13 Oct a Module tag per module, for cost allocation.
  • .github/workflows/dev.yaml — since 20 Oct (#3), on a push to dev and on manual dispatch only: fmt -check, init -lockfile=readonly, validate, plan -out=tfplan, then apply -auto-approve tfplan, the apply only when the ref is dev, so a branch pushed for review runs nothing and a manual run on another branch only plans. One concurrency group for applies on dev, never cancelled, and one per branch for plans; a read-only token, a 60-minute timeout, actions/checkout@v4, Terraform 1.12.1, and static AWS access keys held as secrets of the GitHub Environment dev. The first run, 39475850266 on d230b133, applied (evidence, 2026-10-21: “F38-b — the dev estate, checked”, AC-1). Before #3: it ran on every push to any branch with actions/checkout@v2; on 13 Oct the apply step was commented out for one commit, 17fbf02, and restored by the next, 9eac094; 76fce06, 15 Oct, commented it out again, and its plan failed at validate until #3.
  • overlays/uat and .github/workflows/uat.yaml (13 Nov 2026, the delivery team’s forge-et-iaac#11, F38-h WI-3.1) — the temporary uat instance: dev’s modules and sizes with et-uat-* names, Environment = UAT on every resource, state at et-ops uat/terraform.tfstate, the three options above on, its DNS records through the management-account role dns-uat (overlays/uat/env.auto.tfvars), the database pinned to us-east-1b with the tasks, the lead-ranking queue wired as dev’s, and dev’s Entra and Exa secrets and SES identity read by name; no secrets, SES, ECR, AgentCore or alerts module, and no capture-app hosting yet (.forge/owed.md, “Build forge-et-pwa’s deploy-uat and the capture app’s UAT hosting”; added by forge-et-iaac#23 on 19 Nov, below). uat.yaml plans, applies or destroys it, started by hand on dev only, under svc-github’s key (the dev Environment), with a service input (Rev 7’s three steps). Dev’s plan read no change before the review (run 43021956495) and after the merge (run 43376614685).

The uat instance, as read on 13 Nov 2026 (Claude, for Tharinda; read-only AWS calls as svc-harness). Senuthmi Thimansa applied it the same morning, 14 minutes after #11 merged: uat.yaml plan run 43376819951 (06:20 UTC, service unticked, “Plan: 75 to add, 0 to change, 0 to destroy”) and apply run 43378306661 (06:23–06:29 UTC, “Apply complete! Resources: 75 added, 0 changed, 0 destroyed”) — step 1 of Rev 7’s three, before F38-h wave 4 built deploy-uat (F38-h spec Rev 8; the lead keeps it up). resourcegroupstaggingapi get-resources with Environment=UAT lists: the VPC, two subnets, an internet gateway, a route table and three security groups; the load balancer et-uat-backend (two listeners, one target group, no target); the database et-uat (db.t3.micro, 20 GiB, us-east-1b, single-AZ, available since 06:27 UTC, deletion protection off) with its RDS-managed secret; the cluster et-uat (no service, no task); the log group /ecs/uat/backend; two KMS keys; the buckets et-uat-attachments and et-uat-spa (empty); the CloudFront distribution E3HG4BWVQ8JVD8 for uat.crm.empire.lk; the scheduler queue and the lead-triage, lead-research, lead-ranking and mailbox-ingest queues, each with its DLQ; and the schedule group et-uat with all eight schedules ENABLED. Both certificates, uat.crm.empire.lk and api.uat.crm.empire.lk, are ISSUED, so the DevOps team’s dns-uat role works, and both names resolve. https://uat.crm.empire.lk/ answers 403 (no bundle) and https://api.uat.crm.empire.lk/actuator/health 503 (no task). No task-definition family for UAT is registered, and no repository has a uat GitHub Environment. With nothing consuming it, et-uat-scheduler held 249 messages at 10:01 UTC, the schedules’ ticks since the apply; each waits until the backend starts or expires after the queue’s four-day retention (.forge/owed.md, “The uat instance runs with no backend”). Cost while idle, from AWS’s us-east-1 list prices: about USD 1.3 a day, mostly the load balancer and its public IPv4 addresses and the database; dev’s month to date on 13 Nov was USD 25.66 (Cost Explorer, 1–12 Nov).

The uat instance with its backend, 16–17 Nov 2026 (F38-h WI-5.1; Senuthmi Thimansa deployed, Claude read and ran the failure cases for Tharinda). The evidence is in .forge/plans/F38-h/F38-h-WI-5.1-record.md.

  • The deploy. Senuthmi tagged v0.1.0-rc1 in both app repos on 16 Nov. The backend’s deploy-uat run 43810012491 registered et-uat-backend:1 and stopped. uat.yaml plan run 43809114419 (service unticked, No changes.) and apply run 43810590144 (service ticked, 1 added) created the service. The backend run’s re-run registered and rolled out et-uat-backend:2. The frontend’s run 43814847625 deployed to E3HG4BWVQ8JVD8.
  • What runs. UAT runs et-uat-backend:2: one task in us-east-1b, the same zone as the database, on image sha256:a579880240e6…. That is the image dev built for f392e8e (build-f392e8ec…), tagged v0.1.0-rc1. Both hosts report 0.1.0.
  • The plan after the bring-up. uat.yaml plan run 43949795728 (service ticked, at b175dae) printed No changes.
  • The failure cases. A broken build’s deploy (et-uat-backend:3) was rolled back by the circuit breaker, and :2 kept serving.

The uat instance’s data and sign-in, 17–19 Nov 2026 (F38-h WI-5.2, Tharinda; Claude drove the browser from 19 Nov at Tharinda’s instruction). The evidence is in .forge/plans/F38-h/F38-h-WI-5.2-record.md.

  • First sign-in. CRM01 signed in on UAT through Microsoft on 17 Nov, the first sign-in there. It is the seeded IT Admin with the six Settings page keys (AD #41). Before anything was entered, UAT had only the IT Admin role and crm01@empire.lk.
  • The data (.forge/plans/F38-h/uat-data-checklist.md Parts 1–5, entered through the app on 19 Nov). IT Admin gained the checklist’s account, lead and pricing keys. Seven roles were created: Sales Representative, Sales Manager, R&D / Sample Owner (Taster), Tea Room (Pricing), Executive / Management, Administrator and Smoke test (view only). The users: crm02@empire.lk Sales Representative (manager CRM03), crm03@empire.lk Sales Manager, crm04@empire.lk Tea Room (Pricing), and crm05@empire.lk the smoke account with Smoke test (view only). The SMOKE records are SMOKE Account, SMOKE Lead, SMOKE Standard (BS-000001), sales 2026/98 and 2026/99, and one published price on each. The checklist’s three TEST accounts and five TEST leads were added too.
  • Smoke runs. Against dev, unchanged, the run passed on its second attempt. Against uat it did not run: v0.1.0-rc1 has no uat smoke target (.forge/owed.md, “Build the smoke suite’s uat target in forge-et-fe”).
  • The mailbox check (F38-h decision 11). crm03@empire.lk was connected on dev and UAT at once for about six hours. The result was harmless: each environment kept its own subscription, delta state and stored mail. No renewal of either subscription was exercised, because both expire on 26 Nov. CRM03 was then disconnected on UAT and on dev. Its dev role went back to RETEST 4.2 reader.
  • What stays. Spec Rev 13 keeps the instance for the round, so the data stays on UAT. CRM04 has no Exchange mailbox (§5). Its failed connection 18 stays on dev (.forge/owed.md).

The capture app’s UAT hosting, 19 Nov 2026 (F38-h spec Rev 12, FR-13; built by Senuthmi Thimansa; Claude read it for Tharinda with read-only AWS calls as svc-harness).

  • Applied. forge-et-iaac#23 (approved by Shyamin Ayesh) merged into dev at 05:02 UTC (7e3481a; dev.yaml changed nothing of its own). Senuthmi then ran uat.yaml at 7e3481a: plan run 44463955499, and apply run 44464377454 with Apply complete! Resources: 14 added, 3 changed, 0 destroyed. Eleven of the 14 are the capture hosting: the bucket et-uat-capture (force_destroy), its CloudFront distribution and headers policy (dev’s capture CSP, connect-src naming api.uat), the certificate and the capture.uat.crm.empire.lk record. The rest came from #20, #21 and #22, which had waited on dev since 17 Nov: the card-read queue and its DLQ, the price-broadcast-send schedule, and lead research’s timeout. forge-et-iaac#24 (merged later that day) is not on UAT yet.
  • What it serves. https://capture.uat.crm.empire.lk/ answered 403 at 11:03 UTC: the bucket is empty until forge-et-pwa’s deploy-uat first runs, with the next -rc tag.
  • The deploy side. forge-et-pwa#169 (deploy-uat) and forge-et-be#474 (the capture redirect and CORS origin in td.uat.json) merged at 06:28 and 06:27 UTC. forge-et-pwa has a uat Environment (tag rule v*-rc*; svc-github’s key; AWS_REGION, S3_BUCKET=et-uat-capture, CLOUDFRONT_ALIAS=capture.uat.crm.empire.lk, VITE_API_BASE_URL=https://api.uat.crm.empire.lk, VITE_CSRF_COOKIE_NAME=XSRF-TOKEN-UAT), and forge-et-fe’s has CAPTURE_URL=https://capture.uat.crm.empire.lk, which forge-et-fe#582 (open) reads.
  • An effect of the apply. UAT’s backend still runs v0.1.0-rc1, which has no price-broadcast-send handler, so the new schedule’s minute ticks are refused (UnknownScheduledJobKindException in /ecs/uat/backend) and dead-lettered: et-uat-scheduler-dlq held 341 messages at 11:03 UTC. It stops when an -rc cut from today’s main deploys (.forge/owed.md).
  • The uat Environments. Each holds the v*-rc* tag rule, its variables and svc-github’s key (names read 17 Nov).
  • Left. Every throwaway tag, branch and image tag is removed. The ECR tag v0.0.0-rc6 went on 17 Nov through a temporary inline policy on svc-harness, deleted straight after.
  • UAT is now behind dev (read 17 Nov). forge-et-iaac#20, #21 and #22 (the price-broadcast-send schedule, lead research’s 270 s timeout, the card-read queue) merged to dev after UAT’s last apply at b175dae. The lead decided to apply them only with the next -rc tag, whose backend has F37-a wave 1 and F36-b wave 2 (.forge/owed.md, “Bring UAT up to forge-et-iaac dev with the next -rc tag”).

The backend’s dev deploy, in forge-et-be (written by Shyamin Ayesh in commits pushed straight to main with no PR and so no review record: a first build-only workflow on 5 Oct (87d2bfd5), seven trial changes on 8 Oct (f722532e to f31c4630), and the deploy as it stands on 9 Oct, c0f4582e to 03404e8e): the build workflow (.github/workflows/build.yaml), job dev, runs on every push to main. It builds the image in Docker on the runner (docker build --pull --no-cache-filter runtime, no platform set, so linux/amd64), pushes crm/backend:build-<short-sha>, renders .aws/td.dev.json with the image and deploys it to the backend service in et-dev. Its release job (tags v*.*.*) is written but commented out. The task definition .aws/td.dev.json: family backend, X86_64, 0.5 vCPU, 1 GiB, the two roles above, awslogs to /ecs/dev/backend. Its plain environment values are ET_ENVIRONMENT=dev, the business timezone, the RDS host, port and database, ET_SECURITY_LOGIN_REDIRECT_URI=https://dev.crm.empire.lk, the CORS list (https://dev.crm.empire.lk,https://capture.dev.crm.empire.lk), the CSRF cookie domain dev.crm.empire.lk and name XSRF-TOKEN-DEV (AD #33), ET_SECURITY_SEED_ADMIN_EMAIL=crm01@empire.lk, scanning switched off (ET_FILES_SCAN_PROVIDER=none with ET_FILES_SCAN_OFF_ACKNOWLEDGED=true, F38-j), ET_MESSAGING_SCHEDULER_QUEUE=et-dev-scheduler and ECS-format JSON logs; the database username and password are secrets references to the RDS-managed secret. Since 13 Oct 2026 (a further commit, dfc9be15, also pushed straight to main with no PR) the four ET_SECURITY_ENTRA_* values are secrets references to the base/entra-secrets secrets, not environment entries. Placeholders still, as read from the running task definition backend:25 on 14 Oct: ET_FILES_STORAGE_BUCKET, ET_MAILBOX_PUBLIC_BASE_URL and ET_MAILBOX_TOKEN_KEY_ID; ET_MESSAGING_MAILBOX_INGEST_QUEUE is not set. (Re-read 21 Oct 2026, 21:15 UTC — 22 Oct in Sri Lanka: the service runs backend:51, registered 21 Oct at 09:39 UTC, with the same three placeholders and still no ET_MESSAGING_MAILBOX_INGEST_QUEUE; since forge-et-be#280 and #283 it also sets ET_AI_FEATURES_EMAILDRAFT_MODEL (the Haiku 4.5 profile) and ET_MESSAGING_LEADTRIAGEQUEUE=et-dev-lead-triage. No SES value is missing: the backend has no SES sender yet and register §D has no SES key; the backend’s own sending is F38-e’s (F38 FR-26), not F38-d FR-6’s.) (Re-read 22 Oct 2026, 05:05 UTC: the service runs backend:53, registered 22 Oct at 04:40 UTC, which adds ET_AI_FEATURES_ACCOUNTBRIEFING_MODEL (the Haiku 4.5 profile, forge-et-be#296, F13) and is otherwise the same — the three placeholders and no ingest-queue key.) (Re-read 25 Oct 2026: the placeholders are gone. The service runs backend:62, registered by svc-github on 23 Oct at 11:20 UTC from forge-et-be eccf2992 (rollout COMPLETED, one task running). Since forge-et-be#300 (F38-d WI-4.1, merged 22 Oct) it sets ET_FILES_STORAGE_BUCKET=et-dev-attachments, ET_MAILBOX_PUBLIC_BASE_URL=https://api.dev.crm.empire.lk, ET_MAILBOX_TOKEN_KEY_ID = the id of alias/et-dev-mailbox-token and ET_MESSAGING_MAILBOX_INGEST_QUEUE=et-dev-mailbox-ingest. Since forge-et-be#307 (F11-c, 23 Oct) it also sets ET_MESSAGING_LEADRESEARCHQUEUE=et-dev-lead-research and the lead-research model override, and reads ET_SEARCH_EXA_APIKEY from et-dev-exa-api-key. The three AI overrides (email draft, account briefing, lead research) all name the Haiku 4.5 us. profile while Sonnet 5.5 is refused. F38-d FR-6’s values are set; uploads and the mailbox are still to be proved by use.) Result, checked 14 Oct 2026: https://api.dev.crm.empire.lk/actuator/health/liveness answers 200 UP, http:// redirects to https://, and /oauth2/authorization/microsoft redirects to Microsoft with the real tenant (§2.1). (Update 2026-10-23, F38-d WI-4.2: the workflow described above was replaced by forge-et-be#275 (F38-d wave 2, merged 19 Oct) and trimmed by forge-et-be#286 (20 Oct, F38-d spec Rev 3). As on main at 06062de8: it runs on a push to main that changes more than .github/workflows/**, and by hand (workflow_dispatch); a deploy-dev concurrency group that never cancels a run; job build (ubuntu-latest, no AWS credentials) builds the jar on the runner with setup-gradle and uploads it as an artifact; job deploy (ubuntu-24.04-arm, the dev Environment) prints its AWS identity, skips the build when ECR already holds crm/backend:build-<full commit SHA>, otherwise builds the runtime stage only for linux/arm64 with the jar as a build context and pushes it, then renders .aws/td.dev.json with the image’s digest and updates the service without waiting for the rollout and without an image scan of its own (ECR’s scan on push checks each image). The task definition is ARM64. The workflow’s runs since 19 Oct are recorded in “F38-d — the dev deploy pipelines, checked” below, AC-2 and AC-12.)

The CRM frontend’s dev deploy, in forge-et-fe (added 13 Oct 2026 by Shyamin Ayesh in three ci: commits pushed straight to main, 9c78916 to bf4adc8, with no PR and so no review record): the build workflow (.github/workflows/build.yaml), one job, runs on every push to main and on manual dispatch. It installs with pnpm install --frozen-lockfile, builds with VITE_API_BASE_URL and VITE_CSRF_COOKIE_NAME from the GitHub Environment dev’s variables (AD #22, AD #33), then, with the static AWS key from that environment’s secrets, runs aws s3 sync dist/ s3://et-dev-spa --delete and invalidates /* on the distribution. Result: https://dev.crm.empire.lk serves the app, and on 13 Oct Shyamin and the lead each signed in through Microsoft as CRM01 — the first sign-in on dev (register “Start here” step 20). forge-et-pwa has no deploy yet. (Update 2026-10-26, F38-d WI-7.1: it has one since 25 Oct, forge-et-pwa build.yaml from forge-et-pwa#91; architecture.md § How dev is deployed today, and “F38-d — the dev deploy pipelines, checked”, AC-5.) (Update 2026-10-19, F38-d WI-1.1, DR-34: the single aws s3 sync … --delete above is how the workflow was first written; since forge-et-fe#259 (14 Oct, BUG-136) it syncs in stages — hashed assets with max-age=31536000, immutable, then the other files and the HTML last with max-age=0, must-revalidate, then a --delete --size-only prune — in a concurrency group that never cancels a run. Read at 32c0dea5.) (Update 2026-10-23, F38-d WI-4.2: since forge-et-fe#298 (F38-d wave 2, merged 19 Oct) the workflow has two jobs. build (no AWS credentials; the dev Environment for its two VITE_* variables only) first checks that neither value is blank (check::config), then installs, builds and uploads dist as the artifact web-dist; deploy downloads it, prints its AWS identity and runs the staged sync and the invalidation. A push that changes only .github/workflows/** starts no run, and #298’s own merge started none. Its runs since 19 Oct are recorded in “F38-d — the dev deploy pipelines, checked” below, AC-1 and AC-12.)

Not there yet, as read in the account on 14 Oct 2026 and again on 21 Oct (Shyamin Ayesh’s own list of 9 Oct 2026 is marked †): narrowing the CI user’s IAM † (withdrawn for dev, F38 Rev 9); the backend task definition’s values for uploads and the mailbox (on 21 Oct forge-et-be .aws/td.dev.json, and the running backend:53 on 22 Oct, still hold the placeholders ET_FILES_STORAGE_BUCKET, ET_MAILBOX_TOKEN_KEY_ID and ET_MAILBOX_PUBLIC_BASE_URL — the last is the API’s own address, not a new resource — and has no ET_MESSAGING_MAILBOX_INGEST_QUEUE; F38-d FR-6, its WI-4.1) (update 2026-10-25: done — set by forge-et-be#300 on 22 Oct, read in backend:62; the backend deploy paragraph above); the capture app’s S3 + CloudFront and the capture.dev.crm.empire.lk record (update 2026-10-22, F38-b WI-5.1: written by the delivery team as forge-et-iaac#5, awaiting Shyamin Ayesh’s review — the PR list below) (update 2026-10-22, later: done — #5 merged and applied the same day; the capture-app bullet of “What it holds” above. Still missing for the capture app: forge-et-pwa’s dev Environment and its deploy workflow, F38-d wave 5) (update 2026-10-25: the dev Environment is complete — the delivery team’s two VITE_* variables, and the deploy key’s AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY, added by Shyamin Ayesh on 23 Oct at 11:49 UTC (gh api repos/altrium-io/forge-et-pwa/environments/dev/secrets, names only). Only the deploy workflow is missing: forge-et-pwa has no build workflow yet, F38-d WI-5.1) (update 2026-10-26: done — forge-et-pwa build.yaml (f7d6b3f, F38-d wave 5) deployed on 25 Oct, by hand (run 40224118047) and on a push to main (run 40228839882); the host answers 200) (update 2026-10-26, F38-b WI-4.1: checked — the root, /index.html and a deep path serve the same build with the four security headers, the bucket refuses a direct request, and the certificate and aa answer are in place; “F38-b — the dev estate, checked”, AC-7); AWS Budgets (none in the dev account, aws budgets describe-budgets, 25 Oct), CloudTrail, account S3 Block Public Access, an account password policy, the alarms and SNS topic (update 2026-10-25: the topic et-dev-alerts and the failed-deploy rule exist since forge-et-iaac#7, the base/alerts bullet above; still no CloudWatch alarm), and SES (still in the sandbox, no identities) (update 2026-10-21, F38-a WI-3.1: SES is done for dev — the identity dev.crm.empire.lk is verified and production access was granted on 21 Oct, declared in forge-et-iaac at 0eee623; “F38-a — the development account, checked”, AC-5). Done since 9 Oct: the Entra values in Secrets Manager (13 Oct), the AgentCore Gateway (13 Oct), the CRM frontend’s hosting and record (13 Oct), and, from the delivery team’s forge-et-iaac#3, applied on 20 Oct: the attachments bucket †, the storage and mailbox token KMS keys † (the account had AWS-managed keys only), the mailbox ingest queue † and its dead-letter queue (no job queue — F38 Rev 16), EventBridge Scheduler with its schedules and role †, the task role’s grants for them, and RDS rotation turned off (keeping running tasks working across an RDS password rotation †, replaced by F38 Rev 14) — each read in the account on 21 Oct in “F38-b — the dev estate, checked” below. Both halves of the backend’s cloud startup check (register A-1) are on main, and the backend booted on dev with scanning switched off. GuardDuty is not built at all — neither Malware Protection nor threat detection — by Altrium management’s decision of 9 Oct 2026 (agenda item 8), so scanning stays switched off in every environment.

PRs, as read on 19 Nov 2026, about 17:30 UTC: five open — #17, Senuthmi Thimansa’s overlays/prod (F38-i wave 4, approve at bba8a28, held until production is needed); #25, F36-c’s research queues (Dinil); #26, dev.yaml validating every overlay (the delivery team, the lead’s decision); #27, the capture app’s CSP allowing blob: images (Nanndish); and #28, the delivery team’s fixes from the 14 Nov review of dev (below). #20 to #24 merged between 17 and 19 Nov, each with an approve record; #20 and #24 show no GitHub approval, and #21 and #22 were approved by Senuthmi Thimansa rather than Shyamin Ayesh (on 13 Nov, after 06:30 UTC: one open — #17, at 814beb1; on 12 Nov: #11 and #17 open, #19 merged that day; on 3 Nov, after 07:03 UTC: three open — #11, F38-h’s waves 1 and 3, ready for review and not merged, now also carrying F43’s 6596205; #15, F38-i’s wave 1, ready for review, not merged; and #16, F43’s wave 3, ready for review, not merged; on 29 Oct, after 05:52 UTC: #11 and #15 open, #13 and #14 merged into dev on 28 Oct; earlier on 29 Oct, after 05:07 UTC: #11 open; before that: #11 and #12 open, #12 a draft whose base is #11’s branch; as read on 27 Oct: #11, #12, #13 and #14 open; on 26 Oct in the evening: #11 open, the branch feature/F38-h-wave-3 pushed with no PR; on 26 Oct in the afternoon: #11 open as a draft; in the morning, after #9 merged: none open; on 25 Oct: #9 open; on 22 Oct after #5 merged: none open) (re-check with gh pr list -R altrium-io/forge-et-iaac); the remote branches on 19 Nov are dev and the five open PRs’ heads — feature/F38-i-prod-overlay, feature/F36-c-research-queues, ci/validate-every-overlay, fix/capture-csp-blob-images and fix/ecs-task-tags-and-agentcore (git -C $WS/forge-et-iaac branch -r):

  • PR #1 (the switch to us-east-1) and PR #2 (an ECR module with an et-be repository), both by Senuthmi Thimansa — closed by 17 Oct; both were overtaken by what was pushed to dev (the region switch is applied, and ECR was added as crm/backend).

  • forge-et-iaac#3 (F38-b wave 1), opened on 17 Oct by the delivery team from feature/F38-b-wave-1, approved by Shyamin Ayesh on 20 Oct (11:06 UTC), merged into dev by the lead the same day (18:33 UTC) as d230b133, and applied by its dev.yaml run 39475850266 (the bullets above say what it changed; its last commit, a359b61, was pushed after the approval, and the PR says so). It adds the storage and mailbox token KMS keys, the attachments bucket et-dev-attachments, the mailbox ingest queue and its dead-letter queue, the four EventBridge schedules with their schedule group and role, and the backend task role’s SQS, S3, KMS and Bedrock grants. It runs the ECS service in us-east-1b, the database’s zone, and outputs the three values for the task definition (F38-b FR-9). It also removes 76fce06’s rejected master_user_secret_rotation_days line from base/rds, so terraform validate passes again, and turns the RDS-managed secret’s automatic rotation off (base/rds/rotation.tf, rotation_enabled = false — F38 Rev 14). Added 17–18 Oct (F38-b spec Revs 3–4): it restores the apply step, with dev.yaml running only on a push to dev or by hand and applying only on dev, plus separate apply and plan concurrency, a read-only token, a timeout, a fmt check, a read-only lock file and checkout@v4; the storage key usable only through S3 (kms:ViaService); a .gitignore; the committed provider lock file; 14-day dead-letter retention; a README.md and a CLAUDE.md; and module options for a uat instance in the dev account (defaults are dev’s values). Added 18 Oct (F38-b spec Rev 7, e35554d): a CloudFront response-headers policy on the CRM’s distribution — HSTS, X-Content-Type-Options, X-Frame-Options: SAMEORIGIN and a referrer policy, no CSP. Added 20 Oct (a359b61, the lead’s request; F38-d spec Rev 3): the backend target group’s health check every 10 s, a 5 s timeout, two healthy and three unhealthy checks, instead of the ALB defaults (30 s, five healthy), so a new task takes traffic in about 20 s instead of about 2.5 minutes (aws elbv2 describe-target-groups --names et-dev-backend, 21 Oct: a 10 s interval, a 5 s timeout, two healthy and three unhealthy checks). Its branch commits carry the skip-CI marker, which the old workflow needed; the merge commit does not, so the push to dev ran the apply. (Update 2026-10-21, F38-b WI-2.1: what the apply created, the three FR-9 values and the merge run are recorded in “F38-b — the dev estate, checked” below.)

  • forge-et-iaac#4 (F11-d), opened on 20 Oct by the delivery team from feature/F11-d-lead-triage-queue, stacked on #3 (its base was #3’s branch), approved by Shyamin Ayesh and merged by the lead into #3’s branch on 20 Oct 2026 (59b43d2), a minute before #3 merged, so it reached dev with #3 and was applied by the same run; it needs no further merge. It adds the et-dev-lead-triage queue and its dead-letter queue, the task role’s send and consume rights on it, the lead_triage_queue_name output, and a fifth schedule, lead-triage-sweep at rate(1 hour). The task-definition key ET_MESSAGING_LEADTRIAGEQUEUE = et-dev-lead-triage came with forge-et-be#283, merged and deployed as backend:44 the same evening; the hourly sweep has run on dev, and a lead created on dev is still to be checked (register §B F11-d rows 1 and 2).

  • forge-et-iaac#5 (F38-b wave 5), opened on 22 Oct by the delivery team from feature/F38-b-wave-5, approved by Shyamin Ayesh and merged into dev by the lead with a merge commit on 22 Oct 2026 (5e078c0d, 07:22 UTC); dev.yaml run 39740047356 planned only additions (11) and applied them (the capture-app bullet of “What it holds” above). It adds the capture app’s dev hosting from the CRM’s modules, with name = "capture": the certificate for capture.dev.crm.empire.lk, the bucket et-dev-capture, its CloudFront distribution with its own origin access control and the same security headers as the CRM’s, and the alias record in the crm.empire.lk zone, plus the outputs capture_bucket_id, capture_cloudfront_distribution_id and capture_cloudfront_domain_name (F38-b spec FR-7, Rev 8). No module changes. The PR asks for the branch to be planned before the merge (dev.yaml applies straight after its plan on dev) and for a merge commit; its commits carry no skip-CI marker, which the current dev.yaml does not need. Until F38-d’s capture deploy uploads a build (its WI-7.1), the host answers with an error. (Update 2026-10-26, F38-d WI-7.1: the capture app deploys since 25 Oct and the host answers 200; “F38-d — the dev deploy pipelines, checked”, AC-5.)

  • forge-et-iaac#7 (F38-e wave 1, WI-1.1), opened on 23 Oct by the delivery team from feature/F38-e-wave-1 (head e68a7b6); approved by Shyamin Ayesh on 23 Oct (09:59 UTC), merged into dev by the lead (1590009, 10:41 UTC) and applied by dev.yaml run 39940862972 (“Plan: 8 to add, 0 to change, 0 to destroy”: the topic, its policy, four subscriptions, the rule and its target). The base/alerts bullet above has what it created, as read on 25 Oct, all four subscriptions confirmed. Shyamin’s message approving #6, #7 and #8 also asked for one Secrets Manager module (.claude/rules/forge-et-iaac.md). It adds a new module, base/alerts, called from overlays/dev: the SNS topic et-dev-alerts, not encrypted (it carries alert text only, and CloudWatch and EventBridge cannot publish to a topic under the AWS-managed aws/sns key — F38-e WI-1.1 plan D-2); one email subscription for each of the four developers (alert_email_recipients in env.auto.tfvars); a topic policy whose one statement lets only cloudwatch.amazonaws.com and events.amazonaws.com publish, when aws:SourceAccount is the dev account; and the EventBridge rule et-dev-backend-deployment-failed on SERVICE_DEPLOYMENT_FAILED for et-dev/backend, targeting the topic through an input transformer that sends the service, the deployment id, ECS’s reason and the time as labelled lines; plus the output alert_topic_arn and the alerts row of README.md (F38-e spec FR-4, FR-5). No existing module changes, and no log source, alarm or metric filter is added. fmt -check -recursive, init -backend=false -lockfile=readonly and validate exit 0 on its head (Terraform 1.12.1 in Docker). The PR asked for a merge commit; its commit carries no skip-CI marker. Each of the four recipients had to confirm their subscription from AWS’s email, because SNS deletes an unconfirmed one after a few days; all four had by 25 Oct.

  • forge-et-iaac#8 (F11-c), opened on 23 Oct 2026 by the delivery team from feature/F11-c-lead-research-queue, cut from dev at 5e078c0d; approved by Shyamin Ayesh on its final head 2a21a5ae (10:00 UTC), merged into dev by Dinil (d86d2ca9, 10:07 UTC) and applied by dev.yaml run 39936834361 (“Apply complete! Resources: 4 added, 3 changed, 0 destroyed”). Read as svc-harness the same day: the queue et-dev-lead-research has a 240 s visibility timeout and redrives to et-dev-lead-research-dlq after 3 receives; the schedule lead-research-sweep is rate(1 hour) and ENABLED; the secret et-dev-exa-api-key (ARN suffix -GSWUGX, the AWS-managed key) was created with no value. Its value is set by hand: IAM’s policy simulator allows secretsmanager:PutSecretValue to the dev account’s SSO ADMINISTRATORS role (AdministratorAccess) and denies it to svc-harness. Dinil set the value on 23 Oct 2026 (16:05 Colombo time, one version, AWSCURRENT), as svc-harness under a temporary inline policy temp-set-exa-key allowing only PutSecretValue on this secret, which he added and removed the same afternoon with the lead’s agreement (no SSO administrator was available); CloudTrail records svc-harness as the actor. Read afterwards: svc-harness holds only its spike-invoke-gateway inline policy and is denied PutSecretValue again, and the backend’s execution role et-dev-backend-execution is allowed GetSecretValue on the secret. The same shape as #4: the et-dev-lead-research queue and its dead-letter queue (maxReceiveCount 3, 14-day retention), with a 240 s visibility timeout rather than 30 s because one research run can take about 110 s; the task role’s send and consume rights on it; the lead_research_queue_name output for ET_MESSAGING_LEADRESEARCHQUEUE; and the lead-research-sweep schedule at rate(1 hour). It also adds base/search-secrets (the empty secret, read by the task’s execution role) and the exa_api_key_secret_arn output. The task-definition lines — the queue name, the Haiku 4.5 override and the ET_SEARCH_EXA_APIKEY reference — came with forge-et-be#307, merged on 23 Oct 2026 (eccf2992) and deployed as backend:62; a lead created on dev that afternoon was researched in 6.6 s (.forge/cloud-gated-validation.md §B F11-c). (Update 2026-11-17: the queue’s visibility timeout is now 270 s — forge-et-iaac#21, next bullet.)

  • forge-et-iaac#21 (F36-c), opened on 16 Nov 2026 by the delivery team from feature/F36-c-WI-1.2-research-visibility; approved by Senuthmi Thimansa on 17 Nov (05:00 UTC) and merged into dev by Dinil (41322061, 05:09 UTC), by his decision without Shyamin Ayesh’s approval, which .claude/rules/forge-et-iaac.md asks for (a PR comment records it); applied by dev.yaml run 44022112490 (“Apply complete! Resources: 0 added, 1 changed, 0 destroyed”; its plan was the one in-place change visibility_timeout_seconds 240 → 270 on module.sqs.aws_sqs_queue.lead_research). It raises et-dev-lead-research’s visibility timeout from 240 s to 270 s, because F36-c’s plan-research searcher shares the search rate limiter and lead research’s longest run becomes about 181 s at the 70 s AI pacing cap (F36-c WI-1.2 plan Rev 1). Read as svc-harness the same day: VisibilityTimeout 270.

  • forge-et-iaac#6 (F38-d wave 3, WI-3.1), opened on 22 Oct by the delivery team from feature/F38-d-wave-3; approved by Shyamin Ayesh on 23 Oct (09:59 UTC), merged into dev by the lead (acd1cf7, 10:41 UTC) and applied by dev.yaml run 39940853448, whose plan was the two in-place updates the PR named and nothing else (“Plan: 0 to add, 2 to change, 0 to destroy”; “Apply complete! Resources: 0 added, 2 changed, 0 destroyed”). It makes crm/backend’s tags immutable and sets the backend target group’s deregistration delay to 30 s (the ECR and ALB bullets above, read 25 Oct). The next backend deploy of a commit whose image ECR already holds is still to be seen (F38-d WI-6.1).

  • forge-et-iaac#9 (F38-b wave 6, WI-6.1), opened on 25 Oct 2026 by the delivery team from feature/F38-b-wave-6 (head df13676, cut from dev at 1590009); approved by Shyamin Ayesh on 26 Oct (04:00 UTC), merged into dev by the lead (c728e06b, 04:52 UTC) and applied by dev.yaml run 40271292253, whose plan was the five moves and nothing else (“Plan: 0 to add, 5 to change, 0 to destroy”; “Apply complete! Resources: 0 added, 5 changed, 0 destroyed”; the base/secrets bullet above has the reads afterwards). No plan of the branch was read before the merge: the plan-only manual run was not started, and the lead merged on Shyamin’s approval after the moved blocks, names and recovery window were checked against dev’s code by reading it. The branch was deleted after the merge. The change Shyamin asked for that day: one module, base/secrets, creates the five secrets from one list with for_each and prevent_destroy, no value in Terraform; base/entra-secrets and base/search-secrets are removed; overlays/dev/moved.tf moves each secret to its new address, so none is destroyed; the execution role gets the same six ARNs in the same order; README.md and CLAUDE.md follow (a second commit, df13676, also corrects CLAUDE.md’s “The DevOps team merges” to the rule of 21 Oct, which Shyamin may drop). The expected plan, quoted in the PR: five secrets “will be updated in-place” “(moved from …)”, only the Module tag changing, 0 to add, 5 to change, 0 to destroy; any destroy or create of a secret, or a change to the execution role’s policy, means do not merge (WI-6.1 plan D-7). fmt -check, init -backend=false -lockfile=readonly and validate exit 0 on its head (Terraform 1.12.1 in Docker); gitleaks clean; the push ran no workflow (F38-b spec FR-12, AC-13, Rev 11).

  • forge-et-iaac#10, by the delivery team, changes only README.md (stale statements corrected); merged into dev by the lead on 26 Oct (5401a88, 06:55 UTC); its dev.yaml run, 40282639877, applied nothing to change.

  • forge-et-iaac#11 (F38-h wave 1, WI-1.1), opened on 26 Oct 2026 by the delivery team from feature/F38-h-wave-1 (head 4bab623, cut from dev at 5401a88), ready for review since 26 Oct (Shyamin Ayesh, Senuthmi Thimansa and Dinil asked), not merged. Its plan-only manual dev.yaml run, 40332535104 at 4bab623, read No changes. Your infrastructure matches the configuration. with the apply step skipped (WI-1.1 plan D-10), and read-only AWS CLI listings before and after it found nothing tagged Environment = UAT. It adds three options for a destroyable uat instance, each defaulting to dev’s value, with overlays/dev unchanged: base/s3’s force_destroy (default false), base/kms’s deletion_window_in_days (default 30, 7 to 30 allowed), and base/ecs’s bootstrap_task_definition (default false: a placeholder first revision in the service’s family — busybox 1.37.0 pinned by digest, no task role or secrets, nothing listening on 8080 — and the service created with no running task). The one change in how dev is managed: the ECS service now ignores desired_count as well as task_definition, so Terraform no longer corrects dev’s running count (WI-1.1 plan D-2). With the defaults nothing is created in AWS (the lead’s constraint of 26 Oct: code only until Empire agrees to a UAT environment). fmt -check, init -backend=false -lockfile=readonly and validate exit 0 on its head (Terraform 1.12.1 in Docker); gitleaks clean; the push ran no workflow. (Update 2026-10-29: #11 now carries waves 1 and 3. The lead merged #12 into this branch (1c09397, 05:07 UTC), whose tree is identical to #12’s checked head e78b12e. #11’s title became feat(F38-h): the temporary uat instance — base/ options, the uat overlay and its apply/destroy workflow, and its body was rewritten for both waves. It is still ready for review with no approval. The branch is 4 commits behind dev (#13 and #14) and merges into it cleanly (git merge-tree). The plan run above predates wave 3’s files, so a new one is owed on the final head, with the lead’s approval. Before the merge, UAT’s Haiku 5.5 lines and its CSP are settled: .forge/owed.md, “Finish forge-et-iaac#11, F38-h’s waves 1 and 3”.) (Update 2026-10-29, later: the head is 0742cdd, 0 commits behind dev, after three commits by the delivery team at the lead’s request:

    • 0b52172 merges origin/dev, with no conflicts.
    • c1ef354 gives overlays/uat Haiku 5.5’s Global profile and model in bedrock_invoke_arns, and the CRM’s Content-Security-Policy built from https://api.uat.crm.empire.lk, passed to spa_cloudfront.
    • 0742cdd moves dev.yaml and uat.yaml to checkout@v7, configure-aws-credentials@v6 and setup-terraform@v4 (DR-31).
    • 6596205 (2026-11-03, F43-WI-3.1) gives overlays/uat the three Bedrock routes it lacked — global. Haiku 4.5, us. Sonnet 5.5 and us. Haiku 5.5 — so its bedrock_invoke_arns is the same list as dev’s after forge-et-iaac#16 (F43 spec FR-R2): each model granted as its two profiles and one foundation-model wildcard, Haiku 4.5’s three regional ARNs replaced by its wildcard. The push ran no workflow; a comment on #11 names the commit.

    Checks, attempt 2 on 0742cdd: fmt -check -recursive, init -backend=false -lockfile=readonly and validate in both overlays, actionlint 1.7.7 on both workflows, the lock files’ cmp, and gitleaks 8.30.1 over origin/dev..HEAD, each EXIT=0. The pushes ran no workflow. A review comment on the PR, of the same day, lists every finding of a review of the PR, of dev’s Terraform, of the dev account (read only) and of the app repos’ deploy files, and where each went. Still owed before merge: a dev.yaml plan run of 0742cdd with the lead’s approval, the lead’s comment approving the two workflow changes, and Shyamin Ayesh’s approval.) (Update 2026-10-29, later still: the plan run was made with the lead’s approval. Manual dev.yaml run 40825713309 on 0742cdd concluded success, and its terraform::plan reads No changes. Your infrastructure matches the configuration. terraform::apply was skipped. Its new action versions all ran, and the Node.js 20 warning is gone. Afterwards, read as svc-harness: nothing is tagged Environment = UAT, no task-definition family starts with et-uat, and backend is still on backend:85 with 1 task desired and 1 running. Still owed: the lead’s comment approving the two workflow changes, and Shyamin Ayesh’s approval.) (Update 2026-11-04: b47a33e merges origin/dev (b8489c6, with #15 and #16), with no conflict; fmt, init -backend=false + validate in both overlays and gitleaks over origin/dev..HEAD each EXIT=0, named in a PR comment.) (Update 2026-11-06: f66b7d8 passes availability_zone = var.availability_zone (us-east-1b) to overlays/uat’s module "rds", so UAT’s database is created in the backend tasks’ zone, using the input #15 added to base/rds; fmt -check -recursive, init -backend=false + validate in overlays/uat, git diff --check and gitleaks each EXIT=0; named in a PR comment. The head is f66b7d8, 0 commits behind dev, with no review yet. The plan run of 29 Oct was of 0742cdd, so a new plan-only dev.yaml run of the head comes before the merge, with the lead’s approval; then the lead’s comment on the two workflow changes, and Shyamin Ayesh’s approval. On the DevOps Trello board since 5 Nov as “[ET] Review forge-et-iaac#11”, In Review.) (Update 2026-11-11: 098695b (F38-h spec Rev 7) removes the placeholder task definition — base/ecs/bootstrap.tf, the busybox image and bootstrap_task_definition — after Shyamin Ayesh asked why Terraform writes a task definition when forge-et-be’s .aws/ holds them. base/ecs gains create_service (default true), with a moved block that keeps dev’s service in place as aws_ecs_service.backend[0], and service.tf is otherwise back to dev’s (desired_count = 1, ignore_changes = [task_definition]), so dev’s management is unchanged; overlays/uat gains a create_service variable and uat.yaml a service input, both ticked by default. Each round applies without the service, deploy-uat registers the revision from td.uat.json, and a second apply creates the service. fmt -check -recursive, init -backend=false + validate in both overlays, actionlint on both workflows, the lock files’ cmp, git diff --check and gitleaks over origin/dev..HEAD each EXIT=0; the push ran no workflow. e1edc3d adds a warning to the service input’s description. /forge-review-pr with iaac-reviewer posted an approve record for e1edc3d the same day (comment): no Blocker or Important; the null service_name with the service off is deferred to .forge/owed.md’s production first-apply row. The lead’s plan-only dev.yaml run of e1edc3d (run 43021956495, 2026-11-11) read module.ecs.aws_ecs_service.backend moved to module.ecs.aws_ecs_service.backend[0] and Plan: 0 to add, 0 to change, 0 to destroy., with terraform::apply skipped; a resourcegroupstaggingapi get-resources read afterwards found no resource tagged Environment=UAT in the dev account. The lead approved the two workflow changes with a comment the same day. Still owed: Shyamin Ayesh’s approval.) (Update 2026-11-13: merged. Shyamin Ayesh approved e1edc3d on GitHub (06:07 UTC) and Senuthmi Thimansa merged #11 into dev with a merge commit (b175dae, 06:08 UTC). Its dev.yaml run 43376614685 read the one move (module.ecs.aws_ecs_service.backend to [0]), Plan: 0 to add, 0 to change, 0 to destroy. and “Apply complete! Resources: 0 added, 0 changed, 0 destroyed.” What it added is in “What it holds” above, and the uat instance Senuthmi applied afterwards is under “The uat instance” below it. The branch feature/F38-h-wave-1 is merged and still on origin (forge-et-iaac does not delete a branch on merge), to be deleted by hand.)

  • forge-et-iaac#12 (F38-h wave 3, WI-3.1), from feature/F38-h-wave-3 (pushed on 26 Oct 2026 by the delivery team, head e78b12e, cut from feature/F38-h-wave-1 at 4bab623, because its overlay sets the three options #11 adds), opened on 27 Oct 2026 by the delivery team as a draft with base feature/F38-h-wave-1, so its diff is wave 3’s eleven files only; not ready for review and not merged. A draft PR replaces the branch with no PR, so the work is visible and /et-cleanup keeps both branches (the lead’s decision of 27 Oct; F38-h WI-3.1 plan ## Notes). Once #11 has merged, the PR’s base is changed to dev before feature/F38-h-wave-1 is deleted (deleting a PR’s base branch closes the PR; delete_branch_on_merge is off in this repository), dev is merged into the branch, and the checks and the lead-approved dev.yaml plan run follow before it is marked ready (.forge/owed.md, “Close F38-h’s wave 1 and take wave 3 to its PR”). It adds overlays/uat, the temporary uat instance — dev’s modules and sizes with et-uat-* names, Environment = UAT on every resource, state at et-ops uat/terraform.tfstate, the three options on, records through role/dns-uat (the DevOps team’s role to create, .forge/owed.md), dev’s five secrets and SES identity and configuration set read by name; no secrets, ses, ecr, agentcore or alerts module and no capture-app hosting — and .github/workflows/uat.yaml, which plans, applies or destroys it, started by hand on dev only, under svc-github’s key (the dev Environment); plus README.md and CLAUDE.md. base/, overlays/dev and dev.yaml are unchanged, so nothing is applied to dev by it, and uat.yaml cannot be dispatched until it is on dev. fmt -check, init -backend=false -lockfile=readonly and validate in both overlays, actionlint on uat.yaml, and the lock-file cmp exit 0 on its head (Terraform 1.12.1 and actionlint 1.7.7 in Docker); gitleaks clean; neither the push nor the draft PR ran a workflow (this repository has no pull-request trigger). Its dev.yaml plan, with the lead’s approval, comes after dev is merged in. (Update 2026-10-29: #13 and #14 are now on dev. Its overlays/uat/locals.tf copies dev’s bedrock_invoke_arns from before #13, so Haiku 5.5’s two lines are added there by hand, and it does not set #14’s content_security_policy, so UAT’s CRM would have no CSP until it does — .forge/owed.md, “Finish forge-et-iaac#11, F38-h’s waves 1 and 3”, step 1.) (Update 2026-10-29, later: #12 is merged into feature/F38-h-wave-1 by the lead (1c09397, 05:07 UTC), not into dev. Its base was never changed; its work goes to dev with #11 (the #11 bullet above). feature/F38-h-wave-3 has been deleted from the remote.)

  • forge-et-iaac#13 (F38-b FR-11; .forge/owed.md’s former row “Allow the backend’s task role to call Haiku 5.5”, closed on 2026-11-06 — engagement-log.md), opened on 27 Oct 2026 by the delivery team from feature/F38-b-haiku-5-5-grant (head 3c0ddaf, cut from dev at 5401a88); merged into dev by the lead on 28 Oct (a77370d, 05:04 UTC; GitHub shows no approving review on it; Shyamin Ayesh approved it outside GitHub, the lead confirmed on 2026-11-06) and applied by dev.yaml run 40635870998, whose plan was the expected one (“Plan: 0 to add, 1 to change, 0 to destroy”: module.ecs.aws_iam_role_policy.task_bedrock updated in place); the policy as read afterwards is in the base/ecs bullet above. It adds Haiku 5.5’s Global profile global.anthropic.claude-haiku-5-5 and arn:aws:bedrock:*::foundation-model/anthropic.claude-haiku-5-5 to overlays/dev/locals.tf’s bedrock_invoke_arns, the same shape as Sonnet 5.5’s; aws bedrock get-inference-profile read the profile as ACTIVE, routing to the Region-less model ARN and us-east-1’s. The expected plan is an in-place update of the task role’s inline policy bedrock and nothing else. Read before it as svc-harness: et-dev-backend-task’s bedrock policy allows bedrock:InvokeModel on Haiku 4.5’s us. profile and its three Regions’ models and on Sonnet 5.5’s Global profile and model, nothing else. fmt -check, init -backend=false and validate exit 0 on its head (Terraform 1.12.1 in Docker); gitleaks clean; the push ran no workflow, and no plan-only dev.yaml run was started.

  • forge-et-iaac#14 (F38; .forge/owed.md, “A Content-Security-Policy on both frontends”), opened on 27 Oct 2026 by the delivery team from feature/F38-frontend-csp (head f4f10d4); approved by Shyamin Ayesh, merged into dev by the lead on 28 Oct (876d634, 16:18 UTC) and applied by dev.yaml run 40724734100 (“Plan: 0 to add, 2 to change, 0 to destroy”: module.spa_cloudfront and module.capture_cloudfront’s aws_cloudfront_response_headers_policy.security, updated in place; “Apply complete! Resources: 0 added, 2 changed, 0 destroyed”). base/cloudfront gains an optional content_security_policy (default null, which sends none), and overlays/dev/locals.tf builds one policy per app from api_origin. Read on 29 Oct with curl -sI on https://dev.crm.empire.lk/, https://dev.crm.empire.lk/leads and https://capture.dev.crm.empire.lk/, each with the four headers of #3 and:

    • CRM: default-src 'self'; script-src 'self'; style-src 'self' 'unsafe-inline'; img-src 'self' data:; font-src 'self'; connect-src 'self' https://api.dev.crm.empire.lk; frame-src 'self'; frame-ancestors 'self'; form-action 'self'; base-uri 'none'; object-src 'none'
    • Capture app: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self'; font-src 'self'; connect-src 'self' https://api.dev.crm.empire.lk; worker-src 'self'; manifest-src 'self'; frame-ancestors 'none'; form-action 'self'; base-uri 'none'; object-src 'none' (Update 2026-11-19: this img-src 'self' blocks the card photo, which the capture app shows from a blob: URL; the photo shows its alt text on capture.dev.crm.empire.lk. forge-et-iaac#27, open and reviewed approve, changes the capture app’s img-src to 'self' blob: in overlays/dev and overlays/uat. It is not merged or applied yet. UAT gets it only from a uat.yaml apply, and production only if #17 carries the same line; .forge/owed.md, “Allow blob: images in the capture app’s CSP”.)

    Loaded the same day in the browser, without signing in: the CRM’s sign-in page renders with Inter loaded, its call to https://api.dev.crm.empire.lk answers 401 as expected for no session, and the console holds no CSP violation; the capture app renders, its service worker is activated and controls the page, and the console is empty. The signed-in checks (HTML email in the inbox — register §B F20-c item (e) — attachments, the editor, a price import, dialogs, mailbox connect) and the capture app’s offline reload are still to do (.forge/owed.md). The capture app’s responses still carry X-Frame-Options: SAMEORIGIN beside frame-ancestors 'none'; browsers that support CSP follow frame-ancestors, so the stricter rule applies.

  • forge-et-iaac#16 (F43 wave 3, F43-WI-3.1), opened on 2026-11-03 by the delivery team from feature/F43-wave-3 (head 2b1a96c, cut from dev at 876d634), approved by Shyamin Ayesh, merged on 2026-11-04 at b8489c6 (17:58 UTC), and applied by its dev.yaml run 41882320196: “Apply complete! Resources: 0 added, 1 changed, 0 destroyed”. It adds to overlays/dev/locals.tf’s bedrock_invoke_arns the inference profiles global.anthropic.claude-haiku-4-5-20251001-v1:0, us.anthropic.claude-sonnet-5-5 and us.anthropic.claude-haiku-5-5, and arn:aws:bedrock:*::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0 in place of Haiku 4.5’s three regional foundation-model ARNs, so the task role may call both routes of all three planned models, each granted in the same shape (two profiles and one wildcard) (F43 spec FR-R2; .forge/owed.md, “Allow the backend’s task role to call Haiku 5.5”). Its plan-only manual dev.yaml run, 41592977282, read Plan: 0 to add, 1 to change, 0 to destroy (module.ecs.aws_iam_role_policy.task_bedrock, in place), with the apply step skipped. Checks on 2b1a96c: fmt -check -recursive, init -backend=false -lockfile=readonly and validate in overlays/dev, and gitleaks 8.30.1 over origin/dev..HEAD, each EXIT=0. Register §B F43 item 4.

  • forge-et-iaac#15 (F38-i wave 1, WI-1.1), opened on 29 Oct 2026 by the delivery team from feature/F38-i-wave-1 (head 29c02a7, cut from dev at 876d634), approved by Shyamin Ayesh on 4 Nov 2026 (08:06 UTC, on head 29c02a7), merged into dev by the lead on 4 Nov (37b7210, 17:54 UTC) and applied by dev.yaml run 41881575399: No changes. Your infrastructure matches the configuration., “Apply complete! Resources: 0 added, 0 changed, 0 destroyed.” A review comment of 4 Nov on the PR found no Blocker or Important; its three Nits for production’s overlay are in .forge/owed.md. Before the merge, it was ready for review from 29 Oct: its plan-only manual dev.yaml run, started with the lead’s approval (run 40829800911, on head 29c02a7, conclusion success, terraform::apply skipped), read No changes. Your infrastructure matches the configuration. (WI-1.1 plan D-10), after read-only re-reads of dev’s database and its security group matched the plan; review was then requested from Shyamin Ayesh, Senuthmi Thimansa and Dinil Gamage, and it merges after Shyamin Ayesh’s approval. It adds the module options production needs, each defaulting to dev’s value, with overlays/dev unchanged: base/networking’s private_subnet_cidrs (default []: private subnets in the public subnets’ zones, with their own route table holding no route, no NAT gateway) and the output private_subnet_ids; base/security-groups’s database_ingress_from_backend_only (default false: the database admits 5432 from the VPC range with dev’s descriptions; true admits only the backend’s group); and base/rds’s storage_type ("gp2"), max_allocated_storage (null), backup_retention_period (0), deletion_protection (false), skip_final_snapshot (true; when false, a final snapshot <project>-<environment>-final), storage_encrypted (false; true uses the AWS-managed aws/rds key, no kms_key_id) and availability_zone (null); plus the three rows of README.md’s module table. storage_encrypted and availability_zone would replace the database if they ever differed from dev’s state, so any module.rds line in the plan stops the PR. fmt -check, init -backend=false -lockfile=readonly and validate exit 0 on its head (Terraform 1.12.1 in Docker); gitleaks clean; the push ran no workflow. Nothing is created in production by it: F38-i WI-4.1’s overlays/prod sets the options.

  • forge-et-iaac#17 (F38-i wave 4, WI-4.1), opened on 5 Nov 2026 by Senuthmi Thimansa from feature/F38-i-prod-overlay (head da7a983 on 6 Nov): feat(F38-i): add the production overlay, overlays/prod as code only. It was written from the Trello production card rather than F38-i’s spec. The delivery team’s review comment of 5 Nov (issuecomment-6657172880) lists what it must change to match the spec. The PR is adopted as wave 4’s implementation, and the list it must reach is WI-4.1’s plan (.forge/plans/F38-i/F38-i-WI-4.1-plan.md, D-1, draft on 6 Nov). No approval yet. Merging it creates nothing, because dev.yaml applies only overlays/dev. On the DevOps Trello board since 5 Nov as “[ET] Review forge-et-iaac#17”, In Review. Update 2026-11-09: at Tharinda’s request the delivery team merged dev into the branch (9919a59, bringing #18) and added the lead-ranking queue to overlays/prod (897752d: the queue in consumed_queue_arns and sent_queue_arns, and the lead_ranking_queue_name output); fmt -check -recursive, init -backend=false and validate exit 0. #11 got the same lines for overlays/uat (c7b65a1, 037c55a). Update 2026-11-11: Shyamin Ayesh told Tharinda that #17 does not merge until production’s infrastructure is needed. It stays open at 897752d, not approved. Update 2026-11-13: Senuthmi pushed 814beb1 on 12 Nov, which she says makes change-list items 1–15 (WI-4.1 D-1) with every new base/ option defaulting to dev’s value — a new base/audit module and options in base/ecr, base/networking, base/rds and base/ses — and commented on 13 Nov (04:32 UTC). Her checks: fmt -check -recursive, and init -backend=false and validate in overlays/prod and overlays/dev, each exit 0, with Terraform 1.15.6, not the 1.12.1 dev.yaml pins; not run: gitleaks, the checks with 1.12.1, overlays/uat’s validate, and a plan-only dev.yaml run of the branch. No delta review of 814beb1 is posted yet. dev has since gained #19 and #11 (base/ecs’s create_service), so merging dev in, validating all three overlays and the plan-only run come before its merge (.forge/owed.md, “Bring forge-et-iaac#17 … to F38-i WI-4.1’s change list”). Update 2026-11-13, 11:18 UTC: the first review record on #17 was posted for 814beb1 (issuecomment-6851275427, verdict block, iaac-reviewer). It found D-1 met in full and every base/ default leaving dev and uat unchanged, with fmt, validate in all three overlays and gitleaks clean on Terraform 1.12.1. Two Important items are left for Senuthmi: merge dev in, and create_service for production’s first apply, or leave that to WI-12.1. No plan-only dev.yaml run of the branch yet. Update 2026-11-13, evening: with Senuthmi’s agreement, the delivery team pushed 3fc1add, merging dev at b175dae into #17 (README conflict resolved; fmt and validate in all three overlays exit 0 on Terraform 1.12.1). #17’s head is now 3fc1add, with no conflict with dev. create_service is still open. Update 2026-11-14: a full review of 3fc1add was posted as one comment that replaces the PR’s earlier review comments (issuecomment-6872662518, verdict block, iaac-reviewer). D-1 is met, the merge’s apply to dev is expected to read No changes., and the running uat instance is not affected. One Important item is open: create_service and a fixed service name in the alerts ARN for production’s first apply, or a note leaving it to WI-12.1. Seven nits are optional. Update 2026-11-19: Senuthmi’s e590a2b (16 Nov) adds production’s create_service and the fixed service name in the alerts ARN. At Tharinda’s request the delivery team pushed 4e2f973 (dev merged in at 8b48548, #20–#24) and bba8a28 (#22’s card-read queue wired in overlays/prod). A delta review of the three commits posted approve (issuecomment-7008385059). With the lead’s approval, the plan-only dev.yaml run of bba8a28 (44569651738, apply skipped) planned one in-place change that is not #17’s: the SES configuration set’s max_delivery_seconds 0 → null, which dev’s own applies also make (.forge/owed.md, “dev’s apply changes the SES configuration set on every run”). Merging #17 changes nothing else on dev. Still held for Shyamin’s go-ahead.

  • forge-et-iaac#18 (F11-g), opened on 2026-11-08 by the delivery team from feature/F11-g-lead-ranking-queue, merged into dev by Dinil on 2026-11-09 (8e510ce, 08:14 UTC) and applied by dev.yaml run 42584544763: “Apply complete! Resources: 3 added, 2 changed, 0 destroyed”. It adds to base/sqs the queue et-dev-lead-ranking (visibility timeout 150 s, the backend’s et.lead-ranking.visibility-timeout, because a ranking call may wait up to 60 s for a pacing slot) and its dead-letter queue et-dev-lead-ranking-dlq (maxReceiveCount 3, 14-day retention); to base/scheduler a seventh schedule, lead-ranking-sweep at rate(1 hour); and to overlays/dev the task role’s send and consume rights on the queue and the lead_ranking_queue_name output. The two task-definition values came with forge-et-be#375, merged on 2026-11-09 (98553b72) and deployed as backend:120 (rollout COMPLETED, read with aws ecs describe-services): ET_MESSAGING_LEADRANKINGQUEUE=et-dev-lead-ranking and ET_AI_FEATURES_LEADRANKING_MODEL = Haiku 4.5’s Global profile. Register §B F11-g rows 1–3.

  • forge-et-iaac#19 (F20-g), merged into dev on 2026-11-12 (4b3d883e) and applied by dev.yaml run 43176389232: “Apply complete! Resources: 1 added, 0 changed, 0 destroyed”. It adds to base/scheduler one more schedule, scheduled-email-send at rate(1 minute), UTC, sending the job to et-dev-scheduler; nothing else changed. It merged only after dev’s backend image (build-6edbbdd5) held F20-g wave 1, with an approve review record on its head (38cfc37) and Shyamin Ayesh’s approval given outside GitHub (the lead, 2026-11-14). Read on 2026-11-14: the schedule ENABLED, et-dev-scheduler and its dead-letter queue empty. Register §B F20-g row (a).

  • forge-et-iaac#20 (F37-a), merged into dev by Tharinda on 2026-11-17 (fa20b5e8, 12:40 UTC) and applied by dev.yaml run 44085299528: “Apply complete! Resources: 1 added, 0 changed, 0 destroyed”. It adds to base/scheduler one more schedule, price-broadcast-send at rate(1 minute), UTC, sending the job to et-dev-scheduler; nothing else changed. It merged only after dev’s backend (backend:151, image build-b901e4bb) held F37-a wave 1, with an approve review record on its head (019b4bc) and Shyamin Ayesh’s approval given outside GitHub (the lead, 2026-11-17). Read on 2026-11-17: the schedule ENABLED; five minutes after the apply, no firing refused as an unknown job kind and et-dev-scheduler-dlq at 0. Register §B F37-a row (a).

  • forge-et-iaac#22 (F36-b), opened on 2026-11-16 by the delivery team (Dinil) from feature/F36-b-card-read-queue (head 94ecf78, cut from dev at b175dae): the capture app’s card-read queue et-dev-card-read and its dead-letter queue et-dev-card-read-dlq in base/sqs/card_read.tf (redrive 3, visibility 150 s), and in overlays/dev and overlays/uat the backend task role’s consume and send rights on the queue and the card_read_queue_name output. Its merge’s apply should add 2 queues on dev and change the task role’s two SQS policies in place, destroying nothing; UAT changes only on its next hand-run apply. /forge-review-pr posted approve on 94ecf78; not yet approved by Shyamin Ayesh, not merged. The production overlay gets the queue pair from base/sqs on its first apply, with no rights until it is wired (asked for on #17). Register §B F36-b item (a).

  • forge-et-iaac#28, opened on 2026-11-19 by the delivery team (Claude, for Tharinda, at the lead’s request) from fix/ecs-task-tags-and-agentcore, cut from dev at 8b48548; head 8f77e99. It fixes the two Important findings of the delivery team’s 14 Nov review of the code on dev that merged without a review, and the SES diff above: base/ecs propagates the service’s tags to its tasks and turns on ECS managed tags (on 19 Nov dev’s service read propagateTags NONE and its task had no tags, so UAT’s Fargate cost could not be told from dev’s); base/agentcore, dev’s module "agentcore" and the agentcore_gateway_url output are removed (unused since Exa was chosen; its role allowed InvokeGateway on gateway/*); and base/ses ignores the configuration set’s delivery_options. Terraform 1.12.1: fmt, and validate in overlays/dev and overlays/uat, exit 0; gitleaks exit 0. The lead-approved plan-only run 44572051073 planned 0 to add, 1 to change, 4 to destroy: the ECS service’s two tag settings in place, and the gateway, its target, its role and its policy destroyed (run on 2ba0261; 8f77e99 changes only comments). /forge-review-pr posted approve on 8f77e99 (comment): no Blocker or Important; its three Nits, all comments, are fixed in 8f77e99. The apply should start no ECS deployment, so the tasks get their tags at the next forge-et-be dev deploy. #28 conflicts with #17 in README.md only. Waits for Shyamin Ayesh’s approval. The review’s smaller findings are not in it; they have their own .forge/owed.md row, as does removing svc-harness’s spike-invoke-gateway policy after the apply.

  • forge-et-iaac#24 (F11-d), opened on 2026-11-19 by the delivery team (Dinil) from fix/F11-d-lead-triage-visibility-timeout, cut from dev at 7e3481a: base/sqs/lead_triage.tf sets the et-dev-lead-triage queue’s visibility timeout from 30 s to 150 s, matching et.lead-triage.visibility-timeout, and the card-read, lead-ranking and lead-research comments say “the shared 30 s”. /forge-review-pr posted approve on 4d41916; merged into dev by Dinil on 2026-11-19 (8b48548, 08:30 UTC) with no GitHub approval from Shyamin Ayesh on the PR. dev.yaml run 44488259851 planned and applied “0 added, 2 changed, 0 destroyed”: module.sqs.aws_sqs_queue.lead_triage visibility_timeout_seconds 30 → 150 in place, and module.ses.aws_sesv2_configuration_set.this max_delivery_seconds 0 → null, a change that is not #24’s and that the run before it (44462431613, 7e3481a) also made (.forge/owed.md, “dev’s apply changes the SES configuration set on every run”). UAT gets the new timeout at its next hand-run apply; prod from base/sqs on its first.

  • forge-et-iaac#25 (F36-c), opened on 2026-11-19 by the delivery team (Dinil) from feature/F36-c-research-queues, cut from dev at 8b48548, merged into dev by Dinil on 2026-11-20 (6fca402, 04:22 UTC; approved on GitHub by Tharinda) and applied by dev.yaml run 44673993352: “Apply complete! Resources: 5 added, 3 changed, 0 destroyed” — the four queues and the schedule added, the task role’s two SQS policies and the SES drift changed. Read on 2026-11-20: both queues and their dead-letter queues exist, and capture-research-sweep (group et-dev) is ENABLED. The PR: base/sqs gains et-dev-plan-research (visibility 300 s) and et-dev-capture-research (330 s), each with a dead-letter queue (redrive 3, 14-day retention); base/scheduler gains capture-research-sweep at rate(1 hour); overlays/dev and overlays/uat give the backend task role consume and send rights on both queues and output their names. Its merge’s apply should add 5 resources and change the task role’s two SQS policies in place (plus the SES drift every dev apply shows), destroying nothing; UAT gets the same at its next hand-run apply, one more change it is behind dev. /forge-review-pr posted approve on bd19e78. The td.dev.json keys are in forge-et-be#485, merged only after that apply (register §B F36-c items (a)–©); it merged on 2026-11-20 (d240a830), and dev’s backend:172 carries the five values.

What dev costs — 14 Oct 2026

Measured (Cost Explorer, read on 14 Oct): about USD 2.03 a day, roughly USD 62 a month for 10–12 Oct — ECS Fargate ~$18, the load balancer ~$16, RDS ~$15.50, public IPv4 addresses ~$11 (the load balancer’s two and the task’s one, at $0.005 an hour each), Secrets Manager ~$0.30, rising to about $2 with the four Entra secrets added on 13 Oct. Bedrock and AgentCore use so far is under USD 2 in all (the SP-ARCH-004 run, .forge/design/ai/lead-research-source-options.md §11).

Shyamin Ayesh’s estimate (an HTML page shared on 14 Oct, outside the repository) prices the dev Terraform module by module. The corrections below are for the team’s own tracking and are not raised with Shyamin (the lead, 14 Oct). Its line items for Fargate, the load balancer and RDS match the measured figures, but four things in it are wrong or missing:

  • Its headline “fixed monthly ≈ $67–72” and its own recap “≈ $54–60” disagree.
  • The recap leaves out the public IPv4 charge of about $11 a month — the main gap between the two.
  • It calls the database storage gp3; it is gp2 (same price).
  • It says ECR is bounded by its lifecycle rule; the build-* images never expired until forge-et-iaac#3 added a keep-the-last-100 rule, applied on 20 Oct (row 8).
  • It prices an AgentCore web search at “$0.01–0.05, not published”; the verified price is $7 per 1,000 searches (lead-research-source-options.md §11). This matters only if AgentCore is chosen, which after SP-ARCH-004 stage 2 is unlikely (the lead leans towards Exa, .forge/owed.md); if it is not chosen, the gateway is removed (it has no standing charge — USD 0.0001 in October).

It also says no cost allocation tags are configured. The Module, Project and Environment tags are on the resources since 13 Oct, but a tag is usable in Cost Explorer only once it is activated in the management account’s Billing console, which is not done. (The lead, 18 Oct: billing settings — the AWS Budgets, their alerts and recipients, and activating the cost allocation tags — are handled by Shyamin Ayesh, Kavindu Anthony, Malinda Wellalage and Thilina Abeyrathna, not the delivery team. The delivery team keeps the bill down through its design and architecture choices and its daily use of the account.)

Against the cost annexe (.forge/client/sow-cost-annexe.md, the estimate of Empire Teas’ AWS bill): the annexe puts dev at about $34 a month, assuming no load balancer in dev, a part-time backend ($6) and no public IPv4 charge. Dev as built runs at about $62. The off-hours schedule F38 FR-16 asked for (backend at zero tasks and the database stopped outside working hours) would bring it to roughly $40–45, but the lead deferred it on 14 Oct until development is complete (F38 Rev 10; a post-issue note on the proposal records it): the team runs work against dev overnight. After development, dev is scaled down, scheduled off or decommissioned. The annexe’s production figure also has no line for public IPv4 addresses (about $7–11 a month). Re-deriving the annexe is a question for the lead (.forge/owed.md).

What dev cost in October 2026 — against the annexe (F38-i FR-14)

Read on 2 Nov 2026 as et-dev (svc-harness) in the dev account, with four read-only calls (F38-i-WI-10.1 plan, subtask 2; the output is kept outside the repository until wave 10 merges). Cost Explorer reported October with Estimated: false, so AWS has finalised the month. The dev account is a member of the Organization, so the invoice itself goes to the management account and the dev account sees none; the Cost Explorer service and usage-type lines below stand in for the invoice lines (the lead’s decision, F38-i-WI-10.1 plan D-1). October’s record types are usage USD 51.26 and tax USD 0.00, with no credit or refund, so usage and the net charge are the same. Dev was not up for the whole month (the backend has run since 9 Oct, and resources were added on 13, 20, 22 and 25 Oct), so the last column is the run rate of the estate as built: the daily usage of 25–31 Oct averaged and multiplied by 30.

Compared with .forge/client/sow-cost-annexe.md §3, development column (figures in USD, usage at list price):

Annexe row Annexe dev figure October 2026 Cost Explorer lines Difference
Backend application, including the background worker (Fargate) 6 11.98 Elastic Container Service: Fargate vCPU-hours and GB-hours, x86 until the move to ARM64 (forge-et-be#275, merged 19 Oct) and ARM64 after it +5.98. The annexe assumed a part-time backend; dev runs around the clock (the off-hours schedule is deferred, F38 Rev 10), and October counts only from 9 Oct
Static hosting for both applications (S3 + CloudFront) ~1 0.00 CloudFront 0.00002; the frontend buckets’ S3 cost is inside the S3 line of the next row, because Cost Explorer cannot split S3 by bucket without the cost allocation tags −1. CloudFront’s use is inside its free allowance. The row carries both the CRM and the capture app; annexe §6 also counts the capture app’s static hosting in its “Incremental AWS” line, so the annexe may count it twice
Database (RDS PostgreSQL, single-AZ) 14 13.49 Relational Database Service: db.t3.micro instance-hours 11.51, gp2 storage 1.98 −0.51, because October counts only from the database’s creation; a full month is about 15.50 (the 14 Oct reading above)
Load balancer not used 12.25 Elastic Load Balancing: LoadBalancerUsage 12.24 +12.25. The annexe has no load balancer in dev; dev has one (the 14 Oct reading above)
S3, SQS, ECR, Route 53, Secrets Manager 4 1.83 Secrets Manager 1.42 (the database secret and, from 13 Oct, the four Entra secrets); ECR storage 0.29; S3 0.12; SQS 0.00; Route 53 not billed here, because the crm.empire.lk zone is in the management account (F38-a AC-6) −2.17
CloudWatch and data transfer 4 0.00 CloudWatch, CloudWatch Events and SNS 0.00; no data-transfer line −4. Dev’s logs, the one alert topic and its traffic are inside the free allowances
Threat detection (GuardDuty) 5 0.00 none — GuardDuty is not enabled in dev −5, as the annexe’s own note says (off since 9 Oct)
Not in the development column: public IPv4 addresses — 8.20 Virtual Private Cloud: PublicIPv4:InUseAddress +8.20. The load balancer’s two addresses and the backend task’s one, at USD 0.005 an hour each; the annexe has no line for them in either environment
Not in the development column: KMS keys — 0.72 Key Management Service: keys +0.72. The keys F38-b created on 20 Oct (about USD 2 a month for a full month)
Not in the development column: AI — 2.70 Claude Haiku 4.5 on Bedrock 2.09 (input tokens 1.54, output tokens 0.54); Bedrock AgentCore web search 0.61 +2.70. Annexe §5 prices AI for the estate as a whole, not per environment
Not in the development column: other — 0.10 Cost Explorer API requests 0.10; SES 0.0002 +0.10
Total ≈ 34 usage 51.26; net 51.26 +17.26
Run rate, 25–31 Oct × 30 ≈ 34 63.46 daily usage 2.05–2.44 +29.46

The largest differences are the ones the 14 Oct reading named: the load balancer the annexe left out of dev, the backend running around the clock, and the public IPv4 addresses the annexe has in neither environment. In October they add USD 26.43. The rows the annexe over-estimates (CloudWatch, GuardDuty, the storage services, static hosting and, for a part month, the database) are USD 12.68 below it, and AI, KMS and the Cost Explorer requests, which the development column does not cover, add USD 3.52; together that is October’s USD 17.26 difference, to rounding. The run rate is not split by service. Reported to the lead on 2 Nov 2026, each difference with its Cost Explorer line, in the F38-i-WI-10.1 session. The annexe is the lead’s to change (.forge/owed.md, “The cost annexe’s dev figure is about half of what dev costs”).

Dev’s Entra values as plain values — 6 Nov 2026 (F38-i wave 3)

Dev signs in with the plain Entra values: forge-et-be#374 (F38-i WI-3.1; merge 5a14884b, 6 Nov 2026 03:39 UTC) moved ET_SECURITY_ENTRA_TENANT_ID, _CLIENT_ID and _CERTIFICATE from secrets references to plain environment values in .aws/td.dev.json. The merge’s build run 42157845766 succeeded and deployed backend:111 (image crm/backend@sha256:88cf25ea…); read with describe-services and describe-task-definition as svc-harness (no get-secret-value), the service runs 1 of 1 tasks on backend:111 with its rollout COMPLETED, the revision’s secrets hold only the database user and password, the Entra private key and the Exa key, and its environment holds the three Entra values. pnpm e2e:smoke from forge-et-fe main (4a91ed9c), run on 6 Nov 2026 at about 03:56 UTC with the sign-in made by hand in its Chromium window: all nine checks passed as CRM05@empire.lk (role smoke_test_view_only), 20 s from the end of the sign-in. A first run that morning signed in as crm01@empire.lk from a remembered Microsoft session: its sign-in check passed and its key check stopped the run, as it does for an account holding manage keys. The three et-dev-entra-* secrets for these values still exist in base/secrets; F38-i wave 8 removes them.

A read-only scan of the dev account — 4 Nov 2026

Read on 4 Nov 2026, about 18:00–18:30 UTC, as svc-harness with describe, list and get calls, CloudWatch Logs Insights queries and Cost Explorer reads only (no get-secret-value, no write), after forge-et-iaac#15 and #16 were applied that day. Nothing is broken, and every gap found is a decision already recorded.

  • Backend: service backend on backend:103 (registered 4 Nov 18:00 UTC), 1 task desired and running in us-east-1b, rollout COMPLETED, its target healthy; the image is build-66e6f9fa…, forge-et-be main’s head (forge-et-be#362), with an ECR scan COMPLETE and no findings. The revision’s environment, secrets, CPU, memory, architecture and roles match .aws/td.dev.json on main, except hostPort, which ECS fills in. https://api.dev.crm.empire.lk/actuator/health/liveness answers UP; http:// redirects with 301. 103 revisions of backend are ACTIVE; they cost nothing.
  • Load balancer and target group: ELBSecurityPolicy-TLS13-1-2-2021-06 on 443; health check /actuator/health/liveness every 10 s; 30 s deregistration delay. drop_invalid_header_fields is off (an optional hardening item for production).
  • Database: et-dev, created 5 Oct 2026 and unchanged by #15; PostgreSQL 18.3, db.t3.micro, us-east-1b, gp2, 20 GiB with 18.3 GB free, not public, not encrypted, no deletion protection, backups off (dev decisions, the proposal’s post-issue note of 14 Oct and F38-i decision 11). Its CPU credits are full (288), and its lowest freeable memory over the last day was about 146 MB of 1 GiB. AWS listed an operating-system update (system-update) with no date by which it would apply itself. 6 Nov: it was applied, with the lead’s approval, from 18:59:54 to 19:02:17 UTC on 5 Nov; the engagement log of 2026-11-06 has the details. Same instance, creation time, endpoint, engine and zone; no maintenance pending. The database’s security group admits 5432 from the VPC range, as recorded.
  • Queues and schedules: all eight queues and their dead-letter queues are empty, each with server-side encryption and a redrive after three receives; the six schedules in et-dev are enabled, each sending to et-dev-scheduler.
  • Buckets: et-dev-attachments (KMS, Bucket Keys), et-dev-spa and et-dev-capture (SSE-S3) are versioned, block all public access and are not public; none has a lifecycle rule (.forge/owed.md, the buckets’ non-current versions). The account has no account-level S3 Block Public Access (a dev decision, item 13 below).
  • Frontends: both distributions are Deployed with TLS 1.2 (2021) and HTTPS redirects; https://dev.crm.empire.lk/, /leads and https://capture.dev.crm.empire.lk/ answer 200 with each app’s Content-Security-Policy and the four headers of #3.
  • Identities: IAM users svc-github (AdministratorAccess, two active keys, of 13 and 23 Oct, no console) and svc-harness (one key; a console password without MFA), both recorded dev exceptions; the svc-harness console sign-in of 23 Oct (05:24 UTC) was Nanndish Satgunarajah: the lead shared svc-harness’s credentials with him and set up the AWS CLI on his machine for his testing work (the lead, 6 Nov), as item 8’s decision of 14 Oct allows. svc-harness is therefore used from more than one delivery-team machine, and CloudTrail shows its calls under the one user name. No root access key; no account password policy, CloudTrail trail, GuardDuty detector or AWS budget (dev decisions, item 13 below; budgets are the billing owners’).
  • Secrets, keys and certificates: the RDS-managed secret (rotation off) and the five app secrets, each with an AWSCURRENT version; the two customer-managed KMS keys, enabled with yearly rotation; the three certificates ISSUED and in use, the first ending 25 Apr 2027 and eligible for renewal.
  • Mail and alerts: SES has production access and sends from the verified dev.crm.empire.lk (DKIM and MAIL FROM SUCCESS), with no suppressed address; the alert topic et-dev-alerts has four confirmed email subscriptions, and the rule et-dev-backend-deployment-failed is enabled. There is no CloudWatch alarm, as decided for dev (.forge/owed.md, “Decide the UAT and production alarm sets”).
  • Bedrock and AgentCore: the task role’s bedrock policy allows all six inference profiles of Haiku 4.5, Sonnet 5.5 and Haiku 5.5 (since #16); the quotas are unchanged — Haiku 4.5 at 10 requests a minute on each route, Sonnet 5.5 and Haiku 5.5 at 0. The AgentCore gateway et-dev-gateway is READY, kept with svc-harness’s spike-invoke-gateway policy until the search provider is chosen.
  • Logs: no ERROR line in seven days. The last day’s only WARN lines are the two startup warnings each deploy logs (the OpenAPI document and uploads not scanned, both expected in dev). The RATE_LIMITED lines of 29 Oct and 2 Nov are F42’s pacing proofs.
  • Nothing left over: one VPC (et-dev, no default VPC), no EC2 instance, volume, database snapshot, unattached address or network interface; all 49 tagged resources carry Project = ET, none is untagged, and nothing is tagged Environment = UAT.
  • Cost: USD 2.05, 2.46 and 2.11 on 1–3 Nov (4 Nov partial), about USD 63 a month, as October’s run rate. In order: the load balancer, the database, ECS, public IPv4 and Haiku 4.5 (USD 0.38 over the four days).

Checked against F38 and the lead’s recorded decisions

First checked 9 Oct 2026; updated 14 Oct 2026 (rows 1, 4, 7, 9, 11–15), 17 Oct 2026 (row 16), 19 Oct 2026 (rows 5, 6, 8–10 and 13–15: notes from F38-d’s review, “F38-d — the dev deploy pipelines, checked” below) and 23 Oct 2026 (rows 5, 8, 9 and 15: F38-d WI-4.2, the workflows after F38-d’s wave 2) and 26 Oct 2026 (rows 6, 8 and 14: F38-d WI-6.1, the deliberate checks on dev; rows 5, 9, 10 and 15: F38-d WI-7.1, the closing sweep). What is built, compared with the F38 parent spec, the decisions recorded in the F38-b and F38-d specs’ Open Questions (2026-10-02) and the RDS values sent to Shyamin Ayesh on 2 Oct 2026 (row 4 lists them). F38-b was approved on 17 Oct and F38-d on 18 Oct; this is a list to settle with Shyamin, not a verdict on his work. Matches: us-east-1; no NAT gateway, with the tasks in a public subnet admitting only the load balancer; TLS 1.2 or later on the listener; the liveness health check; database secrets through task-definition secrets references and AWS access through the task role (AD #22); the scheduler queue’s dead-letter queue and redrive; 30-day log retention and JSON logs; the cross-site cookie and CORS values (AD #33); crm01@empire.lk as the seeded admin (AD #34); the deploy key held as a GitHub Environment secret (F38 FR-19); the --pull --no-cache-filter runtime build.

# Item What F38 or the lead’s decision says What is built Urgency
1 Database password rotation Every database secret rotates without breaking a running backend (F38 FR-10, OQ-K) RDS-managed rotation, whose default schedule is every 7 days; ECS reads secrets only when a task starts, and nothing restarts the task. The first rotation of a secret created on 9 Oct falls about 16 Oct, after which the running task cannot open new connections. Shyamin has it on his list. 14 Oct: still on, every 7 days, next run 16 Oct 23:59:59 UTC (17 Oct 05:29:59 in Sri Lanka) 15 Oct: Shyamin lengthened the schedule to every 365 days without rotating immediately; read in the account on 16 Oct — next run 15 Oct 2027 23:59:59 UTC (16 Oct 2027 05:29:59 in Sri Lanka), last rotation still 9 Oct (the secret’s creation). Nothing restarts the task yet. The schedule is not in Terraform: CloudTrail shows it set by the AWS CLI (RotateSecret, rotateImmediately: false, 15 Oct 16:38 Sri Lanka time). The matching commit, forge-et-iaac 76fce06, adds master_user_secret_rotation_days = 365 to base/rds/instance.tf, an argument aws_db_instance does not accept, so terraform plan fails on dev (run 38724927953, “Unsupported argument”); the same commit comments out the apply step in dev.yaml. The line sits in the shared base/rds module, so if it worked it would reach every environment, production included. Replacing the database (item 2’s encrypted restore) creates a new RDS-managed secret on the 7-day default, under a new ARN the task definition must name Before production (F38-i), or before the schedule is shortened — force a new deployment after each successful rotation. The dated risk is gone. Also owed in forge-et-iaac: remove or replace the rejected argument and restore the apply step; if the schedule is kept in code, put it in the dev overlay, not base/rds. After item 2’s restore, set the new secret’s schedule again with --no-rotate-immediately, unless the restart exists by then. 16 Oct, the lead’s decision (F38 Rev 14): no scheduled rotation in any environment — the longest interval (or off) in the shared Terraform module, no restart built, a change by hand with a redeploy when needed (item 1) Applied 20 Oct: forge-et-iaac#3 (the delivery team, F38-b spec Revs 1, 3, 4) removed the rejected line, turned rotation off in the shared base/rds module (rotation_enabled = false) and restored the apply step; merged as d230b133 and applied by run 39475850266. aws secretsmanager describe-secret on the RDS-managed secret, 21 Oct: RotationEnabled: false, no next rotation date, last rotated 9 Oct. The “force a new deployment after each rotation” and “after item 2’s restore” sentences above are superseded: no restart is built (Rev 14), and there is no restore (F38 Rev 17). (Evidence, 2026-10-21: “F38-b — the dev estate, checked”, AC-3.) (Update 2026-10-26, F38-b WI-4.1: re-read — still RotationEnabled: false with no next rotation date, and base/rds/rotation.tf at c728e06b declares it; “F38-b — the dev estate, checked”, AC-3.)
2 Encryption at rest RDS encrypted with a customer-managed KMS key (FR-10, FR-13, NFR-7; proposal §A.4) Not encrypted (storage_encrypted unset). An RDS instance cannot be encrypted in place: it is a snapshot, an encrypted copy and a restore. Decided by the lead, 17 Oct 2026 (F38 Rev 17): dev stays unencrypted, with no database or backups key; production is undecided (Shyamin Ayesh, Thilina Abeyrathna and Malinda Wellalage decide — F38-i, .forge/owed.md) Accepted in dev (F38 Rev 17)
3 Database subnets and access Private subnets; the database admits only the tasks’ security group (FR-9) The two public subnets with publicly_accessible = false; 5432 open to the whole VPC range (Update 2026-10-26, F38-b WI-4.1: unchanged and recorded as found — the subnet group is et-dev-public-0 and et-dev-public-1, routed to the internet gateway, and et-dev-rds admits 5432 from 10.0.0.0/16; a connection from outside AWS times out; “F38-b — the dev estate, checked”, AC-2.) Optional in dev, the DevOps team’s call (F38 Rev 18, 18 Oct); required in production (F38-i)
4 Database settings sent 2 Oct db.t4g.micro, gp3 20 GiB autoscaling to 100, backups 3 days with point-in-time recovery, rds.force_ssl = 1, four CloudWatch alarms db.t3.micro (Shyamin’s commit ea80c3b switched to it), 20 GiB without autoscaling, backup retention not set (the AWS default applies), default parameter group, no alarms. 14 Oct: backups are disabled (retention 0, no snapshots). Decided by the lead, 14 Oct: backups are optional in dev — the DevOps team’s call, not pushed; production keeps 14 days (F38 Rev 11; a post-issue note on the proposal, whose §A.5 and §A.6.3 say 3 days in development). Turning them on costs nothing extra at this size, because RDS includes backup storage up to the database’s own size 18 Oct (F38 Rev 18): gp3 with autoscaling, deletion protection and the final snapshot are optional in dev too; the four CloudWatch alarms are F38-e’s to decide 23 Oct: decided by the lead — no database alarms in dev; UAT and production decide theirs (F38-e spec decision 1, F38 Rev 23; .forge/owed.md “Decide the UAT and production alarm sets”) 22 Oct: rds.force_ssl = 1 is already met — the instance uses default.postgres18, whose rds.force_ssl is 1 (aws rds describe-db-parameters --db-parameter-group-name default.postgres18, source system) (Update 2026-10-26, F38-b WI-4.1: recorded as found — PostgreSQL 18.3 with automatic minor upgrades, db.t3.micro, gp2 20 GiB with no autoscaling, deletion protection off, skip_final_snapshot = true, backup retention 0, single-AZ, rds.force_ssl 1; “F38-b — the dev estate, checked”, AC-3.) Optional in dev; required in production (F38-i)
5 CPU architecture The task runs ARM64 (Graviton), and deploy-dev builds for linux/arm64 (lead, 2026-10-02, F38-b and F38-d Open Questions) X86_64, image built for linux/amd64 (Reviewed 2026-10-19, F38-d WI-1.1: DR-1, fixed by F38-d WI-2.1.) (Update 2026-10-23, F38-d WI-4.2: since forge-et-be#275 (merged 19 Oct) the task definition is ARM64 and the image is built on an arm64 runner for linux/arm64; read on 22 Oct, the running task on backend:58 reports arm64 and its image’s config linux/arm64. DR-1 fixed; evidence: “F38-d — the dev deploy pipelines, checked”, AC-4.) (Checked 2026-10-26, F38-d WI-7.1: current against forge-et-be origin/main 05cadc2a; “F38-d — the dev deploy pipelines, checked”, AC-10.) Low — a cost choice; change the task definition and the build together
6 Task-definition values stopTimeout 60 s; target-group deregistration delay 30 s; JAVA_TOOL_OPTIONS with -XX:MaxRAMPercentage=65 on a 1 GiB task and -XX:+ExitOnOutOfMemoryError (lead, 2026-10-02) None set (the task is 1 GiB) (Reviewed 2026-10-19, F38-d WI-1.1: DR-2 and DR-3. “None set” holds for the task definition and the target group; the image itself sets JAVA_TOOL_OPTIONS="-XX:MaxRAMPercentage=75 -XX:+ExitOnOutOfMemoryError" (forge-et-be Dockerfile), so the running task already exits on an out-of-memory error and only the heap share differs. Fixed by F38-d WI-4.1 (stopTimeout, JAVA_TOOL_OPTIONS) and WI-3.1 (the deregistration delay).) (Closed 2026-10-26: .aws/td.dev.json on forge-et-be main sets stopTimeout 60 and JAVA_TOOL_OPTIONS -XX:MaxRAMPercentage=65 -XX:+ExitOnOutOfMemoryError, and the target group’s deregistration delay is 30 s since forge-et-iaac#6.) (Update 2026-10-26, F38-d WI-6.1: shown on dev — the running backend:71 (and backend:68 before it) has stopTimeout 60 and JAVA_TOOL_OPTIONS, its task logs Picked up JAVA_TOOL_OPTIONS: -XX:MaxRAMPercentage=65 -XX:+ExitOnOutOfMemoryError, deregistration_delay.timeout_seconds reads 30, and four rolling deploys left no load-balancer 5xx; DR-2 and DR-3 fixed. “F38-d — the dev deploy pipelines, checked”, AC-2 and AC-7.) Closed
7 Secrets in the task definition No secret in a task-definition JSON (F38 NFR-2); the Entra private key from Secrets Manager (register §D) ET_SECURITY_ENTRA_PRIVATE_KEY and _CERTIFICATE are plain environment placeholders in the committed .aws/td.dev.json. Done 13 Oct 2026 (forge-et-be dfc9be15): all four Entra values are secrets references to Secrets Manager, and no key is in git Closed
8 ECR tags Immutable tags, and deploy-dev skips the build when ECR already holds the commit’s image (lead, 2026-10-02, F38-d) Mutable tags, no skip; since forge-et-iaac#3 was applied on 20 Oct the lifecycle rule keeps the last 100 build-* images (aws ecr get-lifecycle-policy --repository-name crm/backend, 21 Oct; evidence: “F38-b — the dev estate, checked”, “Other changes in forge-et-iaac#3”) (Reviewed 2026-10-19, F38-d WI-1.1: DR-4 (the short-SHA tag and no skip — F38-d WI-2.1) and DR-5 (MUTABLE at #3’s head — F38-d WI-3.1).) (Update 2026-10-23, F38-d WI-4.2: since forge-et-be#275 each image is tagged build-<full commit SHA> and the deploy job skips the build when ECR already holds that tag, then deploys by digest; DR-4 fixed, and the skip is shown on dev by F38-d WI-6.1. The tags stay MUTABLE until F38-d WI-3.1 (DR-5).) (Closed 2026-10-26: forge-et-iaac#6 (F38-d WI-3.1), applied on 23 Oct, made the tags immutable — aws ecr describe-repositories, read 25 Oct: imageTagMutability IMMUTABLE; DR-5 fixed.) (Update 2026-10-26, F38-d WI-6.1: IMMUTABLE re-read; a manual run of a commit ECR holds skipped the build and deployed that image’s digest (AC-2). ECR refusing a second push under an existing tag is not yet shown, because svc-harness may not push (AC-11; .forge/owed.md).) (Update 2026-11-19: IMMUTABLE re-read with aws ecr describe-repositories; the lead accepted the setting as AC-11’s evidence and AC-11 is ticked, F38-d spec Rev 6.) Closed
9 The dev deploy workflow deploy-dev with build and deploy as separate jobs, only the deploy job holding AWS credentials, per-environment concurrency, timeout-minutes, persist-credentials: false, paths-ignore, manual dispatch (F38 FR-20); the jar built on the runner with setup-gradle (lead, 2026-10-02) One job named dev in the build workflow, holding the key throughout; none of the listed rules; the jar built inside Docker. Checked 14 Oct, the gaps that can cause a real fault: (a) no concurrency group, so when two merges land close together the older run can finish last and deploy the older image; (b) the deploy step does not wait for the service to become stable, so a deploy that ECS rolls back still shows green; © no timeout-minutes, which every other app-repo job sets, so a hung run can use up to six hours of the organisation’s Actions minutes; (d) it runs on every push to main, documentation-only merges included (Reviewed 2026-10-19, F38-d WI-1.1: DR-6 – DR-9 confirm each gap, and the first deploy’s ecs::deploy step took one second, which shows (b); the review adds DR-17 (an inaccurate header comment), DR-18 (a tag trigger with no live job) and DR-20 (no printed identity), all for F38-d WI-2.1, and DR-19 (the commented-out release job — .forge/owed.md).) (Update 2026-10-20: (b) is accepted for dev, not fixed. forge-et-be#275 added the wait; the lead then withdrew it for dev (F38-d spec Rev 3, F38 Rev 22) and forge-et-be#286 removes it. The ECS circuit breaker restores the previous task, and the deployment is recorded FAILED in ECS while the run stays green; deploy-uat and deploy-prod wait.) (Update 2026-10-23, F38-d WI-4.2: since forge-et-be#275 the workflow has a build job with no AWS credentials that builds the jar on the runner, and a deploy job in the dev Environment; a deploy-dev concurrency group that never cancels; timeout-minutes on both jobs; persist-credentials: false; paths-ignore: .github/workflows/** and workflow_dispatch, with the tag trigger gone; an accurate header comment; and an identity::aws step. DR-6 – DR-8, DR-17, DR-18 and DR-20 fixed. The live checks of the manual run and the queued run are F38-d WI-6.1’s; evidence: “F38-d — the dev deploy pipelines, checked”, AC-2 and AC-3.) (Closed 2026-10-26, F38-d WI-7.1: (a) and © are fixed and (d) is fixed for workflow-only pushes, (b) is accepted for dev; WI-6.1 showed the manual run and the queued run on dev, and the file on forge-et-be origin/main 05cadc2a is as described (its only later commits change pr-validation.yml); “F38-d — the dev deploy pipelines, checked”, AC-2 and AC-10.) Closed
10 Image scan ECR’s scan on push for every dev image; trivy image only in deploy-uat and deploy-prod, where it blocks (AD #38 as restated 2026-10-20; F38-d spec Rev 3) No scan in the workflow (ECR’s basic scan on push runs) (Reviewed 2026-10-19, F38-d WI-1.1: DR-10, fixed by F38-d WI-2.1.) (Update 2026-10-20: forge-et-be#275 added a report-only trivy image; the lead then withdrew it for dev and forge-et-be#286 removes it, so dev matches this row again with ECR’s scan alone.) (Checked 2026-10-26, F38-d WI-7.1: current against forge-et-be origin/main 05cadc2a.) None for dev
11 CI identities A deploy service user for the app repos and a separate Terraform service user, each with only what it needs (FR-19) One CI user, to be narrowed (Shyamin’s list). 14 Oct: it is svc-github, with AdministratorAccess; its key was replaced on 13 Oct and the new one put into the dev environment secrets of forge-et-iaac, forge-et-be and forge-et-fe within the same minute, so all three share one administrator key Accepted for dev by the lead, 14 Oct (F38 Rev 9): one svc-github key for the three repositories is kept for now, because asking the DevOps team for a narrowed user per need slows the work. Production follows FR-19 (F38-i). 18 Oct (F38 Rev 18): the 90-day key replacement is dropped for dev — replaced only if it may have leaked; production keeps it
12 Resource tags A team or a role, never a person (agenda item 18 below) Owner = shyamin@altrium.io (still, on 14 Oct; a Module tag per module was added on 13 Oct). Kept as is: the lead, 18 Oct — no change needed Accepted (the lead, 18 Oct)
13 How the deploy changes reached the app repos Before 14 Oct: every change to an app repo’s main is a PR, CI changes need the lead’s approval, and every app-repo PR carries a review record (.claude/CLAUDE.md § Boundaries and § Quality Gates; git-conventions.md). Since 14 Oct the DevOps team’s deploy files are the one exception (decision column) Seventeen commits pushed straight to main with the admin bypass, no PR and no review: fourteen to forge-et-be (87d2bfd5 on 5 Oct, seven on 8 Oct, five on 9 Oct, dfc9be15 on 13 Oct) and three to forge-et-fe on 13 Oct (9c78916, 5ab1e44, bf4adc8). Every one changes only .github/workflows/build.yaml or .aws/td.dev.json; one commit message, 4a14feb1 “Avoid AWS authentication for testing”, has no type, which commitlint would refuse (Reviewed 2026-10-19, F38-d WI-1.1: no further direct pushes since 14 Oct — every commit since then that changes .github/workflows or .aws came through a PR (forge-et-be#263; forge-et-fe#259, #265, #278; forge-et-pwa#82); the four current files are reviewed in “F38-d — the dev deploy pipelines, checked” below, DR-1 – DR-35.) Decided by the lead, 14 Oct: Shyamin is a GitHub organisation administrator and will keep pushing deploy configuration straight to main; that is accepted while a push breaks nothing and keeps to the team’s standards. The delivery team checks each such push after the fact (git-conventions.md § Branch Protection) and raises what it finds with Shyamin or fixes it by PR. All seventeen were checked on 14 Oct: nothing outside the deploy files is touched, no secret is in git, and the eight before 9 Oct are trial versions the current files replace; the gaps in the current files are rows 9 and 15
14 Other task-definition values (added 11 Oct 2026, from .aws/td.dev.json at 03404e8e) Every value in register §D set for dev Placeholders: ET_SECURITY_ENTRA_TENANT_ID and _CLIENT_ID (non-secret; values in §2.1), ET_FILES_STORAGE_BUCKET (no bucket yet), ET_MAILBOX_PUBLIC_BASE_URL (https://api.dev.crm.empire.lk), ET_MAILBOX_TOKEN_KEY_ID (no KMS key yet). ET_MESSAGING_MAILBOX_INGEST_QUEUE is missing, so no mail would be ingested, silently. The backend boots because every placeholder is non-empty, but sign-in, uploads and the mailbox cannot work. 13 Oct: the tenant and client IDs moved to secrets references with the certificate and key (row 7), and sign-in works; the bucket, mailbox and KMS placeholders and the missing ingest queue remain (Reviewed 2026-10-19, F38-d WI-1.1: DR-11 confirms the three placeholders and the missing ingest queue at 5c457091 — F38-d WI-4.1; DR-22 finds every other §D value set and none of the “must stay unset” keys; DR-23 asks whether the two non-secret Entra IDs stay in Secrets Manager — .forge/owed.md.) High — forge-et-iaac#3 (F38-b), applied on 20 Oct, created the bucket et-dev-attachments, the mailbox token key alias/et-dev-mailbox-token and the ingest queue et-dev-mailbox-ingest, and outputs their values (F38-b FR-9); they still have to be written into .aws/td.dev.json, which on 21 Oct still holds the three placeholders and no ingest-queue key. Closed 2026-10-25: forge-et-be#300 (F38-d WI-4.1, merged 22 Oct) wrote all four, and the running backend:62 holds them with none of the “must stay unset” keys (the backend deploy paragraph above); uploads and the mailbox are still to be walked on dev (F38-g) (Update 2026-10-26, F38-d WI-6.1: the running backend:71 sets every key the code requires at forge-et-be e88fef50 and none of the “must stay unset” keys (AC-7), and the first upload on dev succeeded — POST /api/v1/email/attachments answered 201 and the object is in et-dev-attachments; DR-11 fixed. The mailbox is still to be walked (F38-f, F38-g).)
15 The CRM frontend’s dev deploy (added 14 Oct 2026, from forge-et-fe .github/workflows/build.yaml at bf4adc8) deploy-dev with build and deploy as separate jobs, AWS credentials only in the deploy job, per-environment concurrency, timeout-minutes, persist-credentials: false, paths-ignore, manual dispatch (F38 FR-20); the deploy user narrowed to the deploy actions (FR-19); the dev environment restricted to main (dropped for dev, F38 Rev 18) One job (id dev, shown as build), which sets up the AWS key only after the install and build steps, but in the same job; manual dispatch is there; no timeout-minutes, persist-credentials: false or paths-ignore (a concurrency group that never cancels a run was added by forge-et-fe#259, BUG-136); the bucket name and distribution ID written into the workflow; the key is svc-github’s administrator key (row 11); the GitHub Environment dev has no branch rule, in this repository as in forge-et-be (read 14 Oct); no permissions: block for the job’s token. The same faults as row 9 (a), © and (d) apply: two close merges can leave the older bundle in the bucket. (e) One plain aws s3 sync --delete with no cache headers, where architecture.md § CI/CD Pipelines asks for a staged sync — hashed assets first with max-age=31536000, immutable, other files with max-age=0, must-revalidate, index.html last with max-age=0, must-revalidate. Checked 14 Oct: the files are served with no Cache-Control, so a browser may keep an old index.html after a deploy, and because --delete removed the old hashed scripts, that page then loads blank until a hard refresh. (e) fixed by forge-et-fe#259 (BUG-136): the staged sync with those headers, and a last step of its own that deletes. The workflow does not run the browser gate or the unit tests, which is right for a deploy (Reviewed 2026-10-19, F38-d WI-1.1, at 32c0dea5: DR-12 – DR-15 confirm the one job, the missing permissions: block, timeout-minutes, persist-credentials: false, paths-ignore and guard step, and DR-16 that (e) and the concurrency group are fixed; the review adds DR-24 (no printed identity) and DR-25 (no newline at the end of the file), all for F38-d WI-2.2, and DR-27 and DR-28 (names and deploy targets across the repositories — .forge/owed.md).) (Update 2026-10-23, F38-d WI-4.2: since forge-et-fe#298 (merged 19 Oct) the workflow has a build job with no AWS step, whose first step, check::config, fails on a blank VITE_API_BASE_URL or VITE_CSRF_COOKIE_NAME, and a deploy job joined to it by the web-dist artifact; paths-ignore: .github/workflows/**; timeout-minutes on both jobs; persist-credentials: false; a permissions: block per job ({} on the deploy job); an identity::aws step; and a newline at the end of the file. DR-12 – DR-15, DR-24 and DR-25 fixed. The guard failing on a blank value is F38-d WI-6.1’s; DR-27 and DR-28 stay .forge/owed.md rows. One run since, 39510630707 on 20 Oct, hit the deploy job’s 10-minute limit in its first S3 sync — BUG-254.) (Closed 2026-10-26, F38-d WI-7.1: WI-6.1 showed the guard failing on a blank value (AC-5); forge-et-fe’s build.yaml has no commit since forge-et-fe#298 on origin/main b0bebb28; the capture app’s build.yaml follows the same rules (forge-et-pwa#91, AC-3 and AC-5). DR-27 and DR-28 stay .forge/owed.md rows; BUG-254 is in .forge/bugs.md. “F38-d — the dev deploy pipelines, checked”, AC-10.) Closed
16 The tasks’ Availability Zone (added 17 Oct 2026, from forge-et-iaac base/ecs/service.tf and overlays/dev/main.tf at 76fce06) The backend tasks and the database run in one Availability Zone (F38 FR-9; proposal §A.2.1) The ECS service is given both public subnets, so a task may start in either zone; the database is single-AZ, in one of them Low. Decided by the lead, 17 Oct: the delivery team’s F38-b PR gives the service the one subnet in the database’s zone (F38-b FR-2). PR: forge-et-iaac#3 (F38-b wave 1), applied on 20 Oct: on 21 Oct the service has one subnet, in us-east-1b, and the database is in us-east-1b (aws ecs describe-services, aws rds describe-db-instances); recorded with evidence by F38-b’s WI-4.1 (AC-2) (Update 2026-10-26, F38-b WI-4.1: recorded — the service’s one subnet is subnet-02fb26b2351e546d1 in us-east-1b, its running task is there, and the database is in us-east-1b; “F38-b — the dev estate, checked”, AC-2.)

To suggest to Shyamin, as suggestions only — (All of these were done by the delivery team’s forge-et-iaac#3 (F38-b spec Rev 3), merged and applied on 20 Oct 2026, which also added a README and a CLAUDE.md; what stays open is in item 14.) The repository’s conventions beyond the PR-to-dev rule are still on hold (above): a .gitignore so .terraform/, plan files and state cannot be committed; committing .terraform.lock.hcl (neither is in the repository as of b7e923a); applying only from dev rather than from any branch; actions/checkout@v4 or later; a dead-letter queue retention of 14 days rather than the 4-day default. (GitHub OIDC was dropped from this list on 2026-10-02: CI keeps a service user’s key in GitHub secrets — F38 Rev 4. Importing the hosted zone into Terraform is F38-i’s item, before production’s records are written (F38 Rev 13); the zone sits in the ets-org account, §3.) The repository has no branch protection beyond the organisation’s repository-level ruleset (re-checked 9 Oct 2026) and no CODEOWNERS; it has had a README since #3.

Before 2026-09-29 this section read: AD #23 places the IaC in forge-et-be/terraform/; Shyamin Ayesh may stand up a separate Terraform repository, which is not done yet, and AD #23 stands until the separate repo is real and confirmed.

Open with the DevOps team

Drafted 9 Oct 2026; last updated 26 Oct 2026 (item 4: AWS’s answer on the Bedrock quotas, no increase; the Trello board as exported on 26 Oct, last activity 12:31 UTC: every ET card for dev work is in Done except “[ET] Bedrock quota increase…”, which the lead moved to Blocked that day with a comment recording AWS’s answer and the plan; “[ET] One Secrets Manager module (base/secrets)” is in Done; the UAT card is on hold in “Project ET TODOs”, and the production card is there too, untouched since 25 Oct); before that 25 Oct (items 4, 5 and 9: quotas unchanged, the task-definition values set, forge-et-pwa’s key added; the Trello board as exported on 23 Oct); before that 22 Oct (item 4: the quota requests AWS refused on 21 Oct; the Trello board as exported on 21 Oct, the 22 Oct note below; 21 Oct: forge-et-iaac#3 merged and applied on 20 Oct, items 1, 5 and 14). Everything still asked of Shyamin Ayesh (with Senuthmi Thimansa, and Kavindu Anthony where the account holder must act), after the call of 6 Oct 2026 and his work up to b7e923a (re-checked against 9eac094 on 14 Oct). This list is the record; the lead’s working copy is an HTML checklist outside the repository, numbered the same way. “Item N” below means a row of “Checked against F38” above. When something here is done or answered, strike it with the date, and move what the answer decides into the section it belongs to.

(Update 16 Oct 2026 — the lead put the open asks on the DevOps team’s Trello board as three cards and went through them with Shyamin Ayesh the same day; he may take them up the week of 19 Oct. Nothing on them is done yet: read as svc-harness on 16 Oct, the database is still unencrypted with backups off, the RDS secret still rotates every 365 days (next 16 Oct 2027), and forge-et-iaac dev still ends at 76fce06, whose terraform plan fails (run 38724927953) with the apply step commented out. The cards: “[ET] Misc tasks”, twelve items — the database fixes (item 3), uploads and mailbox infrastructure (item 5 and “Checked against F38” row 14), the capture app’s hosting with forge-et-pwa’s dev environment and key (items 5, 9), account safety basics with AWS Budgets (item 13), the cost allocation tags (item 13), restricting deploys to main (items 6, 9), Terraform state (item 10), hand-made resources (item 12), forge-et-iaac housekeeping including closing PRs #1 and #2 (item 14), the svc-github key replacement by 11 Jan 2027 (item 9), svc-harness (item 7) and production (F38-i); “[ET] et-iaac repo fixes” — remove the rejected master_user_secret_rotation_days line from base/rds, set the 365-day schedule on the RDS-managed secret with an aws_secretsmanager_secret_rotation in the dev overlay (or behind a variable only dev sets), turn the dev.yaml apply step back on once plan passes, and, before go-live, restart the backend after each successful rotation and test it with a forced rotation (item 1); and “[ET] Consider dropping DB password rotation”, the lead’s question whether rotation is needed at all, which the lead decided the same day: no scheduled rotation (item 1; F38 Rev 14). The first card lists every point for reference, not as a commitment to do them all (the lead, 16 Oct): some may not be needed and will be skipped, and which are needed, and in what order, is reviewed as the work goes on, mostly after the other two cards. Some of its points are ones this section already records as decided otherwise for dev — uploads and mailbox infrastructure (the delivery team’s PRs, “The delivery team’s side” below), the account safety basics and the Terraform state move (left out of dev by F38 Rev 13), and the hand-made resources list (compiled by the delivery team on 15 Oct, item 12); those decisions are unchanged unless the review changes them. Since 18 Oct (F38 Rev 18) also: restricting deploys to main, the 90-day key replacement, the budgets and the cost allocation tags are dropped for the delivery team, the database fixes are optional in dev, and the svc-harness production check moved to F38-i; the rotation and forge-et-iaac fixes are in the delivery team’s forge-et-iaac#3, merged and applied on 20 Oct.) (Update 22 Oct 2026 — the board as the lead exported it on 21 Oct: the cards above and “[ET] Lengthen the DB password rotation schedule…”, “[ET] Implement end-to-end dev environment workflow…”, “[ET] Expire old build images in ECR”, “[ET] Security headers on the CRM frontend” and “[ET] Set up SES for the CRM’s system emails…” are in Done. “[ET] Misc tasks” no longer holds the twelve items of 16 Oct: its description now lists what forge-et-iaac#3 and #4 added, so the dev items that were left out of dev or made optional (F38 Revs 13, 18) are on no card, and the production ones are on “[ET] Production environment and go-live”. Still open: “[ET] Bedrock quota increase for Sonnet 5.5 and Haiku 5.5 (currently 0), and more than 10 requests a minute for Haiku 4.5” (In Review, item 4), “[ET] forge-et-pwa deployment in dev env” (the capture app’s hosting, item 5), “[ET] UAT environment — temporary, one per acceptance round” (F38-h) and “[ET] Production environment and go-live” (F38-i).)

(Update 13 Oct 2026 — a read-only review of the dev account, recorded in full in .forge/audits/2026-10-13-aws-dev-account-audit.md. Its corrections to the record above: “Checked against F38” row 4’s backups are disabled (retention 0, no snapshots), not the AWS default; the AgentCore Gateway in “Not there yet” is built (et-dev-gateway, READY, 13 Oct); and item 7 is overtaken (below). Its most urgent findings, in order: (1) item 1 — the RDS secret rotates every 7 days and next on 16 Oct 23:59:59 UTC, with nothing restarting the backend; (2) item 7 — svc-harness now holds IAMFullAccess, a console password without MFA, and an access key, so it can make itself an administrator, and the account has no password policy; (3) item 3 — the database has no backups, no encryption and no deletion protection, and encrypting it by changing storage_encrypted would replace it with the data lost, so it is done by snapshot, encrypted copy and restore; (4) item 9 — one CI user, svc-github, with AdministratorAccess, runs both Terraform and the backend deploy; (5) item 13 — no CloudTrail, no account S3 Block Public Access, no budgets, no SNS topic or alarms. Dev applied migration V41 on 12 Oct; sign-in is still blocked by the placeholder Entra values (item 2).)

(Update 14 Oct 2026 — re-read in the account as svc-harness, and in forge-et-iaac at 9eac094, forge-et-be at dfc9be15 and forge-et-fe at bf4adc8. Done on 13 Oct: item 2 (sign-in on dev works; CRM01 signed in on https://dev.crm.empire.lk); the CRM frontend’s hosting and deploy, part of items 5 and 9. Moved but not done: item 4 — the 5.5 model agreements are accepted, but both models have a quota of 0 tokens a minute, so they are still refused. Unchanged and urgent: item 1 (rotation, next 16 Oct 23:59:59 UTC); items 3 and 13 (no encryption, no CloudTrail, no budgets, no alarms; backups on dev are optional, the DevOps team’s call). Found: the frontend deploy sets no cache headers (item 6, row 15 (e)). Accepted for dev by the lead rather than fixed: item 7 (svc-harness’s broad access) and item 9 (one administrator CI key in three repositories — row 11). Malinda tracks the DevOps team’s tasks on a Trello board, list “Project ET TODOs”; the items below are the source for its cards, and this list stays the harness record.) (The lead’s decisions of 14 Oct 2026, applied below and recorded in F38 Rev 9: the DevOps team’s direct pushes to app-repo main are accepted and checked after the fact (item 6); Shyamin raises the Bedrock quota increase (item 4); one administrator svc-github key for the three repositories is kept for dev (item 9); svc-harness keeps its broad access for dev, and Claude uses it for reads only, asking the lead before any call that is not a read, a test invocation included (item 7); backups on dev, and MFA on svc-harness’s console password, are optional — the DevOps team’s call, not pushed — while production keeps both (item 3, item 7; F38 Rev 11).)

  1. Database password rotation Done 20 Oct 2026 (item 1). The delivery team’s forge-et-iaac#3 (F38-b spec Revs 1, 3, 4) removed 76fce06’s rejected line, turned automatic rotation off in base/rds (aws_secretsmanager_secret_rotation with rotation_enabled = false) and restored the apply step; Shyamin approved it, the lead merged it on 20 Oct (d230b133), and its dev.yaml run 39475850266 applied it. Read on 21 Oct: the RDS-managed secret shows RotationEnabled: false and no next rotation date. Nothing is left for the DevOps team. The “Shyamin Ayesh makes the change”, “still owed, before production … force a new ECS deployment” and “Also owed in forge-et-iaac” passages below are superseded — no restart is built (Rev 14). Decided by the lead, 16 Oct 2026 (F38 Rev 14, answering OQ-K) — applied in dev on 20 Oct: the password is not rotated on a schedule in any environment. RDS keeps managing it; its automatic rotation is set in Terraform, in the module every environment uses, to the longest interval Secrets Manager accepts (or off, if AWS allows that for a secret RDS manages); no restart on rotation is built; and the password is changed by hand with a forced redeploy when a leak is suspected or someone who could read it leaves (.forge/guides/aws-setup-explained.md §7). It replaces the restart asked for below and the dev-only schedule and pre-go-live restart on the card “[ET] et-iaac repo fixes”. Shyamin Ayesh makes the change (cards “[ET] Consider dropping DB password rotation” and “[ET] et-iaac repo fixes”); the broken 76fce06 line and the commented-out apply step are fixed with it. As first asked: restart the backend after each rotation. Read the RDS secret’s rotation schedule (done 14 Oct: every 7 days); until a restart is wired, lengthen the schedule well past the milestone (done 15 Oct by Shyamin with the AWS CLI: every 365 days, with no immediate rotation; read in the account on 16 Oct — next run 15 Oct 2027 23:59:59 UTC, last rotation still 9 Oct); still owed, before production (F38-i): force a new ECS deployment when a rotation succeeds, so the schedule can go back to a short interval (F38 FR-10, OQ-K). Production does not copy the dev workaround: a yearly rotation still breaks a backend nothing restarts, once a year, when nobody expects it, and a change made outside Terraform is lost when the database is replaced. Also owed in forge-et-iaac (since 76fce06, 15 Oct): master_user_secret_rotation_days in base/rds/instance.tf is not an argument aws_db_instance accepts, so every terraform plan on dev fails; the delivery team’s F38-b wave-1 PR removes it (the lead, 17 Oct; F38-b spec Rev 1), and the schedule in code stays the DevOps team’s: remove it, or keep the schedule in code in the dev overlay only, for example with an aws_secretsmanager_secret_rotation on the RDS-managed secret (not yet checked with terraform validate); and put back the dev.yaml apply step the same commit commented out, because until then nothing pushed there reaches the dev estate. If a rotation happens before the restart is wired (one run by hand, for instance), restart the backend (aws ecs update-service --cluster et-dev --service backend --force-new-deployment — a change, so Shyamin or the lead runs it; .forge/guides/aws-setup-explained.md §7). Without AWS access, any backend deploy does the same: a merge to forge-et-be main, or running its build workflow by hand on main (workflow_dispatch, added by F38-d on 19 Oct; before that, re-running the last successful run was the way) — the team’s way to pick up a password changed by hand. Done: the password is managed by RDS, never in Terraform’s state. (Update 2026-10-26, F38-b WI-4.1: re-read — rotation still off; “F38-b — the dev estate, checked”, AC-3.)
  2. Sign-in on dev Done 13 Oct 2026 — the four Entra values in Secrets Manager, read through secrets references (forge-et-iaac 259230c, forge-et-be dfc9be15), and CRM01 signed in on dev.crm.empire.lk. As asked: (item 7; register step 15). The non-prod private key and certificate copied from the vault into dev’s Secrets Manager; referenced under secrets in the task definition, never environment, with the execution role allowed to read them; the tenant and client IDs as plain values (§2.1); the production key waits for F38-i.
  3. Database and network fixes (items 2–4). (Decided 18 Oct by the lead, F38 Rev 18: in dev, the private subnets, narrowing 5432 to the backend, gp3 with autoscaling, and deletion protection and the final snapshot per environment are optional, the DevOps team’s call; all are required in production, F38-i. The CloudWatch alarms were F38-e’s to decide: on 23 Oct the lead decided there are none in dev, F38 Rev 23.) Encryption with a customer-managed KMS key (dropped 17 Oct 2026 by the lead: dev stays unencrypted — F38 Rev 17); gp3, 20 GiB growing to 100; backups 14 days on production (on dev and UAT optional — the decision below); 5432 open only to the backend’s security group; private subnets in the backend’s zone; deletion protection and the final snapshot as per-environment settings; the four CloudWatch alarms to the team by SNS. Done: the backend’s security group admits only the load balancer. Settled: db.t3.micro is fine for dev. Backups on dev (decided 14 Oct): optional — the DevOps team’s call, not pushed; production keeps 14 days with point-in-time recovery (F38 Rev 11, a post-issue note on the proposal). (Update 2026-10-26, F38-b WI-4.1: none of the optional dev changes was made, and on 2026-10-26 the lead took the DevOps team as making no more of them in dev (F38-b spec Rev 5); the database and network are recorded as found in “F38-b — the dev estate, checked”, AC-2 and AC-3. The production half stays with F38-i.)
  4. Bedrock (§3; register §C-5). Now (26 Oct 2026): AWS has answered the lead’s reply with no increase — the account keeps new-account limits (Sonnet 5.5 and Haiku 5.5 at 0, Haiku 4.5 at 10 requests a minute) until it has a few billing cycles of Bedrock usage and spend, with no date (the 26 Oct update at the end of this item; what to do about it is .forge/owed.md, “The AI layer under Bedrock’s new-account limits”). As asked on 14 Oct: Shyamin Ayesh requests the quota increase for Sonnet 5.5 and Haiku 5.5 (0 tokens a minute) and for Haiku 4.5’s requests (10 a minute) in dev, and sets up a Service Quotas request template in the management account so accounts created later get the same quotas (the 14 Oct notes below). As first asked: Kavindu Anthony, as account holder, asks AWS to enable Sonnet 5.5 and Haiku 5.5 (a support case or the AWS account team), and who sends it and when is agreed; then one call to each in dev, and in production once it exists; both models’ quotas read in dev and production and raised where low; the team’s address added as an alternate contact on the account; bedrock:InvokeModel on both inference profiles for the backend’s task role when the AI features need it (17 Oct: the delivery team writes this grant in its F38-b PR — F38 Rev 16, F38-b FR-11). Done: the use-case form, and Haiku 4.5 subscribed. (Re-checked 13 Oct 2026, afternoon, as svc-harness: a one-line call to global.anthropic.claude-sonnet-5-5 and to us.anthropic.claude-haiku-5-5 still returns “is not available for this account … contact AWS Sales”.) (Update 14 Oct 2026: Shyamin accepted both models’ agreements, which now read AVAILABLE, but Dinil found the calls still refused, and a re-check the same day found why: Service Quotas gives Sonnet 5.5 and Haiku 5.5 0 tokens a minute on every cross-Region and Global quota. What is left is a quota increase for both models — through Service Quotas or an AWS support case — and Haiku 4.5’s 10 requests a minute raised too. Shyamin Ayesh raises it (the lead, 14 Oct). A higher quota is a ceiling, not a charge: Bedrock bills only the tokens used, so the increase adds no cost by itself (the budget alerts in item 13 guard against runaway use). Quotas are per account and per Region, so an increase in the management account raises only its own; dev needs its own request. AWS Organizations’ quota request templates (Service Quotas, set up in the management account, up to 10 quotas, needing an Organization with all features on) file the listed increases automatically for every account created afterwards — not for existing ones — so a template now would cover the production and UAT accounts when they are made. The Anthropic first-time-use form, by contrast, is filed once at the management account and inherited by every account (AWS “Request access to models”, read 14 Oct).) (Update 15 Oct 2026, F38-a WI-1.1: the quotas before the increase are recorded — 0 tokens a minute for both 5.5 models, 10 requests a minute for Haiku 4.5, and no increase requested in dev yet; Service Quotas has no requests-per-minute quota for either 5.5 model, so only their tokens per minute can be raised (§4 “F38-a — the development account, checked”, AC-4). The increase is still Shyamin Ayesh’s.) (Update 16 Oct 2026, read 05:55 UTC as svc-harness: on 15 Oct at 11:15 UTC Shyamin Ayesh filed six Service Quotas requests in dev (Trello card comment) — Sonnet 5.5 L-94A31E46 and L-31AB82D0 to 6,000,000 tokens a minute, Haiku 5.5 L-47520EB7 and L-F26CA8D1 to 10,000,000, Haiku 4.5 L-CCA5DF70 and L-E5084BBA to 10,000 requests a minute — each at the AWS default, because Service Quotas refused a value below it. All six read CASE_CLOSED, closed between 11:33 and 11:40 UTC the same day, and every quota is unchanged (0, 0, 0, 0, 10, 10), so AWS has not granted the increase and Sonnet 5.5 and Haiku 5.5 are still refused. Case ids 179206308400541, 179206308600465, 179206308600039, 179206308700165, 179206308700734, 179206344900729. AWS’s reason is in each case’s correspondence in the AWS Support Center console, which this identity cannot read (the Support API needs a paid support plan). Next: Shyamin reads the cases’ replies and acts on them — usually answering AWS’s questions or opening a “Service limit increase” case for Bedrock in the Support Center with the use case and expected volume. F38-a WI-2.1 does not run until a quota is raised.) (Update 16 Oct 2026, from Shyamin Ayesh through the lead: AWS’s replies to the six cases never reached him, because no email address was set up where AWS sends them. He has set one up and will resubmit the quota requests; the quotas were still 0, 0, 0, 0 and 10 when read on 16 Oct.) (Update 22 Oct 2026, read as svc-harness on 21 Oct at 21:13 UTC, and from the Trello card “[ET] Bedrock quota increase for Sonnet 5.5 and Haiku 5.5 …” as exported on 21 Oct: AWS refused three more requests on 21 Oct, each filed at the AWS default — Haiku 4.5’s Global requests per minute (L-E5084BBA) to 10,000 at 06:16 UTC, case 179256341700132, answered “We are unable to approve your request at this time … Accessibility is subject to change automatically as time passes”; Haiku 5.5’s Global tokens per minute (L-F26CA8D1) to 10,000,000 at 09:24, case 179257464900828; and Sonnet 5.5’s Global tokens per minute (L-31AB82D0) to 6,000,000 at 09:25, case 179257474300142. All three were closed within 16 minutes, and no quota changed. Every request, with its case number, is in “F38-a — the development account, checked”, AC-4, the 21 Oct 21:13 UTC entry. The lead’s reply, posted on the card at 09:02 UTC for Shyamin Ayesh to send through a Support “Service limit increase” case, asks AWS to reconsider at much smaller values: Haiku 4.5 at 100 requests a minute on its cross-Region and Global quotas (from 10), and Sonnet 5.5 and Haiku 5.5 at 400,000 tokens a minute on both of their quotas (from 0) — or any non-zero quota for the two 5.5 models now. Its justification: the account is Empire Teas’ development account for a CRM used by about 15 staff, with a production account to follow; AI is a first-class part of the CRM, called on Bedrock so the data stays in Empire Teas’ account, and a person approves every output; about 1,500–2,000 calls and USD 30–60 a month in production, with short peaks (the morning briefings, a batch of up to about 200 leads, trade shows); and the application’s cost controls. Two of those controls are not in place today: dev has no AWS budget (aws budgets describe-budgets returns none, 21 Oct), and the monthly AI spending ceiling is not built (.forge/owed.md, “The AI on/off switch and monthly ceiling need a Phase 2 owner”). The asks are production’s W-6 figures, double F38-a FR-7’s dev targets, and include Sonnet 5.5’s cross-Region quota, which FR-7 does not list; AC-4 is met by any quota that lets its two calls answer, so the spec does not change. The two requests of 09:24 and 09:25 were filed after the reply was posted; whether the Support case has been sent, and AWS’s answer, are not recorded yet. F38-a WI-2.1 still waits.) (Update 22 Oct 2026, later — read as svc-harness at 04:59 UTC: no quota has changed, Haiku 4.5’s included (still 10 requests a minute on both its quotas), and Service Quotas holds no request after 21 Oct 09:25 UTC. The lead reports that Shyamin Ayesh asked AWS again on 21 Oct for Sonnet 5.5 and Haiku 5.5; as it is not in Service Quotas’ history it was presumably the Support case, which dev cannot read, and its answer comes through Shyamin. Which quotas matter: the backend calls Haiku 4.5 through its us. profile, whose quota is L-CCA5DF70, and Sonnet 5.5 through its global. profile, L-31AB82D0; WI-2.1’s Haiku 5.5 call uses us., L-47520EB7. The two Haiku requests filed through Service Quotas on 21 Oct were for the Global quotas instead (L-E5084BBA, L-F26CA8D1); the Support-case reply asks for both quotas of each model, so a grant at those values covers what the backend and WI-2.1 call. The table is in AC-4, the 22 Oct 04:59 UTC entry.) (Update 22 Oct 2026, from Shyamin Ayesh through the lead: the “asked again on 21 Oct” above was the two Service Quotas requests of 09:24 and 09:25 UTC, not a Support case. AWS’s emailed answer to the Sonnet 5.5 one — “Global cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5 to 6000000”, so case 179257474300142 — is the same refusal as case 179256341700132’s: “We are unable to approve your request at this time … Accessibility is subject to change automatically as time passes … please contact your Account Manager or Solutions Architect.” Shyamin then sent AWS the lead’s reply, the text on the Trello card. Read as svc-harness at 11:52 UTC: every quota still unchanged and no new Service Quotas request; AWS’s answer to the reply is awaited.) (Re-read 25 Oct 2026 as svc-harness: unchanged — L-CCA5DF70 and L-E5084BBA 10 requests a minute, L-31AB82D0, L-F26CA8D1 and L-47520EB7 0 tokens a minute; Service Quotas’ newest request is still 21 Oct 09:25 UTC; no answer to the reply has been reported. The task role’s Bedrock grant still names only the Haiku 4.5 us. profile and the Sonnet 5.5 global. profile, so Haiku 5.5 would also need a grant.) (Update 26 Oct 2026 — AWS’s answer to the lead’s reply, an AWS Support email reported in a session that day; it resolves the case and names no case number: no increase. AWS says Sonnet 5.5, Haiku 4.5, the other 5.5 models and several newer models need “a consistent usage and spend history” on the account over a few billing cycles; a new account has lower requests-per-minute quotas than the defaults Service Quotas shows; AWS reviews the account automatically and lifts the restriction once usage and spend are enough, after which increases can be requested through the usual channels. It suggests building that history by using the Bedrock models already available, which the delivery team does not do without the lead (AD #19; the spend is Empire Teas’). The same day Shyamin Ayesh filed two more Service Quotas requests: Opus 5.5’s cross-Region tokens a minute L-A4430697 to 30,000,000 at 09:52 UTC, case 179300836700331 (Opus 5.5 is not in the design), and Sonnet 5.5’s cross-Region tokens a minute L-94A31E46 to 6,000,000 at 09:53 UTC, case 179300838800048; both CASE_CLOSED at 10:08 UTC. Read afterwards as svc-harness: L-CCA5DF70 and L-E5084BBA 10 requests a minute; L-31AB82D0, L-94A31E46, L-F26CA8D1, L-47520EB7 and L-A4430697 0 tokens a minute. F38-a WI-2.1 waits with no date.) (Later on 26 Oct: F38-a spec Rev 4 moved WI-2.1’s two calls to the register’s §C-5 row “Sonnet 5.5 and Haiku 5.5 answer one call each in dev”, so F38-a closes without them; every check that waits on the 5.5 models is listed in ai-provider-strategy.md §4d.) (Re-read 2 Nov 2026 as svc-harness, reads only: unchanged — L-CCA5DF70 and L-E5084BBA 10 requests a minute, L-58BE175A and L-9A11C666 5,000,000 tokens a minute; Haiku 4.5’s batch inference quotas exist (L-28612175 100,000 records a job, L-EDFC771A at least 100). Both Haiku 4.5 inference profiles, us. and global.anthropic.claude-haiku-4-5-20251001-v1:0, read ACTIVE. The task role et-dev-backend-task allows the us. Haiku 4.5 profile and the Global Sonnet 5.5 and Haiku 5.5 profiles, not the Global Haiku 4.5 profile, which F43 adds (.forge/specs/F43-ai-quota-backlog-and-busy-answer-spec.md).) (Update 11 Nov 2026, the lead: a new Trello card asks Shyamin Ayesh to (1) set up the quota request template in the management account with dev’s six requests of 15 Oct at the AWS defaults (F38-i decision 28), and (2) make one more dev request, a Support “Service limit increase” case opened from the management account and naming dev, at the 22 Oct reply’s figures, as an exception to the 26 Oct “no further requests”. The template covers only accounts created after it, and AWS still reviews each request, so production should expect the same new-account limits as dev. Neither step can lower a quota or remove model access in dev. Tracked in .forge/owed.md, the row “Sonnet 5.5 and Haiku 5.5 refused for the AWS account”.) (Update 19 Nov 2026, read as svc-harness: on 17 Nov two Service Quotas requests were made in dev for Haiku 4.5 at 10,000 requests a minute, cross-Region L-CCA5DF70 (case 179490479200086) and Global L-E5084BBA (case 179490496300492); Service Quotas shows both CASE_CLOSED within about 20 minutes, and on 19 Nov every quota is unchanged: Haiku 4.5 at 10 requests a minute on each route, Sonnet 5.5 and Haiku 5.5 at 0 tokens a minute on all four of their quotas. No request for the 5.5 models was made, and no Support case at the 22 Oct reply’s figures shows in dev’s history.)
  5. The rest of the dev estate Done for dev 2026-10-26 (F38-b WI-4.1; the note at the end of this item) (“Not there yet” above). (The off-hours schedule, F38 FR-16, is deferred until development is complete — F38 Rev 10.) KMS keys (the storage key and the mailbox token key with rotation; no database or backups key in dev — F38 Rev 17); the attachments bucket (versioned, public access blocked, storage-key encryption, no GuardDuty); the mailbox ingest queue with its dead-letter queue at maxReceiveCount 3 (no job queue — F38 Rev 16); EventBridge Scheduler’s schedules and its role (sqs:SendMessage on the scheduler queue only); the task-role grants as each lands (S3 including ListBucket and object versions, KMS on the mailbox and storage keys, consuming from and sending to the ingest queue); S3 + CloudFront for both frontends with the dev. and capture.dev. records and the bucket names and distribution IDs for the pipelines; the AgentCore Gateway and its URL for Dinil Gamage; when the team may raise its own Terraform PRs, and which parts are the team’s (CloudTrail, certificates, SES, budgets) (settled 14 Oct, “The delivery team’s side” below; budgets are not the delivery team’s — F38 Rev 18). Done: the KMS keys, the bucket, the ingest queue, the scheduler and the task-role grants, from the delivery team’s forge-et-iaac#3, applied on 20 Oct (each read in the account on 21 Oct; “What it holds” above); the load balancer, service, task definition and roles; the backend pipeline’s names (crm/backend, et-dev, backend); the AgentCore Gateway and its URL (13 Oct, et-dev-gateway, from forge-et-iaac base/agentcore/ — checked and used for SP-ARCH-004 stage 2 the same day); S3 + CloudFront and the dev.crm.empire.lk record for the CRM frontend, with its bucket name and distribution ID in forge-et-fe’s workflow (13 Oct). Done 22 Oct: the capture app’s hosting and capture.dev. record, from the delivery team’s forge-et-iaac#5 (F38-b spec Rev 8), with the bucket name and distribution ID as outputs for forge-et-pwa’s workflow (“What it holds” above). Still open: the delivery team still writes the new resources’ values into forge-et-be .aws/td.dev.json (F38-d FR-6, its WI-4.1; “Checked against F38” row 14). Done 22 Oct: forge-et-be#300 wrote them, and the running backend:62 holds them (read 25 Oct; the backend deploy paragraph above). (Update 2026-10-21, F38-b WI-2.1: the keys, the bucket, the ingest queue, the schedules and the grants are recorded as applied, with the three values, in “F38-b — the dev estate, checked” below.) (Update 2026-10-26, F38-b WI-4.1: every ask in this item is done for dev — the capture app’s hosting serves its build since F38-d’s first capture deploy of 25 Oct, with its four headers and its bucket refusing direct requests (“F38-b — the dev estate, checked”, AC-7); the off-hours schedule stays deferred (F38 Rev 10).)
  6. The backend deploy workflow and task definition (items 5, 6, 8–10, 13, 14). Restrict forge-et-be’s dev environment to main (dropped 18 Oct by the lead, F38 Rev 18: no branch rule on the dev environments; a workflow_dispatch trigger on forge-et-be’s build.yaml instead, so dev can be redeployed by hand — added by F38-d, 66c9149d, 19 Oct); a concurrency group per environment, so an older run can never deploy after a newer one (“Checked against F38” row 9 (a)); the deploy step waiting for the service to become stable, so a rolled-back deploy shows red (row 9 (b)) and trivy image, report-only (both withdrawn for dev by the lead, 2026-10-20, F38-d spec Rev 3); full-commit-SHA tags with a skip when ECR has the image, and ECR IMMUTABLE; persist-credentials: false, timeout-minutes, a concurrency group that never cancels a run; separate build and deploy jobs and the name deploy-dev; ARM64 for the task and the build, together; stopTimeout 60, deregistration delay 30 s, JAVA_TOOL_OPTIONS; ET_MESSAGING_MAILBOX_INGEST_QUEUE (the queue exists since 20 Oct); the same workflow rules for forge-et-fe’s deploy, added 13 Oct (“Checked against F38” row 15), and for forge-et-pwa’s when it is written; in forge-et-fe’s, the staged S3 sync with cache headers (done by forge-et-fe#259, BUG-136, which also added a concurrency group that never cancels a run). Done: the trigger on main with a deploy; the dev environment holding the key; the Entra values as secrets references (13 Oct). Settled (the lead, 14 Oct): the DevOps team may keep pushing these files straight to app-repo main; the delivery team checks each push after the fact (“Checked against F38” row 13). (Update 2026-10-19, F38-d WI-1.1: the full review of the four deploy files is done — “F38-d — the dev deploy pipelines, checked” below, DR-1 – DR-35. The fixes in F38-d’s scope are its waves 2–5 (WI-2.1, WI-2.2, WI-3.1, WI-4.1, WI-5.1); dev.yaml’s are recorded, not changed, as items 8–10 of the .forge/owed.md row “DevOps follow-ups from F38-b’s review of forge-et-iaac”; the rest are .forge/owed.md rows.) (Update 2026-10-23, F38-d WI-4.2: wave 2’s fixes merged on 19 Oct — forge-et-be#275 (the concurrency group, timeout-minutes, persist-credentials: false, full-commit-SHA tags with the skip, separate build and deploy jobs, ARM64 for the task and the build) and forge-et-fe#298 (the same rules for the CRM’s deploy, with the guard step) — and their runs on dev are recorded in “F38-d — the dev deploy pipelines, checked”, AC-1 – AC-4 and AC-12. Still open in this item: ECR IMMUTABLE and the 30 s deregistration delay (F38-d WI-3.1); the task-definition values above (F38-d WI-4.1); forge-et-pwa’s deploy (F38-d WI-5.1); and the workflow and step names (DR-27, .forge/owed.md).) (Update 2026-10-26, F38-d WI-6.1: ECR IMMUTABLE and the 30 s deregistration delay were applied by forge-et-iaac#6 (23 Oct), the task-definition values by forge-et-be#300 (22 Oct) and forge-et-be#306 (25 Oct), and forge-et-pwa’s deploy by forge-et-pwa#91 (25 Oct); the deliberate checks on dev — a manual run, a queued run, the skip, two failed deploys rolled back, the CRM guard, the first upload — are recorded in “F38-d — the dev deploy pipelines, checked”. Still open in this item: the capture app’s deploy records (F38-d WI-7.1) and the workflow and step names (DR-27).)
  7. (Update 13 Oct 2026, morning: aws login fails for svc-harness with “Authentication failed – Invalid request” through both the us-east-1 and the eu-north-1 sign-in endpoints, and the user could not create its own access key. Later that day Shyamin Ayesh attached IAMFullAccess to it “so you can create the keys now, don’t create any more users”, and the lead created a key, used as the CLI profile et-dev for read-only checks. Its policies are now ReadOnlyAccess, IAMFullAccess, AmazonBedrockFullAccess and AWSMarketplaceFullAccess, with a console password and no MFA — enough to make itself an administrator. Still asked: remove IAMFullAccess now the key exists, remove the console password or add MFA, narrow the Marketplace policy, and an account password policy; or the Identity Center read-only permission set below, used through aws configure sso, in its place. Review: .forge/audits/2026-10-13-aws-dev-account-audit.md C2.) (Update 13 Oct 2026, afternoon, for SP-ARCH-004 stage 2: svc-harness could not call the AgentCore Gateway — AmazonBedrockFullAccess grants bedrock:*, not bedrock-agentcore:* — so, with Shyamin Ayesh’s agreement (“you can self grant permission”), Dinil added an inline policy spike-invoke-gateway allowing only bedrock-agentcore:InvokeGateway on et-dev-gateway. It is outside Terraform and comes off when the spike closes (owed.md). Also asked: narrow AmazonBedrockFullAccess to invoking models, since it lets svc-harness buy provisioned throughput. Read the same day: the gateway role et-dev-agentcore-gateway’s trust policy carries aws:SourceAccount = 570417736903 and an aws:SourceArn condition on this account’s gateways, which the audit’s Low item says it lacks — re-check before acting on that item.) (Decided by the lead, 14 Oct 2026 (F38 Rev 9): svc-harness keeps its broad access in dev for now — IAMFullAccess included — because going back to the DevOps team for each permission slows the work. The condition: Claude uses it for read-only calls, and asks the lead for explicit permission before any call that is not a read, a test invocation included (.claude/CLAUDE.md § ASK FIRST). So the narrowing asks below are withdrawn for dev; what stays is that it reaches dev only, never production, and that keys are shared through the password vault. MFA on its console password, or removing the password, is optional in dev — Shyamin’s call, not pushed (the lead, 14 Oct; F38 Rev 11); production people’s logins use Identity Center with MFA (FR-2).) svc-harness (§3). What is still asked, after the 14 Oct decision: confirm it reaches dev only, never production (moved to F38-i by F38 Rev 18: svc-harness is a dev-account user, so this is checked when the production account exists — no production role may trust it); share future keys through the password vault, not as a file or in Slack. Settled or withdrawn: Dinil’s failing login (an access key replaced it, 13 Oct); stopgap or permanent (permanent for dev, a named exception to F38 NFR-1 — Rev 9); removing IAMFullAccess, the console password or the Marketplace permissions, narrowing AmazonBedrockFullAccess and an Identity Center user for it (all withdrawn for dev by the decision above; MFA optional, Rev 11). The earlier notes in this item, from 13 Oct, record those asks as they stood.
  8. Logins for the team (§3, F38-a FR-2). (Decided 14 Oct 2026 by the lead, on Shyamin’s suggestion of 13 Oct: in dev the team keeps using svc-harness’s access key through the CLI profile et-dev, shared through the password vault; no per-person Identity Center logins for dev, and no more IAM users. The dev asks below are withdrawn; production’s stay — F38 Rev 9, Rev 11.) Identity Center logins with MFA for the four developers; one permission set that can change dev and one that only reads; production read-only for the team; a date.
  9. CI users (item 11; F38 FR-19). The deploy user narrowed to the deploy actions; a separate Terraform user whose key the app repositories never hold; dev environments and the key in forge-et-fe and forge-et-pwa; each dev environment restricted to main. Done: the key is in forge-et-be’s dev environment secrets, and since 13 Oct in forge-et-fe’s with the two VITE_* variables. Found 14 Oct: the key in all three repositories is the same svc-github key, which holds AdministratorAccess (“Checked against F38” row 11). Decided by the lead, 14 Oct: accepted for dev; the narrowed deploy user and the separate Terraform user are withdrawn for dev and stay required for production (F38-i). Decided by the lead, 25 Oct (F38 Rev 26): the temporary uat instance, which runs in the dev account, uses the same svc-github key in each uat Environment; no UAT deploy or Terraform user is created. What remains for dev: forge-et-pwa’s dev environment and key. Done 23 Oct: forge-et-pwa’s dev environment holds the two VITE_* variables (the delivery team’s) and the key (Shyamin Ayesh, 11:49 UTC), so nothing remains for dev here. (Dropped 18 Oct by the lead, F38 Rev 18: restricting each dev environment to main, and replacing the svc-github key every 90 days in dev; production’s CI users and their key replacement are F38-i’s.)
  10. Terraform state Recorded 15 Oct 2026, F38-a AC-3 (§4 “F38-a — the development account, checked”); production’s state bucket is F38-i’s. As asked: (F38-a FR-5). A state bucket per account in us-east-1 (versioned, customer-managed KMS key, TLS only, public access blocked), dev’s state moved into it, and who does it when. Today it is et-ops, in the management account, in us-west-2. Decided for dev by the lead, 14 Oct (F38 Rev 13): dev’s state stays in et-ops; nothing to ask. Production’s own state bucket is F38-i’s.
  11. Requests AWS approves by hand (register steps 9–10; F38-a FR-7, FR-8). SES domain verification and production access in dev and production — each environment sends from its own domain, confirmed by the lead on 15 Oct (F38-a W-2): dev from noreply@dev.crm.empire.lk (identity dev.crm.empire.lk), production from noreply@crm.empire.lk (identity crm.empire.lk), and no shared mail subdomain such as mail.crm.empire.lk; a subdomain of crm.empire.lk needs no request to Empire IT; dev is still in the SES sandbox with no identity (14 Oct); Bedrock quota requests in production (or the request template, item 4); confirmation that the production account exists and who can reach it. (14 Oct: the lead asked Shyamin Ayesh for dev’s SES set-up, with the database rotation and the Bedrock quotas, and the three are on the DevOps team’s Trello board. Since 15 Oct F38-a covers dev only (F38 Rev 12): dev’s SES identity, with DMARC p=none and no report address, and its production access are F38-a FR-8; production’s SES, quotas and account are F38-i’s. 16 Oct: Shyamin Ayesh asked on the Trello card whether one mail.crm.empire.lk domain would serve every environment; the lead answered there with the per-environment domains above. Later on 16 Oct Shyamin confirmed he goes ahead with noreply@dev.crm.empire.lk and noreply@crm.empire.lk; mail.crm.empire.lk had been shared earlier only as an example, which caused the question. Nothing is set up in SES yet. Which address UAT sends from is F38-h Q-1.) (Update 2026-10-21, F38-a WI-3.1: dev’s SES identity dev.crm.empire.lk and dev’s production access are done — the identity was created on 21 Oct by forge-et-iaac 0eee623 and is verified with 2048-bit Easy DKIM and the MAIL FROM domain bounce.dev.crm.empire.lk, DMARC is p=none with no report address, and production access was granted the same day, case 179256302200929; “F38-a — the development account, checked”, AC-5. The item stays open for production’s SES and quota requests and the production account, F38-i’s.)
  12. Resources made by hand List compiled 15 Oct 2026, F38-a NFR-3 (§4 “F38-a — the development account, checked”, “What was made by hand”); the dns-dev scope confirmation is F38-i’s. As asked: (F38-a NFR-3). A list of what was made outside Terraform (SSO, the crm.empire.lk zone, et-ops, the dns-dev role, and the Entra secret values, set by hand on purpose) and which of it moves into Terraform; confirmation that dns-dev can change only dev.crm.empire.lk names, so dev’s Terraform cannot touch production’s records. (15 Oct: the delivery team compiles the list itself from its reads, F38-a NFR-3; the dns-dev confirmation matters once production’s records exist and moves to F38-i.)
  13. Security basics (F38-a FR-2, FR-4, FR-11; register §C-8). Activate the Project, Environment and Module tags as cost allocation tags in the management account’s Billing console, so Cost Explorer can split the bill by environment and module (“What dev costs” above; added 14 Oct); account-level S3 Block Public Access and an account password policy (neither is set in dev, 14 Oct); security and billing alternate contacts; root access management for the dev and production accounts; CloudTrail per account with its log bucket locked (one year, break-glass role only); AWS Budgets with alerts (none exists, 14 Oct), and who receives them, agreed with Empire Teas (not the delivery team’s since 18 Oct — see the last note). (Decided for dev by the lead, 14–15 Oct (F38 Rev 13; a post-issue note on the proposal): in dev, no CloudTrail trail (AWS’s free 90-day event history is the record), no root protection, password policy, alternate contacts or account-level S3 Block Public Access — none of them is asked for; dev’s budgets and their recipients are the DevOps team’s and Empire Teas’ to arrange, and F38-a only records them. Still asked in dev: activating the cost allocation tags. Production gets every item in this list (F38-i). Checked 15 Oct with IAM’s policy simulator: svc-harness may set the password policy itself but none of the others, and its IAMFullAccess could grant it the rest, which the lead’s read-only condition rules out.) (Update 15 Oct 2026, F38-a WI-1.1: dev still has no AWS budget (§4 “F38-a — the development account, checked”, AC-7); the cost allocation tags are still asked.) (The lead, 18 Oct: the budgets and the cost allocation tags are not the delivery team’s to ask for or track — Shyamin Ayesh, Kavindu Anthony, Malinda Wellalage and Thilina Abeyrathna handle billing settings; the delivery team keeps costs down in its design and daily use.)
  14. forge-et-iaac housekeeping (“To suggest to Shyamin” above; items 8, 12). Kept as is (the lead, 18 Oct): the personal Owner tag; no team or role tag is needed. Done by the delivery team’s forge-et-iaac#3 (F38-b spec Rev 3, Rev 6), merged and applied on 20 Oct: apply only from dev; .gitignore; commit .terraform.lock.hcl; actions/checkout v4; dead-letter queue retention 14 days; an ECR lifecycle rule that keeps the last 100 build-* images (33375bf, 0757750, 18 Oct — the lead moved it from the DevOps team the same day). Done: nothing left running in us-west-2 (only the default VPC remains, audit 13 Oct); close PRs #1 and #2 (closed by 17 Oct). The review of #3 raised more points for the DevOps team (.forge/owed.md, “DevOps follow-ups from F38-b’s review of forge-et-iaac”).
  15. Database access for the team — not urgent; set up the first time it is needed (the lead, 14 Oct 2026). Today nobody outside AWS can reach the dev database: 5432 is open only inside the VPC and nothing passes a connection through. Most checks need no database access (laptops run the same database and migrations; the backend’s logs are in CloudWatch), so it is asked for when a real need comes, for example the F38-f retest’s audit-row checks or a dev-only bug. The chosen way costs nothing: a Session Manager tunnel through the running backend task (ECS Exec), since dev runs around the clock. What Shyamin sets up then: enableExecuteCommand on the backend service; the four ssmmessages permissions (CreateControlChannel, CreateDataChannel, OpenControlChannel, OpenDataChannel) on the task role; ssm:StartSession for svc-harness on the dev cluster’s tasks and the AWS-StartPortForwardingSessionToRemoteHost document; and a read-only PostgreSQL login for people, dev_readonly, with its password in Secrets Manager (a database change, so the lead approves it; a dev_readwrite login only if a real need comes). Not chosen: a small EC2 access host (about USD 7 a month) and a public database limited to our addresses (about USD 3.65 a month, and on the internet). Production gets no standing access — break-glass only (F38-i Open Questions). How a developer, or Claude, connects from a Mac or Windows laptop is .forge/guides/aws-setup-explained.md §5. Update 2026-10-25: the need has come — F38-f builds it (F38-f spec FR-7, decided by the lead the same day; F38 Rev 25). It is built as described here, by a delivery-team forge-et-iaac PR to dev that Shyamin approves (unless he prefers to make it himself — F38-f Q-3). Every part sits behind a variable that defaults to off and is on only in overlays/dev, because base/ecs and the secrets modules are shared with UAT and production. The one new running cost is the dev_readonly secret, about USD 0.40 a month. Shyamin or the lead creates the login and sets its secret’s value; Claude never reads a database secret. Not built yet. Update 2026-10-26: not built. Shyamin agreed to it for dev only, noting that opening a path to the database is not the DevOps team’s usual practice; the lead then chose to read F38-f’s database checks from the backend’s log lines instead (Flyway’s startup lines; the audit service’s DEBUG line, switched on for a short window through /actuator/loggers) — F38-f spec Rev 4, F38 Rev 28. This item is back to “set up the first time it is needed”, for a later dev-only need.

The delivery team’s side — tracked here, not on Trello (the lead, 14 Oct: Trello holds only the DevOps team’s tasks). Each item has its home; this is the index:

  • BUG-136, the frontend deploy’s missing cache headers — fixed by forge-et-fe#259 (.forge/bugs.md, tracker.yaml).
  • The deploy workflow fixes in item 6 that the DevOps team has not done: a concurrency group, timeout-minutes, immutable full-SHA tags, separate build and deploy jobs (the backend’s stable-service wait and trivy image were withdrawn for dev on 2026-10-20, F38-d spec Rev 3). They are F38-d’s scope, and we make them by PR (the DevOps team may push them instead) — F38-d In Scope. (Update 2026-10-23, F38-d WI-4.2: the concurrency group, timeout-minutes, full-SHA tags and separate build and deploy jobs are done by forge-et-be#275 and forge-et-fe#298 (19 Oct). Still open: ECR IMMUTABLE (F38-d WI-3.1) and the deploy workflows’ names (DR-27, .forge/owed.md).)
  • Owner of F38-a, F38-b and F38-d: Tharinda (from 14 Oct; F38-a handed over from Dinil) — tracker.yaml. F38-a is now “check and close” in the dev account; its production half moved to F38-i (F38 Rev 12).
  • Who writes the rest of the dev estate (the lead, 14 Oct). The delivery team writes the application-facing Terraform the backend depends on — the attachments bucket, the storage and mailbox token KMS keys, the mailbox ingest queue with its dead-letter queue (no job queue — F38 Rev 16), EventBridge Scheduler and the task-role grants for them (item 5 above; “Checked against F38” row 14) — as PRs to forge-et-iaac dev, each merged only after Shyamin Ayesh approves it, by anyone with write access (the lead, 21 Oct; .claude/rules/forge-et-iaac.md). Account-level work stays with the DevOps team: SES and its production access, the Bedrock quotas, the production account and its baseline, and anything in the management account. F38-b carries the team’s PRs.
  • The dev database stays unencrypted (the lead, 17 Oct; F38 Rev 17, F38-b spec Rev 1). No encrypted restore and no database or backups key in dev, so our Terraform no longer has to land before a DevOps restore; production’s encryption is decided by Shyamin Ayesh, Thilina Abeyrathna and Malinda Wellalage (.forge/owed.md). F38-b’s wave-1 PR, forge-et-iaac#3, also removed 76fce06’s rejected rotation line, so terraform validate passes on dev again, turned the managed secret’s rotation off and restored the apply step (item 1 above; spec Revs 3–4); it was merged and applied on 20 Oct.
  • The Foundation milestone of Fri 16 Oct is not fully met. It includes the cloud environments and the deploy pipeline; as of 15 Oct the backend and the CRM frontend deploy to dev and sign-in works, but uploads (the attachments bucket and KMS keys), the mailbox (its ingest queue and token key), EventBridge Scheduler, the capture app’s hosting and pipeline, SES, and the Bedrock quotas for Sonnet 5.5 and Haiku 5.5 are not there. The work is tracked as F38-a, F38-b and F38-d. Decided by the lead, 15 Oct: not raised with Empire Teas separately. (Since 20 Oct, forge-et-iaac#3: the attachments bucket, the KMS keys, the ingest queue and EventBridge Scheduler exist; uploads and the mailbox still wait on their values in the task definition, and the capture app, SES and the 5.5 quotas are unchanged.) (Update 2026-10-21, F38-a WI-3.1: SES is done for dev — the identity dev.crm.empire.lk is verified and production access was granted on 21 Oct; “F38-a — the development account, checked”, AC-5. The capture app and the 5.5 quotas are unchanged.) (Update 2026-10-22: the capture app’s hosting exists — forge-et-iaac#5, applied the same day; its pipeline, F38-d wave 5, waits on forge-et-pwa’s dev Environment. The 5.5 quotas are unchanged.) (Update 2026-10-25: uploads’ and the mailbox’s values are in the running task definition since 22 Oct (forge-et-be#300), the schedules run, and forge-et-pwa’s dev Environment is complete since 23 Oct, so the capture pipeline waits only on the delivery team’s F38-d wave 5. Still missing: uploads and the mailbox proved by use on dev, the capture app’s deploy, and the 5.5 quotas, unchanged.)
  • Decisions waiting on the lead: the search provider (Exa; Thilina approved it on 15 Oct, and Empire Teas’ answer is awaited), re-deriving the cost annexe, what happens to dev after development, and the notes from the 6 Oct call — .forge/owed.md. (Confirmed by the lead at F38-a’s approval, 2026-10-15: the SES sending addresses noreply@dev.crm.empire.lk and noreply@crm.empire.lk, and UAT in the dev account — F38 OQ-E and OQ-A.)
  • Our checks of the DevOps team’s work: each direct push to an app repo’s main (.claude/rules/git-conventions.md § Branch Protection), and the reconcile of this section whenever forge-et-iaac changes (.forge/owed.md, “Keep the harness in step”). A full review of the workflows and task definition the DevOps team pushed without a review (forge-et-be and forge-et-fe build.yaml, forge-et-be .aws/td.dev.json, forge-et-iaac dev.yaml) is owed under F38-d, and its findings are recorded here (.forge/owed.md, “Review the GitHub Actions workflows the DevOps team wrote”; the lead, 17 Oct). (Update 2026-10-19, F38-d WI-1.1: the review is done — “F38-d — the dev deploy pipelines, checked” below, DR-1 – DR-35, each with its home; no direct push since 14 Oct.)
  • When the DevOps team has done item 15: the lead approves the dev_readonly database login.

Settled since 6 Oct, not to reopen: the database password is managed by RDS; Multi-AZ is off for now (Empire decides for production, register §C-8); the crm.empire.lk zone stays in the management account, and dev writes its records through dns-dev (so the proposal to move the zone into the production account is dropped — F38 FR-6 already keeps the zone where it was created); GuardDuty is off (agenda item 8 below). Added 14 Oct (the lead): the DevOps team’s direct pushes to app-repo main, checked after the fact; one administrator svc-github key for the three repositories in dev; svc-harness’s broad access in dev, used read-only by Claude; Shyamin raises the Bedrock quota increase. Added 15 Oct (the lead, F38 Rev 12 and Rev 13): F38-a covers dev only, with production’s half in F38-i; in dev no CloudTrail trail, root protection, password policy, alternate contacts or account-level S3 Block Public Access; dev’s state stays in et-ops; dev’s budgets are the DevOps team’s and Empire Teas’. Added 16–18 Oct (the lead, F38 Revs 14, 17, 18): the database password is not rotated on a schedule in any environment (rotation off in dev since forge-et-iaac#3 was applied on 20 Oct); the dev database stays unencrypted (production undecided); in dev the database’s private subnets, the narrowed 5432 rule, gp3 with autoscaling, deletion protection and the final snapshot are optional; no branch rule on the forge-et-be and forge-et-fe dev GitHub Environments (forge-et-iaac’s is the DevOps team’s call — .forge/owed.md, “DevOps follow-ups…” (1)); no 90-day svc-github key replacement in dev; billing settings are Shyamin Ayesh’s, Kavindu Anthony’s, Malinda Wellalage’s and Thilina Abeyrathna’s.

F38-a — the development account, checked

Opened 15 Oct 2026 by F38-a-WI-1.1. This section is F38-a’s evidence for the development account (570417736903): the record its acceptance criteria are ticked against (spec .forge/specs/F38-a-account-baseline-spec.md). The rules: every AWS command is a read — describe, list or get, plus sts get-caller-identity, and, only with the lead’s permission at the time, iam generate-credential-report and cloudtrail lookup-events (each entry names the permission) — made with the CLI profile et-dev (the IAM user svc-harness), in us-east-1 unless the command names another Region; times are UTC; account IDs, ARNs, resource names and case numbers are recorded, but no secret value, no access-key ID and no CloudTrail event body (spec NFR-2; the Secrets boundary at the top of this file). Each acceptance criterion has one #### subsection, in AC order, added by the wave that records it: WI-1.1 wrote AC-1, AC-2, AC-3, AC-4’s preparatory records, AC-6, AC-7 and the NFR-3 list; WI-2.1 adds its entries to AC-4 (written here directly; the separate record file its plan names, .forge/plans/F38-a/F38-a-WI-2.1-record.md, was never created); WI-3.1 inserts an AC-5 subsection between AC-4 and AC-6.

Every entry has the same shape: a bold line with the date and UTC time, who ran the read and for which work item, and the profile; a text block holding $ <command>, the output lines that show the result (trimmed with … where lines were cut, never edited) and EXIT=<n>; then Result: — met, not met — <why>, or reported to the lead — .forge/owed.md. Where one command was run in a loop, the block shows it once and a table gives each item’s output.

AC-1 — Organization and Region

2026-10-15 06:38 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws organizations describe-organization --query 'Organization.[Id,MasterAccountId,FeatureSet]' --output text
o-imaviqd8nu	331342166625	ALL
EXIT=0

Result: met — dev is a member of the Organization o-imaviqd8nu, whose management account is 331342166625, with all features on.

2026-10-15 06:38–06:39 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws ec2 describe-regions --query 'Regions[].RegionName' --output text
ap-south-1	eu-north-1	eu-west-3	eu-west-2	eu-west-1	ap-northeast-3	ap-northeast-2	ap-northeast-1	ca-central-1	sa-east-1	ap-southeast-1	ap-southeast-2	eu-central-1	us-east-1	us-east-2	us-west-1	us-west-2
EXIT=0
$ for r in "${regions[@]}"; do
    aws resourcegroupstaggingapi get-resources --region "$r" --query 'ResourceTagMappingList[].ResourceARN' --output text
    aws ec2 describe-vpcs --region "$r" --query 'Vpcs[].[VpcId,IsDefault]' --output text
  done
EXIT=0 for each of the 34 commands (17 Regions × 2)
Region Tagged resources (get-resources) VPCs (describe-vpcs)
us-east-1 31 — three security groups, two subnets, the VPC, the internet gateway, the route table; ECR crm/backend; RDS et-dev and its subnet group; S3 et-dev-spa; the load balancer et-dev-backend, its two listeners and target group; six Secrets Manager secrets (the four et-dev-entra-*, the RDS-managed rds!db-6a52f694-…, and dev/rds/password — next paragraph); the log group /ecs/dev/backend; SQS et-dev-scheduler and et-dev-scheduler-dlq; the AgentCore gateway et-dev-gateway-…; the two ACM certificates; the CloudFront distribution E2RVC3893O10TJ (a global service); the ECS cluster et-dev and service backend vpc-0ea763d5098b71922 (not default — base/networking); no default VPC
The other 16 Regions none one default VPC each (IsDefault True)
$ aws s3api list-buckets --query 'Buckets[].Name' --output text
et-dev-spa
EXIT=0
$ aws s3api get-bucket-location --bucket et-dev-spa --output text
None
EXIT=0

None is how S3 reports us-east-1. Result: met — no dev resource outside us-east-1 apart from the global services (CloudFront, IAM) and the default VPCs; et-ops, the state bucket in us-west-2, is in the management account, not in dev (AC-3). The limit of this check: the Resource Groups Tagging API returns only resources that carry, or once carried, a tag. The Terraform estate tags every resource by default (overlays/dev/providers.tf, default_tags), and the per-Region describe-vpcs and the bucket reads cover the two kinds most likely to be made by hand; an untagged resource of another kind made by hand would not be found by this read.

Two observations, neither a finding. dev/rds/password is the secret forge-et-iaac created on 5 Oct (9e00dc7) and removed when RDS took over the master password (ac1efdc); it is scheduled for deletion, which AWS completes at the end of its recovery window:

$ aws secretsmanager list-secrets --include-planned-deletion --query 'SecretList[].[Name,CreatedDate,DeletedDate,OwningService]' --output text
dev/rds/password	2026-10-05T13:25:04.258000+05:30	2026-10-09T14:07:02.125000+05:30	None
rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e	2026-10-09T14:02:21.775000+05:30	None	rds
et-dev-entra-certificate	2026-10-13T12:49:48.979000+05:30	None	None
…  (et-dev-entra-private-key, et-dev-entra-tenant-id, et-dev-entra-client-id, the same)
EXIT=0

IAM Identity Center’s home Region is us-west-2 (the path of the role AWSReservedSSO_ADMINISTRATORS_…, /aws-reserved/sso.amazonaws.com/us-west-2/, AC-2). The instance belongs to the management account, not to dev, and the 13 Oct audit already records it (.forge/audits/2026-10-13-aws-dev-account-audit.md §2, the Identity Center row); production’s logins are F38-i’s.

AC-2 — identities, event history, Shield

2026-10-15 06:39 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws iam get-credential-report --query '[GeneratedTime,ReportFormat]' --output text
aws: [ERROR]: An error occurred (ReportNotPresent) when calling the GetCredentialReport operation: Unknown
EXIT=254

No report existed at 06:39, and aws iam generate-credential-report is not a describe, list or get call, so .claude/CLAUDE.md § ASK FIRST needed the lead’s permission first (WI-1.1 plan D-3). It was generated and read once the permission was given (the next entry).

2026-10-15 09:11 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness); permission: the lead, Tharinda Rajapaksha, 2026-10-15 09:10 UTC, given in the orchestrator session and passed on by the orchestrator, for generate-credential-report and lookup-events once each (plan D-3, route A)

$ aws iam generate-credential-report --query '[State,Description]' --output text        (09:11:32)
STARTED	No report exists. Starting a new report generation task
EXIT=0
$ aws iam generate-credential-report --query '[State,Description]' --output text        (09:11:37, polling)
COMPLETE	None
EXIT=0
$ aws iam get-credential-report --query '[GeneratedTime,ReportFormat]' --output text
2026-10-15T09:11:35+00:00	text/csv
EXIT=0
$ aws iam get-credential-report --query 'Content' --output text > <a file outside the repository>
EXIT=0
$ base64 -d … | cut -d, -f1,3,4,8,9,10,11,14,15,16          (plan D-5's columns, decoded outside the repository)
user,user_creation_time,password_enabled,mfa_active,access_key_1_active,access_key_1_last_rotated,access_key_1_last_used_date,access_key_2_active,access_key_2_last_rotated,access_key_2_last_used_date
<root_account>,2026-09-30T05:31:55Z,false,false,false,N/A,N/A,false,N/A,N/A
svc-github,2026-09-30T05:43:23Z,false,false,true,2026-10-13T06:55:37Z,2026-10-15T08:08:00Z,false,N/A,N/A
svc-harness,2026-10-06T09:17:10Z,true,false,true,2026-10-13T05:27:28Z,2026-10-15T06:45:00Z,false,N/A,N/A

Result: met — the report was generated at 09:11:35, after the generate-credential-report call at 09:11:32, so it is this read’s own. Only svc-github and svc-harness have an access key, one each, both active and both created on 13 Oct; neither has a second key. The root user has no access key, no password and no MFA, which is expected for a member account created by Organizations and settled for dev by F38 Rev 13. svc-harness has a console password without MFA, optional in dev (F38 Rev 11). This matches the per-user reads below. Nothing to report to the lead.

2026-10-15 06:39–06:41 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws iam list-users --query 'Users[].[UserName,CreateDate,PasswordLastUsed]' --output text
svc-github	2026-09-30T05:43:23+00:00	None
svc-harness	2026-10-06T09:17:10+00:00	2026-10-13T05:39:02+00:00
EXIT=0
$ aws iam get-account-summary --query 'SummaryMap.[AccountAccessKeysPresent,AccountMFAEnabled,Users,Groups,Roles,Policies]' --output text
0	0	2	0	13	0
EXIT=0
$ for u in svc-github svc-harness; do
    aws iam list-attached-user-policies --user-name "$u" --query 'AttachedPolicies[].PolicyName' --output text
    aws iam list-user-policies --user-name "$u" --query 'PolicyNames' --output text
    aws iam list-groups-for-user --user-name "$u" --query 'Groups[].GroupName' --output text
    aws iam list-access-keys --user-name "$u" --query 'AccessKeyMetadata[].[UserName,Status,CreateDate]' --output text
    aws iam get-login-profile --user-name "$u" --query 'LoginProfile.CreateDate' --output text
    aws iam list-mfa-devices --user-name "$u" --query 'length(MFADevices)' --output text
    aws iam get-access-key-last-used --access-key-id <the key's ID, passed in a shell variable and never written down> \
      --query '[UserName,AccessKeyLastUsed.LastUsedDate,AccessKeyLastUsed.ServiceName,AccessKeyLastUsed.Region]' --output text
  done
EXIT=0 for each, except get-login-profile for svc-github: EXIT=254, NoSuchEntity (no console password)
User Created Attached policies Inline policies Groups Access keys Key last used Console password MFA devices
svc-github 30 Sep 2026 05:43 AdministratorAccess none none one, Active, created 13 Oct 2026 06:55 15 Oct 2026 06:05, ecs, us-east-1 none 0
svc-harness 6 Oct 2026 09:17 AWSMarketplaceFullAccess, IAMFullAccess, ReadOnlyAccess, AmazonBedrockFullAccess spike-invoke-gateway none one, Active, created 13 Oct 2026 05:27 15 Oct 2026 06:38, organizations, us-east-1 (this read) yes, created 6 Oct 2026, last used 13 Oct 2026 05:39 0

Result: met for the identities — the only IAM users are svc-github and svc-harness, each with one key, both named dev exceptions (F38 Rev 9, Rev 11); AccountAccessKeysPresent is 0, so the root user has no access key; svc-harness’s console password without MFA is optional in dev (§4 item 7, F38 Rev 11); its one inline policy is the temporary spike-invoke-gateway (NFR-3 below). Root MFA off (AccountMFAEnabled 0) and no password policy are expected in dev (F38 Rev 13). Nothing to report to the lead. The credential report (the entry above) agrees.

Event history. As expected in dev, there is no trail:

$ aws cloudtrail describe-trails --query 'trailList[].Name' --output text
EXIT=0          (no output: no trail — F38 Rev 13)

2026-10-15 09:11 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness); permission: the lead, Tharinda Rajapaksha, 2026-10-15 09:10 UTC (the same permission as the credential report)

$ aws cloudtrail lookup-events --max-results 1 --query 'Events[0].[EventTime,EventName,EventSource]' --output text
2026-10-15T14:39:36+05:30	DescribeTargetHealth	elasticloadbalancing.amazonaws.com
EXIT=0

Result: met — the 90-day event history answers, with an event from 09:09:36 UTC the same morning. Only its time, name and source are recorded; the event body, which carries the caller’s access-key ID, was not read (plan D-4).

2026-10-15 06:40 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws shield get-subscription-state --output text
INACTIVE
EXIT=0
$ aws guardduty list-detectors --query 'DetectorIds' --output text
EXIT=0          (no output: no detector)
$ aws s3control get-public-access-block --account-id 570417736903 --output text
aws: [ERROR]: An error occurred (NoSuchPublicAccessBlockConfiguration) when calling the GetPublicAccessBlock operation: The public access block configuration was not found
EXIT=254

Result: met. Shield Standard is present: AWS applies it to every account at no charge, and INACTIVE is the state of Shield Advanced, the paid subscription, which is not taken (spec FR-4). GuardDuty has no detector — not in dev by F38 Rev 7. Account-level S3 Block Public Access is not set — not in dev by F38 Rev 13; et-dev-spa blocks public access on the bucket itself (§4 “What it holds”, base/s3), and F38-b checks the buckets it owns.

AC-3 — Terraform state

2026-10-15 06:37–06:41 UTC — Claude (F38-a-WI-1.1), forge-et-iaac read through git, no AWS call

$ git -C forge-et-iaac fetch origin
EXIT=0
$ git -C forge-et-iaac rev-parse --short origin/dev
9eac094
$ git -C forge-et-iaac show origin/dev:overlays/dev/providers.tf
…
  backend "s3" {
    bucket       = "et-ops"
    key          = "dev/terraform.tfstate"
    region       = "us-west-2"
    encrypt      = true
    use_lockfile = true
  }
…
EXIT=0

Result: met — dev’s state is in the bucket et-ops (in the management account), key dev/terraform.tfstate, Region us-west-2, encrypted, locked by S3’s own lock file, at forge-et-iaac dev commit 9eac094. Its staying in et-ops is a named dev exception (F38 Rev 13, spec FR-5); production’s own state bucket is F38-i’s.

AC-4 — Bedrock

WI-1.1 recorded the starting point below — the models, the quotas before the increase, and the one-time steps already taken. The quotas after the increase, each request’s case number and the calls to Sonnet 5.5 and Haiku 5.5 are WI-2.1’s, added here.

2026-10-15 06:41–06:42 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws bedrock list-foundation-models --by-provider anthropic --query 'modelSummaries[].modelId' --output text
EXIT=0          (one tab-separated line; the three models' ids in it, one per line:)
anthropic.claude-haiku-4-5-20251001-v1:0
anthropic.claude-haiku-5-5
anthropic.claude-sonnet-5-5
$ aws bedrock list-inference-profiles --query 'inferenceProfileSummaries[].inferenceProfileId' --output text
EXIT=0          (one tab-separated line; the three models' profiles in it, one per line:)
global.anthropic.claude-haiku-4-5-20251001-v1:0
us.anthropic.claude-haiku-4-5-20251001-v1:0
us.anthropic.claude-sonnet-5-5
us.anthropic.claude-haiku-5-5
global.anthropic.claude-haiku-5-5
global.anthropic.claude-sonnet-5-5
$ aws bedrock get-foundation-model-availability --model-id anthropic.claude-sonnet-5-5 --output json
{
    "modelId": "anthropic.claude-sonnet-5-5",
    "agreementAvailability": {
        "status": "AVAILABLE"
    },
    "authorizationStatus": "AUTHORIZED",
    "entitlementAvailability": "AVAILABLE",
    "regionAvailability": "AVAILABLE"
}
EXIT=0

The same read for anthropic.claude-haiku-5-5 and anthropic.claude-haiku-4-5-20251001-v1:0 returned the same four values (AVAILABLE, AUTHORIZED, AVAILABLE, AVAILABLE), each with EXIT=0.

The quotas before the increase. No increase had been requested in dev when this was read, so these are live values, not defaults:

$ aws service-quotas list-service-quotas --service-code bedrock \
    --query "Quotas[?contains(QuotaName, 'Sonnet 5.5') || contains(QuotaName, 'Haiku 5.5') || contains(QuotaName, 'Haiku 4.5')].[QuotaCode,QuotaName,Value]" --output text
L-CCA5DF70	Cross-region model inference requests per minute for Anthropic Claude Haiku 4.5	10.0
L-47520EB7	Cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5	0.0
L-58BE175A	Cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5	5000000.0
L-9A11C666	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5	5000000.0
…  (six batch-inference quotas for Haiku 4.5, not used)
L-94A31E46	Cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5	0.0
L-E5084BBA	Global cross-region model inference requests per minute for Anthropic Claude Haiku 4.5	10.0
L-31AB82D0	Global cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5	0.0
L-F26CA8D1	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5	0.0
EXIT=0                                                     (read 06:43)
$ aws service-quotas list-requested-service-quota-change-history --service-code bedrock \
    --query 'RequestedQuotas[].[QuotaCode,QuotaName,DesiredValue,Status,Created,CaseId]' --output text
EXIT=0          (no output: no quota increase has ever been requested in dev; read 06:45)
$ aws service-quotas list-aws-default-service-quotas --service-code bedrock \
    --query "Quotas[?contains(QuotaName, 'Sonnet 5.5') || contains(QuotaName, 'Haiku 5.5') || (contains(QuotaName, 'Haiku 4.5') && contains(QuotaName, 'per minute'))].[QuotaCode,QuotaName,Value]" --output text
L-CCA5DF70	Cross-region model inference requests per minute for Anthropic Claude Haiku 4.5	10000.0
L-47520EB7	Cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5	10000000.0
L-58BE175A	Cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5	5000000.0
L-9A11C666	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5	5000000.0
L-94A31E46	Cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5	6000000.0
L-E5084BBA	Global cross-region model inference requests per minute for Anthropic Claude Haiku 4.5	10000.0
L-31AB82D0	Global cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5	6000000.0
L-F26CA8D1	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5	10000000.0
EXIT=0                                                     (read 06:45)
Model Quota (code) Profile it governs (in use by F38-a FR-7 or F12) Dev, 15 Oct 06:43 AWS default
Sonnet 5.5 Global cross-Region tokens per minute (L-31AB82D0) global.anthropic.claude-sonnet-5-5 — in use 0 6,000,000
Sonnet 5.5 Cross-Region tokens per minute (L-94A31E46) us.anthropic.claude-sonnet-5-5 0 6,000,000
Haiku 5.5 Cross-Region tokens per minute (L-47520EB7) us.anthropic.claude-haiku-5-5 — in use 0 10,000,000
Haiku 5.5 Global cross-Region tokens per minute (L-F26CA8D1) global.anthropic.claude-haiku-5-5 — in use 0 10,000,000
Haiku 4.5 Cross-Region requests per minute (L-CCA5DF70) us.anthropic.claude-haiku-4-5-20251001-v1:0 — in use 10 10,000
Haiku 4.5 Cross-Region tokens per minute (L-58BE175A) us.anthropic.claude-haiku-4-5-20251001-v1:0 — in use 5,000,000 5,000,000
Haiku 4.5 Global cross-Region requests per minute (L-E5084BBA) global.anthropic.claude-haiku-4-5-20251001-v1:0 10 10,000
Haiku 4.5 Global cross-Region tokens per minute (L-9A11C666) global.anthropic.claude-haiku-4-5-20251001-v1:0 5,000,000 5,000,000

These match the reading of 14–15 Oct (§3, the 14 Oct update; §4 item 4; the spec’s Context). Two points for the request (spec FR-7, Appendix A.3): Service Quotas has no requests-per-minute quota for Sonnet 5.5 or Haiku 5.5 in dev — only tokens per minute — so the requests-per-minute targets FR-7 gives those two models have no quota to request; and Haiku 4.5’s tokens-per-minute quotas (5,000,000) are already above the target of about 200,000, so only its requests per minute (10) need raising. The dev account’s values sit far below AWS’s defaults, which is AWS’s reduced quota for a new account (the §C-5 model-access row, “new accounts may receive reduced quotas”).

2026-10-15 06:45 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws bedrock get-use-case-for-model-access > /dev/null
EXIT=0          (output not recorded — plan D-4; exit 0 means a use case is on file and readable from dev)
$ aws service-quotas get-association-for-service-quota-template --output text
aws: [ERROR]: An error occurred (AccessDeniedException) when calling the GetAssociationForServiceQuotaTemplate operation: The request was called by a member account. To make this request, you must use the organization's master account.
EXIT=254

The one-time steps already taken (the dates as far as they are known; the DevOps team gave none for this record — plan D-8):

Step State Date Source
Anthropic’s first-time-use form Done — a use case is on file and readable from dev (the read above); filed once at the management account and inherited by every member account Before 9 Oct 2026; the exact date as far as the DevOps team knows: not given §4 item 4 (“Done: the use-case form”); §3
Haiku 4.5 Marketplace subscription Done Before 9 Oct 2026; the exact date: not given §3 (9 Oct: “Claude Haiku 4.5 works and is subscribed”); §4 item 4
Haiku 4.5’s completion A call answered on 9 Oct 2026; no new call was made by WI-1.1 (a test invocation needs the lead’s permission) 9 Oct 2026 §3 (the DevOps channel, 9 Oct); register §C-5
Sonnet 5.5 and Haiku 5.5 agreements Accepted by Shyamin Ayesh; both models read AVAILABLE and AUTHORIZED (the read above) 14 Oct 2026 §3 (the 14 Oct update)
The management account’s quota request template Not readable from dev (AccessDeniedException, a member account); the DevOps team’s answer: not given — spec FR-7 (F38-a records it and does not wait on it)

Result: the preparatory records for AC-4 are complete; AC-4 itself waits on the increase and the two calls (WI-2.1).

2026-10-21 21:13–21:14 UTC (22 Oct in Sri Lanka) — Claude (the lead’s F38 reconcile session), profile et-dev (svc-harness)

$ aws service-quotas list-service-quotas --service-code bedrock --query "Quotas[?QuotaCode=='L-CCA5DF70'||QuotaCode=='L-E5084BBA'||QuotaCode=='L-94A31E46'||QuotaCode=='L-31AB82D0'||QuotaCode=='L-47520EB7'||QuotaCode=='L-F26CA8D1'].[QuotaCode,Value]" --output text
L-CCA5DF70	10.0
L-47520EB7	0.0
L-94A31E46	0.0
L-E5084BBA	10.0
L-31AB82D0	0.0
L-F26CA8D1	0.0
EXIT=0
$ aws service-quotas list-requested-service-quota-change-history --service-code bedrock --query "RequestedQuotas[?starts_with(Created,'2026-10-21')].[QuotaCode,DesiredValue,Status,Created,LastUpdated,CaseId]" --output text
L-31AB82D0	6000000.0	CASE_CLOSED	2026-10-21T14:55:41.874000+05:30	2026-10-21T15:11:34.391000+05:30	179257474300142
L-F26CA8D1	10000000.0	CASE_CLOSED	2026-10-21T14:54:08.394000+05:30	2026-10-21T15:10:02.194000+05:30	179257464900828
L-E5084BBA	10000.0	CASE_CLOSED	2026-10-21T11:46:55.682000+05:30	2026-10-21T12:02:48.025000+05:30	179256341700132
EXIT=0

Every request made in dev, with the six of 15 Oct from §4 item 4 (all filed at the AWS default, because Service Quotas refuses a lower value):

Quota (code) Requested Filed (UTC) Case Closed (UTC) Quota after
Sonnet 5.5 cross-Region tokens per minute (L-94A31E46) 6,000,000 15 Oct 11:15 179206308400541 15 Oct 11:34 0
Sonnet 5.5 Global tokens per minute (L-31AB82D0) 6,000,000 15 Oct 11:15 179206308600465 15 Oct 11:34 0
Haiku 5.5 cross-Region tokens per minute (L-47520EB7) 10,000,000 15 Oct 11:15 179206308600039 15 Oct 11:33 0
Haiku 5.5 Global tokens per minute (L-F26CA8D1) 10,000,000 15 Oct 11:15 179206308700165 15 Oct 11:33 0
Haiku 4.5 cross-Region requests per minute (L-CCA5DF70) 10,000 15 Oct 11:15 179206308700734 15 Oct 11:34 10
Haiku 4.5 Global requests per minute (L-E5084BBA) 10,000 15 Oct 11:15 179206344900729 15 Oct 11:40 10
Haiku 4.5 Global requests per minute (L-E5084BBA) 10,000 21 Oct 06:16 179256341700132 21 Oct 06:32 10
Haiku 5.5 Global tokens per minute (L-F26CA8D1) 10,000,000 21 Oct 09:24 179257464900828 21 Oct 09:40 0
Sonnet 5.5 Global tokens per minute (L-31AB82D0) 6,000,000 21 Oct 09:25 179257474300142 21 Oct 09:41 0

AWS’s answer to case 179256341700132, as Shyamin Ayesh posted it on the Trello card on 21 Oct: “We are unable to approve your request at this time. Accessibility is subject to change automatically as time passes.” The answers to the other cases are not readable from dev (the account’s support plan does not allow the Support API) and were not posted. What happens next is §4 item 4, the 22 Oct update.

Result: not met — every quota is unchanged after nine requests (Sonnet 5.5 and Haiku 5.5 at 0 tokens a minute on all four quotas, Haiku 4.5 at 10 requests a minute on both); the two calls cannot be made until a quota is raised (WI-2.1).

2026-10-22 04:59–05:00 UTC — Claude (the lead’s F38 reconcile session), profile et-dev (svc-harness)

$ aws service-quotas list-service-quotas --service-code bedrock --query "Quotas[?contains(QuotaName,'Haiku 4.5')||contains(QuotaName,'Sonnet 5.5')||contains(QuotaName,'Haiku 5.5')].[QuotaCode,Value,QuotaName]" --output text
…                 (the batch-inference quotas for Haiku 4.5 trimmed)
L-CCA5DF70	10.0	Cross-region model inference requests per minute for Anthropic Claude Haiku 4.5
L-58BE175A	5000000.0	Cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5
L-47520EB7	0.0	Cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5
L-94A31E46	0.0	Cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5
L-E5084BBA	10.0	Global cross-region model inference requests per minute for Anthropic Claude Haiku 4.5
L-9A11C666	5000000.0	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 4.5
L-F26CA8D1	0.0	Global cross-region model inference tokens per minute for Anthropic Claude Haiku 5.5
L-31AB82D0	0.0	Global cross-region model inference tokens per minute for Anthropic Claude Sonnet 5.5
EXIT=0
$ aws service-quotas list-requested-service-quota-change-history --service-code bedrock --query "RequestedQuotas[?Created>='2026-10-21T15'].[QuotaCode,DesiredValue,Status,Created,LastUpdated,CaseId,QuotaName]" --output text
EXIT=0          (no output: no Service Quotas request since 21 Oct 09:30 UTC)

Which quota each call uses. Each model has a cross-Region quota (its us. profile) and a Global one (its global. profile), and a call counts only against the quota of the profile it names. The backend calls Haiku 4.5 through us. for the calls a person waits for and, since F43 (2026-11-04), through global. for lead triage and lead research where the task definition sets it, as dev’s does (forge-et-be .aws/td.dev.json; register §D), and Sonnet 5.5 through global. (forge-et-be application.yml, et.ai.features). The task role on dev may call all six inference profiles of the three planned models since forge-et-iaac#16 was applied on 2026-11-04 (register §B F43 item 4); “F38-b — the dev estate, checked”, AC-11, records the grant of 21 Oct, when it held two; WI-2.1’s proof calls Sonnet 5.5 through global. and Haiku 5.5 through us. (spec FR-7).

Model Quota the backend uses Quota WI-2.1’s proof uses Requested through Service Quotas
Haiku 4.5 L-CCA5DF70 (cross-Region requests per minute) for the calls a person waits for; since F43 (2026-11-04), L-E5084BBA (Global requests per minute) for lead triage and lead research — (its 9 Oct call stands) both on 15 Oct; only L-E5084BBA (Global) on 21 Oct
Sonnet 5.5 L-31AB82D0 (Global tokens per minute) L-31AB82D0 both on 15 Oct; L-31AB82D0 on 21 Oct
Haiku 5.5 none — no feature uses it; the Global profile was chosen on 27 Oct, and the task role’s grant for it came with forge-et-iaac#13, applied on 28 Oct (update 27 Oct: once a feature uses it, L-F26CA8D1; the §C-5 proof call follows the choice) L-47520EB7 (cross-Region tokens per minute) both on 15 Oct; only L-F26CA8D1 (Global) on 21 Oct

Result: not met — every quota is unchanged. The two Haiku requests of 21 Oct were for the Global quotas, which neither the backend nor WI-2.1’s proof uses; the lead’s Support-case reply asks for both quotas of each model, so it covers the ones that matter. When AWS answers, the quota of each row’s middle columns is the one to read.

2026-10-22 11:52 UTC — Claude (the lead’s F38 session), profile et-dev (svc-harness). The same two reads as 04:59: L-CCA5DF70 10, L-E5084BBA 10, L-94A31E46 0, L-31AB82D0 0, L-47520EB7 0, L-F26CA8D1 0 (EXIT=0), and no Service Quotas request created after 21 Oct 15:00 Sri Lanka time (EXIT=0, no output). AWS’s emailed answer to case 179257474300142 (Sonnet 5.5’s Global tokens per minute, 6,000,000), passed on by Shyamin Ayesh: “We are unable to approve your request at this time.”

Result: not met — unchanged; Shyamin Ayesh has sent AWS the lead’s reply (§4 item 4, the last 22 Oct update).

2026-10-26 — Claude (session for the lead), profile et-dev (svc-harness). Read after AWS’s answer to the lead’s reply was reported (§4 item 4, the 26 Oct update). Two more Service Quotas requests, both filed by shyamin@altrium.io through the administrators’ SSO role:

Quota (code) Requested Filed (UTC) Case Closed (UTC) Quota after
Opus 5.5 cross-Region tokens per minute (L-A4430697) 30,000,000 26 Oct 09:52 179300836700331 26 Oct 10:08 0
Sonnet 5.5 cross-Region tokens per minute (L-94A31E46) 6,000,000 26 Oct 09:53 179300838800048 26 Oct 10:08 0

The quotas read with aws service-quotas get-service-quota and list-service-quotas: L-CCA5DF70 10, L-E5084BBA 10, L-31AB82D0 0, L-94A31E46 0, L-F26CA8D1 0, L-47520EB7 0, L-A4430697 0; Haiku 4.5’s tokens per minute L-58BE175A and L-9A11C666 5,000,000.

Result: not met, with no date — AWS keeps the account at new-account limits until it has a few billing cycles of Bedrock usage and spend, and lifts them on its own review. What the team does about it is .forge/owed.md, “The AI layer under Bedrock’s new-account limits”. (Later on 26 Oct, F38-a spec Rev 4: AC-4’s two model calls moved to the register’s §C-5 row “Sonnet 5.5 and Haiku 5.5 answer one call each in dev”, whose result is recorded here when it runs; AC-4’s other bullets, the quota bullet read as the quotas after AWS’s answer, are ticked at F38-a’s close against the entries above.)

AC-5 — SES

Recorded 2026-10-21 by F38-a-WI-3.1. The DevOps team set SES up in dev on 21 Oct 2026: forge-et-iaac commit 0eee623 added the module base/ses and its call in overlays/dev/main.tf, and the dev.yaml run that applied it created the identity; production access was requested and granted the same day. The reads below were made after both.

2026-10-21 08:30 UTC — Claude (F38-a-WI-3.1), profile et-dev (svc-harness)

$ aws sesv2 list-email-identities --query 'EmailIdentities[].[IdentityName,IdentityType,SendingEnabled,VerificationStatus]' --output json
[
    [
        "dev.crm.empire.lk",
        "DOMAIN",
        true,
        "SUCCESS"
    ]
]
EXIT=0
$ aws sesv2 get-email-identity --email-identity dev.crm.empire.lk \
    --query '{IdentityType:IdentityType,VerifiedForSendingStatus:VerifiedForSendingStatus,VerificationStatus:VerificationStatus,Dkim:DkimAttributes,MailFrom:MailFromAttributes,VerificationInfo:VerificationInfo,ConfigurationSetName:ConfigurationSetName}' --output json
{
    "IdentityType": "DOMAIN",
    "VerifiedForSendingStatus": true,
    "VerificationStatus": "SUCCESS",
    "Dkim": {
        "SigningEnabled": true,
        "Status": "SUCCESS",
        "Tokens": [
            "golldexyx5mkqi5ioefrpsphs74isxo6",
            "jgt33g3pzmysgcz7gyunxqz2z3dwbfqb",
            "eotkqktemtloay34bmxfbaw5b3tgbuzk"
        ],
        "SigningHostedZone": "dkim.amazonses.com",
        "SigningAttributesOrigin": "AWS_SES",
        "NextSigningKeyLength": "RSA_2048_BIT",
        "CurrentSigningKeyLength": "RSA_2048_BIT",
        "LastKeyGenerationTimestamp": "2026-10-21T11:33:58.642000+05:30"
    },
    "MailFrom": {
        "MailFromDomain": "bounce.dev.crm.empire.lk",
        "MailFromDomainStatus": "SUCCESS",
        "BehaviorOnMxFailure": "USE_DEFAULT_VALUE"
    },
    "VerificationInfo": {
        "LastCheckedTimestamp": "2026-10-21T11:34:58.947000+05:30",
        "LastSuccessTimestamp": "2026-10-21T11:35:00.171000+05:30"
    },
    "ConfigurationSetName": "et-dev-mail"
}
EXIT=0

Result: met — dev.crm.empire.lk is the only identity in the account; it is verified for sending (VerifiedForSendingStatus true, VerificationStatus SUCCESS), signed by Easy DKIM (SigningAttributesOrigin AWS_SES, Status SUCCESS) with a 2048-bit key (CurrentSigningKeyLength RSA_2048_BIT), and its MAIL FROM domain bounce.dev.crm.empire.lk is SUCCESS. The DKIM key was generated at 06:03:58 UTC on 21 Oct, and SES’s last successful check of the identity was at 06:05:00 UTC. Recorded as returned, not checked by AC-5: the identity’s default configuration set is et-dev-mail, declared in base/ses/identity.tf (the Terraform entry below).

2026-10-21 08:30 UTC — Claude (F38-a-WI-3.1), profile et-dev (svc-harness)

$ aws sesv2 get-account \
    --query '{ProductionAccessEnabled:ProductionAccessEnabled,SendingEnabled:SendingEnabled,EnforcementStatus:EnforcementStatus,SendQuota:SendQuota,SuppressedReasons:SuppressionAttributes.SuppressedReasons,MailType:Details.MailType,Review:Details.ReviewDetails}' --output json
{
    "ProductionAccessEnabled": true,
    "SendingEnabled": true,
    "EnforcementStatus": "HEALTHY",
    "SendQuota": {
        "Max24HourSend": 50000.0,
        "MaxSendRate": 14.0,
        "SentLast24Hours": 0.0
    },
    "SuppressedReasons": [
        "BOUNCE",
        "COMPLAINT"
    ],
    "MailType": "TRANSACTIONAL",
    "Review": {
        "Status": "GRANTED",
        "CaseId": "179256302200929"
    }
}
EXIT=0
$ aws support describe-cases --display-id 179256302200929 --include-resolved-cases --no-include-communications \
    --query 'cases[].[caseId,displayId,status,timeCreated]' --output json
aws: [ERROR]: An error occurred (SubscriptionRequiredException) when calling the DescribeCases operation: Amazon Web Services Premium Support Subscription is required to use this service.
EXIT=254

Result: met — the account is out of the SES sandbox (ProductionAccessEnabled true, review GRANTED on case 179256302200929, requested as TRANSACTIONAL mail), sending is on and HEALTHY, and the account-level suppression list holds BOUNCE and COMPLAINT. The send quota is 50,000 messages a day at 14 a second, with nothing sent in the last 24 hours. The support case itself is not readable from dev (the account has no support plan that allows the Support API), so its request and grant dates come from the DevOps team (the dates table below).

2026-10-21 08:30 UTC — Claude (F38-a-WI-3.1), public DNS

$ dig TXT _dmarc.dev.crm.empire.lk +short
"v=DMARC1; p=none"
EXIT=0
$ dig TXT _dmarc.dev.crm.empire.lk +short @1.1.1.1
"v=DMARC1; p=none"
EXIT=0
$ dig CNAME <token>._domainkey.dev.crm.empire.lk +short          (local resolver, then @1.1.1.1, for each of the three tokens above)
golldexyx5mkqi5ioefrpsphs74isxo6.dkim.amazonses.com.
jgt33g3pzmysgcz7gyunxqz2z3dwbfqb.dkim.amazonses.com.
eotkqktemtloay34bmxfbaw5b3tgbuzk.dkim.amazonses.com.
EXIT=0 for each of the six reads; each token's two answers are the same line, shown once
$ dig MX bounce.dev.crm.empire.lk +short
10 feedback-smtp.us-east-1.amazonses.com.
EXIT=0
$ dig MX bounce.dev.crm.empire.lk +short @1.1.1.1
10 feedback-smtp.us-east-1.amazonses.com.
EXIT=0
$ dig TXT bounce.dev.crm.empire.lk +short
"v=spf1 include:amazonses.com ~all"
EXIT=0
$ dig TXT bounce.dev.crm.empire.lk +short @1.1.1.1
"v=spf1 include:amazonses.com ~all"
EXIT=0

Result: met — _dmarc.dev.crm.empire.lk returns one v=DMARC1 record with p=none and no rua= or ruf= tag, through the local resolver and through Cloudflare; the three DKIM CNAMEs point at SES’s signing zone, and the MAIL FROM domain has SES’s feedback MX for us-east-1 and the SPF record v=spf1 include:amazonses.com ~all. The two resolvers agree on every record, so the zone’s own name server was not read.

2026-10-21 08:31–08:33 UTC — Claude (F38-a-WI-3.1), forge-et-iaac at 0eee623

$ git -C forge-et-iaac grep -n -E 'resource "aws_sesv2_email_identity"|resource "aws_sesv2_email_identity_mail_from_attributes"|resource "aws_ses_domain_identity"|resource "aws_ses_domain_dkim"|resource "aws_ses_domain_mail_from"|next_signing_key_length|_domainkey|_dmarc|amazonses' origin/dev
origin/dev:base/ses/identity.tf:3:resource "aws_sesv2_email_identity" "this" {
origin/dev:base/ses/identity.tf:8:    next_signing_key_length = "RSA_2048_BIT"
origin/dev:base/ses/identity.tf:17:resource "aws_sesv2_email_identity_mail_from_attributes" "this" {
origin/dev:base/ses/records.tf:11:  name    = "${aws_sesv2_email_identity.this.dkim_signing_attributes[0].tokens[count.index]}._domainkey.${var.domain_name}"
origin/dev:base/ses/records.tf:13:  records = ["${aws_sesv2_email_identity.this.dkim_signing_attributes[0].tokens[count.index]}.dkim.amazonses.com"]
origin/dev:base/ses/records.tf:22:  records  = ["10 feedback-smtp.${data.aws_region.current.region}.amazonses.com"]
origin/dev:base/ses/records.tf:31:  records  = ["v=spf1 include:amazonses.com ~all"]
origin/dev:base/ses/records.tf:39:  name     = "_dmarc.${var.domain_name}"
EXIT=0
$ git -C forge-et-iaac show origin/dev:base/ses/identity.tf
…
resource "aws_sesv2_email_identity" "this" {
  email_identity         = var.domain_name
  configuration_set_name = aws_sesv2_configuration_set.this.configuration_set_name

  dkim_signing_attributes {
    next_signing_key_length = "RSA_2048_BIT"
  }
…
resource "aws_sesv2_email_identity_mail_from_attributes" "this" {
  email_identity         = aws_sesv2_email_identity.this.email_identity
  mail_from_domain       = "bounce.${var.domain_name}"
  behavior_on_mx_failure = "USE_DEFAULT_VALUE"
}

resource "aws_sesv2_configuration_set" "this" {
  configuration_set_name = "${var.project}-${var.environment}-mail"
…
  suppression_options {
    suppressed_reasons = ["BOUNCE", "COMPLAINT"]
  }
…
EXIT=0
$ git -C forge-et-iaac show origin/dev:base/ses/records.tf
…
resource "aws_route53_record" "dkim" {
  provider = aws.dns
  count    = 3
…
resource "aws_route53_record" "mail_from_mx" {
  provider = aws.dns
…
resource "aws_route53_record" "mail_from_spf" {
  provider = aws.dns
…
resource "aws_route53_record" "dmarc" {
  provider = aws.dns
  zone_id  = var.zone_id
  name     = "_dmarc.${var.domain_name}"
  type     = "TXT"
  records  = ["v=DMARC1; p=none"]
  ttl      = 600
}
EXIT=0
$ git -C forge-et-iaac grep -n 'base/ses' origin/dev -- overlays/dev
origin/dev:overlays/dev/main.tf:36:  source      = "../../base/ses"
EXIT=0
$ git -C forge-et-iaac show origin/dev:overlays/dev/main.tf          (lines 31–42)
module "ses" {
  project     = var.project
  environment = var.environment
  domain_name = "dev.crm.empire.lk"
  zone_id     = var.dns_zone_id
  source      = "../../base/ses"

  providers = {
    aws     = aws
    aws.dns = aws.dns
  }
}
EXIT=0
$ git -C forge-et-iaac log --reverse --format='%H %cI' -S 'resource "aws_sesv2_email_identity"' origin/dev
0eee6238f19eadd5b7e45666e4906aa479a3b7a9 2026-10-21T11:22:20+05:30
EXIT=0
$ gh run list -R altrium-io/forge-et-iaac --workflow dev.yaml --commit 0eee6238f19eadd5b7e45666e4906aa479a3b7a9 --json databaseId,conclusion,createdAt,headSha,headBranch,event
[{"conclusion":"success","createdAt":"2026-10-21T06:03:21Z","databaseId":39553084687,"event":"push","headBranch":"dev","headSha":"0eee6238f19eadd5b7e45666e4906aa479a3b7a9"}]
EXIT=0
Piece (spec AC-5) Declared at overlays/dev module block Gives dev.crm.empire.lk aws.dns provider
The identity, with 2048-bit Easy DKIM base/ses/identity.tf:3 (aws_sesv2_email_identity.this); key length base/ses/identity.tf:8 (next_signing_key_length = "RSA_2048_BIT") module "ses", overlays/dev/main.tf:31 domain_name = "dev.crm.empire.lk" (literal, overlays/dev/main.tf:34) → var.domain_name n/a — default provider
Its MAIL FROM setting (bounce.dev.crm.empire.lk) base/ses/identity.tf:17 (aws_sesv2_email_identity_mail_from_attributes.this) module "ses" "bounce.${var.domain_name}" n/a — default provider
The three DKIM CNAMEs base/ses/records.tf:6 (aws_route53_record.dkim, count = 3) module "ses" "<token>._domainkey.${var.domain_name}" provider = aws.dns in the resource; aws.dns = aws.dns in the module block
The MAIL FROM MX base/ses/records.tf:17 (aws_route53_record.mail_from_mx) module "ses" the MAIL FROM resource’s mail_from_domain the same
The MAIL FROM SPF TXT base/ses/records.tf:26 (aws_route53_record.mail_from_spf) module "ses" the MAIL FROM resource’s mail_from_domain the same
_dmarc TXT (v=DMARC1; p=none) base/ses/records.tf:36 (aws_route53_record.dmarc) module "ses" "_dmarc.${var.domain_name}" the same

Result: met — at forge-et-iaac origin/dev 0eee623, every piece spec AC-5 names is declared in Terraform: the identity with the key length pinned to RSA_2048_BIT, its MAIL FROM setting, and the five Route 53 resources (three DKIM CNAMEs, the MX, the SPF TXT and _dmarc), each written through aws.dns into the zone var.dns_zone_id (Z0102234ZE7QHA32KA60, overlays/dev/env.auto.tfvars — the dns-dev arrangement, AC-6). 0eee623 (Shyamin Ayesh, 21 Oct) is the first commit that declares the identity, and dev.yaml run 39553084687, started by its push to dev at 06:03:21 UTC, applied it with conclusion success. Nothing SES-related was found only in the account. The same commit also declares, not checked by AC-5: the configuration set et-dev-mail (base/ses/identity.tf:23; reputation metrics on, its own suppression of BOUNCE and COMPLAINT), set as the identity’s default; and the backend task role’s inline policy ses (base/ecs/task_role.tf:126–140), which allows ses:SendEmail and ses:SendRawEmail on the identity and that configuration set only (ses_arns, overlays/dev/main.tf:27).

The dates (spec FR-9):

What Date (UTC) Source
The identity created 2026-10-21, 06:03:58 The dev.yaml run 39553084687 above, started 06:03:21 for 0eee623; LastKeyGenerationTimestamp 06:03:58 (the read above); the DevOps team’s facts as gathered in the lead’s session (CloudTrail: CreateEmailIdentity by svc-github at 06:03:58–59)
Production access requested 2026-10-21, 06:10 The lead’s session, 21 Oct (CloudTrail: PutAccountDetails by Shyamin Ayesh’s console login at 06:10:10); not readable from dev by this WI (describe-cases above)
The gap Same day — about six minutes after the identity existed, as FR-9 asks The two rows above
The delay before both (the lead asked on 14 Oct) Seven days. The DevOps team took the Trello card “[ET] Set up SES for the CRM’s system emails” up on 21 Oct, in the week of 19 Oct as §4’s 16 Oct update expected; no other reason was given The lead’s session, 21 Oct (the DevOps team, through the lead)
Production access granted 2026-10-21 Shyamin Ayesh’s comment of 21 Oct on that Trello card, reported to the lead and passed on in the lead’s session; get-account shows GRANTED at 08:30 UTC the same day
Case number 179256302200929 get-account, Details.ReviewDetails.CaseId (the read above)

A test send, refused — not part of AC-5’s evidence. In the lead’s session, before this work item ran, Claude tried one message through SES from svc-harness. It is the one AWS command in this subsection that is not a read, so it is recorded with its permission:

2026-10-21 08:26:57 UTC — Claude (the lead’s session, before F38-a-WI-3.1), profile et-dev (svc-harness). Permission: the lead, Tharinda Rajapaksha, on 2026-10-21 in that session, for one aws sesv2 send-email to crm01@empire.lk.

$ aws sesv2 send-email --from-email-address noreply@dev.crm.empire.lk --destination ToAddresses=crm01@empire.lk --configuration-set-name et-dev-mail --content 'Simple={Subject={Data="ET dev SES test (F38-e)"},Body={Text={Data="…"}}}' --output json
aws: [ERROR]: An error occurred (AccessDeniedException) when calling the SendEmail operation: User 'arn:aws:iam::570417736903:user/svc-harness' is not authorized to perform 'ses:SendEmail' on resource 'arn:aws:ses:us-east-1:570417736903:identity/dev.crm.empire.lk'
EXIT=254

Result: nothing was sent. svc-harness has no ses:SendEmail permission on the identity, and no other identity or profile was tried. Proving delivery to a mailbox that is not verified in SES, with SPF, DKIM and DMARC pass in the headers, is F38-e’s (spec FR-8; F38 FR-26, AC-26). Who sends the first test message, and how, is the .forge/owed.md row “Send the first test message through dev’s SES”.

The first test message, delivered — not part of AC-5’s evidence. The lead chose to send it from their own terminal (Claude Code’s auto mode refuses an identity widening its own permissions) through a temporary inline policy on svc-harness that allowed only ses:SendEmail from noreply@dev.crm.empire.lk on the identity and et-dev-mail. Three calls that are not reads, each run by the lead, Tharinda Rajapaksha, on 2026-10-21:

2026-10-21 08:40–08:46 UTC — the lead, profile et-dev (svc-harness); times from CloudTrail and the message’s Date header, read by Claude

$ aws iam put-user-policy --user-name svc-harness --policy-name temp-ses-test-send --policy-document file://<scratch>/ses-test-policy.json
                                                       (CloudTrail PutUserPolicy, 08:40:41)
$ aws sesv2 send-email --from-email-address noreply@dev.crm.empire.lk --destination ToAddresses=crm01@empire.lk --configuration-set-name et-dev-mail --content 'Simple={Subject={Data="ET dev SES test (F38-e)"},Body={Text={Data="…"}}}'
{
    "MessageId": "010001a15d9f6e15-19128844-0558-484c-8951-681f6004d26e-000000"
}                                                      (Date header 08:41:31)
$ aws iam delete-user-policy --user-name svc-harness --policy-name temp-ses-test-send
                                                       (CloudTrail DeleteUserPolicy, 08:45:37)
$ aws iam list-user-policies --user-name svc-harness --output text        (Claude, read)
POLICYNAMES	spike-invoke-gateway
EXIT=0

The message reached the CRM01 mailbox’s Inbox (crm01@empire.lk, not verified in SES) at 08:41:34 UTC. Its headers, as the lead copied them from Outlook:

Authentication-Results: mx.microsoft.com 1; spf=pass (sender IP is
 54.240.8.57) smtp.mailfrom=bounce.dev.crm.empire.lk; dkim=pass (signature was
 verified) header.d=dev.crm.empire.lk;dkim=pass (signature was verified)
 header.d=amazonses.com;dmarc=pass action=none
 header.from=dev.crm.empire.lk;compauth=pass reason=100
X-MS-Exchange-Organization-SCL: 1

The DKIM signature for dev.crm.empire.lk uses the selector golldexyx5mkqi5ioefrpsphs74isxo6, one of the identity’s three Easy DKIM tokens.

Result: delivered — SPF passes on the MAIL FROM domain bounce.dev.crm.empire.lk, DKIM passes for dev.crm.empire.lk, and DMARC passes, so the card’s “done when” is met in dev. The temporary policy existed for about five minutes and is gone. This proves SES itself; the backend sending through its task role is still F38-e’s (F38 FR-26, AC-26).

AC-6 — certificates and the zone

2026-10-15 06:41 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness), and public DNS

$ aws acm list-certificates --region us-east-1 --query 'CertificateSummaryList[].[DomainName,Status,NotAfter]' --output text
api.dev.crm.empire.lk	ISSUED	2027-04-25T05:29:59+05:30
dev.crm.empire.lk	ISSUED	2027-04-29T05:29:59+05:30
EXIT=0
$ dig NS crm.empire.lk +short
ns-1313.awsdns-36.org.
ns-879.awsdns-45.net.
ns-344.awsdns-43.com.
ns-1983.awsdns-55.co.uk.
EXIT=0
$ dig NS crm.empire.lk @1.1.1.1 +short
ns-1313.awsdns-36.org.
ns-1983.awsdns-55.co.uk.
ns-344.awsdns-43.com.
ns-879.awsdns-45.net.
EXIT=0
$ aws route53 list-hosted-zones --query 'HostedZones[].Name' --output text
EXIT=0          (no output: no hosted zone in the dev account)
$ git -C forge-et-iaac show origin/dev:overlays/dev/providers.tf      (at 9eac094)
…
# Route53 records for dev.crm.empire.lk live in the ets-org zone; this role is scoped to those names.
provider "aws" {
  alias  = "dns"
  region = var.region
  assume_role {
    role_arn = var.dns_role_arn
  }
}
$ git -C forge-et-iaac show origin/dev:overlays/dev/env.auto.tfvars   (at 9eac094)
…
dns_role_arn = "arn:aws:iam::331342166625:role/dns-dev"
dns_zone_id  = "Z0102234ZE7QHA32KA60"

Result: met — both dev certificates are ISSUED in us-east-1 (they expire on 24 and 28 Apr 2027 UTC; ACM renews DNS-validated certificates on its own while the validation records stay); crm.empire.lk resolves to the zone’s four name servers (§1.1), through the local resolver and through Cloudflare; the dev account holds no hosted zone. The dns-dev arrangement: the crm.empire.lk zone (Z0102234ZE7QHA32KA60) is in the management account 331342166625, and dev’s Terraform writes its dev. records into it through the provider alias dns, which assumes the role arn:aws:iam::331342166625:role/dns-dev; the comment in providers.tf says the role is scoped to the dev.crm.empire.lk names. Confirming that scope, importing the zone into Terraform and the CAA record are F38-i’s (F38 Rev 13); capture.dev.crm.empire.lk’s certificate is F38-b’s, with the capture app’s hosting.

AC-7 — budgets

2026-10-15 06:41 UTC — Claude (F38-a-WI-1.1), profile et-dev (svc-harness)

$ aws budgets describe-budgets --account-id 570417736903 --query 'Budgets[].[BudgetName,BudgetLimit.Amount,BudgetLimit.Unit,TimeUnit]' --output text
None
EXIT=0
$ aws budgets describe-budgets --account-id 570417736903 --output json
EXIT=0          (no output: no budget)

Result: recorded as found — dev has no AWS budget on 15 Oct 2026, so there is no amount and no recipient to record. Dev’s budgets and who receives their alerts are the DevOps team’s and Empire Teas’ to set (spec FR-11, F38 Rev 13), so nothing is reported; §4 item 13 still lists them.

What was made by hand (NFR-3)

Compiled 15 Oct 2026 by F38-a-WI-1.1 from the reads above and forge-et-iaac at 9eac094. What exists for dev that forge-et-iaac does not create, in spec NFR-3’s five groups, with its owner and why it may stay so. “The DevOps team” is Shyamin Ayesh, with Senuthmi Thimansa.

Item Group Owner Why it may stay, or when it moves Found by
The accounts and the Organization — o-imaviqd8nu, management account 331342166625, dev 570417736903 (with the role OrganizationAccountAccessRole that Organizations creates in a member account, 30 Sep 2026) The parent’s NFR-3 exceptions The DevOps team; Empire Teas holds the accounts Accounts and an Organization are created outside Terraform (F38 NFR-3) AC-1; iam list-roles
The IAM Identity Center set-up — in dev, the permission-set role AWSReservedSSO_ADMINISTRATORS_… (30 Sep 2026); the instance is the management account’s, home Region us-west-2 The parent’s NFR-3 exceptions The DevOps team Identity Center is set up in the management account (F38 NFR-3); production’s personal logins are F38-i’s iam list-roles (06:41)
Anthropic’s first-time-use form and the Sonnet 5.5 and Haiku 5.5 agreements The parent’s NFR-3 exceptions The DevOps team One-time account actions, not resources AC-4
The Bedrock quota requests and the SES production-access request The parent’s NFR-3 exceptions The DevOps team Requests, not resources; none had been filed on 15 Oct (AC-4’s empty change history; SES is WI-3.1’s) Update 2026-10-21 (F38-a WI-3.1): SES production access requested 2026-10-21, case 179256302200929, granted 2026-10-21 (AC-5). AC-4; AC-5
The four NS records named crm in the empire.lk zone, in Cloudflare The parent’s NFR-3 exceptions Kavindu Anthony (Empire IT) empire.lk is Empire’s own zone, outside AWS (§1.1) AC-6 (dig NS)
The role dns-dev in the management account Named dev exceptions to AD #23 The DevOps team F38 Rev 13; dev’s Terraform assumes it to write dev. records (overlays/dev/env.auto.tfvars) AC-6
The state bucket et-ops (management account, us-west-2) Named dev exceptions to AD #23 The DevOps team F38 Rev 13; dev’s state stays there (spec FR-5) AC-3
The IAM users svc-github and svc-harness, their keys and svc-harness’s console password Named dev exceptions to AD #23 The DevOps team F38 Rev 9, Rev 11, Rev 13; a key made in Terraform would sit in the state AC-2
The hosted zone crm.empire.lk (Z0102234ZE7QHA32KA60, management account) Moves into Terraform later F38-i Imported before production’s records are written (F38 Rev 13) AC-6
The values of the four et-dev-entra-* secrets By hand on purpose The DevOps team Terraform creates the secrets (base/entra-secrets — base/secrets since F38-b wave 6); a value set by hand passes through no plan, apply log or state (§4 “What it holds”) AC-1 (get-resources)
The inline policy spike-invoke-gateway on svc-harness Temporary Dinil Gamage Removed when the search provider is chosen (.forge/owed.md) AC-2 (list-user-policies)

Nothing further was found. Each read that could have found another hand-made item came back clean: all 31 tagged resources in us-east-1 carry the Project tag (the read below, 06:41); of the 13 IAM roles, nine are service-linked (path /aws-service-role/), three are declared in forge-et-iaac (et-dev-agentcore-gateway in base/agentcore/iam.tf, et-dev-backend-execution and et-dev-backend-task in base/ecs/), and the other two are in the table above; spike-invoke-gateway is the only inline policy. Two things are outside Terraform by design and are not hand-made: the ECS task-definition revisions, which forge-et-be’s build workflow registers from the committed .aws/td.dev.json (the service ignores task-definition changes in Terraform — §4 “What it holds”, base/ecs); and the secret dev/rds/password, which Terraform created and then removed, and which waits out its deletion window (AC-1). No item is reported to the lead.

$ aws resourcegroupstaggingapi get-resources --region us-east-1 \
    --query 'ResourceTagMappingList[?!(Tags[?Key==`Project`])].[ResourceARN,join(`,`,Tags[].Key)]' --output text
EXIT=0          (no output: every tagged resource carries Project)
$ aws resourcegroupstaggingapi get-resources --region us-east-1 --query 'length(ResourceTagMappingList)' --output text
31
EXIT=0
$ aws iam list-roles --query 'Roles[].[RoleName,Path,CreateDate]' --output text
AWSReservedSSO_ADMINISTRATORS_05d2e5facfe0b76e	/aws-reserved/sso.amazonaws.com/us-west-2/	2026-09-30T05:41:22+00:00
AWSServiceRoleForECS	/aws-service-role/ecs.amazonaws.com/	2026-09-30T06:34:52+00:00
AWSServiceRoleForElasticLoadBalancing	/aws-service-role/elasticloadbalancing.amazonaws.com/	2026-10-09T08:51:44+00:00
AWSServiceRoleForOrganizations	/aws-service-role/organizations.amazonaws.com/	2026-09-30T05:31:57+00:00
AWSServiceRoleForRDS	/aws-service-role/rds.amazonaws.com/	2026-10-05T08:00:21+00:00
AWSServiceRoleForResourceExplorer	/aws-service-role/resource-explorer-2.amazonaws.com/	2026-09-30T05:42:48+00:00
AWSServiceRoleForSSO	/aws-service-role/sso.amazonaws.com/	2026-09-30T05:35:27+00:00
AWSServiceRoleForSupport	/aws-service-role/support.amazonaws.com/	2026-09-30T05:31:55+00:00
AWSServiceRoleForTrustedAdvisor	/aws-service-role/trustedadvisor.amazonaws.com/	2026-09-30T05:31:55+00:00
et-dev-agentcore-gateway	/	2026-10-13T05:31:00+00:00
et-dev-backend-execution	/	2026-10-09T07:44:37+00:00
et-dev-backend-task	/	2026-10-09T07:47:53+00:00
OrganizationAccountAccessRole	/	2026-09-30T05:31:55+00:00
EXIT=0
$ git -C forge-et-iaac grep -n 'resource "aws_iam_role"' origin/dev      (at 9eac094)
origin/dev:base/agentcore/iam.tf:1:resource "aws_iam_role" "gateway" {
origin/dev:base/ecs/execution_role.tf:12:resource "aws_iam_role" "execution" {
origin/dev:base/ecs/task_role.tf:3:resource "aws_iam_role" "task" {

F38-b — the dev estate, checked

Opened 2026-10-21 by F38-b-WI-2.1. This section is F38-b’s evidence for the dev estate that forge-et-iaac#3 applied on 20 Oct 2026: the record its acceptance criteria are ticked against (spec .forge/specs/F38-b-dev-estate-spec.md). The rules: every AWS command is a read — a describe, list, get or head call, cloudwatch list-metrics or get-metric-statistics, or sts get-caller-identity — made with the CLI profile et-dev (the IAM user svc-harness) in us-east-1; never secretsmanager get-secret-value, a send or receive on a queue, an object read or write, a presigned URL, a Bedrock call or a terraform command. The checks that write (AC-4’s malformed messages, AC-5’s upload and signed GET) are F38-b-WI-3.1’s; they run under spec NFR-4’s temporary policy, and each of their entries names the lead’s approval. GitHub is read with gh, git only through fetch, log, show, rev-list and merge-base on the remote branches, and the public CRM host with curl. Times are UTC; the AWS CLI prints metric timestamps in the reading machine’s zone (+05:30), and the entries give them in UTC. Identifiers are recorded — account IDs, ARNs, key IDs, run IDs, commit SHAs, resource names — but never a secret value, an access-key ID or a person’s email address (spec NFR-3; the Secrets boundary at the top of this file). Each command’s whole output is kept outside the repositories, and the blocks quote the lines that show the result.

Every entry has one of three shapes. A read: a bold line **<YYYY-MM-DD HH:MM> UTC — Claude (F38-b-WI-<x.y>), <source>**, where the source is profile et-dev (svc-harness) for AWS, gh as <login> for GitHub, curl for the public host, or the repository and commit for git; a text block holding $ <command>, the output lines that show the result (trimmed with … where lines were cut, never edited) and EXIT=<n>; then Result: — met, not met — <why>, recorded as found — <AC> is <WI>'s, not yet — <why>, or reported to the lead — .forge/owed.md. Where one command ran once per resource, the block shows it once and a table gives each one’s result. A workflow run: the same bold line, then Run: with the run’s URL, trigger, head SHA and conclusion, then the quoted log lines in a text block, then Result:. A register note: *Update <YYYY-MM-DD> (F38-b WI-<x.y>): <one-sentence result>; evidence: environments-and-identity.md §4 "F38-b — the dev estate, checked", AC-<n>.*, appended to the row’s “Validate / finish when cloud is up” cell.

The subsections, in AC order, each written only when it has content: AC-1 — applied, with no drift (WI-2.1 opens it with the merge run; WI-4.1 adds the no-change plan, the code check that no resource writes a secret’s value into the state — in place of the DevOps team’s state search, F38-b spec Rev 12 — and the resource sweep); AC-2 — the network (WI-4.1); AC-3 — the database (WI-2.1 opens it with the rotation read; WI-4.1 adds the rest); AC-4 — queues and schedules (WI-2.1 opens it with the configuration reads; WI-3.1 adds the dead-letter checks, quoting its record file .forge/plans/F38-b/F38-b-WI-3.1-record.md); AC-5 — the attachments bucket (WI-2.1 opens it; WI-3.1 adds the upload, the 403 and 200 and the removal); AC-6 — the KMS keys (WI-2.1); AC-7 — the frontends' hosting (WI-2.1 opens it with the security headers, spec Rev 7; WI-4.1 adds the direct S3 403, the records and the capture app); AC-8 — the AgentCore Gateway (WI-2.1); AC-9 — the values for F38-d (WI-2.1); AC-10 — the closing sweep (the last of WI-3.1 and WI-4.1 to merge); AC-11 — the Bedrock grant and least privilege (WI-2.1); AC-12 — the temporary policy removed (WI-3.1); then Other changes in forge-et-iaac#3 (spec Revs 3, 6) (WI-2.1 opens it; WI-4.1 may add). A work item that opens AC-<n> inserts it immediately before the first #### AC-<m> heading here with m greater than n, or, when there is none, immediately before #### Other changes in forge-et-iaac#3. Within a subsection, each work item’s entries follow one italic line, *Recorded <date> by F38-b-WI-<x.y>.*, after any earlier work item’s entries.

What the account holds beyond the F38-b spec, and why. forge-et-iaac#4 (F11-d) was merged into #3’s branch before #3 merged (59b43d2), so the same apply created the lead-triage queue et-dev-lead-triage and its dead-letter queue et-dev-lead-triage-dlq, added the lead-triage queue to the task role’s sqs and sqs-send policies, and added a fifth schedule to the group et-dev, lead-triage-sweep (rate(1 hour), the backend’s scheduler contract on forge-et-be main at 9097a68b). forge-et-iaac 0eee623 (F38-a’s SES identity, applied on 21 Oct by dev.yaml run 39553084687) added the task role’s inline ses policy (“F38-a — the development account, checked”, AC-5). The entries below record each as found and say where it comes from. Each matches its Terraform (git -C forge-et-iaac show 7c93e65 -- base overlays; git -C forge-et-iaac show origin/dev:base/ecs/task_role.tf), so none is a finding for the lead.

AC-1 — applied, with no drift

Recorded 2026-10-21 by F38-b-WI-2.1. The merge of #3 and the dev.yaml run that applied it. The no-change plan, the DevOps team’s state search (replaced on 2026-10-25 by a code check, F38-b spec Rev 12) and the resource sweep are WI-4.1’s.

2026-10-21 21:11 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws sts get-caller-identity --profile et-dev --region us-east-1 --query Arn --output text
arn:aws:iam::570417736903:user/svc-harness
EXIT=0

Result: met — every AWS read in this section is made as svc-harness in the dev account.

2026-10-21 21:11 UTC — Claude (F38-b-WI-2.1), gh as tharinda-altrium; git: forge-et-iaac origin/dev at 0eee623

$ gh pr view 3 --repo altrium-io/forge-et-iaac --json state,baseRefName,mergedAt,mergeCommit
{"baseRefName":"dev","mergeCommit":{"oid":"d230b133a4e9f41e14066c80afd9d648d7b20241"},"mergedAt":"2026-10-20T18:33:54Z","state":"MERGED"}
EXIT=0
$ git -C forge-et-iaac rev-list --parents -n 1 d230b133a4e9f41e14066c80afd9d648d7b20241
d230b133a4e9f41e14066c80afd9d648d7b20241 76fce06c05ca21dfa2980ce6da78c24fcd313263 59b43d2e31d79a84df952a04bd85f25be0cebeb0
EXIT=0
$ git -C forge-et-iaac log --oneline d230b133..origin/dev
0eee623 feat(F38-a): add the SES sending identity for dev.crm.empire.lk
EXIT=0
$ gh pr list -R altrium-io/forge-et-iaac --state all --limit 10 --json number,title,state,headRefName,mergedAt
4 MERGED feature/F11-d-lead-triage-queue 2026-10-20T18:32:47Z feat(F11-d): add the lead-triage queue, its DLQ, the task role's rights and the hourly sweep
3 MERGED feature/F38-b-wave-1 2026-10-20T18:33:54Z feat(F38-b): add the KMS keys, attachments bucket, ingest queue, schedules and task-role grants
2 CLOSED add-ecr-et-be null Add ECR repository for et-be image
1 CLOSED switch-region-us-east-1 null Switch dev region from us-west-2 to us-east-1
EXIT=0

(The gh pr list lines are the saved JSON printed one PR per line.)

Result: met — #3 merged into dev on 2026-10-20 18:33:54 UTC as a merge commit with two parents: 76fce06, the dev tip before it, and 59b43d2, #3’s branch head, which is the merge of forge-et-iaac#4 into that branch. One commit has landed on dev since, 0eee623 (F38-a), and no forge-et-iaac PR is open.

2026-10-21 21:12 UTC — Claude (F38-b-WI-2.1), gh as tharinda-altrium

Run: 39475850266 — trigger push, head d230b133a4e9f41e14066c80afd9d648d7b20241 (the merge commit itself), conclusion success, started 2026-10-20 18:33:56 UTC. It is the only dev.yaml run between the merge and the first successful apply; the next run, 39553084687 on 0eee623 (2026-10-21 06:03 UTC), came after it.

$ gh run list -R altrium-io/forge-et-iaac --branch dev --workflow dev.yaml --limit 30 --json databaseId,headSha,event,status,conclusion,createdAt,url
39553084687 0eee6238f push completed success 2026-10-21T06:03:21Z
39475850266 d230b133a push completed success 2026-10-20T18:33:56Z
38724927953 76fce06c0 push completed failure 2026-10-15T11:06:00Z
…
EXIT=0
$ gh run view 39475850266 -R altrium-io/forge-et-iaac --json jobs --jq '.jobs[].steps[] | "\(.name) \(.conclusion)"'
…
terraform::fmt success
terraform init success
terraform::validate success
terraform::plan success
terraform::apply success
…
EXIT=0
$ gh run view 39475850266 -R altrium-io/forge-et-iaac --log      (saved whole; the lines below are taken from the saved file)
terraform::plan   2026-10-20T18:34:27.6662093Z   # module.ecr.aws_ecr_lifecycle_policy.backend must be replaced
terraform::plan   2026-10-20T18:34:27.7308346Z Plan: 28 to add, 5 to change, 1 to destroy.
terraform::apply  2026-10-20T18:35:24.8657690Z Apply complete! Resources: 28 added, 5 changed, 1 destroyed.
EXIT=0

(The gh run list lines are the saved JSON printed one run per line.)

Result: met — the plan and the apply after the merge both succeeded (AC-1’s merge-run half). The one resource destroyed is the ECR lifecycle policy, module.ecr.aws_ecr_lifecycle_policy.backend, replaced by the policy with three rules (spec Rev 6; “Other changes in forge-et-iaac#3” below), so it is not a finding. The no-change plan, the state search and the sweep are WI-4.1’s.

Recorded 2026-10-26 by F38-b-WI-4.1. The readiness run, the no-change plan, the code check that no resource writes a secret’s value into the state (spec Rev 12), the fmt and validate results and the resource sweep. The start condition (Decomposition Plan § Notes, “Waves 2 and 4 wait…”): the DevOps team made no written report. On 2026-10-26 the lead took the Trello board as exported on 25 Oct (every dev card in Done except the Bedrock quotas and base/secrets; the database fixes optional in dev and on no card — “Open with the DevOps team” above) and the absence of any DevOps commit on dev after 0eee623 (21 Oct) as the DevOps team making no further optional dev changes (spec Rev 5). The reference commit (plan D-3, R-4) is c728e06b, the merge of forge-et-iaac#9, the latest F38-b merge on dev; the dispatch facts gave no later DevOps commit. Wave 3’s temporary policy f38-b-write-checks was removed on 23 Oct (.forge/plans/F38-b/F38-b-WI-3.1-record.md R-7, on feature/F38-b-wave-3), so the sweep below expects none.

2026-10-26 05:34 UTC — Claude (F38-b-WI-4.1), gh as tharinda-altrium; git: forge-et-iaac origin/dev at c728e06b

$ gh pr view 3 --repo altrium-io/forge-et-iaac --json state,baseRefName,mergedAt,mergeCommit
{"baseRefName":"dev","mergeCommit":{"oid":"d230b133a4e9f41e14066c80afd9d648d7b20241"},"mergedAt":"2026-10-20T18:33:54Z","state":"MERGED"}
EXIT=0
$ gh pr list --repo altrium-io/forge-et-iaac --state merged --base dev --search "F38-b in:title" --json number,mergeCommit,mergedAt,title
#9 c728e06b 2026-10-26T04:52:34Z refactor(F38-b): declare the Secrets Manager secrets in one module
#5 5e078c0d 2026-10-22T07:22:28Z feat(F38-b): host the capture app on capture.dev.crm.empire.lk
#3 d230b133 2026-10-20T18:33:54Z feat(F38-b): add the KMS keys, attachments bucket, ingest queue, schedules and task-role grants
EXIT=0
$ gh pr list --repo altrium-io/forge-et-harness --head feature/F38-b-wave-2 --state all --json number,state,mergedAt
[{"mergedAt":"2026-10-21T21:39:05Z","number":850,"state":"MERGED"}]
EXIT=0
$ git -C forge-et-iaac rev-parse origin/dev
c728e06b6549c4d9a26823c9d90594d993d467bc
EXIT=0
$ git -C forge-et-iaac log --oneline d230b133..origin/dev
c728e06 Merge pull request #9 from altrium-io/feature/F38-b-wave-6
df13676 docs(F38-b): say who merges, as the team's rule does
057fa73 refactor(F38-b): declare the Secrets Manager secrets in one module
1590009 Merge pull request #7 from altrium-io/feature/F38-e-wave-1
acd1cf7 Merge pull request #6 from altrium-io/feature/F38-d-wave-3
d86d2ca Merge pull request #8 from altrium-io/feature/F11-c-lead-research-queue
2a21a5a feat(F11-c): add an empty Secrets Manager secret for dev's Exa API key
68a9d52 feat(F11-c): add the lead-research queue, its DLQ, the task role's rights and the hourly sweep
e68a7b6 feat(F38-e): add the dev alert topic and the failed backend deploy email
4f61374 feat(F38-d): drain the backend target group for 30 s
bc65bc6 feat(F38-d): make the backend's ECR tags immutable
5e078c0 Merge pull request #5 from altrium-io/feature/F38-b-wave-5
88feb28 feat(F38-b): host the capture app on capture.dev.crm.empire.lk
0eee623 feat(F38-a): add the SES sending identity for dev.crm.empire.lk
EXIT=0
$ git -C forge-et-iaac log --oneline d230b133..origin/dev -- base/rds base/cloudfront
EXIT=0      (nothing printed)
$ gh pr list -R altrium-io/forge-et-iaac --state open --json number,title
[]
EXIT=0

(The gh pr list lines for forge-et-iaac are the saved JSON printed one PR per line.)

Result: met — readiness reads R-1 to R-4: #3 is merged into dev, wave 2’s harness PR is merged, and the reference commit c728e06b is dev’s tip. Since #3’s merge every commit on dev is a delivery-team PR except 0eee623, Shyamin Ayesh’s SES push of 21 Oct, already recorded in “What it holds” above; nothing under base/rds or base/cloudfront changed, so WI-2.1’s rotation and header entries stand without a re-read (plan D-11). No forge-et-iaac PR is open.

2026-10-26 05:34 UTC — Claude (F38-b-WI-4.1), gh as tharinda-altrium

Run: 40271292253 — trigger push, head c728e06b6549c4d9a26823c9d90594d993d467bc (the merge of #9, dev’s tip), conclusion success, started 2026-10-26 04:52:37 UTC. It is the latest dev.yaml run on dev, and the readiness run of plan D-3 R-5; it is not WI-2.1’s qualifying run (39475850266, above). Every run on dev since that one concluded success: 39553084687 (0eee6238), 39740047356 (5e078c0d), 39936834361 (d86d2ca9), 39940853448 (acd1cf7b), 39940862972 (1590009f) and this one.

$ gh run view 40271292253 -R altrium-io/forge-et-iaac --json jobs --jq '.jobs[].steps[] | "\(.name) \(.conclusion)"'
…
terraform::fmt success
terraform init success
terraform::validate success
terraform::plan success
terraform::apply success
EXIT=0
$ gh run view 40271292253 -R altrium-io/forge-et-iaac --log      (saved whole; the lines below are taken from the saved file)
terraform::plan   2026-10-26T04:53:06.8400816Z Plan: 0 to add, 5 to change, 0 to destroy.
terraform::apply  2026-10-26T04:53:12.6864911Z Apply complete! Resources: 0 added, 5 changed, 0 destroyed.
EXIT=0

Result: met — readiness read R-5: the latest run is on dev’s tip and applied successfully. Its five changes are #9’s five secrets moved into base/secrets, each changing only its Module tag (F38-b WI-6.1; AC-13), so they are not drift.

The no-change plan. As first recorded: not yet — no dev.yaml run with a no-change plan after c728e06b: the only run at or after the reference commit was 40271292253, whose plan had five changes (above). WI-4.1 started no run (plan D-5), because a workflow_dispatch of dev.yaml on dev applies with svc-github’s administrator key; .forge/owed.md asked for one. Recorded 2026-10-26, after the lead asked for the run in session:

2026-10-26 06:10 UTC — Claude (F38-b-WI-4.1, follow-up), gh as tharinda-altrium

Run: 40278225094 — trigger workflow_dispatch on dev, started by Claude at the lead’s request (gh workflow run dev.yaml --repo altrium-io/forge-et-iaac --ref dev, 06:09:58 UTC), head c728e06b6549c4d9a26823c9d90594d993d467bc (the reference commit, still dev’s tip; no run was in progress and no forge-et-iaac PR open), conclusion success.

$ gh run view 40278225094 -R altrium-io/forge-et-iaac --json jobs --jq '.jobs[].steps[] | "\(.name) \(.conclusion)"'
…
terraform::fmt success
terraform init success
terraform::validate success
terraform::plan success
terraform::apply success
EXIT=0
$ gh run view 40278225094 -R altrium-io/forge-et-iaac --log      (saved whole; the lines below are taken from the saved file)
terraform::plan   2026-10-26T06:10:27.3726153Z No changes. Your infrastructure matches the configuration.
terraform::apply  2026-10-26T06:10:31.5787272Z Apply complete! Resources: 0 added, 0 changed, 0 destroyed.
EXIT=0
$ grep -ci 'changed outside of Terraform\|Objects have changed' <the saved log>
0
EXIT=1

Result: met — a run on dev at the reference commit planned no changes, and its apply changed nothing; the refresh reported no object changed outside Terraform.

2026-10-26 05:37 UTC — Claude (F38-b-WI-4.1), git: forge-et-iaac at c728e06b

$ git -C forge-et-iaac grep -n -E 'resource "(aws_secretsmanager_secret_version|random_[a-z_]+|aws_iam_access_key)"|secret_string *=|(^|[^_])password *=' c728e06b -- '*.tf'
EXIT=1      (nothing printed)
$ git -C forge-et-iaac grep -n manage_master_user_password c728e06b -- base/rds
c728e06b:base/rds/instance.tf:13:  manage_master_user_password = true
EXIT=0

Result: met — no resource at the commit read writes a secret’s value into the state: no aws_secretsmanager_secret_version, no random_* resource, no aws_iam_access_key, no password or secret_string argument; the database’s password is RDS-managed. The state is not read (spec Rev 12).

2026-10-26 05:38 UTC — Claude (F38-b-WI-4.1), gh as tharinda-altrium

$ gh pr view <n> --repo altrium-io/forge-et-iaac --json body,headRefOid      (each of #3, #5 and #9; saved, the check lines read from the saved files)
EXIT=0
PR Head the checks ran on fmt -check -recursive init -backend=false validate Where recorded
forge-et-iaac#3 (wave 1) 0ff63c5, the branch head before forge-et-iaac#4 (F11-d) was merged into it as 59b43d2 EXIT=0 EXIT=0 EXIT=0 #3’s body, “Checks”; .forge/plans/F38-b/F38-b-WI-1.1-plan.md ## Notes
forge-et-iaac#5 (wave 5) 88feb28 (its head) EXIT=0 EXIT=0 EXIT=0 #5’s body; .forge/plans/F38-b/F38-b-WI-5.1-plan.md
forge-et-iaac#9 (wave 6) df13676 (its head) EXIT=0 EXIT=0 EXIT=0 #9’s body; .forge/plans/F38-b/F38-b-WI-6.1-plan.md

Each was run with Terraform 1.12.1, the version dev.yaml pins. Every dev.yaml run on dev since #3’s merge also ran terraform::fmt and terraform::validate on the merged tree and concluded success (the run list above).

Result: met — every delivery-team F38-b PR head passes fmt -check and validate.

Parent AC-8’s state-bucket clause. Result: met — it does not apply in dev: dev’s state stays in et-ops (parent Rev 13), and F38-a AC-3 records its encryption (“F38-a — the development account, checked”, AC-3).

2026-10-26 05:38–05:39 UTC — Claude (F38-b-WI-4.1), profile et-dev (svc-harness); git: forge-et-iaac at c728e06b

$ aws resourcegroupstaggingapi get-resources --query 'ResourceTagMappingList[].[ResourceARN,Tags[?Key==`Project`]|[0].Value,Tags[?Key==`Module`]|[0].Value]' --output text --profile et-dev --region us-east-1
…  (49 resources, every one tagged Project = ET; counted by Module from the saved file)
EXIT=0
$ aws iam list-roles --query 'Roles[].[RoleName,Path,CreateDate]' --output text --profile et-dev --region us-east-1
AWSReservedSSO_ADMINISTRATORS_05d2e5facfe0b76e  /aws-reserved/sso.amazonaws.com/us-west-2/
AWSServiceRoleForECS, …ElasticLoadBalancing, …Organizations, …RDS, …ResourceExplorer, …ServiceQuotas, …SSO, …Support, …TrustedAdvisor   /aws-service-role/…
et-dev-agentcore-gateway  /
et-dev-backend-execution  /
et-dev-backend-task       /
et-dev-scheduler          /
OrganizationAccountAccessRole  /
EXIT=0
$ aws iam list-policies --scope Local --query 'Policies[].[PolicyName,CreateDate]' --output text --profile et-dev --region us-east-1
EXIT=0      (nothing printed)
$ aws iam list-user-policies --user-name svc-harness --output text --profile et-dev --region us-east-1
POLICYNAMES	spike-invoke-gateway
EXIT=0
$ aws iam list-user-policies --user-name svc-github --output text --profile et-dev --region us-east-1
EXIT=0      (nothing printed)
$ git -C forge-et-iaac grep -n 'source *= *"../../base/' c728e06b -- overlays/dev
…  networking, ecs, ses, secrets, security-groups, rds, ecr, alb, domain, sqs, kms, attachments, scheduler, agentcore, alerts, acm (twice), s3 (twice), cloudfront (twice)
EXIT=0
$ git -C forge-et-iaac grep -n -A2 'resource "aws_iam_role"' c728e06b
base/agentcore/iam.tf: name = "${var.project}-${var.environment}-agentcore-gateway"
base/ecs/execution_role.tf: name = "${var.project}-${var.environment}-backend-execution"
base/ecs/task_role.tf: name = "${var.project}-${var.environment}-backend-task"
base/scheduler/iam.tf: name = "${var.project}-${var.environment}-scheduler"
EXIT=0
$ aws secretsmanager describe-secret --secret-id 'dev/rds/password' --query '{Name:Name,DeletedDate:DeletedDate,Created:CreatedDate}' --output json --profile et-dev --region us-east-1
{"Name": "dev/rds/password", "DeletedDate": "2026-10-09T14:07:02.125000+05:30", "Created": "2026-10-05T13:25:04.258000+05:30"}
EXIT=0
What the sweep found Declared by, or allowed because
45 tagged resources whose Module is acm (2), agentcore, alb (2), alerts (2), attachments, cloudfront (2), domain, ecr, ecs (3), kms (2), networking (5), rds (2), s3 (2), scheduler, secrets (5), security-groups (3), ses (2), sqs (8) each value names a base/<m> that overlays/dev/main.tf calls
The two load-balancer listeners on et-dev-backend (Project tag, no Module tag) base/alb/alb.tf, aws_lb_listener https and http
rds!db-6a52f694-…-YSNFQA (Project tag, no Module tag) the secret RDS creates and owns (OwningService: rds) for the instance base/rds declares with manage_master_user_password = true
dev/rds/password (Project tag, no Module tag) made by forge-et-iaac 9e00dc7 on 5 Oct and destroyed by Terraform when ac1efdc moved the password to RDS; scheduled for deletion since 2026-10-09 08:37 UTC, which AWS completes at the end of its recovery window — as F38-a recorded (“F38-a — the development account, checked”, AC-1)
Roles et-dev-agentcore-gateway, et-dev-backend-execution, et-dev-backend-task, et-dev-scheduler each aws_iam_role block’s name resolves to it with project = et, environment = dev
Nine service-linked roles, AWSReservedSSO_ADMINISTRATORS_…, OrganizationAccountAccessRole roles AWS creates itself (spec FR-1)
No customer-managed policy; svc-github has no inline policy —
spike-invoke-gateway on svc-harness F38-a’s NFR-3 list (“What was made by hand”)
No f38-b-write-checks on svc-harness wave 3 removed it on 23 Oct (spec NFR-4)

The tagging API returns only tagged resources and no IAM roles, which is why the IAM lists are read as well; an untagged resource made by hand would not show here.

Result: met — nothing outside Terraform beyond what FR-1 allows. Together with the entries above, AC-1 is met: its no-change plan is run 40278225094, recorded above.

AC-2 — the network

Recorded 2026-10-26 by F38-b-WI-4.1.

2026-10-26 05:35 UTC — Claude (F38-b-WI-4.1), profile et-dev (svc-harness)

$ aws ec2 describe-nat-gateways --filter Name=state,Values=pending,available --query 'NatGateways[].[NatGatewayId,State]' --output json --profile et-dev --region us-east-1
[]
EXIT=0
$ aws ecs describe-services --cluster et-dev --services backend --query 'services[0].networkConfiguration.awsvpcConfiguration' --output json --profile et-dev --region us-east-1
{"subnets": ["subnet-02fb26b2351e546d1"], "securityGroups": ["sg-0097bf895557b7716"], "assignPublicIp": "ENABLED"}
EXIT=0
$ aws ecs describe-tasks --cluster et-dev --tasks e2a59cfd13c346509b87021d8b3e460b --query 'tasks[0].{AZ:availabilityZone,TaskDef:taskDefinitionArn,Last:lastStatus,Subnet:attachments[0].details[?name==`subnetId`].value|[0]}' --output json --profile et-dev --region us-east-1
{"AZ": "us-east-1b", "TaskDef": "…:task-definition/backend:67", "Last": "RUNNING", "Subnet": "subnet-02fb26b2351e546d1"}
EXIT=0
$ aws rds describe-db-instances --db-instance-identifier et-dev --query 'DBInstances[0].{…}' --output json --profile et-dev --region us-east-1      (D-9's query; the network fields)
"AZ": "us-east-1b", "Public": false, "SubnetGroup": "et-dev", "Subnets": [["subnet-08191744044d950b6","us-east-1a"],["subnet-02fb26b2351e546d1","us-east-1b"]], "SecurityGroups": ["sg-0d0ce1e3c1bd0b568"], "Endpoint": "et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com"
EXIT=0
$ aws ec2 describe-subnets --subnet-ids subnet-02fb26b2351e546d1 subnet-08191744044d950b6 --query 'Subnets[].[SubnetId,AvailabilityZone,VpcId,CidrBlock,MapPublicIpOnLaunch,Tags[?Key==`Name`]|[0].Value]' --output text --profile et-dev --region us-east-1
subnet-02fb26b2351e546d1  us-east-1b  vpc-0ea763d5098b71922  10.0.1.0/24  True  et-dev-public-1
subnet-08191744044d950b6  us-east-1a  vpc-0ea763d5098b71922  10.0.0.0/24  True  et-dev-public-0
EXIT=0
$ aws ec2 describe-route-tables --filters Name=association.subnet-id,Values=subnet-02fb26b2351e546d1,subnet-08191744044d950b6 --query 'RouteTables[].{Id:RouteTableId,Assoc:Associations[].SubnetId,Routes:Routes[].[DestinationCidrBlock,GatewayId,NatGatewayId]}' --output json --profile et-dev --region us-east-1
rtb-0cb8657318ee04d66 — both subnets; routes 10.0.0.0/16 → local, 0.0.0.0/0 → igw-089bfde042eb8701a
EXIT=0
$ aws ec2 describe-security-groups --group-ids sg-0097bf895557b7716 sg-0d0ce1e3c1bd0b568 --query 'SecurityGroups[].{Id:GroupId,Name:GroupName,In:IpPermissions[].{From:FromPort,To:ToPort,Proto:IpProtocol,Cidr:IpRanges[].CidrIp,Groups:UserIdGroupPairs[].GroupId}}' --output json --profile et-dev --region us-east-1
et-dev-backend (sg-0097bf895557b7716): tcp 8080 from sg-0cdf10021add79056 only
et-dev-rds (sg-0d0ce1e3c1bd0b568): tcp 5432 from 10.0.0.0/16
EXIT=0
$ aws elbv2 describe-load-balancers --names et-dev-backend --query 'LoadBalancers[0].{Arn:LoadBalancerArn,SG:SecurityGroups,Scheme:Scheme}' --output json --profile et-dev --region us-east-1
"SG": ["sg-0cdf10021add79056"], "Scheme": "internet-facing"
EXIT=0
$ aws elbv2 describe-listeners --load-balancer-arn arn:aws:elasticloadbalancing:us-east-1:570417736903:loadbalancer/app/et-dev-backend/82461f891f52a937 --query 'Listeners[].[Port,Protocol,SslPolicy]' --output text --profile et-dev --region us-east-1
80   HTTP   None
443  HTTPS  ELBSecurityPolicy-TLS13-1-2-2021-06
EXIT=0
$ aws elbv2 describe-ssl-policies --names ELBSecurityPolicy-TLS13-1-2-2021-06 --query 'SslPolicies[].SslProtocols' --output text --profile et-dev --region us-east-1
TLSv1.2  TLSv1.3
EXIT=0
$ aws cloudfront list-distributions --query 'DistributionList.Items[].[Id,Aliases.Items|join(`,`,@),Origins.Items[0].DomainName,ViewerCertificate.MinimumProtocolVersion]' --output text --profile et-dev --region us-east-1
E2RVC3893O10TJ  dev.crm.empire.lk          et-dev-spa.s3.us-east-1.amazonaws.com      TLSv1.2_2021
EZ06T8FTKF7UZ   capture.dev.crm.empire.lk  et-dev-capture.s3.us-east-1.amazonaws.com  TLSv1.2_2021
EXIT=0
$ aws cloudfront get-distribution-config --id <id> --query 'DistributionConfig.{Aliases:…,MinTLS:ViewerCertificate.MinimumProtocolVersion,…}' --output json --profile et-dev --region us-east-1
EXIT=0      (each of the two; MinTLS "TLSv1.2_2021" on both)

(The security-group, route-table and load-balancer lines summarise the saved JSON; nothing in it was left out.)

Result: met — no NAT gateway; the service lists one subnet, subnet-02fb26b2351e546d1 in us-east-1b, the database instance’s zone, and its running task is there; the instance is not publicly accessible; the backend’s group admits 8080 only from the load balancer’s group sg-0cdf10021add79056; the HTTPS listener’s policy allows TLS 1.2 and 1.3 only (port 80 redirects to HTTPS — “What it holds”, base/alb), and both distributions require TLSv1.2_2021. Recorded as found (optional in dev — parent Rev 18, spec Rev 5): the database’s subnet group is the two public subnets et-dev-public-0 and et-dev-public-1, whose route table sends 0.0.0.0/0 to the internet gateway, and et-dev-rds admits 5432 from the whole VPC (10.0.0.0/16) rather than from the backend’s group; the service runs with a public IP, which the tasks need without a NAT gateway.

2026-10-26 05:36 UTC — Claude (F38-b-WI-4.1), curl (dig, nc and openssl from the developer’s machine)

$ dig +short et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com
10.0.1.55
EXIT=0
$ nc -vz -G 10 -w 10 et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com 5432
nc: connectx to et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com port 5432 (tcp) failed: Operation timed out
EXIT=1
$ openssl version
OpenSSL 4.0.3 29 Sep 2026 (Library: OpenSSL 4.0.3 29 Sep 2026)
EXIT=0
$ openssl s_client -connect api.dev.crm.empire.lk:443 -servername api.dev.crm.empire.lk -tls1_1 -cipher 'DEFAULT@SECLEVEL=0' -msg </dev/null
>>> TLS 1.1, Handshake [length 0086], ClientHello
<<< TLS 1.1, Alert [length 0002], fatal protocol_version
…:error:0A00042E:SSL routines:ssl3_read_bytes:tlsv1 alert protocol version:…:SSL alert number 70
EXIT=1
$ openssl s_client -connect dev.crm.empire.lk:443 -servername dev.crm.empire.lk -tls1_1 -cipher 'DEFAULT@SECLEVEL=0' -msg </dev/null
>>> TLS 1.1, Handshake [length 0082], ClientHello
<<< TLS 1.1, Alert [length 0002], fatal protocol_version
…:error:0A00042E:SSL routines:ssl3_read_bytes:tlsv1 alert protocol version:…:SSL alert number 70
EXIT=1

(Neither saved s_client output has a ServerHello line.)

Result: met — the endpoint resolves to a private address and a connection from outside AWS to 5432 times out; the client sent a TLS 1.1 ClientHello to each host, and each server answered with a fatal protocol_version alert and no ServerHello.

AC-3 — the database

Recorded 2026-10-21 by F38-b-WI-2.1. The rotation setting only, as found; the rest of AC-3 is WI-4.1’s.

2026-10-21 21:17 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws rds describe-db-instances --query 'DBInstances[].MasterUserSecret.SecretArn' --output text --profile et-dev --region us-east-1
arn:aws:secretsmanager:us-east-1:570417736903:secret:rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e-YSNFQA
EXIT=0
$ aws secretsmanager describe-secret --secret-id 'arn:aws:secretsmanager:us-east-1:570417736903:secret:rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e-YSNFQA' --query '{Name:Name,RotationEnabled:RotationEnabled,LastRotatedDate:LastRotatedDate,NextRotationDate:NextRotationDate}' --output json --profile et-dev --region us-east-1
{
    "Name": "rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e",
    "RotationEnabled": false,
    "LastRotatedDate": "2026-10-09T14:03:25.497000+05:30",
    "NextRotationDate": null
}
EXIT=0

Result: recorded as found — AC-3 is WI-4.1’s. The RDS-managed secret’s automatic rotation is off, with no next rotation date, as base/rds/rotation.tf sets it (spec Rev 4); the last rotation was on 9 Oct 2026 (08:33 UTC).

Recorded 2026-10-26 by F38-b-WI-4.1. The instance, its parameter, the Terraform declarations, btree_gist, and the rotation re-read (plan D-9: a hand-run rotate-secret would turn rotation back on without a commit, so it is read again rather than cited).

2026-10-26 05:35–05:36 UTC — Claude (F38-b-WI-4.1), profile et-dev (svc-harness)

$ aws rds describe-db-instances --db-instance-identifier et-dev --query 'DBInstances[0].{Engine:Engine,Version:EngineVersion,AutoMinor:AutoMinorVersionUpgrade,Class:DBInstanceClass,AZ:AvailabilityZone,MultiAZ:MultiAZ,Public:PubliclyAccessible,StorageType:StorageType,Allocated:AllocatedStorage,MaxAllocated:MaxAllocatedStorage,DeletionProtection:DeletionProtection,BackupRetention:BackupRetentionPeriod,ParameterGroups:DBParameterGroups,…,Encrypted:StorageEncrypted}' --output json --profile et-dev --region us-east-1
"Engine": "postgres", "Version": "18.3", "AutoMinor": true, "Class": "db.t3.micro", "AZ": "us-east-1b", "MultiAZ": false, "Public": false,
"StorageType": "gp2", "Allocated": 20, "MaxAllocated": null, "DeletionProtection": false, "BackupRetention": 0,
"ParameterGroups": [{"DBParameterGroupName": "default.postgres18", "ParameterApplyStatus": "in-sync"}], "Encrypted": false
EXIT=0
$ aws rds describe-db-parameters --db-parameter-group-name default.postgres18 --query "Parameters[?ParameterName=='rds.force_ssl'].[ParameterValue,Source]" --output text --profile et-dev --region us-east-1
1	system
EXIT=0
$ aws secretsmanager describe-secret --secret-id 'arn:aws:secretsmanager:us-east-1:570417736903:secret:rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e-YSNFQA' --query '{Name:Name,RotationEnabled:RotationEnabled,RotationRules:RotationRules,LastRotatedDate:LastRotatedDate,NextRotationDate:NextRotationDate,OwningService:OwningService}' --output json --profile et-dev --region us-east-1
{"Name": "rds!db-6a52f694-ad18-40b1-b42b-52138a0bb71e", "RotationEnabled": false, "RotationRules": {"AutomaticallyAfterDays": 365}, "LastRotatedDate": "2026-10-09T14:03:25.497000+05:30", "NextRotationDate": null, "OwningService": "rds"}
EXIT=0

2026-10-26 05:36 UTC — Claude (F38-b-WI-4.1), git: forge-et-iaac at c728e06b, forge-et-be origin/main at 30fd601c

$ git -C forge-et-iaac show c728e06b:base/rds/variables.tf      (saved; the lines below)
variable "engine_version" {
  type     = string
  nullable = false
  default  = "18"
EXIT=0
$ git -C forge-et-iaac grep -n engine_version c728e06b -- overlays/dev
EXIT=1      (nothing printed)
$ git -C forge-et-iaac grep -n auto_minor c728e06b -- base/rds overlays/dev
EXIT=1      (nothing printed)
$ git -C forge-et-iaac show c728e06b:base/rds/instance.tf      (saved; the lines below)
4:  engine_version = var.engine_version
13:  manage_master_user_password = true
19:  publicly_accessible = false
21:  skip_final_snapshot = true
22:  deletion_protection = false
EXIT=0
$ git -C forge-et-iaac grep -n rotation_enabled c728e06b -- base/rds
c728e06b:base/rds/rotation.tf:11:  rotation_enabled = false
EXIT=0
$ git -C forge-et-be show origin/main:src/main/resources/db/migration/V1__extensions.sql      (saved; the line below)
10:CREATE EXTENSION IF NOT EXISTS btree_gist;
EXIT=0

2026-10-26 05:36 UTC — Claude (F38-b-WI-4.1), profile et-dev (svc-harness)

“F38-d — the dev deploy pipelines, checked”, AC-9 (“Dev’s first startup log”) shows dev’s first task migrating an empty schema through 37 migrations but does not quote version 1, so plan D-9’s one get call on the stream it names:

$ aws logs get-log-events --log-group-name /ecs/dev/backend --log-stream-name backend/backend/3499fbf2eb41491eb361db62b0302254 --start-from-head --limit 2000 --query 'events[].message' --output json --profile et-dev --region us-east-1      (saved whole; the line below)
2026-10-09T09:06:26.822235233Z org.flywaydb.core.internal.command.DbMigrate Migrating schema "public" to version "1 - extensions"
EXIT=0

Result: met — engine postgres 18.3, pinned in base/rds/variables.tf (default = "18") with no override in overlays/dev; AutoMinorVersionUpgrade is true, the AWS provider’s default, since no forge-et-iaac file sets auto_minor_version_upgrade; rds.force_ssl is 1 (the PostgreSQL 18 default parameter group’s system value); rotation of the secret the backend reads is off (RotationEnabled: false, no next rotation date, unchanged since WI-2.1’s read), as base/rds/rotation.tf declares; and V1__extensions.sql, which creates btree_gist, ran on dev’s first start on 9 Oct, as part of the 37 migrations F38-d recorded. Recorded as found (optional in dev — parent Revs 11, 17 and 18, spec Rev 5): db.t3.micro, gp2, 20 GiB with no storage autoscaling (MaxAllocatedStorage unset), deletion protection off, skip_final_snapshot = true in base/rds (the RDS API does not return it), backup retention 0, Multi-AZ off, not encrypted at rest.

AC-4 — queues and schedules

Recorded 2026-10-21 by F38-b-WI-2.1. The configuration reads and the schedules’ first firings, for WI-3.1 to cite. The malformed messages that prove each redrive are WI-3.1’s.

2026-10-21 21:14 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws sqs get-queue-url --queue-name et-dev-mailbox-ingest --profile et-dev --region us-east-1
{
    "QueueUrl": "https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-mailbox-ingest"
}
EXIT=0
$ aws sqs get-queue-attributes --queue-url https://sqs.us-east-1.amazonaws.com/570417736903/<queue> --attribute-names QueueArn VisibilityTimeout MessageRetentionPeriod RedrivePolicy SqsManagedSseEnabled ApproximateNumberOfMessages ApproximateNumberOfMessagesNotVisible --output json --profile et-dev --region us-east-1
EXIT=0      (each of the six queues below)
Queue VisibilityTimeout MessageRetentionPeriod RedrivePolicy SqsManagedSseEnabled Messages (visible / in flight) From
et-dev-mailbox-ingest 30 345600 et-dev-mailbox-ingest-dlq, maxReceiveCount 3 true 0 / 0 #3
et-dev-mailbox-ingest-dlq 30 1209600 — true 0 / 0 #3
et-dev-scheduler 30 345600 et-dev-scheduler-dlq, maxReceiveCount 3 true 0 / 0 built 9 Oct
et-dev-scheduler-dlq 30 1209600 — true 0 / 0 built 9 Oct; 14-day retention from #3
et-dev-lead-triage 30 345600 et-dev-lead-triage-dlq, maxReceiveCount 3 true 0 / 0 #4 (F11-d)
et-dev-lead-triage-dlq 30 1209600 — true 0 / 0 #4 (F11-d)

Result: recorded for WI-3.1 — AC-4 is WI-3.1’s. Every item read is as spec FR-4 states: the ingest queue redrives to its dead-letter queue at 3 receives, with a 30-second visibility timeout and SQS-managed encryption, and both F38-b dead-letter queues keep messages 14 days (1,209,600 seconds, spec Rev 3). All three dead-letter queues are empty, so no scheduled tick has failed into them since the apply. The lead-triage pair is forge-et-iaac#4’s and has the same shape.

2026-10-21 21:13 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness) — the task role’s queue policies

$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name sqs --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":["sqs:ReceiveMessage","sqs:GetQueueUrl","sqs:GetQueueAttributes","sqs:DeleteMessage","sqs:ChangeMessageVisibility"],"Effect":"Allow","Resource":["arn:aws:sqs:us-east-1:570417736903:et-dev-scheduler","arn:aws:sqs:us-east-1:570417736903:et-dev-mailbox-ingest","arn:aws:sqs:us-east-1:570417736903:et-dev-lead-triage"]}]}
EXIT=0
$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name sqs-send --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":"sqs:SendMessage","Effect":"Allow","Resource":["arn:aws:sqs:us-east-1:570417736903:et-dev-mailbox-ingest","arn:aws:sqs:us-east-1:570417736903:et-dev-lead-triage"]}]}
EXIT=0

(Policy documents printed compactly from the saved JSON.)

Result: recorded for WI-3.1 — the task role consumes from the scheduler, ingest and lead-triage queues and sends to the ingest and lead-triage queues, naming only their ARNs. The F38-b part (the scheduler and ingest queues) is as spec FR-4 states; the lead-triage ARN in both policies is forge-et-iaac#4’s (overlays/dev/main.tf’s consumed_queue_arns and sent_queue_arns). The task role cannot send to the scheduler queue, so the backend adds no message to it; the message counts under the first firings below equal the schedules’ own invocation attempts.

2026-10-21 21:14 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness) — the schedule group, the schedules and the scheduler role

$ aws scheduler get-schedule-group --name et-dev --output json --profile et-dev --region us-east-1
{"Name":"et-dev","Arn":"arn:aws:scheduler:us-east-1:570417736903:schedule-group/et-dev","State":"ACTIVE","CreationDate":"2026-10-21T00:04:32.705000+05:30"}
EXIT=0
$ aws scheduler list-schedules --group-name et-dev --output json --profile et-dev --region us-east-1
mailbox-subscription-renewal ENABLED
mailbox-delta-poll ENABLED
email-upload-cleanup ENABLED
lead-triage-sweep ENABLED
price-expiry-sweep ENABLED
EXIT=0
$ aws scheduler get-schedule --group-name et-dev --name <schedule> --output json --profile et-dev --region us-east-1
EXIT=0      (each of the five schedules below)

Every schedule has State ENABLED, FlexibleTimeWindow {"Mode":"OFF"}, Target.Arn arn:aws:sqs:us-east-1:570417736903:et-dev-scheduler, Target.RoleArn arn:aws:iam::570417736903:role/et-dev-scheduler and CreationDate 2026-10-20 18:34:58 UTC. Target.Input is {"jobKind":"<name>","idempotencyKey":"<name>#<aws.scheduler.scheduled-time>"}, with < and > unescaped, which is the contract’s payload for each job (forge-et-be docs/messaging/scheduler-cloud-contract.md at 9097a68b).

Schedule Expression Time zone Contract’s expression and zone From
mailbox-subscription-renewal rate(6 hours) UTC rate(6 hours) #3
mailbox-delta-poll rate(15 minutes) UTC rate(15 minutes) #3
price-expiry-sweep cron(5 0 * * ? *) Asia/Colombo cron(5 0 * * ? *), Asia/Colombo #3
email-upload-cleanup rate(1 hour) UTC rate(1 hour) #3
lead-triage-sweep rate(1 hour) UTC rate(1 hour) (added to the contract by F11-d) #4 (F11-d)
$ aws iam get-role --role-name et-dev-scheduler --query Role.AssumeRolePolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Effect":"Allow","Principal":{"Service":"scheduler.amazonaws.com"},"Action":"sts:AssumeRole","Condition":{"StringEquals":{"aws:SourceAccount":"570417736903"},"ArnEquals":{"aws:SourceArn":"arn:aws:scheduler:us-east-1:570417736903:schedule-group/et-dev"}}}]}
EXIT=0
$ aws iam get-role-policy --role-name et-dev-scheduler --policy-name sqs-send --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":"sqs:SendMessage","Effect":"Allow","Resource":"arn:aws:sqs:us-east-1:570417736903:et-dev-scheduler"}]}
EXIT=0

(The list-schedules lines are the saved JSON printed one schedule per line; the other documents are printed compactly.)

Result: recorded for WI-3.1 — the four F38-b schedules match the contract’s expression, time zone, target and Input; the scheduler role may only send to the scheduler queue, and its trust policy admits scheduler.amazonaws.com limited by aws:SourceAccount and by aws:SourceArn to the schedule group et-dev (spec FR-4, Rev 2). The fifth schedule, lead-triage-sweep, is forge-et-iaac#4’s and matches the contract too. All five share the one scheduler queue (FS-9 OQ-5).

2026-10-21 21:15 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness) — the first firings

$ aws cloudwatch list-metrics --namespace AWS/Scheduler --output json --profile et-dev --region us-east-1
{"MetricName":"InvocationAttemptCount","Dimensions":[{"Name":"ScheduleGroup","Value":"et-dev"}]}
{"MetricName":"InvocationAttemptCount","Dimensions":[]}
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/Scheduler --metric-name InvocationAttemptCount --dimensions Name=ScheduleGroup,Value=et-dev --start-time 2026-10-20T18:30:00Z --end-time 2026-10-21T21:15:00Z --period 300 --statistics Sum --output json --profile et-dev --region us-east-1
107 datapoints, Sum 168 in all
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/Scheduler --metric-name TargetErrorCount --dimensions Name=ScheduleGroup,Value=et-dev --start-time 2026-10-20T18:30:00Z --end-time 2026-10-21T21:15:00Z --period 300 --statistics Sum --output json --profile et-dev --region us-east-1
"Datapoints": []
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/Scheduler --metric-name TargetErrorCount --start-time 2026-10-20T18:30:00Z --end-time 2026-10-21T21:15:00Z --period 300 --statistics Sum --output json --profile et-dev --region us-east-1
"Datapoints": []
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/SQS --metric-name NumberOfMessagesSent --dimensions Name=QueueName,Value=et-dev-scheduler --start-time 2026-10-20T18:30:00Z --end-time 2026-10-21T21:15:00Z --period 300 --statistics Sum --output json --profile et-dev --region us-east-1
Sum 168 in all; each 5-minute period's sum equals InvocationAttemptCount's for the same period
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/Scheduler --metric-name InvocationAttemptCount --dimensions Name=ScheduleGroup,Value=et-dev --start-time 2026-10-20T18:30:00Z --end-time 2026-10-20T19:40:00Z --period 60 --statistics Sum --output json --profile et-dev --region us-east-1
18:35 5.0   18:50 1.0   19:05 1.0   19:20 1.0   19:35 3.0      (UTC, 20 Oct)
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/SQS --metric-name NumberOfMessagesSent --dimensions Name=QueueName,Value=et-dev-scheduler --start-time 2026-10-20T18:30:00Z --end-time 2026-10-20T19:40:00Z --period 60 --statistics Sum --output json --profile et-dev --region us-east-1
18:35 5.0   18:50 1.0   19:05 1.0   19:20 1.0   19:35 3.0      (UTC, 20 Oct)
EXIT=0
$ aws cloudwatch get-metric-statistics --namespace AWS/Scheduler --metric-name InvocationAttemptCount --dimensions Name=ScheduleGroup,Value=et-dev --start-time 2026-10-21T18:30:00Z --end-time 2026-10-21T18:40:00Z --period 60 --statistics Sum --output json --profile et-dev --region us-east-1
18:35 5.0      (UTC, 21 Oct)
EXIT=0

(Datapoint lists are summarised from the saved JSON: counts, sums, or each non-empty period in UTC.)

The finest dimension AWS/Scheduler publishes here is the schedule group, so a count is attributed to every schedule due in its period, not split. EventBridge Scheduler published no TargetErrorCount metric at all, and both reads of it return no datapoints — no target error published, at the group or the account level, over the whole window. The metrics show each rate() schedule firing once when it was created and then every interval after: the minute 18:35 UTC on 20 Oct, two seconds after the schedules were created, holds five attempts and five messages, which is one first firing for each of the five schedules (price-expiry-sweep’s cron fell due at 18:35:00 UTC, 00:05 in Colombo); after that each 15-minute period holds one (the delta poll), each hour’s :35 minute three (with the two hourly schedules), each six-hour mark four (with the renewal), and 18:35 UTC on 21 Oct five again (with the cron). That pattern gives 168 firings from 18:35 on 20 Oct to 21:05 on 21 Oct, which is the total both metrics show.

Schedule First expected firing (UTC) Invocation attempts that minute (group) Target errors NumberOfMessagesSent on et-dev-scheduler that minute Result
mailbox-subscription-renewal 2026-10-20 18:35 5 no datapoints 5 fired; message reached the queue
mailbox-delta-poll 2026-10-20 18:35 5 no datapoints 5 fired; message reached the queue
price-expiry-sweep 2026-10-20 18:35 5 no datapoints 5 fired; message reached the queue (again 2026-10-21 18:35: 5 attempts)
email-upload-cleanup 2026-10-20 18:35 5 no datapoints 5 fired; message reached the queue
lead-triage-sweep (#4) 2026-10-20 18:35 5 no datapoints 5 fired; message reached the queue

Result: recorded for WI-3.1 — every schedule has fired with no target error, and the scheduler queue received as many messages as the schedules sent, in every period. The scheduler dead-letter queue is empty (first entry above). No schedule is left for WI-3.1 step 2.1d. The worker’s log line for each tick is F38-f’s (parent AC-40).

Recorded 2026-10-23 by F38-b-WI-3.1. The malformed messages that prove each redrive, and their removal. Each write ran under the temporary policy f38-b-write-checks (AC-12) with the lead’s approval at the time. The running task definition was backend:60, which sets ET_MESSAGING_SCHEDULER_QUEUE=et-dev-scheduler and ET_MESSAGING_MAILBOX_INGEST_QUEUE=et-dev-mailbox-ingest, so the running backend consumes both queues (the run sheet’s ingest branch B). Both dead-letter queues held 0 messages before the run. The full record is .forge/plans/F38-b/F38-b-WI-3.1-record.md, R-0 and R-3 to R-5.

2026-10-23 09:29 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 09:28 UTC — the scheduler queue

$ aws sqs send-message --queue-url https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-scheduler --message-body '{"jobKind":"scheduled-demo-job","idempotencyKey":"scheduled-demo-job#F38-b-WI-3.1-20261023T055341Z"}' --query '[MessageId,MD5OfMessageBody]' --output text --profile et-dev --region us-east-1
e0ad1e5d-5dba-464b-9052-6dc855a041c8	68aab2f46f8e9ea4253f9b1414587f1b
EXIT=0
$ aws sqs get-queue-attributes --queue-url https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-scheduler-dlq --attribute-names ApproximateNumberOfMessages … (every 30 seconds)
09:29:53Z ApproximateNumberOfMessages=0
…
09:31:35Z ApproximateNumberOfMessages=1
EXIT=0
$ aws logs filter-log-events --log-group-name /ecs/dev/backend --filter-pattern '"scheduled-demo-job"' … (from the send)
3 lines, at 09:29:46, 09:30:16 and 09:30:46 UTC, each ERROR from io.awspring.cloud.sqs.listener.sink.AbstractMessageProcessingPipelineSink: "Error processing message e0ad1e5d-5dba-464b-9052-6dc855a041c8"
EXIT=0

Result: met — the message sent at 09:29:41 UTC was received and rejected three times by the running backend, 30 seconds apart (the visibility timeout), and reached et-dev-scheduler-dlq at 09:31:35 UTC; the message deleted from it below has the same MessageId.

2026-10-23 09:41 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 09:37 UTC — the mailbox ingest queue

$ aws sqs send-message --queue-url https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-mailbox-ingest --message-body 'F38-b WI-3.1 malformed test message F38-b-WI-3.1-20261023T055341Z' --query '[MessageId,MD5OfMessageBody]' --output text --profile et-dev --region us-east-1
a9f72d99-0e68-46e6-8db4-3802f18d3b6a	95210bd020ec1b7bd5222a1b18f192d9
EXIT=0
$ aws sqs get-queue-attributes --queue-url https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-mailbox-ingest-dlq --attribute-names ApproximateNumberOfMessages … (every 30 seconds)
09:41:44Z ApproximateNumberOfMessages=0
…
09:43:31Z ApproximateNumberOfMessages=1
EXIT=0

Result: met — no CLI receive was made (branch B): the running backend received the plain-text message three times and failed on it, and it reached et-dev-mailbox-ingest-dlq at 09:43:31 UTC, about two minutes after the send at 09:41:28 UTC; the message deleted from it below has the same MessageId.

2026-10-23 09:50 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 09:49 UTC — the two test messages removed

$ aws sqs receive-message --queue-url <each dead-letter queue> --max-number-of-messages 10 --visibility-timeout 60 --wait-time-seconds 10 --query "Messages[?contains(Body, 'F38-b-WI-3.1-20261023T055341Z')] | [0].[MessageId,ReceiptHandle]" --output text …, then aws sqs delete-message with that message's receipt handle (not recorded)
et-dev-scheduler-dlq       MessageId=e0ad1e5d-5dba-464b-9052-6dc855a041c8  RECEIVE_EXIT=0  DELETE_EXIT=0
et-dev-mailbox-ingest-dlq  MessageId=a9f72d99-0e68-46e6-8db4-3802f18d3b6a  RECEIVE_EXIT=0  DELETE_EXIT=0
$ aws sqs get-queue-attributes --queue-url <each dead-letter queue> --attribute-names ApproximateNumberOfMessages ApproximateNumberOfMessagesNotVisible … (60 seconds later)
et-dev-scheduler-dlq       0 / 0
et-dev-mailbox-ingest-dlq  0 / 0
EXIT=0

Result: met — exactly the two test messages, found by the run’s marker, were deleted, and both dead-letter queues are back to their counts before the run.

Result: AC-4 met — a malformed message sent to each of et-dev-scheduler and et-dev-mailbox-ingest reached its dead-letter queue after three receives by the running backend, and both test messages were removed; with WI-2.1’s configuration reads and first firings above. Record: .forge/plans/F38-b/F38-b-WI-3.1-record.md, R-3 to R-5.

AC-5 — the attachments bucket

Recorded 2026-10-21 by F38-b-WI-2.1. The configuration reads, for WI-3.1 to cite. The upload, its unsigned 403 and signed 200, and its removal are WI-3.1’s.

2026-10-21 21:15 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws s3api get-bucket-versioning --bucket et-dev-attachments --output json --profile et-dev --region us-east-1
{"Status":"Enabled"}
EXIT=0
$ aws s3api get-public-access-block --bucket et-dev-attachments --output json --profile et-dev --region us-east-1
{"PublicAccessBlockConfiguration":{"BlockPublicAcls":true,"IgnorePublicAcls":true,"BlockPublicPolicy":true,"RestrictPublicBuckets":true}}
EXIT=0
$ aws s3api get-bucket-encryption --bucket et-dev-attachments --profile et-dev --region us-east-1
{"ServerSideEncryptionConfiguration":{"Rules":[{"ApplyServerSideEncryptionByDefault":{"SSEAlgorithm":"aws:kms","KMSMasterKeyID":"arn:aws:kms:us-east-1:570417736903:key/676c9326-b585-48f1-88d3-dbd2fd9d4d43"},"BucketKeyEnabled":true,"BlockedEncryptionTypes":{"EncryptionType":["SSE-C"]}}]}}
EXIT=0
$ aws s3api get-bucket-policy --bucket et-dev-attachments --query Policy --output text --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Sid":"DenyInsecureTransport","Effect":"Deny","Principal":"*","Action":"s3:*","Resource":["arn:aws:s3:::et-dev-attachments","arn:aws:s3:::et-dev-attachments/*"],"Condition":{"Bool":{"aws:SecureTransport":"false"}}},{"Sid":"DenyTlsBelow12","Effect":"Deny","Principal":"*","Action":"s3:*","Resource":["arn:aws:s3:::et-dev-attachments","arn:aws:s3:::et-dev-attachments/*"],"Condition":{"NumericLessThan":{"s3:TlsVersion":"1.2"}}}]}
EXIT=0
$ aws s3api get-bucket-lifecycle-configuration --bucket et-dev-attachments --profile et-dev --region us-east-1
aws: [ERROR]: An error occurred (NoSuchLifecycleConfiguration) when calling the GetBucketLifecycleConfiguration operation: The lifecycle configuration does not exist
EXIT=254
$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name s3 --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":["s3:ListBucketVersions","s3:ListBucket"],"Effect":"Allow","Resource":"arn:aws:s3:::et-dev-attachments"},{"Action":["s3:PutObject","s3:GetObject","s3:DeleteObjectVersion","s3:DeleteObject"],"Effect":"Allow","Resource":"arn:aws:s3:::et-dev-attachments/*"}]}
EXIT=0
$ aws guardduty list-malware-protection-plans --output json --profile et-dev --region us-east-1
{"MalwareProtectionPlans":[]}
EXIT=0
$ aws ecs describe-services --cluster et-dev --services backend --query 'services[0].taskDefinition' --output text --profile et-dev --region us-east-1
arn:aws:ecs:us-east-1:570417736903:task-definition/backend:51
EXIT=0
$ aws ecs describe-task-definition --task-definition backend:51 --query "taskDefinition.containerDefinitions[].environment[?name=='ET_FILES_SCAN_PROVIDER' || name=='ET_FILES_SCAN_OFF_ACKNOWLEDGED']" --output json --profile et-dev --region us-east-1
[[{"name":"ET_FILES_SCAN_OFF_ACKNOWLEDGED","value":"true"},{"name":"ET_FILES_SCAN_PROVIDER","value":"none"}]]
EXIT=0

Result: recorded for WI-3.1 — AC-5 is WI-3.1’s. Each item read is as spec FR-5 states: versioning Enabled; the four public-access blocks true; default encryption aws:kms with the storage key and Bucket Keys on (SSE-C is also refused, which #3’s base/attachments sets); a bucket policy with two deny statements for every principal on the bucket and its objects, one on aws:SecureTransport false and one on s3:TlsVersion below 1.2; no lifecycle configuration (NoSuchLifecycleConfiguration is the expected answer, so EXIT 254 here is the result, not a failure — register §B F20-h row (a)); the task role’s s3 policy exactly s3:ListBucket and s3:ListBucketVersions on the bucket and s3:PutObject, s3:GetObject, s3:DeleteObject and s3:DeleteObjectVersion on its objects, which covers the seven S3 calls of S3ObjectStore on forge-et-be main at 9097a68b (put, list versions, delete in batches, delete, get, head, signed GET); no GuardDuty protection plan; and the running task definition backend:51 with ET_FILES_SCAN_PROVIDER=none and ET_FILES_SCAN_OFF_ACKNOWLEDGED=true.

Recorded 2026-10-23 by F38-b-WI-3.1. The upload, its unsigned 403 and signed 200, and its removal, under the temporary policy f38-b-write-checks (AC-12). The full record is .forge/plans/F38-b/F38-b-WI-3.1-record.md, R-6.

2026-10-23 09:56 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 09:54 UTC

$ aws s3api put-object --bucket et-dev-attachments --key f38b-wi31-test/F38-b-WI-3.1-20261023T055341Z.txt --body <a one-line file naming WI-3.1 and the marker> --query '[VersionId,ServerSideEncryption,SSEKMSKeyId,BucketKeyEnabled]' --output text --profile et-dev --region us-east-1
agNr_dmdgwF5i_lvnMi8VCeiYuOFhQng	aws:kms	arn:aws:kms:us-east-1:570417736903:key/676c9326-b585-48f1-88d3-dbd2fd9d4d43	True
EXIT=0
$ curl -sS -o /dev/null -w 'HTTP=%{http_code}\n' https://et-dev-attachments.s3.us-east-1.amazonaws.com/f38b-wi31-test/F38-b-WI-3.1-20261023T055341Z.txt
HTTP=403
EXIT=0
$ aws s3 presign s3://et-dev-attachments/f38b-wi31-test/F38-b-WI-3.1-20261023T055341Z.txt --expires-in 300 …, then curl the URL (the URL is not recorded)
PRESIGN_EXIT=0
X-Amz-Algorithm=AWS4-HMAC-SHA256 (1)
X-Amz-Expires=300
HTTP=200
CURL_EXIT=0
CMP_EXIT=0   (the body read back is the body uploaded)
$ aws s3api delete-object --bucket et-dev-attachments --key f38b-wi31-test/F38-b-WI-3.1-20261023T055341Z.txt --version-id agNr_dmdgwF5i_lvnMi8VCeiYuOFhQng --output json …
{"VersionId": "agNr_dmdgwF5i_lvnMi8VCeiYuOFhQng"}
EXIT=0
$ aws s3api list-object-versions --bucket et-dev-attachments --prefix f38b-wi31-test/F38-b-WI-3.1-20261023T055341Z.txt --query '{Versions: Versions, DeleteMarkers: DeleteMarkers}' --output json …
{"Versions": null, "DeleteMarkers": null}
EXIT=0

Result: AC-5 met — an object uploaded with no encryption argument was stored under the storage key with Bucket Keys on; an unsigned GET was refused (403) and a five-minute SigV4 signed URL served it (200) with the same content; the version was deleted, leaving no version and no delete marker; with WI-2.1’s configuration reads above.

AC-6 — the KMS keys

Recorded 2026-10-21 by F38-b-WI-2.1.

2026-10-21 21:12 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws kms list-aliases --profile et-dev --region us-east-1 --query "Aliases[?AliasName=='alias/et-dev-storage' || AliasName=='alias/et-dev-mailbox-token'].[AliasName,TargetKeyId]" --output text
alias/et-dev-mailbox-token	ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040
alias/et-dev-storage	676c9326-b585-48f1-88d3-dbd2fd9d4d43
EXIT=0
$ aws kms describe-key --key-id alias/<alias> --query 'KeyMetadata.[KeyId,Arn,KeyState,KeyManager,KeySpec,KeyUsage]' --output text --profile et-dev --region us-east-1
EXIT=0      (each key below)
$ aws kms get-key-rotation-status --key-id <key id> --profile et-dev --region us-east-1
EXIT=0      (each key below)
$ aws kms get-key-policy --key-id <key id> --policy-name default --output text --profile et-dev --region us-east-1
EXIT=0      (each key below)
Alias Key id KeyState KeyManager Spec and usage Rotation Key policy
alias/et-dev-storage 676c9326-b585-48f1-88d3-dbd2fd9d4d43 Enabled CUSTOMER SYMMETRIC_DEFAULT, ENCRYPT_DECRYPT KeyRotationEnabled true, 365 days (as found; the spec requires it only on the mailbox token key) one statement, EnableIamPolicies: arn:aws:iam::570417736903:root allowed kms:* on *
alias/et-dev-mailbox-token ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040 Enabled CUSTOMER SYMMETRIC_DEFAULT, ENCRYPT_DECRYPT KeyRotationEnabled true, 365 days one statement, EnableIamPolicies: arn:aws:iam::570417736903:root allowed kms:* on *

2026-10-21 21:13 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness) — the task role’s KMS grant

$ aws iam list-role-policies --role-name et-dev-backend-task --output text --profile et-dev --region us-east-1
POLICYNAMES	bedrock
POLICYNAMES	kms
POLICYNAMES	s3
POLICYNAMES	ses
POLICYNAMES	sqs
POLICYNAMES	sqs-send
EXIT=0
$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name kms --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":["kms:GenerateDataKey","kms:Decrypt"],"Effect":"Allow","Resource":"arn:aws:kms:us-east-1:570417736903:key/ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040","Sid":"DirectKeys"},{"Action":["kms:GenerateDataKey","kms:Decrypt"],"Condition":{"StringEquals":{"kms:ViaService":"s3.us-east-1.amazonaws.com"}},"Effect":"Allow","Resource":"arn:aws:kms:us-east-1:570417736903:key/676c9326-b585-48f1-88d3-dbd2fd9d4d43","Sid":"KeysThroughS3"}]}
EXIT=0
$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name <sqs | sqs-send | s3 | bedrock | ses> --query PolicyDocument --output json --profile et-dev --region us-east-1
EXIT=0      (each; quoted under AC-4, AC-5 and AC-11 — none names a KMS key)

Result: met — the two customer-managed keys exist with their aliases, both Enabled; rotation is on for the mailbox token key; each key policy holds only the account-root delegation statement, with no S3 service statement (spec FR-6, Rev 2); the task role’s kms policy allows kms:GenerateDataKey and kms:Decrypt on the mailbox token key directly and on the storage key only with kms:ViaService = s3.us-east-1.amazonaws.com (Rev 4), and no other inline policy of the task role names a KMS key; and the attachments bucket’s default encryption names the storage key (AC-5 above).

AC-7 — the frontends’ hosting

Recorded 2026-10-21 by F38-b-WI-2.1. The security headers (spec Rev 7), as found; the direct S3 403, the records and the capture app are WI-4.1’s.

2026-10-21 21:16 UTC — Claude (F38-b-WI-2.1), curl

$ curl -sI https://dev.crm.empire.lk/
HTTP/2 200
content-type: text/html
x-cache: RefreshHit from cloudfront
x-frame-options: SAMEORIGIN
referrer-policy: strict-origin-when-cross-origin
x-content-type-options: nosniff
strict-transport-security: max-age=31536000; includeSubDomains
EXIT=0
$ curl -sI https://dev.crm.empire.lk/leads
HTTP/2 200
content-type: text/html
x-cache: Error from cloudfront
x-frame-options: SAMEORIGIN
referrer-policy: strict-origin-when-cross-origin
x-content-type-options: nosniff
strict-transport-security: max-age=31536000; includeSubDomains
EXIT=0
$ curl -s https://dev.crm.empire.lk/      (saved; its script and stylesheet tags listed from the saved file)
<script type="module" crossorigin src="/assets/index-hY2ghhOe.js">
<link rel="modulepreload" crossorigin href="/assets/rolldown-runtime-WkwZ5CU-.js">
… six more modulepreload links …
<link rel="stylesheet" crossorigin href="/assets/ui-D8CHnqLA.css">
<link rel="stylesheet" crossorigin href="/assets/index-LUUQTCBE.css">
EXIT=0
$ curl -sI https://dev.crm.empire.lk/assets/<file>
EXIT=0      (each of the ten files below)

(Header lines picked from each saved response; neither response has a content-security-policy line.)

Files Status content-type x-content-type-options
index-hY2ghhOe.js, rolldown-runtime-WkwZ5CU-.js, ui-DKwIVres.js, query-DtIBILKk.js, dist-BAH3iopo.js, dist-Cs8OWG2r.js, dist-duYbzv0e.js, dist-CJdm8Q5X.js 200 text/javascript nosniff
ui-D8CHnqLA.css, index-LUUQTCBE.css 200 text/css nosniff

2026-10-21 21:17 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws cloudfront get-distribution-config --id E2RVC3893O10TJ --query 'DistributionConfig.[DefaultCacheBehavior.ResponseHeadersPolicyId,CustomErrorResponses]' --output json --profile et-dev --region us-east-1
["6b769820-5e38-45cf-b918-a3d601e35c6f",{"Quantity":2,"Items":[{"ErrorCode":403,"ResponsePagePath":"/index.html","ResponseCode":"200","ErrorCachingMinTTL":0},{"ErrorCode":404,"ResponsePagePath":"/index.html","ResponseCode":"200","ErrorCachingMinTTL":0}]}]
EXIT=0
$ aws cloudfront get-response-headers-policy --id 6b769820-5e38-45cf-b918-a3d601e35c6f --query 'ResponseHeadersPolicy.ResponseHeadersPolicyConfig.[Name,SecurityHeadersConfig]' --output json --profile et-dev --region us-east-1
["et-dev-spa-security-headers",{"XSSProtection":{},"FrameOptions":{"Override":true,"FrameOption":"SAMEORIGIN"},"ReferrerPolicy":{"Override":true,"ReferrerPolicy":"strict-origin-when-cross-origin"},"ContentSecurityPolicy":{},"ContentTypeOptions":{"Override":true},"StrictTransportSecurity":{"Override":true,"IncludeSubdomains":true,"Preload":false,"AccessControlMaxAgeSec":31536000}}]
EXIT=0

Result: recorded as found — AC-7 is WI-4.1’s. The response-headers policy et-dev-spa-security-headers is attached to the default cache behaviour and sets the four headers (HSTS one year with includeSubDomains and no preload, nosniff, SAMEORIGIN, strict-origin-when-cross-origin) and no CSP. The deep path /leads is answered by the 403 → /index.html custom error response (x-cache: Error from cloudfront) and carries the same four headers, which answers spec Rev 7’s open question: a custom error response carries the policy’s headers. Every script and stylesheet index.html loads answers with a JavaScript or CSS content-type, so nosniff has nothing to block. Not yet — the browser load (plan D-10): the agent that wrote this entry had no browser tool, so loading https://dev.crm.empire.lk/ and reading the console for a refused script or stylesheet is still to do, and register §B FS-1’s security-headers row stays open until it is recorded here.

Recorded 2026-10-26 by F38-b-WI-4.1. What WI-2.1’s entries do not read: the CRM’s bodies, the direct S3 requests, the DNS answers, the browser load WI-2.1 left open, and the capture app. Nothing under base/cloudfront changed since #3 (AC-1 above), so WI-2.1’s CRM header reads stand. The capture bucket holds a build (F38-d’s capture deploy, since 25 Oct), so every capture read is made and none is deferred (plan D-16).

2026-10-26 05:37 UTC — Claude (F38-b-WI-4.1), curl

$ curl -sS -D <saved headers> -o <saved body> -w '%{http_code} %{content_type}\n' https://<host><path>
EXIT=0      (each host and path in the table below)
$ shasum -a 256 <the three saved bodies of each host>
EXIT=0
$ curl -sS -o /dev/null -w '%{http_code}\n' https://et-dev-spa.s3.us-east-1.amazonaws.com/index.html
403
EXIT=0
$ curl -sS -o /dev/null -w '%{http_code}\n' https://et-dev-capture.s3.us-east-1.amazonaws.com/index.html
403
EXIT=0
Host Path Status, content-type Body SHA-256 x-cache The four headers; CSP
dev.crm.empire.lk / 200 text/html 3c2bd288…6c5ef2 Miss from cloudfront all four; none
dev.crm.empire.lk /index.html 200 text/html 3c2bd288…6c5ef2 — —
dev.crm.empire.lk /leads/f38-b-wi-4-1-check 200 text/html 3c2bd288…6c5ef2 Error from cloudfront all four; none
capture.dev.crm.empire.lk / 200 text/html 6e44b7ac…8679a1 RefreshHit from cloudfront all four; none
capture.dev.crm.empire.lk /index.html 200 text/html 6e44b7ac…8679a1 — —
capture.dev.crm.empire.lk /leads/f38-b-wi-4-1-check 200 text/html 6e44b7ac…8679a1 Error from cloudfront all four; none

“All four” is strict-transport-security: max-age=31536000; includeSubDomains, x-content-type-options: nosniff, x-frame-options: SAMEORIGIN and referrer-policy: strict-origin-when-cross-origin, picked with grep -i -E '^(strict-transport-security|x-content-type-options|x-frame-options|referrer-policy|content-security-policy|x-cache):' from each saved header file; no file has a content-security-policy line. The full hashes are 3c2bd28833e7eb20244321717d9ae96856a8aed789101c7f8f87ce35df6c5ef2 and 6e44b7ac64db4845bbfe497190cf01030942a189a6892860f449ef81418679a1.

2026-10-26 05:37 UTC — Claude (F38-b-WI-4.1), profile et-dev (svc-harness); git: forge-et-iaac at c728e06b

$ aws s3api list-objects-v2 --bucket et-dev-capture --prefix index.html --max-keys 1 --query 'KeyCount' --profile et-dev --region us-east-1
1
EXIT=0
$ aws cloudfront get-distribution-config --id EZ06T8FTKF7UZ --query 'DistributionConfig.{Aliases:Aliases.Items,MinTLS:ViewerCertificate.MinimumProtocolVersion,Cert:ViewerCertificate.ACMCertificateArn,HeadersPolicy:DefaultCacheBehavior.ResponseHeadersPolicyId,Origin:Origins.Items[0].DomainName,OAC:Origins.Items[0].OriginAccessControlId,Errors:CustomErrorResponses.Items[].[ErrorCode,ResponseCode,ResponsePagePath]}' --output json --profile et-dev --region us-east-1
"Aliases": ["capture.dev.crm.empire.lk"], "MinTLS": "TLSv1.2_2021", "Cert": "…:certificate/5536e4c9-c078-4698-af50-63f5087b427d",
"HeadersPolicy": "dc85f29f-e39d-4902-bc96-6d9b524165b0", "Origin": "et-dev-capture.s3.us-east-1.amazonaws.com", "OAC": "EZG9AI4S3ODHU",
"Errors": [[403, "200", "/index.html"], [404, "200", "/index.html"]]
EXIT=0
$ aws cloudfront get-response-headers-policy --id dc85f29f-e39d-4902-bc96-6d9b524165b0 --query 'ResponseHeadersPolicy.ResponseHeadersPolicyConfig.{Name:Name,Sec:SecurityHeadersConfig}' --output json --profile et-dev --region us-east-1
"Name": "et-dev-capture-security-headers", "Sec": {"XSSProtection": {}, "FrameOptions": {"Override": true, "FrameOption": "SAMEORIGIN"}, "ReferrerPolicy": {"Override": true, "ReferrerPolicy": "strict-origin-when-cross-origin"}, "ContentSecurityPolicy": {}, "ContentTypeOptions": {"Override": true}, "StrictTransportSecurity": {"Override": true, "IncludeSubdomains": true, "Preload": false, "AccessControlMaxAgeSec": 31536000}}
EXIT=0
$ aws acm list-certificates --query "CertificateSummaryList[?DomainName=='capture.dev.crm.empire.lk'].CertificateArn" --output text --profile et-dev --region us-east-1
arn:aws:acm:us-east-1:570417736903:certificate/5536e4c9-c078-4698-af50-63f5087b427d
EXIT=0
$ aws acm describe-certificate --certificate-arn arn:aws:acm:us-east-1:570417736903:certificate/5536e4c9-c078-4698-af50-63f5087b427d --query 'Certificate.{Domain:DomainName,Status:Status,InUseBy:InUseBy,NotAfter:NotAfter}' --output json --profile et-dev --region us-east-1
{"Domain": "capture.dev.crm.empire.lk", "Status": "ISSUED", "InUseBy": ["arn:aws:cloudfront::570417736903:distribution/EZ06T8FTKF7UZ"], "NotAfter": "2027-05-08T05:29:59+05:30"}
EXIT=0
$ git -C forge-et-iaac grep -n 'name *= *"capture"' c728e06b -- overlays/dev
c728e06b:overlays/dev/main.tf:195:  name        = "capture"
c728e06b:overlays/dev/main.tf:202:  name                        = "capture"
EXIT=0

2026-10-26 05:37 UTC — Claude (F38-b-WI-4.1), curl (dig from the developer’s machine)

$ dig NS crm.empire.lk +short
ns-1983.awsdns-55.co.uk.
ns-879.awsdns-45.net.
ns-1313.awsdns-36.org.
ns-344.awsdns-43.com.
EXIT=0
$ dig +tcp +norec <host> A @ns-1983.awsdns-55.co.uk.
EXIT=0      (each of the three hosts below)
Host Flags Answer
dev.crm.empire.lk qr aa four A records (CloudFront, e.g. 3.169.64.84)
api.dev.crm.empire.lk qr aa two A records (the load balancer, 44.217.192.253, 44.209.91.2)
capture.dev.crm.empire.lk qr aa four A records (CloudFront, e.g. 3.169.64.17)

2026-10-26 05:40 UTC — Claude (F38-b-WI-4.1), browser (the Claude desktop app’s built-in browser, signed out)

open https://dev.crm.empire.lk/   → redirected to /login, the sign-in screen rendered ("Sign in", "Sign in with Microsoft")
console: [error] Failed to load resource: the server responded with a status of 401 ()      (the only message: the fetch of https://api.dev.crm.empire.lk/api/v1/me)
performance.getEntriesByType('resource'), the .js and .css files:
  /assets/index-ciRwHs5A.js, LoginRoute-DmkMM6uO.js, traceId-P1Psirjy.js, usePageRestore-Bla3Wplb.js (script)
  rolldown-runtime-WkwZ5CU-.js, ui-CqTkyQEs.js, query-DLAHpSBe.js and five dist-*.js (modulepreload)
  ui-FS1kcD4K.css, index-BLw8hiUN.css, LoginRoute-DYyI5b-o.css (stylesheet)      — every one status 200
document.styleSheets.length = 4

Result: met — both hosts answer the root, /index.html and a deep path with the same index.html (the deep path through the 403 → /index.html custom error response), each response carrying the four headers and no CSP; a direct unsigned request to either bucket’s index.html is refused with 403; every dev host is answered authoritatively from the crm.empire.lk zone. The capture app’s certificate is ISSUED and used only by its distribution, which requires TLSv1.2_2021 and attaches et-dev-capture-security-headers (the same four headers, no CSP), from bucket et-dev-capture (name = "capture" in overlays/dev). In the browser, every script and stylesheet the CRM’s sign-in page loads arrived with 200 and the page rendered, with no refusal under nosniff in the console; the one console error is the signed-out session check answering 401, which is expected. This closes WI-2.1’s open browser load, so register §B FS-1’s security-headers row is validated.

AC-8 — the AgentCore Gateway

Recorded 2026-10-21 by F38-b-WI-2.1.

2026-10-21 21:16 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws bedrock-agentcore-control list-gateways --query 'items[].[name,status]' --output text --profile et-dev --region us-east-1
et-dev-gateway	READY
EXIT=0

Result: met — the SP-ARCH-004 stage-2 run of 13 Oct 2026 is the call through the dev gateway’s web-search target returning results: et-dev-gateway, its one target web-search on the managed web-search connector, called as svc-harness, with one signed tools/list and one test search returning results, and then the comparison’s searches for Empire Teas’ customers (.forge/design/ai/lead-research-source-options.md §11, “Access, as it was set up” and “Results”). The gateway still exists and is READY. No provider has been chosen finally: the lead’s decision of 20 Oct builds F11-c on Exa’s free credit, with the choice provisional until Empire Teas answers (§11; .forge/owed.md, “Choose the lead-research search provider”), so FR-15 is not retired and the gateway stays (spec W-7).

AC-9 — the values for F38-d

Recorded 2026-10-21 by F38-b-WI-2.1. The three overlays/dev outputs of spec FR-9, read from the Outputs: block that terraform apply printed in run 39475850266’s log from the state it had just written, and each cross-checked in the account.

2026-10-21 21:13 UTC — Claude (F38-b-WI-2.1), gh as tharinda-altrium; profile et-dev (svc-harness)

$ gh run view 39475850266 -R altrium-io/forge-et-iaac --log      (saved whole; the terraform::apply lines below are taken from the saved file)
terraform::apply  2026-10-20T18:35:24.8657690Z Apply complete! Resources: 28 added, 5 changed, 1 destroyed.
terraform::apply  2026-10-20T18:35:24.8751472Z Outputs:
…
terraform::apply  2026-10-20T18:35:24.8755329Z attachments_bucket_name = "et-dev-attachments"
terraform::apply  2026-10-20T18:35:24.8756275Z lead_triage_queue_name = "et-dev-lead-triage"
terraform::apply  2026-10-20T18:35:24.8757838Z mailbox_ingest_queue_name = "et-dev-mailbox-ingest"
terraform::apply  2026-10-20T18:35:24.8758814Z mailbox_token_key_id = "ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040"
…
EXIT=0
$ aws s3api head-bucket --bucket et-dev-attachments --profile et-dev --region us-east-1
{"BucketArn": "arn:aws:s3:::et-dev-attachments", "BucketRegion": "us-east-1", "AccessPointAlias": false}
EXIT=0
$ aws kms describe-key --key-id alias/et-dev-mailbox-token --query KeyMetadata.KeyId --output text --profile et-dev --region us-east-1
ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040
EXIT=0
$ aws sqs get-queue-url --queue-name et-dev-mailbox-ingest --profile et-dev --region us-east-1
{"QueueUrl": "https://sqs.us-east-1.amazonaws.com/570417736903/et-dev-mailbox-ingest"}
EXIT=0
Output Value Task-definition variable Outputs: line AWS cross-check Result
attachments_bucket_name et-dev-attachments ET_FILES_STORAGE_BUCKET attachments_bucket_name = "et-dev-attachments" head-bucket EXIT 0, us-east-1 met
mailbox_token_key_id ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040 ET_MAILBOX_TOKEN_KEY_ID mailbox_token_key_id = "ce1dbe4f-711e-4dd3-ab6e-9ce1930e9040" describe-key on the alias gives the same id met
mailbox_ingest_queue_name et-dev-mailbox-ingest ET_MESSAGING_MAILBOX_INGEST_QUEUE mailbox_ingest_queue_name = "et-dev-mailbox-ingest" get-queue-url EXIT 0 met

Result: met — the apply log’s Outputs: block is Terraform printing the root outputs from the state it had just written, so it stands for terraform output, which nobody on the delivery team can run (plan D-4; the fallback was not needed). Register §D carries the three values. The fourth output, lead_triage_queue_name = et-dev-lead-triage, is forge-et-iaac#4’s, and forge-et-be’s .aws/td.dev.json already sets it as ET_MESSAGING_LEADTRIAGEQUEUE (forge-et-be#283). Writing the three F38-b values into the task definition is F38-d’s (FR-6, its WI-4.1).

AC-10 — the closing sweep

Recorded 2026-10-26 by the /forge-deliver close of wave 3 (Claude, for Tharinda), the second of waves 3 and 4 to merge, on feature/F38-b-wave-3 after main (with wave 4) was merged in.

2026-10-26 06:30 UTC — Claude (F38-b wave 3 close), repository forge-et-harness, the merged tree of feature/F38-b-wave-3

The acceptance criteria, confirmed against their evidence: AC-1, AC-2, AC-3 and AC-7 against WI-4.1’s entries above (AC-1’s no-change run 40278225094); AC-6, AC-8, AC-9 and AC-11 against WI-2.1’s Result: lines; AC-4, AC-5 and AC-12 against .forge/plans/F38-b/F38-b-WI-3.1-record.md (R-3 to R-7, PASS; confirmed with the lead at M2); AC-13 against §4 “What it holds”, the base/secrets bullet (forge-et-harness#933).

The register rows FR-10 lists, as they read on this tree:

Row (.forge/cloud-gated-validation.md) Status Evidence or reason
§B FS-2 — RDS engine version pinned validated AC-3 (WI-4.1)
§B FS-2 — the migration role can create btree_gist validated AC-3 (WI-4.1)
§B FS-7 — real SQS main queue + DLQ with redrive validated AC-4 (WI-3.1, 2026-10-23)
§B FS-9 — EventBridge Scheduler schedules + their role validated (producer half; the worker’s log line by F38-f WI-1.1) AC-4 (WI-2.1)
§B FS-9 — shared or dedicated scheduler queue validated AC-4 (WI-2.1)
§B FS-11 — S3 bucket with public-access block and versioning validated AC-5 (WI-3.1, 2026-10-23)
§B FS-11 — task-role S3 permissions open — the policy is recorded (AC-5, AC-6); the presigned-URL and real-call proof is F38-f’s the row’s 2026-10-21 note
§B F20-a — customer-managed KMS key with rotation; task role granted open — the key, rotation and grant are recorded (AC-6); the real KMS-wrapped refresh-token round trip is F20-a’s own live check the row’s 2026-10-21 note
§B F20-b — the ingest queue and ET_MESSAGING_MAILBOX_INGEST_QUEUE open — the queue, its redrive (AC-4) and the running task’s key are recorded; the back-fill and stop-the-worker checks are F38-f’s the row’s 2026-10-23 note
§B F20-b — the renewal and delta-poll schedules validated (producer half; the worker’s log lines by F38-f WI-1.1) AC-4 (WI-2.1)
§B F20-h — (d) the email-upload-cleanup schedule validated (producer half and the worker’s log line) AC-4 (WI-2.1), F38-f WI-1.1
§B F26-g — the price-expiry-sweep schedule validated (producer half and the worker’s log line) AC-4 (WI-2.1), F38-f WI-1.1
§B F12 row 1 — the task role’s Bedrock policy open — the policy half is met (AC-11); the real call from the task is F38-f’s the row’s 2026-10-21 note

The register’s ## Log carries F38-b’s dated lines (WI-2.1 on 2026-10-21, WI-3.1 on 2026-10-23, WI-4.1 and its follow-up on 2026-10-26), and this section carries one subsection per acceptance criterion.

Result: met — every FR-10 row reads validated with evidence or open with a stated reason, and every other acceptance criterion is ticked against its evidence.

AC-11 — the Bedrock grant and least privilege

Recorded 2026-10-21 by F38-b-WI-2.1. The configured models were read from forge-et-be origin/main at 9097a68b (src/main/resources/application.yml, et.ai.features): inquiry-extract and lead-triage on us.anthropic.claude-haiku-4-5-20251001-v1:0, and email-draft on global.anthropic.claude-sonnet-5-5 — only the two profiles AD #19 pins.

2026-10-21 21:13 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name bedrock --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":"bedrock:InvokeModel","Effect":"Allow","Resource":["arn:aws:bedrock:us-west-2::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0","arn:aws:bedrock:us-east-2::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0","arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0","arn:aws:bedrock:us-east-1:570417736903:inference-profile/us.anthropic.claude-haiku-4-5-20251001-v1:0","arn:aws:bedrock:us-east-1:570417736903:inference-profile/global.anthropic.claude-sonnet-5-5","arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-5-5"]}]}
EXIT=0
$ aws iam get-role-policy --role-name et-dev-backend-task --policy-name ses --query PolicyDocument --output json --profile et-dev --region us-east-1
{"Version":"2012-10-17","Statement":[{"Action":["ses:SendRawEmail","ses:SendEmail"],"Effect":"Allow","Resource":["arn:aws:ses:us-east-1:570417736903:identity/dev.crm.empire.lk","arn:aws:ses:us-east-1:570417736903:configuration-set/et-dev-mail"]}]}
EXIT=0
$ aws iam list-role-policies --role-name et-dev-scheduler --output text --profile et-dev --region us-east-1
POLICYNAMES	sqs-send
EXIT=0
$ aws iam list-attached-role-policies --role-name et-dev-backend-task --output json --profile et-dev --region us-east-1
{"AttachedPolicies":[]}
EXIT=0
$ aws iam list-attached-role-policies --role-name et-dev-scheduler --output json --profile et-dev --region us-east-1
{"AttachedPolicies":[]}
EXIT=0

Every inline policy of both roles (the task role’s list is under AC-6; the sqs and sqs-send documents under AC-4; s3 under AC-5; kms under AC-6; the scheduler role’s sqs-send under AC-4):

Role Policy Actions Resources Conditions From
et-dev-backend-task bedrock bedrock:InvokeModel the us. Haiku 4.5 and global. Sonnet 5.5 inference profiles in us-east-1; Haiku 4.5’s foundation model in us-east-1, us-east-2 and us-west-2; Sonnet 5.5’s foundation model in any Region (arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-5-5) — #3
et-dev-backend-task kms kms:GenerateDataKey, kms:Decrypt the mailbox token key; the storage key the storage key only with kms:ViaService = s3.us-east-1.amazonaws.com #3
et-dev-backend-task s3 s3:ListBucket, s3:ListBucketVersions; s3:PutObject, s3:GetObject, s3:DeleteObject, s3:DeleteObjectVersion et-dev-attachments; et-dev-attachments/* — #3
et-dev-backend-task sqs sqs:ReceiveMessage, sqs:GetQueueUrl, sqs:GetQueueAttributes, sqs:DeleteMessage, sqs:ChangeMessageVisibility et-dev-scheduler, et-dev-mailbox-ingest, et-dev-lead-triage — built 9 Oct; ingest from #3; lead triage from #4
et-dev-backend-task sqs-send sqs:SendMessage et-dev-mailbox-ingest, et-dev-lead-triage — #3; lead triage from #4
et-dev-backend-task ses ses:SendEmail, ses:SendRawEmail the identity dev.crm.empire.lk; the configuration set et-dev-mail — 0eee623 (F38-a)
et-dev-scheduler sqs-send sqs:SendMessage et-dev-scheduler — #3

Result: met — the task role’s bedrock policy allows only bedrock:InvokeModel, on exactly the two inference profiles and the foundation models they route to (spec FR-11), and every model configured on forge-et-be main uses one of those two profiles, so no configured model lacks a grant. Neither role has an AWS-managed (or any attached) policy. The only * in any Resource of either role is the Region of Sonnet 5.5’s foundation-model ARN, which NFR-2 allows; the key policies’ Resource: "*" is the key itself (AC-6). The ses policy is F38-a’s and names its two ARNs only. The call from the task role is F38-f’s (parent AC-29).

AC-12 — the temporary policy removed

Opened 2026-10-23 by F38-b-WI-3.1. The temporary inline policy on svc-harness for WI-3.1’s write checks (spec NFR-4), from its creation to its removal. The full record is .forge/plans/F38-b/F38-b-WI-3.1-record.md, R-2 and R-7.

2026-10-23 08:39 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 08:38 UTC

$ aws iam put-user-policy --user-name svc-harness --policy-name f38-b-write-checks --policy-document file://<save folder>/R-2-3.2-2.json --profile et-dev
EXIT=0
$ aws iam list-user-policies --user-name svc-harness --output text --profile et-dev
POLICYNAMES	f38-b-write-checks
POLICYNAMES	spike-invoke-gateway
EXIT=0
$ aws iam get-user-policy --user-name svc-harness --policy-name f38-b-write-checks --query PolicyDocument --output json --profile et-dev
… (the document as created, five statements) …
EXIT=0

The policy f38-b-write-checks has five statements: sqs:SendMessage on et-dev-scheduler and et-dev-mailbox-ingest; sqs:ReceiveMessage on et-dev-mailbox-ingest; sqs:ReceiveMessage and sqs:DeleteMessage on et-dev-scheduler-dlq and et-dev-mailbox-ingest-dlq; s3:PutObject, s3:GetObject, s3:DeleteObject and s3:DeleteObjectVersion on arn:aws:s3:::et-dev-attachments/*; and kms:GenerateDataKey and kms:Decrypt on the storage key arn:aws:kms:us-east-1:570417736903:key/676c9326-b585-48f1-88d3-dbd2fd9d4d43.

Result: not yet — the policy exists; it is removed after AC-4 and AC-5 are recorded (.forge/owed.md).

2026-10-23 09:58 UTC — Nanndish Satgunarajah (F38-b-WI-3.1), profile et-dev (svc-harness), approved by Tharinda Rajapaksha at 09:57 UTC — the policy removed

$ aws iam delete-user-policy --user-name svc-harness --policy-name f38-b-write-checks --profile et-dev
EXIT=0
$ aws iam list-user-policies --user-name svc-harness --output text --profile et-dev   (15 seconds later)
POLICYNAMES	spike-invoke-gateway
EXIT=0
$ aws iam simulate-principal-policy --policy-source-arn arn:aws:iam::570417736903:user/svc-harness --action-names <the SQS, S3 and KMS actions> --resource-arns <the four queues | the test object | the storage key> …
sqs:SendMessage         implicitDeny on all four queues
sqs:ReceiveMessage      allowed on all four queues
sqs:DeleteMessage       implicitDeny on all four queues
s3:PutObject            implicitDeny
s3:GetObject            allowed
s3:DeleteObject         implicitDeny
s3:DeleteObjectVersion  implicitDeny
kms:GenerateDataKey     implicitDeny
kms:Decrypt             implicitDeny
EXIT=0

Result: AC-12 met — f38-b-write-checks was removed at 09:58 UTC on 2026-10-23, about 80 minutes after it was created, and is no longer listed; the six actions the temporary policy granted beyond svc-harness’s own policies read implicitDeny again, the same as before it was created (sqs:ReceiveMessage and s3:GetObject come from ReadOnlyAccess). The .forge/owed.md row for it is removed. Record: .forge/plans/F38-b/F38-b-WI-3.1-record.md, R-2 and R-7.

Other changes in forge-et-iaac#3 (spec Revs 3, 6)

Recorded 2026-10-21 by F38-b-WI-2.1. The 14-day dead-letter retention (Rev 3) is recorded under AC-4.

2026-10-21 21:16 UTC — Claude (F38-b-WI-2.1), profile et-dev (svc-harness)

$ aws ecr get-lifecycle-policy --repository-name crm/backend --query lifecyclePolicyText --output text --profile et-dev --region us-east-1
{"rulePriority":1,"description":"Expire untagged images after 7 days","selection":{"tagStatus":"untagged","countType":"sinceImagePushed","countUnit":"days","countNumber":7},"action":{"type":"expire"}}
{"rulePriority":2,"description":"Keep last 10 release images","selection":{"tagStatus":"tagged","tagPrefixList":["v"],"countType":"imageCountMoreThan","countNumber":10},"action":{"type":"expire"}}
{"rulePriority":3,"description":"Keep last 100 build images","selection":{"tagStatus":"tagged","tagPrefixList":["build-"],"countType":"imageCountMoreThan","countNumber":100},"action":{"type":"expire"}}
EXIT=0

(The policy’s three rules printed one per line from the saved JSON.)

Result: met — the policy that replaced the old one at the apply (AC-1) holds three rules: untagged images expire after 7 days, the last 10 images tagged v* are kept, and the last 100 images tagged build-* are kept (spec Rev 6).

F38-d — the dev deploy pipelines, checked

Opened 2026-10-19 by F38-d-WI-1.1. This section is F38-d’s evidence for the dev deploy pipelines: the record its acceptance criteria are ticked against (spec .forge/specs/F38-d-pipelines-first-deploy-spec.md). The rules: every AWS command is a read — a describe, list or get call, or sts get-caller-identity — made with the CLI profile et-dev (the IAM user svc-harness) in us-east-1; any other AWS call is made only with the lead’s approval at the time, and its entry names that approval. GitHub is read with gh, and git only through fetch, show, log, grep, ls-tree, cat-file and merge-base on the remote trunks. (Update 2026-10-26, F38-d WI-6.1: WI-6.1’s two throwaway forge-et-be branches were committed, pushed and deleted, and its manual runs, the CRM variable change and the ECR push were made, each with the lead’s approval at the time of the step; each entry names that approval.) Times are UTC. Identifiers are recorded — account IDs, ARNs, run IDs, commit SHAs, resource names — but never a secret value, an access-key ID or a person’s email address (spec NFR-2 and NFR-6; the Secrets boundary at the top of this file).

Every entry has one of three shapes. A read: a bold line **<YYYY-MM-DD HH:MM> UTC — Claude (F38-d-WI-<x.y>), <source>**, where the source is profile et-dev (svc-harness) for AWS, gh as <login> for GitHub, or the repository and commit for git; a text block holding $ <command>, the output lines that show the result (trimmed with … where lines were cut, never edited) and EXIT=<n>; then Result: — met, not met — <why>, not readable — <why>, or reported to the lead — .forge/owed.md. Where one command ran once per repository or user, the block shows it once and a table gives each one’s output. A workflow run: the same bold line, then Run: with the run’s URL, its trigger (push, workflow_dispatch), head SHA, conclusion and billed minutes (gh api repos/altrium-io/<repo>/actions/runs/<id>/timing), then the quoted log lines in a text block, then Result:. A register note: *Update <YYYY-MM-DD> (F38-d WI-<x.y>): <one-sentence result>; evidence: environments-and-identity.md §4 "F38-d — the dev deploy pipelines, checked", <subsection>.*, appended to the row’s “Validate / finish when cloud is up” cell; a row becomes validated only when everything in it is proved on dev. (Update 2026-10-23, F38-d WI-4.2: the billed minutes are each job’s completedAt − startedAt, rounded up to a whole minute and summed per run, because the run’s timing endpoint (…/actions/runs/<id>/timing) reports total_ms 0 for these private-repository runs; git is also read with rev-parse and diff --name-only, to compare file blobs and Gradle inputs between commits.)

The subsections, in this order: The review of the DevOps team's deploy files (FR-10) (WI-1.1), then one per spec acceptance criterion, in AC order — AC-1 — CI's AWS access (WI-1.1 opens it; WI-4.2, WI-6.1 and WI-7.1 add to it), AC-2 — the backend workflow on dev (WI-4.2 opens; WI-6.1 adds), AC-3 — the FR-20 checklists (WI-4.2 opens; WI-7.1 adds), AC-4 — ARM64 (WI-4.2), AC-5 — the frontends' deploys (WI-6.1 opens; WI-7.1 adds), AC-6 — dev.yaml and the required checks (WI-1.1 opens; WI-7.1 adds), AC-7 — task-definition values and secrets (WI-6.1 opens; WI-7.1 adds), AC-8 — the database certificate (WI-6.1), AC-9 — first deploy, sign-in and versions (WI-1.1 opens; WI-6.1 and WI-7.1 add), AC-10 — the closing sweep (WI-7.1), AC-11 — one image per commit (WI-6.1) and AC-12 — billed minutes and PR records (WI-4.2 opens it with a minutes table — run, repository, billed minutes, the WI plan’s estimate — and WI-6.1 and WI-7.1 add rows). A subsection is written only when it has content. A work item that opens AC-<n> inserts it immediately before the first #### AC-<m> heading here with m greater than n, or, when there is none, immediately before ### Agenda for the meeting with Shyamin Ayesh. Within a subsection, each work item’s entries follow one italic line, *Recorded <date> by F38-d-WI-<x.y>.*, after any earlier work item’s entries. Work items later than WI-1.1 also note what they change in “Checked against F38” above (WI-4.2 rows 5, 8, 9, 10 and 15; WI-6.1 rows 6 and 14; WI-7.1 checks all of them at AC-10) and in architecture.md § CI/CD Pipelines › How dev is deployed today, with a ### Rev <N> entry in that file’s ## Revisions.

The review of the DevOps team’s deploy files (FR-10)

Recorded 2026-10-19 by F38-d-WI-1.1. The four files were read in full at the commits below and checked against the items the .forge/owed.md row “Review the GitHub Actions workflows the DevOps team wrote” lists; the team’s standards (coding-standards.md §7, architecture.md § CI/CD Pipelines › Per repo, › How dev is deployed today and › Cross-cutting, AD #22, AD #24, AD #38, the F38 parent spec’s FR-19 – FR-23, and this feature’s FR-1 – FR-9 and NFR-1 – NFR-5); for mistakes and things missed; for inconsistencies between the repositories; and against the harness documents that describe them (this section’s “What it holds”, the two “dev deploy” paragraphs and “Checked against F38” rows 5–15, architecture.md § How dev is deployed today, register §B FS-1 and §D, which is compared as written).

2026-10-18 18:59 UTC — Claude (F38-d-WI-1.1), git: forge-et-be origin/main 5c457091, forge-et-fe origin/main 32c0dea5, forge-et-iaac forge-et-iaac#3 head 0ff63c5

$ gh pr view 3 --repo altrium-io/forge-et-iaac --json state,headRefOid,mergedAt,mergeCommit
{"headRefName":"feature/F38-b-wave-1","headRefOid":"0ff63c570567734c6830cd6795a85986ccf3fc31","mergeCommit":null,"mergedAt":null,"state":"OPEN"}
EXIT=0
$ git -C forge-et-be show origin/main:.github/workflows/build.yaml      # 102 lines
EXIT=0
$ git -C forge-et-be show origin/main:.aws/td.dev.json                  # 61 lines
EXIT=0
$ git -C forge-et-fe show origin/main:.github/workflows/build.yaml      # 71 lines, the last with no newline
EXIT=0
$ git -C forge-et-iaac show 0ff63c5:.github/workflows/dev.yaml          # 65 lines
EXIT=0
$ git -C <repo> log --format='%h %cI %an %s' <commit> -- <path>         # one per file
EXIT=0 for each of the four
Repository File Commit read (its date) Last changed in
forge-et-be .github/workflows/build.yaml origin/main 5c457091 (2026-10-17) 03404e8e, 2026-10-09, Shyamin Ayesh, pushed straight to main (“ci: deploy each dev revision to the backend service”)
forge-et-be .aws/td.dev.json origin/main 5c457091 (2026-10-17) dfc9be15, 2026-10-13, Shyamin Ayesh, pushed straight to main (“ci: read the Entra sign-in config from Secrets Manager in dev”)
forge-et-fe .github/workflows/build.yaml origin/main 32c0dea5 (2026-10-18) 56e18b4d, 2026-10-14, Dependabot, by forge-et-fe#265 (after 74f4feef and 1295ec20, forge-et-fe#259, BUG-136)
forge-et-iaac .github/workflows/dev.yaml forge-et-iaac#3 head 0ff63c5 (2026-10-18; #3 open, not merged) 0757750, 2026-10-18, the delivery team in #3; on origin/dev (76fce06) the file was last changed by 76fce06, Shyamin Ayesh, 2026-10-15

Direct pushes since 14 Oct (.claude/rules/git-conventions.md § Branch Protection): every commit since 14 Oct that changes .github/workflows or .aws on an app repository’s main, with the PR that carried it.

2026-10-18 19:00 UTC — Claude (F38-d-WI-1.1), forge-et-be origin/main 5c457091, gh as tharinda-altrium

$ git -C forge-et-be log --format='%H %cI %an %s' --since=2026-10-14 origin/main -- .github/workflows .aws
6f1c2fe89ab4ff9df38046f8627390767ddae123 2026-10-15T14:46:04+05:30 Tharinda Rajapaksha ci: raise commitlint's header-max-length to 100
EXIT=0
$ gh api repos/altrium-io/forge-et-be/commits/6f1c2fe8…/pulls --jq '[.[] | [.number,.base.ref,.merged_at]]'
[[263,"main","2026-10-15T10:40:30Z"]]
EXIT=0

Result: met — no direct push; the one commit came through forge-et-be#263.

2026-10-18 19:00 UTC — Claude (F38-d-WI-1.1), forge-et-fe origin/main 32c0dea5, gh as tharinda-altrium

$ git -C forge-et-fe log --format='%H %cI %an %s' --since=2026-10-14 origin/main -- .github/workflows .aws
1e60b7727bed6348fd90cfbdd6ac1f83778246b8 2026-10-15T14:45:10+05:30 Tharinda Rajapaksha ci: raise commitlint's header-max-length to 100
56e18b4d1515ef58607e0513cfb4c0e6357f02fa 2026-10-14T19:44:11Z dependabot[bot] ci(deps): Bump the github-actions group with 3 updates
74f4feef6660f3d15d014aaf1c0bb22227123aae 2026-10-14T10:29:24+05:30 Tharinda Rajapaksha ci(BUG-136): run one dev deploy at a time
1295ec20f4c4f2866afaab308dccc247fee7a173 2026-10-14T10:23:28+05:30 Tharinda Rajapaksha ci(BUG-136): stage the dev S3 sync with cache headers
EXIT=0
$ gh api repos/altrium-io/forge-et-fe/commits/<sha>/pulls --jq '[.[] | [.number,.base.ref,.merged_at]]'   # once per commit
EXIT=0 for each of the four
Commit PR (base, merged at)
1e60b772 forge-et-fe#278 (main, 2026-10-15T10:27:42Z)
56e18b4d forge-et-fe#265 (main, 2026-10-15T04:56:02Z)
74f4feef, 1295ec20 forge-et-fe#259 (main, 2026-10-14T05:06:19Z)

Result: met — no direct push.

2026-10-18 19:00 UTC — Claude (F38-d-WI-1.1), forge-et-pwa origin/main 96d48ba, gh as tharinda-altrium

$ git -C forge-et-pwa log --format='%H %cI %an %s' --since=2026-10-14 origin/main -- .github/workflows .aws
1b1f027777c1296c6d734f926f551267c9e16f6b 2026-10-15T15:59:41+05:30 Tharinda Rajapaksha ci: raise commitlint's header-max-length to 120
EXIT=0
$ gh api repos/altrium-io/forge-et-pwa/commits/1b1f0277…/pulls --jq '[.[] | [.number,.base.ref,.merged_at]]'
[[82,"main","2026-10-15T10:41:33Z"]]
EXIT=0

Result: met — no direct push; the one commit came through forge-et-pwa#82.

The findings. DR-<n> numbers are this section’s own series, in the order written, never renumbered; a withdrawn finding is struck through with the reason. Under is the spec requirement a finding falls under, or “outside FR-1 – FR-9”. Home is the F38-d work item that fixes it, “recorded only (FR-5)” for dev.yaml with the .forge/owed.md “DevOps follow-ups from F38-b’s review of forge-et-iaac” item it adds to, a .forge/owed.md row by title, or “corrected here”. A later work item replaces open with fixed — <PR>, <date> or recorded — <subsection>. DR-1 – DR-16 confirm the gaps “Checked against F38” rows 5, 6, 8, 9, 10, 14 and 15 record; DR-17 onwards are new.

# File and line Finding Kind Under Home Status
DR-1 forge-et-be .aws/td.dev.json:6; build.yaml:57 Row 5, confirmed: the task runs X86_64, and docker build sets no platform on the ubuntu-latest runner, so the image is linux/amd64. missing FR-3 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by run 39228765669’s docker build --platform linux/arm64 line, the image config arm64 and the task ARM64 (AC-3, AC-4 below)
DR-2 forge-et-be .aws/td.dev.json:14–58 Row 6, confirmed for the task definition: no stopTimeout (ECS’s default, 30 s, applies) and no JAVA_TOOL_OPTIONS. Row 6’s “None set” leaves out the image’s own default: Dockerfile:91 sets JAVA_TOOL_OPTIONS="-XX:MaxRAMPercentage=75 -XX:+ExitOnOutOfMemoryError", so the running task already exits on an out-of-memory error, and only the heap share differs from the lead’s 65 %. missing FR-6 F38-d-WI-4.1 fixed — forge-et-be#306, 2026-10-25; shown by the AC-7 task-definition read (F38-d WI-6.1)
DR-3 forge-et-iaac base/alb at 0ff63c5 Row 6, confirmed: no deregistration_delay on the backend’s target group (git grep deregistration_delay finds none), so the 300 s default applies. missing FR-6 F38-d-WI-3.1 fixed — forge-et-iaac#6, 2026-10-23; shown by R-4 and R-5 (F38-d WI-6.1)
DR-4 forge-et-be build.yaml:43, 56–60 Row 8, confirmed: the tag is build-$(git rev-parse --short HEAD), a short SHA, and every run builds and pushes, with no skip when ECR already holds the commit’s image. missing FR-2 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by the build-1e91f65f… tag in run 39228765669’s log and in aws ecr describe-images (AC-3 below); the skip itself is shown by WI-6.1’s manual run (AC-2)
DR-5 forge-et-iaac base/ecr/repository.tf:3, 53 at 0ff63c5 Row 8, confirmed: image_tag_mutability = "MUTABLE", and the lifecycle comment names build-<short-sha>. missing NFR-3 F38-d-WI-3.1 fixed — forge-et-iaac#6, 2026-10-23; shown by R-4 and R-5 (F38-d WI-6.1)
DR-6 forge-et-be build.yaml:32–75 Row 9, confirmed: one job, dev, configures the key (45–50) before docker build (56–57), so the image build — which compiles the jar inside Docker, not on the runner with setup-gradle — runs with AWS credentials in the job; no separate build job and no artifact. missing FR-2 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by run 39228765669’s two jobs, build with no AWS step, and its [context jar] load from client line with no [build line (AC-2, AC-3 below)
DR-7 forge-et-be build.yaml:20–25 Row 9, confirmed: triggered by a push to main or a v*.*.* tag only — no workflow_dispatch and no paths-ignore, so a workflow-only or documentation-only merge deploys too (row 9 (d)). missing FR-2 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by the workflow_dispatch run 39226901426, by yq '.on' on the file at 1e91f65f, and by forge-et-be#286’s and forge-et-fe#298’s workflow-only merges starting no run (AC-3 below)
DR-8 forge-et-be build.yaml:27–39 Row 9 (a) and ©, confirmed: no concurrency group, no timeout-minutes, and the checkout (38–39) without persist-credentials: false. missing FR-2, NFR-1 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by yq of .concurrency, each job’s timeout-minutes and each checkout’s persist-credentials at 1e91f65f (AC-3 below); the queued run is WI-6.1’s (AC-2)
DR-9 forge-et-be build.yaml:70–75 Row 9 (b), confirmed: amazon-ecs-deploy-task-definition@v2 without wait-for-service-stability. The first deploy run’s ecs::deploy step took one second (09:05:29–09:05:30 UTC, run 37908749494 — AC-9 below), so the run ends before ECS has started the task. Withdrawn 2026-10-20: the lead accepted a run that ends before the rollout in dev (F38-d spec Rev 3); forge-et-be#275 added the wait and forge-et-be#286 removes it. missing FR-2, NFR-1 F38-d-WI-2.1 withdrawn — F38-d spec Rev 3, 2026-10-20
DR-10 forge-et-be build.yaml:56–60 Row 10, confirmed: no trivy image step; only ECR’s scan on push runs. Withdrawn 2026-10-20: ECR’s scan on push is the dev image scan (AD #38 as restated, F38-d spec Rev 3); forge-et-be#275 added trivy image and forge-et-be#286 removes it. missing FR-2 F38-d-WI-2.1 withdrawn — F38-d spec Rev 3, 2026-10-20
DR-11 forge-et-be .aws/td.dev.json:43, 46, 47 Row 14, confirmed: ET_FILES_STORAGE_BUCKET, ET_MAILBOX_PUBLIC_BASE_URL and ET_MAILBOX_TOKEN_KEY_ID hold PLACEHOLDER-… values, and ET_MESSAGING_MAILBOX_INGEST_QUEUE is absent, so no mail is ingested. missing FR-6 F38-d-WI-4.1 fixed — forge-et-be#300, 2026-10-22 (WI-4.1 plan Rev 1); shown by the AC-7 task-definition read (F38-d WI-6.1)
DR-12 forge-et-fe build.yaml:19–71 Row 15, confirmed: one job (id dev, shown as build) that declares environment: dev and configures the key (46–51) after the install and build (37–44) but in the same job; no build artifact. missing FR-4 F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by run 39226907639’s jobs build and deploy and its artifact::upload and artifact::download steps (AC-12 below)
DR-13 forge-et-fe build.yaml:1–17 Row 15, confirmed: no permissions: block, so the job’s token has the repository’s default permissions; forge-et-be’s file sets contents: read. missing FR-4 F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by yq '.permissions, .jobs[].permissions' at 68461e50 and by the web-dist download succeeding under permissions: {} in run 39226907639 (AC-3 below)
DR-14 forge-et-fe build.yaml:3–6, 19–26 Row 15, confirmed: no paths-ignore, no timeout-minutes, and the checkout without persist-credentials: false. missing FR-4 F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by yq of paths-ignore, timeout-minutes and persist-credentials at 68461e50, and by #298’s own merge starting no run (AC-3 below)
DR-15 forge-et-fe build.yaml:40–44 Row 15 and architecture.md § Cross-cutting, confirmed: no guard step; the two VITE_* values are read in the build step’s own env, not declared once at job level. missing FR-4 F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by check::config success before pnpm::build in run 39226907639 (AC-12 below); the guard’s failure on a blank value is WI-6.1’s (AC-5)
DR-16 forge-et-fe build.yaml:8–12, 53–68 no finding — row 15 (e) and its concurrency note: the concurrency group deploy-dev never cancels a run, and the staged sync (hashed assets immutable, the other files and the HTML last with max-age=0, must-revalidate, then a --delete --size-only prune) is as architecture.md § Cross-cutting asks (forge-et-fe#259, BUG-136). no finding FR-4 — —
DR-17 forge-et-be build.yaml:1–16 The header comment describes jobs dev-build and release-build (“one trigger, two jobs … never both”) and the version build-<short-sha>; the file has a job dev and a commented-out release. inaccurate FR-2 (lines WI-2.1 rewrites) F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by the header comment at 1e91f65f (git show 1e91f65f:.github/workflows/build.yaml | sed -n '1,30p'), which describes jobs build and deploy and the tag build-<full commit SHA>
DR-18 forge-et-be build.yaml:24–25, 33 The v*.*.* tag trigger is still on while the only job for tags is commented out, so a release tag starts a build run whose dev job is skipped. FR-2’s triggers are a push to main and a manual dispatch. improvement FR-2 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by yq '.on | keys' at 1e91f65f → ["push","workflow_dispatch"] (AC-3 below)
DR-19 forge-et-be build.yaml:77–102 A commented-out release job, which coding-standards.md §7.5 says to delete (“Commented-out code is deleted, always”); it also names older majors than the live job (checkout@v4, configure-aws-credentials@v4 against @v7, @v6) and the short-SHA tag. The spec leaves the release job to F38-i. inconsistent outside FR-1 – FR-9 .forge/owed.md “Delete or replace the commented-out release job in forge-et-be’s build.yaml” open
DR-20 forge-et-be build.yaml:32–75 No aws sts get-caller-identity step, which AC-1’s printed identity needs. missing FR-1 F38-d-WI-2.1 fixed — forge-et-be#275, 2026-10-19; shown by the arn:aws:iam::570417736903:user/svc-github line in run 39228765669 (AC-1 below)
DR-21 forge-et-be build.yaml no finding — secrets and credentials: the static key comes from the dev Environment’s secrets (accepted for dev, F38 Rev 9; AC-1 below), no run: step prints a secret, and the workflow’s token is contents: read; pinned action versions: major tags, as architecture.md § Cross-cutting asks, kept current by Dependabot’s github-actions group (it changed this file in 4be56768); deploys from a branch other than the trunk: none — a push to main only (DR-7 covers the missing manual run). no finding FR-1, FR-2 — —
DR-22 forge-et-be .aws/td.dev.json no finding — health check: no container healthCheck, as the Dockerfile:82 comment explains (the load balancer checks /actuator/health/liveness); register §D: every “must set” key is present except the placeholders and the ingest queue (DR-11), ET_MESSAGING_SCAN_RESULT_QUEUE is rightly absent with ET_FILES_SCAN_PROVIDER=none, and none of the “must stay unset” keys is set; secrets: the database and Entra values are secrets references, and no secret value is in the file. no finding FR-6, NFR-2 — —
DR-23 forge-et-be .aws/td.dev.json:54–55 ET_SECURITY_ENTRA_TENANT_ID and _CLIENT_ID are secrets references although both are non-secret identifiers (§2.1; register §D names only the private key as coming from Secrets Manager). Each secret costs about USD 0.40 a month, and the two values cannot be read from the task definition. improvement outside FR-1 – FR-9 .forge/owed.md “Decide whether the Entra tenant and client IDs stay in Secrets Manager” open
DR-24 forge-et-fe build.yaml:19–71 No aws sts get-caller-identity step, which AC-1’s printed identity needs. missing FR-1 F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by the identity line in run 39226907639 (AC-1 below)
DR-25 forge-et-fe build.yaml:71 The file ends without a newline. mistake FR-4 (the file WI-2.2 rewrites) F38-d-WI-2.2 fixed — forge-et-fe#298, 2026-10-19; shown by git show 68461e50:.github/workflows/build.yaml | tail -c1 | od -c → \n (AC-3 below)
DR-26 forge-et-fe build.yaml no finding — secrets and credentials: the key comes from the dev Environment’s secrets and is configured only after the install and build; no step prints it; the two VITE_* values are public by design (AD #22); pinned action versions: major tags kept current by Dependabot (56e18b4d); deploys from a branch other than the trunk: a manual run can deploy any branch to dev, as AD #24 (refined 2026-10-18) and F38 Rev 18 allow; stable wait and trivy image: not applicable to a static bundle. no finding FR-1, FR-4 — —
DR-27 forge-et-be build.yaml:18, 38–75; forge-et-fe build.yaml:1, 20, 25–71; forge-et-iaac dev.yaml:36, 43, 50 The names differ from coding-standards.md §7.1 (sentence case for workflow, job and step names) and between the repositories: the workflow is build in both app repositories; steps are area::action with the order reversed between them (awscli::configure in forge-et-be and dev.yaml, configure::aws in forge-et-fe); forge-et-be’s job dev has no name and forge-et-fe’s job dev is named build; dev.yaml has terraform init beside terraform::validate and one unnamed step. The file names (build.yaml) stay, as the spec decides. inconsistent outside FR-1 – FR-9 .forge/owed.md “Agree how the deploy workflows name their steps” open
DR-28 forge-et-be build.yaml:48, 74–75; forge-et-fe build.yaml:14–15, 51, 57–71 The deploy targets are written into the workflows — the region in the step in forge-et-be but in a workflow-level AWS_REGION in forge-et-fe, and the cluster, service, bucket and distribution as literals. architecture.md § How dev is deployed today describes this accurately; F38-h’s deploy-uat needs the same steps with UAT’s names, which the GitHub Environment’s variables could supply. improvement outside FR-1 – FR-9 .forge/owed.md “Decide where the deploy workflows read their AWS target names from” open
DR-29 forge-et-iaac dev.yaml:22–65 No terraform plan posted on a PR, and the apply is not behind an Environment approval: the dev Environment has no protection rules (AC-6 below). missing FR-5 recorded only (FR-5) — .forge/owed.md “DevOps follow-ups…” item 8 open
DR-30 forge-et-iaac dev.yaml:7–10, 22–65 One job runs fmt, init, validate, plan and apply with the key, and there is no paths-ignore. Terraform’s plan itself needs the key (it reads the state and refreshes), so FR-20’s “credentials only in the deploy job” cannot apply as written; the nearest form is a plan job and an apply job joined by the saved plan file. missing FR-5 recorded only (FR-5) — .forge/owed.md “DevOps follow-ups…” item 9 open
DR-31 forge-et-iaac dev.yaml:32, 37, 43 actions/checkout@v4, configure-aws-credentials@v4 and hashicorp/setup-terraform@v3 are older majors than the app repositories’ (@v7, @v6) and than architecture.md § Cross-cutting’s list (@v6, @v6, @v4), and forge-et-iaac has no .github/dependabot.yml at 0ff63c5 to move them. #3 set checkout@v4 itself. inconsistent FR-5 recorded only (FR-5) — .forge/owed.md “DevOps follow-ups…” item 10 open (2026-10-29: fixed on forge-et-iaac#11, 0742cdd — checkout@v7, configure-aws-credentials@v6, setup-terraform@v4 in dev.yaml and uat.yaml, after every run began warning that the old versions target the deprecated Node.js 20; closes when #11 merges)
DR-32 forge-et-iaac dev.yaml:7–10, 26 A manual run on any branch runs terraform plan with svc-github’s administrator key on that branch’s code, because the dev Environment has no branch rule (AC-6 below). missing FR-5 recorded only (FR-5) — .forge/owed.md “DevOps follow-ups…” item 1 (already there) open
DR-33 forge-et-iaac dev.yaml:7–20, 27, 33–34 no finding — what triggers it, and from which branches: a push to dev or a manual run, applying only on dev; concurrency and timeout-minutes: one apply group that never cancels, plans grouped per branch, 60 minutes; persist-credentials: false; a contents: read token. no finding FR-5 — —
DR-34 this file, §4 “The CRM frontend’s dev deploy” paragraph It still describes one aws s3 sync dist/ s3://et-dev-spa --delete; since forge-et-fe#259 (14 Oct) the sync is staged with cache headers, followed by a pruning step, in a concurrency group. inaccurate outside FR-1 – FR-9 corrected here — a dated note on that paragraph corrected 2026-10-19 (F38-d WI-1.1)
DR-35 .forge/cloud-gated-validation.md §B FS-1, the deploy-workflows row’s 2026-10-14 update It says neither app workflow has a concurrency group and the frontend’s S3 sync is not staged; forge-et-fe’s has both since forge-et-fe#259. inaccurate outside FR-1 – FR-9 corrected here — a dated update on that row corrected 2026-10-19 (F38-d WI-1.1)

AC-1 — CI’s AWS access

Recorded 2026-10-19 by F38-d-WI-1.1. The settings half of AC-1 for forge-et-be and forge-et-fe, the two repositories with a dev Environment: where the key is held, the Environment’s rules, and the two IAM users as they stand.

2026-10-18 19:02 UTC — Claude (F38-d-WI-1.1), gh as tharinda-altrium

$ gh secret list -R altrium-io/<repo>                                        # repository secrets
$ gh secret list -R altrium-io/<repo> --env dev                              # the dev Environment's secrets
$ gh api repos/altrium-io/<repo>/actions/organization-secrets --jq '[.total_count, [.secrets[].name]]'
$ gh api repos/altrium-io/<repo>/environments/dev --jq '{protection_rules: [.protection_rules[]?.type], deployment_branch_policy}'
EXIT=0 for each command in each repository
$ gh variable list -R altrium-io/forge-et-fe --env dev
VITE_API_BASE_URL	https://api.dev.crm.empire.lk	2026-10-13T07:05:35Z
VITE_CSRF_COOKIE_NAME	XSRF-TOKEN-DEV	2026-10-13T07:05:47Z
EXIT=0
Repository Repository secrets Organisation secrets dev Environment secrets (updated at) Protection rules Deployment-branch policy
forge-et-be none (empty list) [0,[]] — none AWS_ACCESS_KEY_ID (2026-10-13T06:55:51Z), AWS_SECRET_ACCESS_KEY (2026-10-13T06:56:16Z) [] null
forge-et-fe none (empty list) [0,[]] — none AWS_ACCESS_KEY_ID (2026-10-13T06:56:09Z), AWS_SECRET_ACCESS_KEY (2026-10-13T06:56:25Z) [] null

Result: met — in both repositories the key is a secret of the dev Environment only, with no repository or organisation secret of that name, so a job that does not declare environment: dev does not receive it; the Environment has no protection rule and no branch rule (F38 Rev 18). forge-et-fe’s two build values are Environment variables, public by design (AD #22).

2026-10-18 19:02 UTC — Claude (F38-d-WI-1.1), profile et-dev (svc-harness)

$ aws iam list-attached-user-policies --user-name <user> --query 'AttachedPolicies[].PolicyArn'
$ aws iam list-user-policies --user-name <user> --query PolicyNames
$ aws iam list-groups-for-user --user-name <user> --query 'Groups[].GroupName'
EXIT=0 for each command for each user
$ aws iam list-access-keys --user-name svc-github --query 'AccessKeyMetadata[].[Status,CreateDate]' --output text
Active	2026-10-13T06:55:37+00:00
EXIT=0
$ aws iam get-login-profile --user-name svc-github
aws: [ERROR]: An error occurred (NoSuchEntity) when calling the GetLoginProfile operation: Login Profile for User svc-github cannot be found.
EXIT=254
User Attached policies Inline policies Groups
svc-github arn:aws:iam::aws:policy/AdministratorAccess none none
svc-harness AWSMarketplaceFullAccess, IAMFullAccess, ReadOnlyAccess, AmazonBedrockFullAccess (each arn:aws:iam::aws:policy/…) spike-invoke-gateway none

Result: met — svc-github has one active key, created 2026-10-13 06:55:37 UTC, which is 14 to 48 seconds before the four Environment secrets above were last updated, so both repositories hold that key; it has no console password (NoSuchEntity); its only policy is AdministratorAccess (accepted for dev, F38 Rev 9). svc-harness’s policies are as “Open with the DevOps team” item 7 records, with the inline spike-invoke-gateway still attached (.forge/owed.md). The key’s creation date is in the register’s §B FS-1 “CI’s AWS credentials” row, with no replacement due date in dev (F38 Rev 18).

The printed identity, the dev Environment screenshots and forge-et-pwa: WI-4.2, WI-6.1 and WI-7.1.

Recorded 2026-10-23 by F38-d-WI-4.2. The printed identity: the line the identity::aws step (aws sts get-caller-identity --query Arn --output text) writes in each deploy job, from one run per repository and trigger since wave 2’s PRs (forge-et-be#275, forge-et-fe#298), plus the run that registered the backend’s running task definition (AC-4).

2026-10-22 19:22 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium

$ gh run view --job <deploy job id> --repo altrium-io/<repo> --log > <saved file>      # once per row below
EXIT=0 for each of the five
$ grep -n 'arn:aws:iam::' <saved file>                                                  # a search of the saved log
Repository Run Trigger, branch, head Deploy job The line, as printed
forge-et-be 39226901426 workflow_dispatch, feature/F38-d-wave-2, 5ee0564a (#275’s head) 117975648742 2026-10-19T09:02:03.8069506Z arn:aws:iam::570417736903:user/svc-github
forge-et-be 39228765669 push, main, 1e91f65f (#275’s merge) 117981725186 2026-10-19T09:17:13.7916401Z arn:aws:iam::570417736903:user/svc-github
forge-et-be 39801944235 push, main, 06062de8 (registered backend:58) 119905786491 2026-10-22T15:10:16.1926651Z arn:aws:iam::570417736903:user/svc-github
forge-et-fe 39226907639 workflow_dispatch, feature/F38-d-wave-2, 70cfb875 (#298’s head) 117975511059 2026-10-19T09:01:39.0605739Z arn:aws:iam::570417736903:user/svc-github
forge-et-fe 39287294321 push, main, 8c0bc93a (forge-et-fe#301, the CRM’s first deploy after wave 2) 118178144595 2026-10-19T16:40:31.5293036Z arn:aws:iam::570417736903:user/svc-github

Result: met for forge-et-be and forge-et-fe — every deploy job printed the IAM user svc-github in the dev account 570417736903, from both triggers. The screenshots and forge-et-pwa: WI-7.1.

Recorded 2026-10-26 by F38-d-WI-6.1. The identity line of each run this work item started that has a deploy job reaching identity::aws (W-1, W-2, W-4, W-7 — AC-2 and AC-8 below). The forge-et-fe run W-9 stopped at the guard, so its deploy job was skipped and printed no identity (AC-5 below).

2026-10-26 14:50 UTC — Claude (F38-d-WI-6.1), gh as tharinda-altrium

$ gh run view --job <deploy job id> --repo altrium-io/forge-et-be --log | grep -m1 -oE '[0-9T:.-]+Z arn:aws:iam::[0-9]+:user/[a-z-]+'      # once per row below
EXIT=0 for each of the four
Run Trigger, branch, head Deploy job The line, as printed
40331078887 (W-1) workflow_dispatch, main, e88fef50 121600703266 2026-10-26T13:31:01.2516074Z arn:aws:iam::570417736903:user/svc-github
40331100359 (W-2) workflow_dispatch, main, e88fef50 121604681676 2026-10-26T13:38:31.1709412Z arn:aws:iam::570417736903:user/svc-github
40334043628 (W-4) workflow_dispatch, chore/F38-d-WI-6.1-failing-health-check, 864e2f34 121611100266 2026-10-26T13:50:49.3253658Z arn:aws:iam::570417736903:user/svc-github
40337010507 (W-7) workflow_dispatch, chore/F38-d-WI-6.1-wrong-host, 3c9e13f0 121621623828 2026-10-26T14:10:30.2941160Z arn:aws:iam::570417736903:user/svc-github

Result: met — the same user, svc-github, in a manual run on main and in manual runs on a branch other than main.

Recorded 2026-10-26 by F38-d-WI-7.1. forge-et-pwa’s settings half, the svc-github keys as they stand, the three dev Environment pages, and the capture app’s printed identity.

2026-10-26, before 15:42 UTC (the minute was not recorded) — Claude (F38-d-WI-7.1), gh as tharinda-altrium; profile et-dev (svc-harness)

$ gh secret list -R altrium-io/forge-et-pwa
EXIT=0      # empty: no repository secret
$ gh secret list -R altrium-io/forge-et-pwa --env dev
AWS_ACCESS_KEY_ID	2026-10-23T11:49:21Z
AWS_SECRET_ACCESS_KEY	2026-10-23T11:49:40Z
EXIT=0
$ gh api repos/altrium-io/forge-et-pwa/actions/organization-secrets --jq '[.total_count, [.secrets[].name]]'
[0,[]]
EXIT=0
$ gh api repos/altrium-io/<repo>/environments/dev --jq '{protection_rules: [.protection_rules[]?.type], deployment_branch_policy, updated_at}'      # once per repository
forge-et-be   {"deployment_branch_policy":null,"protection_rules":[],"updated_at":"2026-10-09T07:16:07Z"}
forge-et-fe   {"deployment_branch_policy":null,"protection_rules":[],"updated_at":"2026-10-13T06:55:21Z"}
forge-et-pwa  {"deployment_branch_policy":null,"protection_rules":[],"updated_at":"2026-10-22T16:15:20Z"}
EXIT=0 for each
$ gh api repos/altrium-io/forge-et-pwa/environments/dev/variables --jq '.variables[] | [.name,.value,.updated_at]'
["VITE_API_BASE_URL","https://api.dev.crm.empire.lk","2026-10-22T16:15:23Z"]
["VITE_CSRF_COOKIE_NAME","XSRF-TOKEN-DEV","2026-10-22T16:15:24Z"]
EXIT=0
$ aws iam list-access-keys --user-name svc-github --profile et-dev --region us-east-1 --query 'AccessKeyMetadata[].[Status,CreateDate]' --output text
Active	2026-10-23T11:49:08+00:00
Active	2026-10-13T06:55:37+00:00
EXIT=0
$ gh api repos/altrium-io/<repo>/environments/dev/secrets --jq '.secrets[] | [.name,.updated_at]'      # forge-et-be and forge-et-fe, re-read
forge-et-be   AWS_ACCESS_KEY_ID 2026-10-13T06:55:51Z, AWS_SECRET_ACCESS_KEY 2026-10-13T06:56:16Z
forge-et-fe   AWS_ACCESS_KEY_ID 2026-10-13T06:56:09Z, AWS_SECRET_ACCESS_KEY 2026-10-13T06:56:25Z
EXIT=0
Repository Repository secrets Organisation secrets dev Environment secrets (updated at) Protection rules Deployment-branch policy
forge-et-pwa none (empty list) [0,[]] — none AWS_ACCESS_KEY_ID (2026-10-23T11:49:21Z), AWS_SECRET_ACCESS_KEY (2026-10-23T11:49:40Z) [] null

Result: met for forge-et-pwa — the key is a secret of the dev Environment only, with no repository or organisation secret of that name, and the Environment has no protection rule and no branch rule (F38 Rev 18); its two build values are Environment variables, public by design (AD #22). svc-github now has two active keys. The second was created 2026-10-23 11:49:08 UTC, 13 to 32 s before forge-et-pwa’s two secrets were set, so forge-et-pwa holds the new key; forge-et-be and forge-et-fe still hold the key of 2026-10-13 (their secrets’ times are unchanged since WI-1.1’s reading). Both keys are recorded in the register’s §B FS-1 “CI’s AWS credentials” row, so a response to a leak looks at both.

2026-10-26, about 15:50 UTC — Claude (F38-d-WI-7.1), the browser pane on github.com, signed in by the lead (Claude entered no credentials)

Each repository’s https://github.com/altrium-io/<repo>/settings/environments/<id>/edit page for dev was opened in the browser pane, a screenshot was taken for the lead to see in the session (no image file is kept — WI-7.1 plan D-7, the lead’s Q-2 answer), and the page’s text was read (get_page_text):

Repository (Environment id) “Deployment branches and tags” “Environment secrets” “Environment variables” Protection rules
forge-et-be (23854236185) No restriction AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY (last updated “2 weeks ago”) “This environment has no variables.” no section shown on the page
forge-et-fe (24148686320) No restriction AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY (“2 weeks ago”) VITE_API_BASE_URL https://api.dev.crm.empire.lk, VITE_CSRF_COOKIE_NAME XSRF-TOKEN-DEV no section shown on the page
forge-et-pwa (25044556164) No restriction AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY (“3 days ago”) VITE_API_BASE_URL https://api.dev.crm.empire.lk, VITE_CSRF_COOKIE_NAME XSRF-TOKEN-DEV no section shown on the page

Result: met — each page shows what the gh api reads above show: no branch rule (“No restriction”), the key as the Environment’s two secrets, and no protection rule.

2026-10-26 15:40–16:05 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium

$ gh run view --job <deploy job id> --repo altrium-io/forge-et-pwa --log | grep -m1 -oE '[0-9T:.-]+Z arn:aws:iam::[0-9]+:user/[a-z-]+'      # once per row below
EXIT=0 for each of the three
Run Trigger, branch, head Deploy job The line, as printed
40224118047 workflow_dispatch, main, 777fdc4b (forge-et-pwa#91’s merge) 121261873702 2026-10-25T19:03:13.9305205Z arn:aws:iam::570417736903:user/svc-github
40228839882 push, main, c4b5db1e (forge-et-pwa#93) 121276076865 2026-10-25T20:02:26.5886626Z arn:aws:iam::570417736903:user/svc-github
40279386333 push, main, 0ecd1322 (forge-et-pwa#94) 121429170903 2026-10-26T06:22:25.5426148Z arn:aws:iam::570417736903:user/svc-github

Result: met — the capture app’s deploy job prints svc-github in the dev account, from both triggers. With WI-1.1’s, WI-4.2’s and WI-6.1’s entries above, AC-1 is met in every clause for all three app repositories.

AC-2 — the backend workflow on dev

Recorded 2026-10-23 by F38-d-WI-4.2. The automatic half of AC-2: a merge to main deploys dev. The runs from forge-et-be#275’s merge until forge-et-be#286 show #275’s workflow as it ran, with a wait for the rollout and a trivy image scan; F38-d spec Rev 3 removed both from the dev deploy (forge-et-be#286, merged 2026-10-20 19:26:18 UTC as f2a2f443), and AC-2 now asks for “no image scan and no wait”. The wait and the scan in the earlier runs are recorded as facts of those runs, not as AC-2 evidence. Runs are found with gh run list --repo altrium-io/forge-et-be --workflow build.yaml --branch main --json databaseId,headSha,event,createdAt,conclusion (the list and every run’s minutes are in AC-12 below); only the runs WI-4.2’s plan D-4 selects get an entry here.

The pre-merge manual run (cited, not the merge’s run).

2026-10-22 19:16 UTC — Claude (F38-d-WI-4.2), forge-et-be at 5ee0564a and 1e91f65f, gh as tharinda-altrium

Run: 39226901426, workflow_dispatch on feature/F38-d-wave-2, head 5ee0564a72783d8c7c4444dc364165dccd3f7251 (forge-et-be#275’s head), success, 7 billed minutes (build 52 s, deploy 331 s).

$ git -C forge-et-be rev-parse --short=10 <commit>:<path>          # 5ee0564a (#275's head), then 1e91f65f (its merge commit)
.github/workflows/build.yaml   72cdd527b9   72cdd527b9
.aws/td.dev.json               214f500dd1   214f500dd1
Dockerfile                     cfe63a76a4   cfe63a76a4
EXIT=0 for each

Result: cited — the run was dispatched before the merge, on the PR’s head, and ran the same three files that merged; it is not counted as the merge’s deploy.

The merge’s deploy run.

2026-10-22 19:22 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium

Run: 39228765669, push on main, head 1e91f65f1056dccedd8d200415598d73ccb80d37 (forge-et-be#275’s merge), success, 7 billed minutes (build 49 s, deploy 311 s).

$ gh run view 39228765669 --repo altrium-io/forge-et-be --json jobs --jq '.jobs[] | .name as $j | .steps[] | [$j,.number,.name,.conclusion,.startedAt,.completedAt] | @tsv'
build    checkout · gradle::validate · java::setup · gradle::setup · gradle::build · artifact::upload — each success
deploy   checkout · awscli::configure · identity::aws · ecr::login · ecr::find · artifact::download · docker::build · trivy::day · trivy::cache · trivy::setup · trivy::scan · docker::push · ecr::digest · ecs::render · ecs::deploy — each success
deploy   ecs::deploy   success   2026-10-19T09:18:41Z   2026-10-19T09:22:06Z
EXIT=0
$ gh run view --job 117981725186 --repo altrium-io/forge-et-be --log > <saved file>      # deploy; searched by content (every line's step column reads UNKNOWN STEP)
EXIT=0
2026-10-19T09:18:30.6433328Z Report Summary
2026-10-19T09:18:30.6449614Z │                                      Target                                      │  Type  │ Vulnerabilities │
2026-10-19T09:18:30.6451832Z │ 570417736903.dkr.ecr.us-east-1.amazonaws.com/crm/backend:build-1e91f65f1056dcce- │ ubuntu │        0        │
2026-10-19T09:18:30.6454956Z │ app/lib/HdrHistogram-2.2.2.jar                                                   │  jar   │        0        │
2026-10-19T09:18:30.6456873Z │ app/lib/HikariCP-7.0.2.jar                                                       │  jar   │        0        │
…  (one `ubuntu` row and 190 `jar` rows in all; rows whose count is not 0: 0)
2026-10-19T09:18:42.5631881Z Deployment started. Watch this deployment's progress in the Amazon ECS console: …
2026-10-19T09:22:06.0708590Z Deployment ecs-svc/8236213561887077312 verified: rolloutState=COMPLETED
$ gh run view --job 117981373566 --repo altrium-io/forge-et-be --log > <saved file>      # build
EXIT=0
2026-10-19T09:16:15.6218840Z Cache hit for restore-key: gradle-home-v2|Linux-X64|fast[a3ffae5017bf055c747c9d8463f8b19c]-5c45709105b8fd49b7eec7065fb4924fe30ce071
2026-10-19T09:16:17.8032853Z Restored cache entry with key gradle-home-v2|Linux-X64|build[debe7e2089933f89baa724f718db4c25]-1e91f65f1056dccedd8d200415598d73ccb80d37 in 2341ms
2026-10-19T09:16:38.1925812Z > Task :compileJava FROM-CACHE
2026-10-19T09:16:40.0965380Z 5 actionable tasks: 4 executed, 1 from cache

Result: met — the merge deployed dev: build ran with no AWS step, and every deploy step, from ecr::find to ecs::deploy, succeeded. As #275 ran it, the scan found no HIGH or CRITICAL vulnerability in any target (no finding was present) and the run waited 205 s for ECS to report rolloutState=COMPLETED; both steps were removed by forge-et-be#286. The build job restored the Gradle cache through a restore key from the Fast lane’s run on 5c457091 and took :compileJava from the build cache.

The first later run with no Gradle input change (the cache clause, for WI-6.1).

2026-10-22 19:23 UTC — Claude (F38-d-WI-4.2), forge-et-be at 1e91f65f and 0e0a766a, gh as tharinda-altrium

Run: 39235519408, push on main, head 0e0a766ac657b206bbaa70231a89b4b46fabf31a (forge-et-be#276, F11-h), success, 7 billed minutes (build 102 s, deploy 252 s).

$ git -C forge-et-be diff --name-only 1e91f65f 0e0a766a -- build.gradle gradle.lockfile settings.gradle gradle.properties gradle/
EXIT=0                                                     (no output: no Gradle input changed)
$ gh run view --job 118003518179 --repo altrium-io/forge-et-be --log > <saved file>      # build
EXIT=0
2026-10-19T10:10:37.5782632Z Cache hit for restore-key: gradle-home-v2|Linux-X64|build[debe7e2089933f89baa724f718db4c25]-1e91f65f1056dccedd8d200415598d73ccb80d37
2026-10-19T10:10:40.3499689Z Restored cache entry with key gradle-home-v2|Linux-X64|build[debe7e2089933f89baa724f718db4c25]-0e0a766ac657b206bbaa70231a89b4b46fabf31a in 2996ms
2026-10-19T10:11:33.0483685Z > Task :compileJava
2026-10-19T10:11:55.0643796Z 5 actionable tasks: 5 executed

Result: recorded for WI-6.1 — the run restored the cache the merge run saved (the Cache hit for restore-key line names 1e91f65f’s key; the Restored cache entry line beside it is the run’s own key) and compiled, because forge-et-be#276 changed Java sources. A compilation skipped from the cache is shown only by the merge run above.

The first run after forge-et-be#286, and the run that registered the running task definition.

2026-10-22 19:23 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium; profile et-dev (svc-harness)

Run: 39484112323, push on main, head a02ef4850fd34b39e7894aad3d760c79f3c50ebe (forge-et-be#285, BUG-199; git merge-base --is-ancestor f2a2f443 a02ef485 → EXIT=0), success, 3 billed minutes (build 115 s, deploy 43 s). Run: 39801944235, push on main, head 06062de8fed6cecb4c9776041b50c69457142993 (origin/main when read), success, 3 billed minutes (build 107 s, deploy 43 s).

$ gh run view <id> --repo altrium-io/forge-et-be --json jobs --jq '…steps…'              # both runs
deploy   checkout · awscli::configure · identity::aws · ecr::login · ecr::find · artifact::download · docker::build · docker::push · ecr::digest · ecs::render · ecs::deploy — each success, no trivy::* step
deploy   ecs::deploy   success   2026-10-20T19:33:41Z   2026-10-20T19:33:42Z               (39484112323)
deploy   ecs::deploy   success   2026-10-22T15:10:46Z   2026-10-22T15:10:47Z               (39801944235)
EXIT=0 for each
$ gh run view --job 118840919940 --repo altrium-io/forge-et-be --log > <saved file>      # 39484112323's deploy
EXIT=0
2026-10-20T19:33:42.6860755Z Deployment started. Watch this deployment's progress in the Amazon ECS console: …
2026-10-20T19:33:42.6993607Z Post job cleanup.
$ grep -ci trivy <saved file>
0
$ gh run view --job 119905786491 --repo altrium-io/forge-et-be --log > <saved file>      # 39801944235's deploy
EXIT=0
2026-10-22T15:10:47.6249496Z Deployment started. Watch this deployment's progress in the Amazon ECS console: …
2026-10-22T15:10:47.6605087Z Post job cleanup.
$ grep -ci trivy <saved file>
0
$ aws ecs list-service-deployments --cluster et-dev --service backend --created-at before=2026-10-20T19:40:00Z,after=2026-10-20T19:33:00Z
"startedAt": "2026-10-21T01:05:34.435000+05:30", "createdAt": "2026-10-21T01:05:34.360000+05:30", "finishedAt": "2026-10-21T01:09:04.133000+05:30", … "status": "SUCCESSFUL"
"startedAt": "2026-10-21T01:03:50.971000+05:30", "createdAt": "2026-10-21T01:03:50.920000+05:30", "finishedAt": "2026-10-21T01:05:34.603000+05:30", … "status": "STOPPED", "statusReason": "Replaced by a new service deployment."
EXIT=0
$ aws ecs list-service-deployments --cluster et-dev --service backend --created-at after=2026-10-19T09:00:00Z --max-results 100      # then jq '.serviceDeployments | group_by(.status) | map("\(.[0].status)=\(length)")'
STOPPED=1 SUCCESSFUL=22
EXIT=0
$ aws ecs describe-services --cluster et-dev --services backend --query 'services[0].{td:taskDefinition,deployments:deployments[].{id:id,status:status,rollout:rolloutState,td:taskDefinition,created:createdAt,updated:updatedAt,running:runningCount}}'
"td": "arn:aws:ecs:us-east-1:570417736903:task-definition/backend:58"
{"id": "ecs-svc/6601867314795104299", "status": "PRIMARY", "rollout": "COMPLETED", "td": "…/backend:58", "created": "2026-10-22T20:40:47.526000+05:30", "updated": "2026-10-22T20:43:59.289000+05:30", "running": 1}
EXIT=0

Result: met for “no image scan and no wait” — since forge-et-be#286 the deploy job has no trivy::* step and ecs::deploy ends a second after it starts, and ECS then completes the rollout on its own: run 39801944235’s deployment (created 15:10:47 UTC) was COMPLETED at 15:13:59 UTC, and every forge-et-be push run after #286 (AC-12 below) has the same step list and a one-second ecs::deploy. The first run after #286 left its deployment STOPPED because the next merge’s run deployed 1 min 44 s later, and that deployment succeeded; no deployment since wave 2 failed. #286’s own merge, which changed only .github/workflows/build.yaml, started no run (AC-3 below).

Still WI-6.1’s (AC-2): a manual run on main that deploys without a new commit; a second run started during one that waits; the tick of the cache clause (the restore is shown above); a run whose commit ECR already holds skipping the build; and a deploy whose task cannot start, recorded by ECS as failed while the previous task keeps serving. (Update 2026-10-26, F38-b WI-6.1: two of these were shown by one run. gh workflow run build.yaml --repo altrium-io/forge-et-be --ref main, started at the lead’s request to prove the deploy after forge-et-iaac#9, ran as 40274424861 (workflow_dispatch, 30fd601c, the commit backend:66 already ran). Its ecr::find step logged “ECR holds crm/backend:build-30fd601c…; the build and push are skipped”, docker::build and docker::push were skipped, and it deployed crm/backend@sha256:42544037…, the digest backend:66 ran. backend:67 rolled out COMPLETED, one task, the target healthy. The build job (the Gradle jar) still ran before the skip. WI-6.1 records and ticks AC-2.)

Recorded 2026-10-26 by F38-d-WI-6.1. The clauses only a run started on purpose can show. Before W-1, a teammate’s merge (forge-et-be#314, e88fef50) deployed through push run 40330092962 (success), so W-1 waited until backend:69 had rolled out (13:28:30 UTC) and ECR held build-e88fef50… (sha256:8cf884eb…). Every write below was asked of the lead with AskUserQuestion immediately before it ran (W-1 and W-2 in one call, two questions).

2026-10-26 13:29 UTC — Claude (F38-d-WI-6.1), gh as tharinda-altrium; approved by the lead at 13:29 UTC

Approval: AskUserQuestion “W-1: may I run gh workflow run build.yaml --repo altrium-io/forge-et-be --ref main?” → “Yes, run W-1”; “W-2: once W-1 reads in_progress, may I run the same command a second time …?” → “Yes, run W-2”.

$ aws ecr describe-images --repository-name crm/backend --image-ids imageTag=build-e88fef50d88abd1f84697075e9db37546d6d7aff --query 'imageDetails[0].imageDigest' --output text
sha256:8cf884eb7560af552bbd7ff7a0603984f802c5b2115a8dac4f262d1641cee640
EXIT=0
$ gh workflow run build.yaml --repo altrium-io/forge-et-be --ref main          # W-1, 13:29:51
https://github.com/altrium-io/forge-et-be/actions/runs/40331078887
EXIT=0
$ gh run view 40331078887 --json status      # 13:30:00
in_progress
$ gh workflow run build.yaml --repo altrium-io/forge-et-be --ref main          # W-2, 13:30:00
EXIT=0
$ gh run list --repo altrium-io/forge-et-be --workflow build.yaml --limit 3 --json databaseId,createdAt,event,status,conclusion      # 13:30:11
40331100359 2026-10-26T13:30:03Z workflow_dispatch queued
40331078887 2026-10-26T13:29:54Z workflow_dispatch in_progress
40330092962 2026-10-26T13:23:00Z push completed success
EXIT=0

Run: 40331078887 (W-1) — workflow_dispatch, main, e88fef50, success, 2 billed minutes. Run: 40331100359 (W-2) — workflow_dispatch, main, e88fef50, success, 2 billed minutes; queued while W-1 ran and still queued after W-1 ended at 13:31:12; its jobs were created at 13:37:19.

build   gradle::setup   2026-10-26T13:30:09.3503219Z Restored cache entry with key gradle-home-v2|Linux-X64|build[debe7e2089933f89baa724f718db4c25]-e88fef50d88abd1f84697075e9db37546d6d7aff in 2029ms
build   gradle::build   2026-10-26T13:30:22.7697307Z > Task :compileJava FROM-CACHE
build   gradle::build   2026-10-26T13:30:24.4772515Z 5 actionable tasks: 4 executed, 1 from cache
deploy  ecr::find       2026-10-26T13:31:02.9383381Z ECR holds crm/backend:build-e88fef50d88abd1f84697075e9db37546d6d7aff; the build and push are skipped
deploy  ecr::digest     2026-10-26T13:31:03.6807891Z Deploying crm/backend@sha256:8cf884eb7560af552bbd7ff7a0603984f802c5b2115a8dac4f262d1641cee640 (build-e88fef50d88abd1f84697075e9db37546d6d7aff)
deploy  ecs::deploy     2026-10-26T13:31:04.9006154Z Deployment started. …

Both runs’ deploy jobs ran ecr::find, ecr::digest, ecs::render and ecs::deploy, with artifact::download, docker::build and docker::push skipped, and no image-scan step. Then the ECS poll (aws ecs describe-services every 30 s and list-service-deployments, reads only):

W-1  deployment Z5Ue_vQYBfyu36WJQE-UM  started 13:31:11 · task 543f949e started 13:31:21, registered 13:31:52 · task 2c3e5c68 stopped 13:33:14, deregistered 13:33:24 · SUCCESSFUL 13:33:38 · "deployment completed" 13:34:17
W-2  deployment q-SCj-8vg6pRQFulQUZIw  started 13:38:36 · task 92072687 started 13:38:45, registered 13:39:06 · task 543f949e stopped 13:40:29, deregistered 13:40:39 · SUCCESSFUL 13:40:41 · "deployment completed" 13:41:41
13:42:09  [{"status":"PRIMARY","rollout":"COMPLETED","td":"…/backend:71","running":1}]
EXIT=0 for each read

Result: met — a manual run on main deployed with no new commit; a second run started while the first was in_progress waited (queued) and neither was cancelled; ECR held the commit’s image, so the build and push were skipped and that image’s digest deployed; and the build job restored setup-gradle’s cache and took :compileJava from the build cache. W-1 restored by its exact key (the push run had just built the same commit), so the restore of an earlier run’s cache for a new commit with no Gradle input change rests on run 39235519408’s Cache hit for restore-key: …build[…]-1e91f65f… line, recorded above by WI-4.2. The light-load check during these two rollouts is in AC-7 below.

2026-10-26 13:49 UTC — Claude (F38-d-WI-6.1), forge-et-be 864e2f34 and gh as tharinda-altrium; approved by the lead at 13:49 UTC

Approval: AskUserQuestion “Quiet window: have you told the team …?” → “Yes, team told”; “W-3: may I push the throwaway branch chore/F38-d-WI-6.1-failing-health-check …?” → “Yes, push it”; “W-4: after W-3, may I run gh workflow run build.yaml --repo altrium-io/forge-et-be --ref chore/F38-d-WI-6.1-failing-health-check? …” → “Yes, run W-4”.

The branch was cut from origin/main e88fef50 and adds one line to .aws/td.dev.json: "healthCheck": { "command": ["CMD-SHELL", "exit 1"], "interval": 5, "timeout": 2, "retries": 1, "startPeriod": 0 } (git diff --stat: 1 file, 1 insertion). Before W-4 the service ran backend:71, PRIMARY, COMPLETED, task 920726873c4c4bd183b9e66a7298fb49.

$ git push -u origin chore/F38-d-WI-6.1-failing-health-check          # W-3, 13:49:35
EXIT=0
$ gh workflow run build.yaml --repo altrium-io/forge-et-be --ref chore/F38-d-WI-6.1-failing-health-check      # W-4, 13:49:51
https://github.com/altrium-io/forge-et-be/actions/runs/40334043628
EXIT=0

Run: 40334043628 (W-4) — workflow_dispatch, chore/F38-d-WI-6.1-failing-health-check, 864e2f34, success (the run ends after the service update, F38-d spec Rev 3), 2 billed minutes. ECR had no image for the commit, so docker::build and docker::push ran: “Deploying crm/backend@sha256:97a429cc… (build-864e2f34b950f5c59857260ca055f4e706d1241b)”. The ECS poll, every 30 s from 13:51:38 (reads only):

13:51:38  PRIMARY/IN_PROGRESS/backend:72/running=0/failed=0   ACTIVE/COMPLETED/backend:71/running=1
13:55:22  PRIMARY/IN_PROGRESS/backend:72/running=0/failed=1   ACTIVE/COMPLETED/backend:71/running=1
14:00:41  PRIMARY/IN_PROGRESS/backend:72/running=0/failed=2   ACTIVE/COMPLETED/backend:71/running=1
14:06:01  PRIMARY/IN_PROGRESS/backend:71/running=1/failed=0   ACTIVE/FAILED/backend:72/running=0/failed=3
14:06:33  PRIMARY/IN_PROGRESS/backend:71/running=1/failed=0   DRAINING/FAILED/backend:72/running=0/failed=3
14:07:06  PRIMARY/COMPLETED/backend:71/running=1
$ aws ecs describe-services --cluster et-dev --services backend --query 'services[0].events[:14].[createdAt,message]' --output text      # times converted to UTC
14:00:28  (service backend) (task 68f23e1602a443269b833e075ccf02e4) failed container health checks.
14:05:38  (service backend) (task c52ceed7d705472691379465655371bb) failed container health checks.
14:05:56  (service backend) (deployment ecs-svc/4268833645685348749) deployment failed: tasks failed to start.
14:05:56  (service backend) rolling back to deployment ecs-svc/4217488972732024061.
14:06:52  (service backend) (deployment ecs-svc/4217488972732024061) deployment completed.
EXIT=0
$ aws ecs list-service-deployments --cluster et-dev --service backend --created-at after=2026-10-26T13:49:00Z
ROLLBACK_SUCCESSFUL  …/service-revision/et-dev/backend/4268833645685348749  13:51:31  14:06:12
EXIT=0
$ aws ecs list-tasks --cluster et-dev --service-name backend --desired-status RUNNING --query taskArns --output text      # 14:07:06
arn:aws:ecs:us-east-1:570417736903:task/et-dev/920726873c4c4bd183b9e66a7298fb49
EXIT=0

Liveness was called about every two seconds from the laptop from 13:49:50 to 14:07:17: 470 of 471 calls answered 200; the one other, at 14:06:51, got no answer within 5 s (000), after list-service-deployments’ rollback end (14:06:12) and while the failed task set was still draining. The load balancer recorded no HTTPCode_Target_5XX_Count or HTTPCode_ELB_5XX_Count datapoint from 13:49 to 14:10. The failure also reached the delivery team’s alert topic: a subscriber of et-dev-alerts received the email “A deployment of the backend service on dev failed. ECS’s deployment circuit breaker stopped it.” for deployment ecs-svc/4268833645685348749, reason “ECS deployment circuit breaker: tasks failed to start.”, time 2026-10-26T14:05:56Z, and showed it in the session (F38-e, register §B).

2026-10-26 14:08 UTC — Claude (F38-d-WI-6.1), forge-et-be; approved by the lead at 14:08 UTC

Approval: AskUserQuestion “W-5: may I delete the throwaway branch? …” → “Yes, delete it”.

$ git push origin --delete chore/F38-d-WI-6.1-failing-health-check
EXIT=0
$ git worktree remove <the branch's worktree>; git branch -D chore/F38-d-WI-6.1-failing-health-check
Deleted branch chore/F38-d-WI-6.1-failing-health-check (was 864e2f34).
EXIT=0
$ git ls-remote --heads origin chore/F38-d-WI-6.1-failing-health-check      # 14:08:57; prints nothing
EXIT=0

Result: met — a task definition that fails its health check, deployed from a branch by hand, was rolled back by the circuit breaker after three failed tasks; ECS recorded the deployment FAILED while the rollback ran, named the failure and the rollback in its events, and lists it ROLLBACK_SUCCESSFUL; the task that served before the deploy (920726873c…) served throughout; and the branch was deleted. Each failing task registered in the target group and took traffic for about three minutes before ECS stopped it, because the load balancer’s health check (/actuator/health/liveness) passed and only the container’s check failed; the application in it was healthy, and the load balancer recorded no 5xx. Undone by: the circuit breaker (the service), W-5 (the branch). AC-2 is met in every clause (the merge deploys and “no image scan and no wait”: WI-4.2’s entries above).

AC-3 — the FR-20 checklists

Recorded 2026-10-23 by F38-d-WI-4.2. The log checks of AC-3 for forge-et-be, read from the merge’s deploy run and from the two post-#286 runs above, and the read that shows a workflow-only push starts no deploy. The per-work-item checklists are in the PR bodies, forge-et-be#275 and forge-et-fe#298, and are not copied here; WI-7.1 ticks AC-3.

2026-10-22 19:23 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium; profile et-dev (svc-harness)

Run: 39228765669, push on main, head 1e91f65f1056dccedd8d200415598d73ccb80d37, success, 7 billed minutes.

$ grep -nE 'Run docker build |\[context jar\]|\[build |\[runtime 3/8\]|#16 DONE|writing image sha256|naming to' <deploy job 117981725186's saved log>
2026-10-19T09:17:18.3801735Z ##[group]Run docker build --platform linux/arm64 --pull --no-cache-filter runtime --provenance=false --build-context jar=build/image -t "$IMAGE" .
2026-10-19T09:17:20.0582295Z #5 [context jar] load .dockerignore
2026-10-19T09:17:25.9678101Z #11 [context jar] load from client
2026-10-19T09:17:34.6244651Z #16 [runtime 3/8] RUN apt-get update  && DEBIAN_FRONTEND=noninteractive apt-get upgrade -y       -o Dpkg::Options::=--force-confdef -o Dpkg::Options::=--force-confold  && rm -rf /var/lib/apt/lists/*
2026-10-19T09:17:38.8956416Z #16 DONE 7.9s
2026-10-19T09:17:40.5279635Z #22 writing image sha256:bf7bb351f6d9268805d391ebc7f9af2c19ce67285f9c2cfa1eb9f5457d13ecee done
2026-10-19T09:17:40.5280988Z #22 naming to 570417736903.dkr.ecr.us-east-1.amazonaws.com/crm/backend:build-1e91f65f1056dccedd8d200415598d73ccb80d37 done
$ grep -c '\[build ' <saved log>
0
$ grep -c 'CACHED' <saved log>
0
$ aws ecr describe-images --repository-name crm/backend --query 'sort_by(imageDetails,&imagePushedAt)[?imagePushedAt>=`2026-10-19`].[imageTags[0],imageDigest,imageManifestMediaType,imagePushedAt]' --output text
build-1e91f65f1056dccedd8d200415598d73ccb80d37  sha256:b6d1bee3c64ea471cf0f2fc5eec099531a3b716c198a6d52541785b4165824b8  application/vnd.docker.distribution.manifest.v2+json  2026-10-19T14:48:40.460000+05:30
…  (one `build-<full sha>` tag for each of the 22 `forge-et-be` runs in AC-12, the last `build-06062de8fed6cecb4c9776041b50c69457142993`  sha256:0e135dca217639240645c58c364a1a598db727375ca63b7e7c4173f6eb5677a8)
EXIT=0

The two post-#286 runs show the same: in 39484112323’s deploy log, #11 [context jar] load from client, #15 [runtime 3/8] RUN apt-get update … then #15 DONE 10.0s, and naming to …/crm/backend:build-a02ef4850fd34b39e7894aad3d760c79f3c50ebe; in 39801944235’s, #15 DONE 8.8s and naming to …/crm/backend:build-06062de8fed6cecb4c9776041b50c69457142993; in both, grep -c '\[build ' and grep -c 'CACHED' print 0.

Result: met for forge-et-be’s log checks — the jar came from the runner ([context jar] load from client, no [build stage), the runtime stage’s upgrade step ran in every run read (a DONE line, no CACHED line), the image is tagged build-<full commit SHA>, and ECR holds that tag with its digest.

2026-10-22 19:24 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium

$ gh pr view 298 --repo altrium-io/forge-et-fe --json state,mergedAt,mergeCommit,headRefOid,files
{"state":"MERGED","mergedAt":"2026-10-19T09:16:07Z","mergeCommit":"68461e50b59cf10ed5b748b22b3f1322251fb660", … "files":[".github/workflows/build.yaml"]}
EXIT=0
$ gh run list --repo altrium-io/forge-et-fe --workflow build.yaml --branch main --limit 100 --json databaseId,headSha,event,createdAt
…  (no run whose headSha is 68461e50…; the first run after it is 39287294321 on 8c0bc93a, created 2026-10-19T16:39:53Z)
EXIT=0
$ gh pr view 286 --repo altrium-io/forge-et-be --json state,mergedAt,mergeCommit,headRefOid,files
{"state":"MERGED","mergedAt":"2026-10-20T19:26:18Z","mergeCommit":"f2a2f4435b2f928ad9ef4d28e134a040cd303a25", … "files":[".github/workflows/build.yaml"]}
EXIT=0
$ gh run list --repo altrium-io/forge-et-be --workflow build.yaml --branch main --limit 100 --json databaseId,headSha,event,createdAt,conclusion,status
…  (no run whose headSha is f2a2f443…; the next run is 39484112323 on a02ef485, created 2026-10-20T19:30:56Z)
EXIT=0

Result: met for AC-3’s last clause — forge-et-fe#298’s merge, which changed only .github/workflows/build.yaml, started no deploy, and neither did forge-et-be#286’s in forge-et-be.

The FR-20 rules in the merged files, read with yq.

2026-10-22 19:25 UTC — Claude (F38-d-WI-4.2), forge-et-be at 1e91f65f and origin/main 06062de8, forge-et-fe at 68461e50

$ git -C <repo> show <commit>:.github/workflows/build.yaml | yq -o=json -I=0 '{"on": .on, "concurrency": .concurrency, "permissions": .permissions, "jobs": (.jobs | to_entries | map({"id": .key, "runs-on": .value."runs-on", "timeout": .value."timeout-minutes", "env": .value.environment, "perm": .value.permissions, "checkout_pc": ([.value.steps[] | select(.uses // "" | test("checkout")) | .with."persist-credentials"])}))}'
EXIT=0 for each
forge-et-be at 1e91f65f (origin/main 06062de8 where it differs) forge-et-fe at 68461e50 (also origin/main 8e59d076, blob 15ce07074b)
on push to main with paths-ignore: [".github/workflows/**"]; workflow_dispatch the same
concurrency deploy-dev, cancel-in-progress: false the same
workflow permissions {} {}
build job ubuntu-latest, timeout-minutes: 10, no Environment, contents: read, checkout persist-credentials: false ubuntu-latest, timeout-minutes: 10, environment: dev (its VITE_* variables only; no AWS step), contents: read, checkout persist-credentials: false
deploy job ubuntu-24.04-arm, timeout-minutes: 30 (10 since forge-et-be#286), environment: dev, contents: read, checkout persist-credentials: false ubuntu-latest, timeout-minutes: 10, environment: dev, permissions: {}, no checkout
last byte — \n (tail -c1 | od -c)

Result: recorded for WI-7.1’s AC-3 tick — each rule is present in both merged files. The frontends’ guard step and the staged sync with HTML last are shown by the CRM runs in AC-12 (check::config before pnpm::build; sync::assets, sync::static, sync::html, sync::prune, cache::invalidate, each success).

Still WI-7.1’s: ticking AC-3, and forge-et-pwa’s checklist.

Recorded 2026-10-26 by F38-d-WI-7.1. The three workflow PRs’ checklists, and forge-et-pwa’s workflow-only push. The checklists are in the PR bodies and are not copied here: forge-et-be#275 (“AC-3 checklist — FR-20 rules, forge-et-be”), forge-et-fe#298 (“AC-3 — the FR-20 rules, one by one (CRM)”) and forge-et-pwa#91 (“AC-3 — the FR-20 rules, one by one (capture app)”), each ticked rule by rule.

2026-10-26 15:40 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium

$ gh pr view 91 --repo altrium-io/forge-et-pwa --json files,mergeCommit --jq '[[.files[].path], .mergeCommit.oid]'
[[".github/workflows/build.yaml"], "777fdc4beb202c415b4da4df64d078d473d0a89e"]
EXIT=0
$ gh run list --repo altrium-io/forge-et-pwa --workflow build.yaml --commit 777fdc4beb202c415b4da4df64d078d473d0a89e --json databaseId,event
[[40224118047,"workflow_dispatch"]]
EXIT=0
$ gh run view 40228839882 --repo altrium-io/forge-et-pwa --json jobs      # forge-et-pwa#93's merge, the steps in order
build:  check::config, checkout, enable::corepack, setup::node, pnpm::dependencies, pnpm::build, artifact::upload — each success
deploy: artifact::download, configure::aws, identity::aws, sync::assets, sync::static, sync::manifest, sync::html, sync::prune, cache::invalidate — each success
EXIT=0

Result: met — forge-et-pwa#91’s merge (777fdc4b) changed only .github/workflows/build.yaml and started no push run; the only run on that commit is the lead-approved manual one. The capture app’s runs show the guard first and the staged sync with the HTML after the assets, the static files and the manifest, then the prune and the invalidation. With WI-4.2’s entries above, each workflow PR’s checklist covers every FR-20 rule, and AC-3 is met.

AC-4 — ARM64

Recorded 2026-10-23 by F38-d-WI-4.2.

2026-10-22 19:20–19:21 UTC — Claude (F38-d-WI-4.2), profile et-dev (svc-harness)

$ aws ecs list-tasks --cluster et-dev --service-name backend --query taskArns --output text
arn:aws:ecs:us-east-1:570417736903:task/et-dev/0d1262eae096409cac08e47873b2619b
EXIT=0
$ aws ecs describe-tasks --cluster et-dev --tasks <arn> --query 'tasks[].{arn:taskArn,td:taskDefinitionArn,status:lastStatus,started:startedAt,arch:attributes[?name==`ecs.cpu-architecture`].value,digest:containers[0].imageDigest,image:containers[0].image}'
"td": "arn:aws:ecs:us-east-1:570417736903:task-definition/backend:58", "status": "RUNNING", "started": "2026-10-22T20:41:29.165000+05:30",
"arch": ["arm64"], "digest": "sha256:0e135dca217639240645c58c364a1a598db727375ca63b7e7c4173f6eb5677a8"
EXIT=0
$ aws ecs describe-task-definition --task-definition backend:58 --query 'taskDefinition.{rev:revision,rp:runtimePlatform,image:containerDefinitions[0].image,registeredAt:registeredAt}'
"rev": 58, "rp": {"cpuArchitecture": "ARM64", "operatingSystemFamily": "LINUX"},
"image": "570417736903.dkr.ecr.us-east-1.amazonaws.com/crm/backend@sha256:0e135dca217639240645c58c364a1a598db727375ca63b7e7c4173f6eb5677a8", "registeredAt": "2026-10-22T20:40:47.028000+05:30"
EXIT=0
$ aws ecr batch-get-image --repository-name crm/backend --image-ids imageDigest=<digest> --query 'images[0].imageManifest' --output text      # then jq '{mediaType, config: .config.digest, layers: (.layers|length)}'
sha256:b6d1bee3… (build-1e91f65f…, #275's merge):  {"mediaType":"application/vnd.docker.distribution.manifest.v2+json","config":"sha256:bf7bb351f6d9268805d391ebc7f9af2c19ce67285f9c2cfa1eb9f5457d13ecee","layers":12}
sha256:0e135dca… (build-06062de8…, running):       {"mediaType":"application/vnd.docker.distribution.manifest.v2+json","config":"sha256:d971126f0a5c7bb9970067435d8dd0a70eb017603df0a824f5c9cc7054378036","layers":12}
EXIT=0 for each
$ curl -s "$(aws ecr get-download-url-for-layer --repository-name crm/backend --layer-digest <config digest> --query downloadUrl --output text)" | jq -c '{os,architecture}'      # the signed link is never printed
sha256:bf7bb351…:  {"os":"linux","architecture":"arm64"}
sha256:d971126f…:  {"os":"linux","architecture":"arm64"}
EXIT=0 for each
$ curl -s -w '\nHTTP=%{http_code}\n' https://api.dev.crm.empire.lk/actuator/health/liveness
{"components":{"livenessState":{"status":"UP"}},"status":"UP"}
HTTP=200
EXIT=0

The running revision backend:58 was registered at 15:10:47.028 UTC, inside run 39801944235’s ecs::render–ecs::deploy steps (15:10:46–15:10:47 UTC, AC-2 above); its deploy log’s writing image sha256:d971126f… and build-06062de8…: digest: sha256:0e135dca… lines match the config and the task’s image digest, and the merge run’s writing image sha256:bf7bb351… (AC-3 above) matches build-1e91f65f…'s config. The PR: forge-et-be#275, merged 2026-10-19 09:16:02 UTC as 1e91f65f, whose gh pr view 275 --json files lists .aws/td.dev.json, .github/workflows/build.yaml and Dockerfile — the ARM64 task definition and the linux/arm64 build in one PR.

Result: met — the running task reports ecs.cpu-architecture arm64 on revision backend:58, whose runtimePlatform.cpuArchitecture is ARM64; the merge’s image and the running image are single-platform linux/arm64; both landed in forge-et-be#275; and liveness answers UP.

AC-5 — the frontends’ deploys

Recorded 2026-10-26 by F38-d-WI-6.1. The CRM’s guard failing on a blank value. W-11 (a manual CRM deploy) was not run: the CRM’s deploy from a main push is already recorded, run 39287294321 (AC-1 and AC-12, WI-4.2); the latest such run is 40279583139 (push, 326c481a, 26 Oct 06:23 UTC, success — AC-9 below).

2026-10-26 14:31 UTC — Claude (F38-d-WI-6.1), gh as tharinda-altrium; approved by the lead at 14:31 UTC

Approval: AskUserQuestion “… have you told the team that for about 10 minutes a forge-et-fe merge’s dev deploy will fail at its config check?” → “Yes, team told”; “W-8: may I run gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe --body ' ' …?” → “Yes, set it”; “W-9: may I run gh workflow run build.yaml --repo altrium-io/forge-et-fe --ref main …?” → “Yes, run W-9”; “W-10: may I restore it straight after W-9 ends, whatever W-9’s result …?” → “Yes, restore it”.

$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe      # before W-8
XSRF-TOKEN-DEV
EXIT=0
$ gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe --body ' '      # W-8, 14:31:30
EXIT=0
$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe | od -c
0000000       \n
EXIT=0
$ gh workflow run build.yaml --repo altrium-io/forge-et-fe --ref main      # W-9, 14:31:43
https://github.com/altrium-io/forge-et-fe/actions/runs/40340419206
EXIT=0
$ gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe --body 'XSRF-TOKEN-DEV'      # W-10, 14:32:07
EXIT=0
$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-fe
XSRF-TOKEN-DEV
EXIT=0

Run: 40340419206 (W-9) — workflow_dispatch, main, 326c481a, failure, 1 billed minute. The build job ran 14:31:50–14:31:55: check::config failed; checkout, enable::corepack, setup::node, pnpm::dependencies, pnpm::build and artifact::upload were skipped; the deploy job was skipped.

2026-10-26T14:31:54.0206786Z   VITE_CSRF_COOKIE_NAME:      # the log shows a blank value here (trailing spaces trimmed in this record)
2026-10-26T14:31:54.0453312Z ##[error]VITE_CSRF_COOKIE_NAME is empty or blank in the dev GitHub Environment's variables
2026-10-26T14:31:54.0468080Z ##[error]Process completed with exit code 1.

Result: met for forge-et-fe — a variable set to a single space failed the guard before the install and build, nothing was deployed, and the variable was restored 37 s after it was blanked. Undone by: W-10. The capture app’s guard, its merge-started deploy and its cache headers, and the tick: WI-7.1.

Recorded 2026-10-26 by F38-d-WI-7.1. The capture app’s deploys, its served headers and its guard. W-1 (a manual capture deploy) was not run: two push runs on main had already deployed the capture app.

2026-10-26 15:38–15:40 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium

Run: 40228839882 — push on main, head c4b5db1e75b7e264ef7970c69cc5b116cfa07ec5 (forge-et-pwa#93, which changed CLAUDE.md, src/app/telemetry.ts, src/main.tsx and src/ui/SHARED-WITH-FE.sha256), success, 2 billed minutes (build 15 s, deploy 11 s). This is the capture app’s first merge-started deploy. The next, 40279386333 (push, 0ecd1322, forge-et-pwa#94, documentation only), also succeeded, as did the first deploy, the lead-approved manual run 40224118047 on 777fdc4b (Decomposition Plan ## Notes, “Wave 5 — merged, and the capture app’s first deploy”).

$ gh run view --job 121276076865 --repo altrium-io/forge-et-pwa --log > <saved file>; grep -nE 'create-invalidation|"Id": "I|E2RVC3893O10TJ' <saved file>
2026-10-25T20:02:31.3590149Z ##[group]Run aws cloudfront create-invalidation --distribution-id "EZ06T8FTKF7UZ" --paths "/*"
2026-10-25T20:02:32.3553105Z         "Id": "I5NVA3AHFSSNZURCAGUPJJ6LXM",
EXIT=0
$ grep -c E2RVC3893O10TJ <saved file>      # the CRM's distribution; no match expected
0

The latest run’s log (job 121429170903) names the same distribution, invalidation I6KVUOP9K6LW0LSJDLOSQ5C2HH, and no E2RVC3893O10TJ. The CRM’s merge-started deploy is WI-4.2’s run 39287294321 above, whose cache::invalidate names E2RVC3893O10TJ (AC-12).

2026-10-26 15:40 UTC — Claude (F38-d-WI-7.1), curl through https://capture.dev.crm.empire.lk/

$ curl -s -o /dev/null -w '%{http_code}\n' https://capture.dev.crm.empire.lk/
200
$ for p in sw.js manifest.webmanifest icons/icon-192.png icons/icon-512.png icons/icon-maskable-512.png apple-touch-icon.png favicon.ico favicon-32.png; do curl -sI "https://capture.dev.crm.empire.lk/$p" | grep -iE '^(HTTP|cache-control|content-type|age|x-cache)'; done
File Status content-type cache-control x-cache
sw.js 200 text/javascript max-age=0, must-revalidate RefreshHit
manifest.webmanifest 200 application/manifest+json max-age=0, must-revalidate RefreshHit
icons/icon-192.png 200 image/png max-age=0, must-revalidate Miss
icons/icon-512.png 200 image/png max-age=0, must-revalidate Miss
icons/icon-maskable-512.png 200 image/png max-age=0, must-revalidate Miss
apple-touch-icon.png 200 image/png max-age=0, must-revalidate Miss
favicon.ico 200 image/vnd.microsoft.icon max-age=0, must-revalidate Miss
favicon-32.png 200 image/png max-age=0, must-revalidate Miss

2026-10-26 15:42 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium; approved by the lead at 15:42 UTC

Approval: AskUserQuestion “Before W-2: have you told the team that for about 10 minutes a forge-et-pwa merge’s dev deploy will fail at its config check (check::config)?” → “Yes, team told”; “W-2: may I run gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa --body ' ' …?” → “Yes, set it”; “W-3: may I run gh workflow run build.yaml --repo altrium-io/forge-et-pwa --ref main? …” → “Yes, run W-3”; “W-4: may I restore it … straight after W-3 ends, whatever W-3’s result, and read it back?” → “Yes, restore it”.

$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa      # before W-2, 15:42:01
XSRF-TOKEN-DEV
EXIT=0
$ gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa --body ' '      # W-2, 15:42:02
EXIT=0
$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa | od -c
0000000       \n
EXIT=0
$ gh workflow run build.yaml --repo altrium-io/forge-et-pwa --ref main      # W-3, 15:42:05
https://github.com/altrium-io/forge-et-pwa/actions/runs/40351896602
EXIT=0
$ gh variable set VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa --body 'XSRF-TOKEN-DEV'      # W-4, 15:42:29, after the run ended at 15:42:17
EXIT=0
$ gh variable get VITE_CSRF_COOKIE_NAME --env dev --repo altrium-io/forge-et-pwa
XSRF-TOKEN-DEV
EXIT=0

Run: 40351896602 (W-3) — workflow_dispatch, main, 0ecd1322, failure, 1 billed minute. The build job ran 15:42:12–15:42:17: check::config failed; checkout, enable::corepack, setup::node, pnpm::dependencies, pnpm::build and artifact::upload were skipped; the deploy job was skipped.

2026-10-26T15:42:15.1111257Z   VITE_CSRF_COOKIE_NAME:      # the log shows a blank value here (trailing spaces trimmed in this record)
2026-10-26T15:42:15.1340170Z ##[error]VITE_CSRF_COOKIE_NAME is empty or blank in the dev GitHub Environment's variables
2026-10-26T15:42:15.1357320Z ##[error]Process completed with exit code 1.

No push run started on forge-et-pwa main between W-2 and W-4 (gh run list --workflow build.yaml --branch main --event push lists 40279386333 as the latest), so no teammate’s merge failed at the guard and W-6 was not needed.

Result: met — the capture app deployed from a merge to main and invalidated its own distribution EZ06T8FTKF7UZ, never the CRM’s; sw.js, the manifest (as application/manifest+json), the install icons, the touch icon and the favicons are served with max-age=0, must-revalidate; and a variable set to a single space failed the capture guard before the build, deploying nothing, and was restored 27 s after it was blanked. Undone by: W-4. With WI-4.2’s and WI-6.1’s entries for the CRM, AC-5 is met in every clause.

AC-6 — dev.yaml and the required checks

Recorded 2026-10-19 by F38-d-WI-1.1. The first reading: dev.yaml as forge-et-iaac#3 leaves it (#3 open, not merged), the app repositories’ required check names, and their workflows. Spec FR-5: F38-d records dev.yaml and changes nothing in it.

2026-10-18 18:59–19:02 UTC — Claude (F38-d-WI-1.1), forge-et-iaac at forge-et-iaac#3 head 0ff63c5, gh as tharinda-altrium

$ git -C forge-et-iaac fetch origin feature/F38-b-wave-1
EXIT=0
$ git -C forge-et-iaac show 0ff63c570567734c6830cd6795a85986ccf3fc31:.github/workflows/dev.yaml
EXIT=0
$ gh api repos/altrium-io/forge-et-iaac/environments/dev --jq '{protection_rules: [.protection_rules[]?.type], deployment_branch_policy}'
{"deployment_branch_policy":null,"protection_rules":[]}
EXIT=0
FR-20 rule dev.yaml at 0ff63c5 Result
Runs in the dev GitHub Environment environment: dev (line 26) met
Per-environment concurrency that never cancels a run in progress one group terraform-dev-apply for dev and one per other branch, cancel-in-progress: false (18–20) met
Build and deploy as separate jobs joined by an artifact, AWS credentials only in the deploy job one job, dispatch, holds the key for fmt, init, validate, plan and apply (24–65); the plan itself needs the key not met — DR-30, .forge/owed.md “DevOps follow-ups…” item 9
paths-ignore on workflow-only changes none (7–10) not met — DR-30, item 9
persist-credentials: false on the checkout (33–34) met
timeout-minutes on every job 60 (27) met
A plan posted on each PR none; a branch pushed for review runs nothing (3–6) not met — DR-29, item 8
The apply behind an Environment approval the apply runs when the ref is dev (62–65), and the dev Environment has no protection rules (above) not met — DR-29, item 8
The staged S3 sync and the VITE_* guard — n/a — no frontend

2026-10-18 19:02–19:03 UTC — Claude (F38-d-WI-1.1), gh as tharinda-altrium

$ gh api repos/altrium-io/<repo>/rulesets --jq '.[] | [.id,.name,.source_type] | @tsv'
EXIT=0 in each of forge-et-be, forge-et-fe, forge-et-pwa
$ gh api repos/altrium-io/<repo>/rulesets/<id> --jq '[.name, (.conditions.ref_name.include|join(",")), ([.rules[] | select(.type=="required_status_checks") | .parameters.required_status_checks[].context] | join(","))] | @tsv'
EXIT=0 for each repository ruleset; EXIT=1 for the organisation ruleset 16326145, which has no `conditions` (jq: join cannot be applied to null), re-read as below
$ gh api repos/altrium-io/<repo>/rulesets/16326145 --jq '[.name, .target, .enforcement, ([.rules[] | select(.type=="required_status_checks") | .parameters.required_status_checks[].context] | join(",")), ([.rules[].type]|join(","))] | @tsv'
REPO-CTL	repository	active		repository_create,repository_transfer,repository_delete
EXIT=0 in each repository
Repository Ruleset (id, source) Target Required check contexts Result
forge-et-be protect-main (21058125, Repository) refs/heads/main Fast lane, Slow lane, Forge review, Test record met
forge-et-be protect-releases (21056001, Repository) refs/heads/release/* Fast lane, Slow lane, Forge review, Test record met
forge-et-fe protect-main (21058126, Repository) refs/heads/main Fast lane, Forge review, Slow lane, Test record met
forge-et-fe protect-releases (21056003, Repository) refs/heads/release/* Fast lane, Forge review, Slow lane, Test record met
forge-et-pwa protect-main (21058127, Repository) refs/heads/main Fast lane, Forge review, Slow lane, Test record met
forge-et-pwa protect-releases (21056005, Repository) refs/heads/release/* Fast lane, Forge review, Slow lane, Test record met
all three REPO-CTL (16326145, Organization) repositories (creation, transfer, deletion) none — not a branch ruleset —

2026-10-18 19:03 UTC — Claude (F38-d-WI-1.1), gh as tharinda-altrium; each repository at origin/main

$ git -C <repo> ls-tree --name-only origin/main .github/workflows/
$ gh workflow list -R altrium-io/<repo> --all
EXIT=0 for each command in each repository
Repository Workflow files on origin/main Workflows (gh workflow list --all)
forge-et-be (5c457091) build.yaml, forge-review-dependabot.yml, forge-review-docs-only.yml, pr-validation.yml build, Forge review — Dependabot, Forge review — docs only, PR validation, Dependabot Updates
forge-et-fe (32c0dea5) build.yaml, forge-review-dependabot.yml, forge-review-docs-only.yml, pr-validation.yml build, Forge review — Dependabot, Forge review — docs only, PR validation, Dependabot Updates
forge-et-pwa (96d48ba) forge-review-dependabot.yml, forge-review-docs-only.yml, pr-validation.yml Forge review — Dependabot, Forge review — docs only, PR validation, Dependabot Updates

Result: the required check names are Fast lane, Slow lane, Test record and Forge review on main and release/* in all three app repositories, and no app repository has a deploy-prod workflow. dev.yaml meets the FR-20 rules on its environment, concurrency, checkout and timeout, and not the four in the table, each covered by an item of the .forge/owed.md row “DevOps follow-ups from F38-b’s review of forge-et-iaac”. Read again as merged, and ticked, by WI-7.1.

Recorded 2026-10-26 by F38-d-WI-7.1. dev.yaml as merged, and the rulesets and workflow files read again.

2026-10-26 15:40 UTC — Claude (F38-d-WI-7.1), forge-et-iaac origin/dev 5401a88, gh as tharinda-altrium; each app repository at origin/main

$ git -C forge-et-iaac fetch origin
EXIT=0
$ git -C forge-et-iaac diff --stat 0ff63c570567734c6830cd6795a85986ccf3fc31 origin/dev -- .github/workflows/dev.yaml
EXIT=0      # no output: the file is unchanged
$ git -C forge-et-iaac rev-parse 0ff63c570567734c6830cd6795a85986ccf3fc31:.github/workflows/dev.yaml origin/dev:.github/workflows/dev.yaml
7eac9d528e3a0c5f1bd5fc538393af3d57c9dc9d
7eac9d528e3a0c5f1bd5fc538393af3d57c9dc9d
EXIT=0
$ gh api repos/altrium-io/<repo>/rulesets/<id> --jq '[.name, (.conditions.ref_name.include|join(",")), ([.rules[] | select(.type=="required_status_checks") | .parameters.required_status_checks[].context] | join(","))] | @tsv'      # protect-main and protect-releases in each app repository
EXIT=0 for each of the six
$ git -C <repo> ls-tree --name-only origin/main .github/workflows/; git -C <repo> ls-tree --name-only origin/main .github/workflows/ | grep -c deploy-prod
0      # in each app repository (exit 1, as grep -c exits when nothing matches)
Repository (origin/main) protect-main protect-releases Workflow files
forge-et-be (05cadc2a) 21058125: Fast lane, Slow lane, Forge review, Test record 21056001: the same build.yaml, forge-review-dependabot.yml, forge-review-docs-only.yml, pr-validation.yml
forge-et-fe (b0bebb28) 21058126: Fast lane, Forge review, Slow lane, Test record 21056003: the same the same four
forge-et-pwa (0ecd132) 21058127: Fast lane, Forge review, Slow lane, Test record 21056005: the same the same four (build.yaml since forge-et-pwa#91)

Result: met — dev.yaml as merged into dev is the blob WI-1.1 reviewed at forge-et-iaac#3’s head (7eac9d52), so its table above stands unchanged; the required check names are unchanged in all three app repositories; and none has a deploy-prod workflow. AC-6 is met.

AC-7 — task-definition values and secrets

Recorded 2026-10-26 by F38-d-WI-6.1. Everything but the capture bundle’s search, which is WI-7.1’s.

The values, against the code read that day. The list was re-derived on 2026-10-26 from forge-et-be origin/main e88fef50: the required fields of EtProperties (et.environment, the datasource’s host, database, username and password, et.business-timezone, the login redirect URI, the seed admin email, the CSRF cookie domain and name), FilePipelineProperties (the storage bucket), and the keys register §D lists as required by a guard in a cloud environment (the four Entra values, the scheduler, mailbox-ingest, lead-triage and lead-research queues, the mailbox base URL and token key, the Exa key, ET_DATASOURCE_SSL_MODE, JSON logs). The scan-result queue is not needed, because scanning is switched off as F38-j allows.

2026-10-26 14:55 UTC — Claude (F38-d-WI-6.1), profile et-dev (svc-harness)

$ aws ecs describe-task-definition --task-definition backend:71 --query 'taskDefinition.containerDefinitions[0].[environment[].name, secrets[].name]' --output text      # sorted, then each required key looked for
ET_AI_FEATURES_ACCOUNTBRIEFING_MODEL ET_AI_FEATURES_EMAILDRAFT_MODEL ET_AI_FEATURES_LEADRESEARCH_MODEL ET_BUSINESS_TIMEZONE ET_DATASOURCE_DATABASE ET_DATASOURCE_HOST ET_DATASOURCE_PASSWORD ET_DATASOURCE_PORT ET_DATASOURCE_SSL_MODE ET_DATASOURCE_USERNAME ET_ENVIRONMENT ET_FILES_SCAN_OFF_ACKNOWLEDGED ET_FILES_SCAN_PROVIDER ET_FILES_STORAGE_BUCKET ET_MAILBOX_PUBLIC_BASE_URL ET_MAILBOX_TOKEN_KEY_ID ET_MESSAGING_LEADRESEARCHQUEUE ET_MESSAGING_LEADTRIAGEQUEUE ET_MESSAGING_MAILBOX_INGEST_QUEUE ET_MESSAGING_SCHEDULER_QUEUE ET_SEARCH_EXA_APIKEY ET_SECURITY_CORS_ALLOWED_ORIGINS ET_SECURITY_CSRF_COOKIE_DOMAIN ET_SECURITY_CSRF_COOKIE_NAME ET_SECURITY_ENTRA_CERTIFICATE ET_SECURITY_ENTRA_CLIENT_ID ET_SECURITY_ENTRA_PRIVATE_KEY ET_SECURITY_ENTRA_TENANT_ID ET_SECURITY_LOGIN_REDIRECT_URI ET_SECURITY_SEED_ADMIN_EMAIL JAVA_TOOL_OPTIONS LOGGING_STRUCTURED_FORMAT_CONSOLE
EXIT=0
required keys present: 25 of 25 · "must stay unset" keys present: 0 (no endpoint override, SQS endpoint, AWS credentials, demo queues, ingest DLQ, listener auto-startup or Spring profile)
ET_ENVIRONMENT=dev  ET_DATASOURCE_SSL_MODE=verify-full  ET_FILES_STORAGE_BUCKET=et-dev-attachments  ET_FILES_SCAN_PROVIDER=none  ET_FILES_SCAN_OFF_ACKNOWLEDGED=true  ET_SECURITY_CSRF_COOKIE_NAME=XSRF-TOKEN-DEV  LOGGING_STRUCTURED_FORMAT_CONSOLE=ecs
$ aws ecs describe-task-definition --task-definition backend:68 …      # the revision running at the readiness read; backend:71 the same
rev=68 arch=ARM64 stopTimeout=60
JAVA_TOOL_OPTIONS=-XX:MaxRAMPercentage=65 -XX:+ExitOnOutOfMemoryError
PLACEHOLDER- values: 0
EXIT=0
$ aws logs get-log-events --log-group-name /ecs/dev/backend --log-stream-name backend/backend/96917096d8b546f3b37f498c78eae9a2 --start-from-head --limit 300      # the first line of the task's stream
Picked up JAVA_TOOL_OPTIONS: -XX:MaxRAMPercentage=65 -XX:+ExitOnOutOfMemoryError
EXIT=0
$ aws elbv2 describe-target-group-attributes --target-group-arn <et-dev-backend> --query "Attributes[?Key=='deregistration_delay.timeout_seconds'].Value" --output text
30
EXIT=0

The upload (W-12). 2026-10-26 14:40 UTC — Claude (F38-d-WI-6.1), the browser pane as CRM01 and profile et-dev (svc-harness); approved by the lead at 14:40 UTC. Approval: AskUserQuestion “W-12 changed: the account page has no attachments panel, so the upload goes through Compose email → Attach files (POST /api/v1/email/attachments …). No email is sent … Go ahead?” → “Yes, upload it”. The account page (SMOKE Account, the lead’s choice) has no attachments panel; the only upload the CRM offers is Compose email → Attach files, whose picker accepts PDF, images and Office files. The lead chose f38d-wi61-upload-check.png (a 16×16 PNG, 79 bytes) in the file dialog; the compose window was then discarded (“Discard this email? Nothing is saved.”) and no email was sent.

> performance.getEntriesByType('resource').filter(e => /attachments/.test(e.name))      # the pane's request log did not list the XHR
{"url": "https://api.dev.crm.empire.lk/api/v1/email/attachments", "type": "xmlhttprequest", "status": 201, "start": "2026-10-26T14:42:07.097Z", "ms": 2183}
$ aws s3api list-objects-v2 --bucket et-dev-attachments --query 'sort_by(Contents,&LastModified)[-5:].{k:Key,s:Size,t:LastModified}'
attachments/email_upload/1/562c4811-cdbd-4130-802c-a9df817187dd  79  2026-10-26T14:42:09+00:00
EXIT=0

The object is the only one in the bucket, so this was the first upload on dev; it stays there as an unsent draft attachment (the lead kept it). Undone by: nothing — kept.

No 5xx across the rolling deploys. The load balancer app/et-dev-backend/82461f891f52a937, read for each window from a minute before the run to three minutes after the rollout or rollback ended:

$ aws cloudwatch get-metric-statistics --namespace AWS/ApplicationELB --metric-name <m> --dimensions Name=LoadBalancer,Value=app/et-dev-backend/82461f891f52a937 --start-time <t0> --end-time <t1> --period 60 --statistics Sum --query Datapoints      # m = HTTPCode_Target_5XX_Count, HTTPCode_ELB_5XX_Count
13:28–13:46 (W-1, W-2)  []  []
13:49–14:10 (W-4)       []  []
14:08–14:22 (W-7)       []  []
EXIT=0 for each
$ aws cloudwatch get-metric-statistics … --metric-name TargetConnectionErrorCount … 13:31–13:38      # no datapoint; RequestCount 21–29 a minute, the laptop's calls

An empty Datapoints list means no 5xx (the load balancer publishes no zero). From the laptop, liveness was called about every two seconds through each window: 1,044 of 1,048 calls answered 200, and four got no answer within 5 s (000) — 13:32:20 (W-1’s rollout), 13:36:15 (no deployment running), 14:06:51 (W-4’s failed task set draining) and 14:12:11 (W-7’s first failing task starting). None reached the load balancer as an error. .forge/owed.md “Look at the four liveness calls that got no answer within 5 s on 26 Oct” checks them from a second client.

The text search. One grep -E pattern set, from a file: -----BEGIN [A-Z ]*PRIVATE KEY-----, AKIA[0-9A-Z]{16}, ASIA[0-9A-Z]{16}, aws_secret_access_key, (password|passwd|secret)["']?\s*[:=]\s*["']?[^"'\s$]{6,}, matched without case.

2026-10-26 14:45–15:15 UTC — Claude (F38-d-WI-6.1), profile et-dev (svc-harness)

$ grep -cEi -f <patterns> <backend:71 task definition JSON>
7      # every match is "secret:" inside a secrets[] entry's Secrets Manager ARN
$ jq '.taskDefinition.containerDefinitions[0].secrets[]' …      # all seven: a valueFrom ARN (arn:aws:secretsmanager:…), no value
$ aws ecr batch-get-image --repository-name crm/backend --image-ids imageDigest=sha256:8cf884eb… --query 'images[0].imageManifest'      # the running image: 13 layers
$ curl -s "$(aws ecr get-download-url-for-layer --repository-name crm/backend --layer-digest <d> --query downloadUrl --output text)" -o <n>.tgz      # once per layer; the URL never printed
EXIT=0 for each of the 13
$ tar -tzf <n>.tgz | grep -E '\.(yml|yaml|properties|json|env|pem|key|txt)$' …; tar -xzOf <n>.tgz <file> | grep -aEci -f <patterns>      # each layer's text files
layer 3  opt/java/openjdk/conf/management/management.properties: 2
layer 9  app/rds-ca/global-bundle.pem: 1
every other text file: 0
$ tar -xzOf <n>.tgz | grep -aEc -e '-----BEGIN [A-Z ]*PRIVATE KEY-----'      # every file in each layer
layer 1  6   (usr/lib/aarch64-linux-gnu/libgnutls.so.30.37.1)
layer 2  1   (usr/lib/aarch64-linux-gnu/libssh.so.4.9.6)
every other layer: 0
$ unzip -p app/app.jar <each of its 62 .yml/.properties/.json/.sql/… files> | grep -aEci -f <patterns>      # the application jar, layer 13
0 for each; PRIVATE KEY marker anywhere in the jar: 0
$ aws s3 cp --recursive s3://et-dev-spa <folder>; grep -rEnoi -f <patterns> <folder>      # the CRM bundle, 313 files
assets/RichTextBody-BiuqxII-.js:91   password===t.password&&e.host===t.host …
assets/index-ciRwHs5A.js:9           password:!0,range:!0,search:!0 …
EXIT=0

Each match was opened and judged: the task definition’s are secret names inside ARNs, with no value; in libgnutls and libssh the marker is a string the library uses to parse PEM files, not a key; management.properties has the JDK’s commented placeholders <keystore-password> and <truststore-password>; the CA bundle’s match is base64 certificate text that reads AKia… under a case-insensitive search, in a public certificate; and the two bundle matches are library code (a URL comparison and a map of input types). Limit: the search cannot look for the real secret values, because reading them needs get-secret-value, which no check may call; it finds key material and value-shaped assignments only. The dependency jars in layer 10 are third-party libraries and were not opened. The downloaded layers, jar and bundle were deleted after this entry was written.

Result: met for everything but the capture bundle — the task definition carries every value the code requires on dev and none it must not; stopTimeout 60, JAVA_TOOL_OPTIONS with both flags, and the 30 s delay; the first multipart upload succeeded; no load-balancer 5xx across four rolling deploys; and no private key, password or secret value in the task definition, the running image or the CRM bundle. The capture bundle’s search and the tick: WI-7.1.

Recorded 2026-10-26 by F38-d-WI-7.1. The capture bundle’s text search, with WI-6.1’s pattern set above.

2026-10-26 15:41 UTC — Claude (F38-d-WI-7.1), profile et-dev (svc-harness)

$ aws s3 cp --recursive s3://et-dev-capture <folder> --only-show-errors; find <folder> -type f | wc -l
17
EXIT=0
$ grep -rEnoi -f <patterns> <folder>
assets/index-DXqM6BDq.js:8   password:!0,range:!0,search:!0,tel:!0,text:!0,time:!0,url:!0,week:!0};function
EXIT=0
$ grep -rlaE -e '-----BEGIN [A-Z ]*PRIVATE KEY-----' <folder>
EXIT=1      # no file

The 17 objects: index.html, sw.js, manifest.webmanifest, the three icons/*.png, apple-touch-icon.png, favicon.ico, favicon-32.png, and in assets/ three scripts with their source maps, one stylesheet and one font. The one match is library code, the same map of input types the CRM bundle has. Limit: as above, the search finds key material and value-shaped assignments only, not the real secret values. The download was deleted after the search.

Result: met — no private key, password or secret value in the capture bundle. With WI-6.1’s entries above, AC-7 is met in every clause.

AC-8 — the database certificate

Recorded 2026-10-26 by F38-d-WI-6.1.

2026-10-26 12:55 UTC — Claude (F38-d-WI-6.1), profile et-dev (svc-harness)

$ aws ecs describe-task-definition --task-definition backend:68 --query 'taskDefinition.containerDefinitions[0].environment[?name==`ET_DATASOURCE_SSL_MODE`]'
ET_DATASOURCE_SSL_MODE=verify-full
EXIT=0
$ aws logs get-log-events --log-group-name /ecs/dev/backend --log-stream-name backend/backend/96917096d8b546f3b37f498c78eae9a2 --start-from-head --limit 300      # the running task's start, 26 Oct 07:13 UTC; messages quoted
com.zaxxer.hikari.HikariDataSource et-pool - Starting...
com.zaxxer.hikari.pool.HikariPool et-pool - Added connection org.postgresql.jdbc.PgConnection@1bcdd302
com.zaxxer.hikari.HikariDataSource et-pool - Start completed.
org.flywaydb.core.FlywayExecutor Database: jdbc:postgresql://et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com:5432/et (PostgreSQL 18.3)
org.flywaydb.core.internal.command.DbValidate Successfully validated 52 migrations (execution time 00:00.614s)
com.empireteas.EtApplication Started EtApplication in 79.197 seconds (process running for 80.541)
EXIT=0
$ curl -s -w '\nHTTP=%{http_code}\n' https://api.dev.crm.empire.lk/actuator/health/liveness
{"components":{"livenessState":{"status":"UP"}},"status":"UP"}
HTTP=200
EXIT=0

The pool is named et-pool, so its start line reads et-pool - Start completed. rather than the plan’s HikariPool-1 - …. backend:68 was registered by push run 40284243505 on 25e61c91, which contains WI-4.1’s merge 8513d77d (git merge-base --is-ancestor → EXIT=0); every later revision carries the same value (backend:71, AC-7 below).

2026-10-26 14:09 UTC — Claude (F38-d-WI-6.1), forge-et-be 3c9e13f0 and gh as tharinda-altrium; approved by the lead at 14:08 UTC

Approval: AskUserQuestion “W-6: may I push a second throwaway branch chore/F38-d-WI-6.1-wrong-host … sets ET_DATASOURCE_HOST … to the RDS instance’s IP address (10.0.1.55) …?” → “Yes, push it”; “W-7: after W-6, may I run gh workflow run build.yaml --repo altrium-io/forge-et-be --ref chore/F38-d-WI-6.1-wrong-host … then delete that branch the same way as W-5?” → “Yes, run W-7”.

$ aws rds describe-db-instances --db-instance-identifier et-dev --query 'DBInstances[0].Endpoint.Address' --output text
et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com
EXIT=0
$ dig +short A et-dev.c8d6c2gekzm3.us-east-1.rds.amazonaws.com | tail -1      # a first attempt printed nothing
10.0.1.55
EXIT=0
$ git diff --stat origin/main chore/F38-d-WI-6.1-wrong-host
 .aws/td.dev.json | 2 +-
$ git push -u origin chore/F38-d-WI-6.1-wrong-host          # W-6, 14:09:15
EXIT=0
$ gh workflow run build.yaml --repo altrium-io/forge-et-be --ref chore/F38-d-WI-6.1-wrong-host      # W-7, 14:09:27
https://github.com/altrium-io/forge-et-be/actions/runs/40337010507
EXIT=0

Run: 40337010507 (W-7) — workflow_dispatch, chore/F38-d-WI-6.1-wrong-host, 3c9e13f0, success (the run ends after the service update), 2 billed minutes; “Deploying crm/backend@sha256:93e8dc3c… (build-3c9e13f02e5958ca28ac6ba60904969305e0dda9)”. The ECS poll and the reads after it (reads only; times UTC):

14:11:13  PRIMARY/IN_PROGRESS/backend:73/running=0/failed=0   ACTIVE/COMPLETED/backend:71/running=1
14:13:21  PRIMARY/IN_PROGRESS/backend:73/running=0/failed=1   ACTIVE/COMPLETED/backend:71/running=1
14:15:30  PRIMARY/IN_PROGRESS/backend:73/running=1/failed=2   ACTIVE/COMPLETED/backend:71/running=1
14:17:07  PRIMARY/IN_PROGRESS/backend:73/running=0/failed=3   ACTIVE/COMPLETED/backend:71/running=1
14:17:40  PRIMARY/IN_PROGRESS/backend:71/running=1/failed=0   ACTIVE/FAILED/backend:73/running=0/failed=3
14:18:44  PRIMARY/COMPLETED/backend:71/running=1
$ aws ecs describe-tasks …      # the stopped backend:73 tasks
2c7995b830df4b1aaa4f702126f7d2cf  Essential container in task exited  exitCode 1  14:13:08
178f236f8a174bfaa8428b0a2a65a160  Essential container in task exited  exitCode 1  14:14:49
3eda98198ca34b2e8c6161cefe16e66b  Essential container in task exited  exitCode 1  14:16:36
2099d37d7dee4ae98ff409c6fad4a961  Scaling activity initiated by (deployment ecs-svc/1719102153073437308)  14:18:22
$ aws ecs describe-services … events
14:17:24  (service backend) (deployment ecs-svc/1719102153073437308) deployment failed: tasks failed to start.
14:17:24  (service backend) rolling back to deployment ecs-svc/4217488972732024061.
14:18:25  (service backend) (deployment ecs-svc/4217488972732024061) deployment completed.
$ aws ecs list-service-deployments --cluster et-dev --service backend --created-at after=2026-10-26T14:09:00Z
ROLLBACK_SUCCESSFUL  …/service-revision/et-dev/backend/1719102153073437308  14:11:11  14:17:40
$ aws logs get-log-events --log-group-name /ecs/dev/backend --log-stream-name backend/backend/2c7995b830df4b1aaa4f702126f7d2cf --start-from-head --limit 300      # 15 events; messages quoted
Server name validation failed: certificate for host 10.0.1.55 dNSName entries subjectAltName, but none of them match. Assuming server name validation failed
… The hostname 10.0.1.55 could not be verified by hostnameverifier PgjdbcHostnameVerifier.
Exception encountered during context initialization - cancelling refresh attempt: org.springframework.beans.factory.BeanCreationException: Error creating bean with name 'entityManagerFactory' …
Application run failed
$ aws ecs list-tasks --cluster et-dev --service-name backend --desired-status RUNNING --query taskArns --output text      # 14:18:44
arn:aws:ecs:us-east-1:570417736903:task/et-dev/920726873c4c4bd183b9e66a7298fb49
EXIT=0 for each read
$ git push origin --delete chore/F38-d-WI-6.1-wrong-host; git worktree remove <its worktree>; git branch -D chore/F38-d-WI-6.1-wrong-host      # W-7's last part, 14:20
Deleted branch chore/F38-d-WI-6.1-wrong-host (was 3c9e13f0).
EXIT=0
$ git ls-remote --heads origin 'chore/F38-d-WI-6.1-*'      # 14:20:13; prints nothing
EXIT=0

Liveness from the laptop, 14:09:27–14:19:47: 254 of 255 calls answered 200; one, at 14:12:11, got no answer within 5 s (000). The load balancer recorded no 5xx datapoint from 14:08 to 14:22.

“The same image”. The workflow tags each image with its commit, so a branch run builds a new image (sha256:93e8dc3c…, not main’s 8cf884eb…). AC-8’s “same image” is met by the same application: the branch differs from main only in .aws/td.dev.json, which the image’s .dockerignore allowlist leaves out, so the jar was built from identical sources, with only the database host changed (the lead’s answer at the WI-6.1 plan review, Q-4).

Locally, the backend connects as before: DatasourceConfigSpec (the default prefer) and DatasourceSslModeIT (verify-full against a container with a certificate) are on forge-et-be main since forge-et-be#306, whose build record on d28b5ff passed with 10,179 tests and 0 failed (F38-d Decomposition Plan § Notes, “Wave 4 — what /forge-deliver F38-d did”).

Result: met — on dev the backend connects with verify-full and starts; the same application pointed at the database’s IP address, which the RDS certificate does not cover, is refused by pgjdbc’s host-name check and never starts, and the circuit breaker rolled back to backend:71 while the previous task kept serving; the branch was deleted. Undone by: the circuit breaker (the service), W-7’s last part (the branch).

AC-9 — first deploy, sign-in and versions

Recorded 2026-10-19 by F38-d-WI-1.1. The first-deploy check (spec FR-8) and dev’s first startup log. The deploy came before the check, so each condition is shown to have been on main in the commit dev first ran (spec W-4).

The first dev deploy.

2026-10-18 19:00 UTC — Claude (F38-d-WI-1.1), gh as tharinda-altrium; profile et-dev (svc-harness)

$ gh run list -R altrium-io/forge-et-be --workflow build.yaml --created 2026-10-05..2026-10-09 --json databaseId,headSha,event,conclusion,createdAt --limit 100
37745865691  010110de  push  cancelled  2026-10-08T07:49:51Z
…            (seven more runs on 8 and 9 Oct, from before `78377e6b` added the ECS steps)
37901728427  78377e6b  push  success    2026-10-09T07:54:31Z
37906247651  adb10b87  push  success    2026-10-09T08:38:54Z
37908749494  03404e8e  push  success    2026-10-09T09:02:49Z
37939488834  3dc5f4b9  push  success    2026-10-09T13:48:25Z
37989507547  86f5cee9  push  failure    2026-10-09T20:49:10Z
EXIT=0
$ gh run view <id> -R altrium-io/forge-et-be --json jobs --jq '.jobs[] | .steps[] | [.name,.conclusion] | @tsv'
37901728427: … docker::push success · ecs::render success · ecs::register success (07:56:47Z)
37906247651: … docker::push success · ecs::render success · ecs::register success (08:41:56Z)
37908749494: … docker::push success · ecs::render success · ecs::deploy success (09:05:29Z–09:05:30Z)
EXIT=0 for each
$ gh run view 37908749494 -R altrium-io/forge-et-be --json headSha,createdAt,conclusion,event,url,headBranch
{"conclusion":"success","createdAt":"2026-10-09T09:02:49Z","event":"push","headBranch":"main","headSha":"03404e8ed0ca76389b6d0f7ac3df3390f237ce77","url":"https://github.com/altrium-io/forge-et-be/actions/runs/37908749494"}
EXIT=0
$ aws ecs list-task-definitions --family-prefix backend --sort ASC --max-items 3
backend:1  backend:2  backend:3
EXIT=0
$ aws ecs describe-task-definition --task-definition backend:1 --query 'taskDefinition.registeredAt'
"2026-10-09T13:26:47.372000+05:30"
EXIT=0
$ aws ecs describe-task-definition --task-definition backend:3 --query 'taskDefinition.[revision,registeredAt]'
[3, "2026-10-09T14:35:29.804000+05:30"]
EXIT=0

Result: the first deploy is run 37908749494, a push of 03404e8ed0ca76389b6d0f7ac3df3390f237ce77 (“ci: deploy each dev revision to the backend service”) created 2026-10-09 09:02:49 UTC, whose ecs::deploy step succeeded at 09:05:30 UTC with backend:3 (registered 09:05:29 UTC). The two runs before it registered backend:1 (07:56:47 UTC, run 37901728427) and backend:2 (08:41:56 UTC, run 37906247651) without deploying them — registered, not deployed. The earliest log stream was created at 09:06:02 UTC, after that run started (the startup log below), so 03404e8e is the first commit dev ran.

The four conditions. Each PR is found from a file it added: the commit on origin/main that added the key file (git log --format=%H --diff-filter=A -1 origin/main -- <key file>), the PR that commit belongs to (gh api repos/altrium-io/<repo>/commits/<sha>/pulls), and that PR read with gh pr view <n> -R altrium-io/<repo> --json number,baseRefName,mergedAt,mergeCommit. A PR closed unmerged, or merged into a wave branch, is not evidence. For forge-et-be, git merge-base --is-ancestor <merge commit> 03404e8e exits 0 when the PR was on main in the first deployed commit; a forge-et-fe PR is shown by its merge time before the first deploy run’s start, 2026-10-09 09:02:49 UTC.

2026-10-18 19:00–19:02 UTC — Claude (F38-d-WI-1.1), gh as tharinda-altrium; forge-et-be and forge-et-fe at origin/main

$ git -C <repo> log --format=%H --diff-filter=A -1 origin/main -- <key file>      # once per row below
EXIT=0 for each of the 13
$ gh api repos/altrium-io/<repo>/commits/<sha>/pulls --jq '.[] | [.number,.base.ref,.merged_at] | @tsv'
EXIT=0 for each of the 13 (each row's PR, base `main`, except the two `forge-et-fe` rows noted below)
$ gh pr view <n> -R altrium-io/<repo> --json number,state,baseRefName,mergedAt,mergeCommit
EXIT=0 for each of the 13 (each MERGED into `main`)
$ git -C forge-et-be merge-base --is-ancestor <merge commit> 03404e8ed0ca76389b6d0f7ac3df3390f237ce77; echo "EXIT=$?"
EXIT=0 for each of the eight forge-et-be rows
Condition PR Key file (adding commit) Merged at (UTC) Merge commit On the first deployed commit Result
A-1 — F5-f, the real RecordVisibilityCheck forge-et-be#153 src/main/java/com/empireteas/attachment/service/RoutingRecordVisibilityCheck.java (3de34391) 2026-10-02 16:24:30 68a2a8d6 EXIT=0 met
A-1 — F5-f forge-et-be#160 src/main/java/com/empireteas/auth/access/RecordSharing.java (7660e897) 2026-10-03 18:31:42 e42d8507 EXIT=0 met
A-1 — F38-c, the GuardDuty scanner forge-et-be#82 src/main/java/com/empireteas/platform/scan/GuardDutyFileScanProvider.java (7aca53d2) 2026-09-27 18:22:36 c43e47be EXIT=0 met
A-1 — F38-c forge-et-be#94 src/main/java/com/empireteas/attachment/service/AttachmentMetrics.java (75e54348) 2026-09-28 06:21:02 ad45d038 EXIT=0 met
F5-j — the cross-site CSRF cookie forge-et-be#78 src/test/groovy/com/empireteas/platform/config/CsrfPropertiesBindingSpec.groovy (8bf86363) 2026-09-27 07:48:56 7d4eb42e EXIT=0 met
F5-j forge-et-fe#59 src/lib/config/cloudBuildEnv.ts (a13e1169) 2026-09-27 07:49:08 b251f051 merged before 2026-10-09 09:02:49 met
F5-k — the seeded IT Admin forge-et-be#79 src/main/resources/db/migration/V28__access_model_realignment.sql (bee5240d) 2026-09-27 19:12:50 462a8452 EXIT=0 met
F5-k forge-et-be#124 src/test/groovy/com/empireteas/auth/LastAdminRaceIT.groovy (34e823f6) 2026-09-30 14:39:31 02069dd2 EXIT=0 met
F5-k forge-et-fe#67 e2e/seam/access.seam.spec.ts (f350780f) 2026-09-28 05:09:35 40b8c8bf merged before 2026-10-09 09:02:49 met
F5-k forge-et-fe#95 src/app/RoutePermissionGuard.tsx (6c8b1f81) 2026-10-02 06:48:52 2a5daa96 merged before 2026-10-09 09:02:49 met
F5-k forge-et-fe#121 e2e/browser/access/route-guards.spec.ts (e6d0bfdd) 2026-10-03 10:16:56 426e5ebb merged before 2026-10-09 09:02:49 met
A-3 — F20-b wave 2 (closed 2026-09-25, register §A-3) forge-et-be#57 src/main/java/com/empireteas/mailbox/service/MailboxRenewalService.java (345c7fb4) 2026-09-25 04:27:58 41acdef2 EXIT=0 met
A-3 forge-et-fe#55 src/features/email/SyncIndicator.tsx (bdae7f15) 2026-09-25 04:28:17 17a2d4ce merged before 2026-10-09 09:02:49 met

For forge-et-fe#67 and forge-et-fe#121, the commits API names only the per-work-item PR that first carried the key file into its wave branch (forge-et-fe#64 into feature/F5-k-wave-1, forge-et-fe#120 into feature/F5-k-wave-3), which is not evidence on its own. The wave PRs into main are shown to have brought the file instead: the adding commit is an ancestor of the wave PR’s merge commit and not of that merge’s first parent (the main it was merged into).

$ git -C forge-et-fe merge-base --is-ancestor f350780f 40b8c8bf; echo "EXIT=$?"      # in #67's merge
EXIT=0
$ git -C forge-et-fe merge-base --is-ancestor f350780f 40b8c8bf^1; echo "EXIT=$?"    # not in the main it merged into
EXIT=1
$ git -C forge-et-fe merge-base --is-ancestor e6d0bfdd 426e5ebb; echo "EXIT=$?"      # in #121's merge
EXIT=0
$ git -C forge-et-fe merge-base --is-ancestor e6d0bfdd 426e5ebb^1; echo "EXIT=$?"
EXIT=1

Result: met — all four conditions held on main before the first deploy on 2026-10-09; none was waived.

Dev’s first startup log. Read with get calls only (no filter-log-events), from the stream with the lowest creation time.

2026-10-18 18:58–18:59 UTC — Claude (F38-d-WI-1.1), profile et-dev (svc-harness)

$ aws logs describe-log-groups --log-group-name-prefix /ecs/dev/backend --query 'logGroups[].[logGroupName,retentionInDays]'
[["/ecs/dev/backend", 30]]
EXIT=0
$ aws logs describe-log-streams --log-group-name /ecs/dev/backend --query 'sort_by(logStreams,&creationTime)[0:10].[logStreamName,creationTime,firstEventTimestamp,lastEventTimestamp]' --output json
["backend/backend/3499fbf2eb41491eb361db62b0302254", 1791536762592, 1791536767329, 1791554420403],
["backend/backend/b4fee809a557400aa16289f49ccdb4e1", 1791553882264, 1791553888651, 1791604288737],
…  (eight later streams)
EXIT=0
$ aws logs get-log-events --log-group-name /ecs/dev/backend --log-stream-name backend/backend/3499fbf2eb41491eb361db62b0302254 --start-from-head --limit 2000 --query '{next:nextForwardToken,events:events[].[timestamp,message]}' --output json
…  (92 events, saved whole outside the repository; the matched messages below)
EXIT=0

The earliest stream, backend/backend/3499fbf2eb41491eb361db62b0302254, was created 2026-10-09 09:06:02 UTC (first event 09:06:07, last 14:00:20 — the task the next deploy replaced). The first page held every line needed, so no further page or stream was read. The messages, quoted as returned (each the message field of the ECS-format JSON line, with its @timestamp and logger):

2026-10-09T09:06:10.113994322Z com.empireteas.EtApplication No active profile set, falling back to 1 default profile: "default"
2026-10-09T09:06:26.804682564Z org.flywaydb.core.internal.command.DbMigrate Current version of schema "public": << Empty Schema >>
2026-10-09T09:06:29.136267369Z org.flywaydb.core.internal.command.DbMigrate Migrating schema "public" to version "28 - access model realignment"
2026-10-09T09:06:29.581575118Z org.flywaydb.core.internal.command.DbMigrate Successfully applied 37 migrations to schema "public", now at version v37 (execution time 00:01.613s)
2026-10-09T09:06:55.713237559Z com.empireteas.EtApplication Started EtApplication in 47.392 seconds (process running for 48.383)

Result: met — dev’s first task ran with no profile set (Spring’s default, so neither local nor test), migrated an empty schema, and applied F5-k’s seed migration V28__access_model_realignment.sql (logged as version “28 - access model realignment”) among the 37 migrations, then started. Read on 2026-10-18 (UTC), inside the log group’s 30-day retention, which removes the 9 Oct events about 8 Nov.

Sign-in as CRM01 and the frontends’ versions: WI-6.1 and WI-7.1.

Recorded 2026-10-26 by F38-d-WI-6.1. The CRM01 sign-in and the CRM’s version.

2026-10-26 14:36 UTC — Claude (F38-d-WI-6.1), the browser pane on https://dev.crm.empire.lk; approved by the lead before 14:36 UTC (the minute was not recorded)

Approval: AskUserQuestion “W-14: may I open https://dev.crm.empire.lk in the browser pane for you to sign in as CRM01 (you enter the credentials), then read /api/v1/me and the page’s data-app-version from that session?” → “Yes, open it”. The page opened on /login (“Sign in with Microsoft”); the lead completed the Microsoft sign-in as CRM01 (Claude entered no credentials), and the pane landed on /accounts. Read from that page, the API call made with the session’s own cookie:

> const r = await fetch('https://api.dev.crm.empire.lk/api/v1/me', {credentials: 'include'}); const j = await r.json();
> ({status: r.status, keys: Object.keys(j), role: j.role, permissions: j.permissions, appVersion: document.documentElement.dataset.appVersion})
status 200 · keys user, role, roleName, permissions (the user object was not read) · role "IT Admin"
permissions: access.manage, account.all, account.manage, account.view, buying_standard.manage, buying_standard.view, email.manage, lead.all, lead.manage, lead.view, price.publish, quote.approve_sub_floor, settings.manage
appVersion "0.1.0"

2026-10-26 14:11 UTC — Claude (F38-d-WI-6.1), gh as tharinda-altrium, forge-et-fe 326c481a, profile et-dev (svc-harness)

$ gh run list --repo altrium-io/forge-et-fe --workflow build.yaml --json databaseId …; gh run view <id> --json jobs      # newest run whose deploy job succeeded
40279583139 push 326c481af2e999df80d2c7b3db30e020372f0163 2026-10-26T06:23:57Z deploy success, sync::html 2026-10-26T06:24:37Z-2026-10-26T06:24:39Z
EXIT=0
$ git -C forge-et-fe show 326c481af2e999df80d2c7b3db30e020372f0163:package.json | jq -r .version
0.1.0
EXIT=0
$ aws s3api head-object --bucket et-dev-spa --key index.html --query '{LastModified:LastModified,CacheControl:CacheControl}'
{"LastModified": "2026-10-26T06:24:39+00:00", "CacheControl": "max-age=0, must-revalidate"}
EXIT=0

Result: met for the CRM — CRM01 signed in through Microsoft on https://dev.crm.empire.lk and landed in the app; /api/v1/me returned 200 with access.manage among the IT Admin role’s permissions; the page’s data-app-version (0.1.0, set by the app at runtime, not present in the static HTML) equals package.json’s version at 326c481a, the commit of the latest CRM deploy run; and index.html’s LastModified (06:24:39) is inside that run’s sync::html step (06:24:37–06:24:39). The lead signs out at the end of the session. The capture app’s version and the tick: WI-7.1.

Recorded 2026-10-26 by F38-d-WI-7.1. The capture app’s version.

2026-10-26 15:40 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium, forge-et-pwa 0ecd1322, profile et-dev (svc-harness)

$ curl -s https://capture.dev.crm.empire.lk/ | grep -o '<meta name="app-version"[^>]*>'
<meta name="app-version" content="0.1.0">
$ git -C forge-et-pwa show 0ecd13224bef4d111b6a7b9b8ff787e214d17108:package.json | jq -r .version      # the head of the latest run whose deploy job succeeded, 40279386333
0.1.0
EXIT=0
$ aws s3api head-object --bucket et-dev-capture --key index.html --query '{LastModified:LastModified,CacheControl:CacheControl}'
{"LastModified": "2026-10-26T06:22:31+00:00", "CacheControl": "max-age=0, must-revalidate"}
EXIT=0
$ gh api repos/altrium-io/forge-et-pwa/actions/jobs/121429170903 --jq '.steps[] | select(.name=="sync::html") | [.started_at,.completed_at]'
["2026-10-26T06:22:29Z","2026-10-26T06:22:30Z"]
EXIT=0
$ grep -nE 'index\.html' <the job's saved log>
2026-10-26T06:22:30.2486137Z upload: dist/index.html to s3://et-dev-capture/index.html

Result: met — the served <meta name="app-version"> (0.1.0) equals package.json’s version at 0ecd1322, the latest successful deploy’s commit, and index.html was uploaded in that run’s sync::html step: the log’s upload line is 06:22:30.25 and S3 records LastModified 06:22:31, one second later on S3’s clock; no other step uploaded it. With WI-1.1’s and WI-6.1’s entries above, AC-9 is met in every clause.

AC-10 — the closing sweep

Recorded 2026-10-26 by F38-d-WI-7.1. What the feature’s acceptance criteria rest on, and what the sweep checked and changed.

The twelve criteria. AC-1 — met (WI-1.1, WI-4.2, WI-6.1 and WI-7.1 above). AC-2 — met (WI-4.2, WI-6.1; ticked at wave 6). AC-3 — met (WI-4.2, WI-7.1). AC-4 — met (WI-4.2; ticked at wave 4). AC-5 — met (WI-4.2, WI-6.1, WI-7.1). AC-6 — met (WI-1.1, WI-7.1). AC-7 — met (WI-6.1, WI-7.1). AC-8 — met (WI-6.1; ticked at wave 6). AC-9 — met (WI-1.1, WI-6.1, WI-7.1). AC-10 — met (this subsection). AC-11 — not met: ECR’s refusal of a second push under an existing tag is not shown, and the lead chose on 2026-10-26 to leave it open (.forge/owed.md, “F38-d AC-11: show ECR refusing a second push under an existing tag”). AC-12 — met (WI-4.2, WI-6.1, WI-7.1 below).

The review’s findings (DR-1 – DR-35). Each row’s Status is fixed — <PR>, <date>, withdrawn — F38-d spec Rev 3, no finding, corrected 2026-10-19, or open with a home. The open rows and their homes, each found by grep -c -F "**<title>" .forge/owed.md → 1: DR-19 “Delete or replace the commented-out release job in forge-et-be’s build.yaml”; DR-23 “Decide whether the Entra tenant and client IDs stay in Secrets Manager”; DR-27 “Agree how the deploy workflows name their steps”; DR-28 “Decide where the deploy workflows read their AWS target names from”; DR-29 – DR-32 “DevOps follow-ups from F38-b’s review of forge-et-iaac” (items 8, 9, 10 and 1). Met.

“Checked against F38” rows 5, 6, 8, 9, 10, 14 and 15, against the trunks. Read 2026-10-26: git -C <repo> log --since=2026-10-22 origin/main -- .github/workflows .aws lists, in forge-et-be, WI-4.1’s and F11-c’s .aws/td.dev.json commits (already in rows 6 and 14) and three changes to pr-validation.yml only (2cfb9446, 85070261, dd7db6ed, F38-f’s comment corrections); in forge-et-fe, nothing; in forge-et-pwa, f7d6b3f (wave 5’s build.yaml). forge-et-iaac’s dev.yaml is unchanged (AC-6). Rows 6, 8 and 14 were closed against the trunks by WI-6.1 and stand. Rows 5, 9, 10 and 15 describe the files as they are, and each got a dated note in that table, rows 9 and 15 marked closed. Met.

architecture.md § CI/CD Pipelines › How dev is deployed today. Compared with the four deploy files on the trunks: the backend, CRM, infrastructure and capture-app bullets match. Changed: the heading’s read date (26 Oct), and the Environments bullet now names forge-et-pwa among the repositories whose dev Environment has no branch rule (Rev 77). Met.

§4’s other statements about the capture app. Found with grep -n -E 'capture|forge-et-pwa' outside this section. Given a dated update: the capture-app bullet of “What it holds” (the empty bucket and the 403), forge-et-iaac#5’s PR bullet (“the host answers with an error”), and “The CRM frontend’s dev deploy” paragraph (“forge-et-pwa has no deploy yet”). Already current: “Not there yet” (26 Oct update), “Open with the DevOps team” items 5, 6 and 9, the 16 Oct and 22 Oct updates (each a dated record of that day), the summary of the dev estate at the end of §4, and the URL tables. Met.

The register’s §B F38-d block (.forge/cloud-gated-validation.md): “The image runs on Fargate ARM64” — validated; “Task-definition values for the container” — validated; “trivy image in deploy-dev” — withdrawn (spec Rev 3); “The jar is built on the runner” — validated; “One image per commit” — open, the refused re-push (AC-11, the lead’s choice) and the digest UAT promotes (F38-h), owner F38-d / forge-et-iaac; “The backend verifies the RDS certificate” — validated; “A deploy that ECS rolls back” — validated; “The CRM deploy’s new build job on dev” — validated; “The capture app’s deploy on dev” — validated by this work item. Met — every row is validated, withdrawn, or open with its reason and owner.

.forge/owed.md. “Settle with Shyamin Ayesh the other differences between the dev estate and F38” removed: rows 5, 6, 8, 9 and 15 are built and row 12 is accepted (its only pointer, the F38-d spec’s ## Input Sources, is a point-in-time list and stays). “Review the GitHub Actions workflows the DevOps team wrote” kept with a dated note: every finding is fixed, withdrawn or homed, and F38-h and F38-i still point at it to carry the findings into deploy-uat and deploy-prod. “Say in git-conventions.md that forge-et-pwa deploys to dev” was already removed, and § Branching already says so (git grep -n "has none yet" origin/main -- .claude .forge/design prints nothing).

AC-11 — one image per commit

Recorded 2026-10-26 by F38-d-WI-6.1.

2026-10-26 12:55–13:00 UTC — Claude (F38-d-WI-6.1), profile et-dev (svc-harness), gh as tharinda-altrium

$ aws ecr describe-repositories --repository-names crm/backend --query 'repositories[0].imageTagMutability' --output text
IMMUTABLE
EXIT=0
$ gh pr view 275 --repo altrium-io/forge-et-be --json mergedAt --jq .mergedAt      # FR-2's skip rule (WI-2.1)
2026-10-19T09:16:02Z
EXIT=0
$ gh pr list --repo altrium-io/forge-et-iaac --head feature/F38-d-wave-3 --state all --json number,state,mergedAt      # the immutability PR (WI-3.1)
[{"mergedAt":"2026-10-23T10:41:13Z","number":6,"state":"MERGED"}]
EXIT=0
$ aws ecr get-lifecycle-policy --repository-name crm/backend --query lifecyclePolicyText --output text | jq -c '.rules[] | {p:.rulePriority,d:.description,s:.selection}'
{"p":1,"d":"Expire untagged images after 7 days","s":{"tagStatus":"untagged","countType":"sinceImagePushed","countUnit":"days","countNumber":7}}
{"p":2,"d":"Keep last 10 release images","s":{"tagStatus":"tagged","tagPrefixList":["v"],"countType":"imageCountMoreThan","countNumber":10}}
{"p":3,"d":"Keep last 100 build images","s":{"tagStatus":"tagged","tagPrefixList":["build-"],"countType":"imageCountMoreThan","countNumber":100}}
EXIT=0
$ aws ecr describe-images --repository-name crm/backend --image-ids imageDigest=sha256:5a27c7e6… --query 'imageDetails[0].{tags:imageTags,pushed:imagePushedAt}'      # the image backend:68 ran
{"tags": ["build-25e61c91858a2135256ef1bec731a09c159635fc"], "pushed": "2026-10-26T07:13:02.758000+00:00"}
EXIT=0

The image the service runs at the end of this work item, backend:71’s sha256:8cf884eb…, is tagged build-e88fef50d88abd1f84697075e9db37546d6d7aff (AC-2 above).

2026-10-26 14:43 UTC — Claude (F38-d-WI-6.1), profile et-dev (svc-harness); approved by the lead at 14:43 UTC

Approval: AskUserQuestion “W-13: may I ask ECR to point the existing tag build-236e990 (digest ca3358c6…, a 9 Oct image nothing runs) at a different image’s manifest (build-78377e6) with aws ecr put-image? …” → “Yes, run W-13”. Before it, tag A’s digest and manifest were saved (describe-images, batch-get-image; media type application/vnd.docker.distribution.manifest.v2+json for both images).

$ M=$(aws ecr describe-repositories --repository-names crm/backend --query 'repositories[0].imageTagMutability' --output text); echo "MUTABILITY=$M"; if [ "$M" = IMMUTABLE ]; then aws ecr put-image --repository-name crm/backend --image-tag build-236e990 --image-manifest "file://<saved manifest of build-78377e6>" --image-manifest-media-type application/vnd.docker.distribution.manifest.v2+json; echo "EXIT=$?"; fi
MUTABILITY=IMMUTABLE
aws: [ERROR]: An error occurred (AccessDeniedException) when calling the PutImage operation: User: arn:aws:iam::570417736903:user/svc-harness is not authorized to perform: ecr:PutImage on resource: arn:aws:ecr:us-east-1:570417736903:repository/crm/backend because no identity-based policy allows the ecr:PutImage action
EXIT=254
$ aws ecr describe-images --repository-name crm/backend --image-ids imageTag=build-236e990 --query 'imageDetails[0].imageDigest' --output text
sha256:ca3358c67352789f842503436c4f1457487e68f03ba7504fcf35b75dc307d92e
EXIT=0

Result: not met — the push was refused by IAM before ECR’s tag check ran: none of svc-harness’s policies allows ecr:PutImage (its policies: AC-1 above), so the refusal does not show immutability. Nothing changed (tag A’s digest is the one saved before). Shown: the repository is IMMUTABLE; the immutability PR (forge-et-iaac#6, 2026-10-23T10:41:13Z) merged after the skip rule reached main (forge-et-be#275, 2026-10-19T09:16:02Z); rule 3’s tagPrefixList is ["build-"]; and the deployed images are tagged build-<full sha>. The refused re-push needs an identity that may push, and is .forge/owed.md’s “F38-d AC-11: show ECR refusing a second push under an existing tag”. AC-11 stays unticked.

AC-12 — billed minutes and PR records

Recorded 2026-10-23 by F38-d-WI-4.2. The billed minutes of every build run in forge-et-be and forge-et-fe from wave 2’s pre-merge manual runs to the time of reading, beside the estimate of the plan that changed the workflow.

How the minutes are counted. GitHub bills each job rounded up to a whole minute. A run’s billed minutes here are each job’s completedAt − startedAt, rounded up and summed: gh run view <id> --repo altrium-io/<repo> --json jobs --jq '[.jobs[] | {name, s: ((.completedAt|fromdateiso8601) - (.startedAt|fromdateiso8601))} | .m = (((.s + 59) / 60) | floor)] | {jobs: ., billed: (map(.m) | add)}'. The run’s timing endpoint cannot be used, because it reports 0 for these private-repository runs:

2026-10-22 19:25 UTC — Claude (F38-d-WI-4.2), gh as tharinda-altrium

$ gh api repos/altrium-io/forge-et-be/actions/runs/39228765669/timing
{"billable":{"UBUNTU":{"total_ms":0,"jobs":2,"job_runs":[{"job_id":117981373566,"duration_ms":0},{"job_id":117981725186,"duration_ms":0}]}},"run_duration_ms":368000}
EXIT=0
$ gh run list --repo altrium-io/<repo> --workflow build.yaml --branch main --limit 100 --json databaseId,headSha,event,createdAt,conclusion,status
EXIT=0 for each repository
$ gh run view <id> --repo altrium-io/<repo> --json status,conclusion,headSha,headBranch,event,createdAt,url,jobs      # once per row below
EXIT=0 for each of the 44

Result: recorded. The arm64 deploy job is listed under UBUNTU, which is all the API shows; the rate at which GitHub bills an arm64 minute is on the organisation’s billing page, which this record does not read (the lead, 2026-10-19).

Estimates: forge-et-be before forge-et-be#286, WI-2.1 D-12 (about 15–16 a run until WI-3.1’s 30 s deregistration delay is applied, about 11 after; WI-3.1 had not merged when these runs ran); forge-et-be after #286, forge-et-be#286’s PR body (2–3); forge-et-fe, WI-2.2 D-8 (2). Each job’s seconds are followed by its billed minutes in brackets.

Run Repository Trigger Head Conclusion Jobs Billed minutes Estimate
39226901426 forge-et-be workflow_dispatch 5ee0564a success build 52 s (1), deploy 331 s (6) 7 15–16 (WI-2.1 D-12)
39226907639 forge-et-fe workflow_dispatch 70cfb875 success build 28 s (1), deploy 17 s (1) 2 2 (WI-2.2 D-8)
39228765669 forge-et-be push 1e91f65f success build 49 s (1), deploy 311 s (6) 7 15–16 (WI-2.1 D-12)
39235519408 forge-et-be push 0e0a766a success build 102 s (2), deploy 252 s (5) 7 15–16 (WI-2.1 D-12)
39287294321 forge-et-fe push 8c0bc93a success build 27 s (1), deploy 16 s (1) 2 2 (WI-2.2 D-8)
39366500940 forge-et-fe push 83b8ad79 success build 26 s (1), deploy 23 s (1) 2 2 (WI-2.2 D-8)
39366531250 forge-et-be push bb7a2e6f success build 104 s (2), deploy 280 s (5) 7 15–16 (WI-2.1 D-12)
39366541171 forge-et-fe push d9450ece success build 23 s (1), deploy 26 s (1) 2 2 (WI-2.2 D-8)
39371315660 forge-et-fe push d903794e success build 40 s (1), deploy 22 s (1) 2 2 (WI-2.2 D-8)
39374212302 forge-et-be push 21b29f9e success build 35 s (1), deploy 293 s (5) 6 15–16 (WI-2.1 D-12)
39377895770 forge-et-be push 5d1ecd0c success build 97 s (2), deploy 776 s (13) 15 15–16 (WI-2.1 D-12)
39405207108 forge-et-fe push b576aaf2 success build 29 s (1), deploy 15 s (1) 2 2 (WI-2.2 D-8)
39423585119 forge-et-fe push 41852183 success build 38 s (1), deploy 17 s (1) 2 2 (WI-2.2 D-8)
39478869105 forge-et-fe push 6cc34d31 success build 23 s (1), deploy 14 s (1) 2 2 (WI-2.2 D-8)
39484112323 forge-et-be push a02ef485 success build 115 s (2), deploy 43 s (1) 3 2–3 (forge-et-be#286)
39484353197 forge-et-be push e9a18c0e success build 58 s (1), deploy 45 s (1) 2 2–3 (forge-et-be#286)
39503214129 forge-et-be push 41dc61e7 success build 105 s (2), deploy 39 s (1) 3 2–3 (forge-et-be#286)
39504367626 forge-et-fe push d9d6c844 success build 24 s (1), deploy 13 s (1) 2 2 (WI-2.2 D-8)
39510630707 forge-et-fe push 5b4a0138 cancelled — the deploy job exceeded its 10-minute timeout-minutes in sync::assets (BUG-254) build 28 s (1), deploy 900 s (15) 16 2 (WI-2.2 D-8)
39521556006 forge-et-be push aa987d73 success build 94 s (2), deploy 40 s (1) 3 2–3 (forge-et-be#286)
39521584931 forge-et-fe push b2dc3ee7 success build 27 s (1), deploy 13 s (1) 2 2 (WI-2.2 D-8)
39544798890 forge-et-be push efe80546 success build 99 s (2), deploy 55 s (1) 3 2–3 (forge-et-be#286)
39550541762 forge-et-be push fe88fdb1 success build 56 s (1), deploy 40 s (1) 2 2–3 (forge-et-be#286)
39557716213 forge-et-fe push 3b307cc7 success build 31 s (1), deploy 15 s (1) 2 2 (WI-2.2 D-8)
39559085905 forge-et-be push ae470144 success build 51 s (1), deploy 39 s (1) 2 2–3 (forge-et-be#286)
39562867998 forge-et-be push 718a07b3 success build 100 s (2), deploy 40 s (1) 3 2–3 (forge-et-be#286)
39571466115 forge-et-fe push 21288b2c success build 30 s (1), deploy 17 s (1) 2 2 (WI-2.2 D-8)
39578382589 forge-et-be push 9097a68b success build 69 s (2), deploy 53 s (1) 3 2–3 (forge-et-be#286)
39688233354 forge-et-fe push 029fcc76 success build 28 s (1), deploy 20 s (1) 2 2 (WI-2.2 D-8)
39693917080 forge-et-fe push 10696c8d success build 24 s (1), deploy 24 s (1) 2 2 (WI-2.2 D-8)
39708805585 forge-et-be push 38755865 success build 130 s (3), deploy 44 s (1) 4 2–3 (forge-et-be#286)
39714930703 forge-et-fe push 1592be74 success build 27 s (1), deploy 23 s (1) 2 2 (WI-2.2 D-8)
39723444029 forge-et-be push 3ef0064f success build 111 s (2), deploy 51 s (1) 3 2–3 (forge-et-be#286)
39727038992 forge-et-be push e841198b success build 56 s (1), deploy 41 s (1) 2 2–3 (forge-et-be#286)
39739022004 forge-et-fe push 1d97cd6f success build 24 s (1), deploy 25 s (1) 2 2 (WI-2.2 D-8)
39743884203 forge-et-be push e5503562 success build 52 s (1), deploy 48 s (1) 2 2–3 (forge-et-be#286)
39757087734 forge-et-fe push 1992577d success build 36 s (1), deploy 16 s (1) 2 2 (WI-2.2 D-8)
39767578461 forge-et-fe push 77d2ddda success build 31 s (1), deploy 23 s (1) 2 2 (WI-2.2 D-8)
39769056042 forge-et-be push bb434334 success build 50 s (1), deploy 41 s (1) 2 2–3 (forge-et-be#286)
39781741479 forge-et-be push 16b301b5 success build 43 s (1), deploy 41 s (1) 2 2–3 (forge-et-be#286)
39801944235 forge-et-be push 06062de8 success build 107 s (2), deploy 43 s (1) 3 2–3 (forge-et-be#286)
39802123578 forge-et-fe push 3b390a6d success build 27 s (1), deploy 22 s (1) 2 2 (WI-2.2 D-8)
39803044883 forge-et-fe push 9c5f0d85 success build 53 s (1), deploy 20 s (1) 2 2 (WI-2.2 D-8)
39819754231 forge-et-fe push 8e59d076 success build 29 s (1), deploy 17 s (1) 2 2 (WI-2.2 D-8)

Against the estimates. The backend’s runs with forge-et-be#275’s workflow billed 6 to 15 minutes, 7 in most of them, about half of WI-2.1 D-12’s 15–16, because the wait for a stable service took 170–693 s rather than the 9 minutes estimated; after forge-et-be#286 they billed 2 to 4, against #286’s 2–3 (the one 4 is run 39708805585, whose build job took 130 s). The CRM’s runs billed 2, as WI-2.2 estimated, except run 39510630707, which billed 16.

The cancelled CRM run. Run 39510630707 (5b4a0138, 2026-10-20 22:47 UTC) built and uploaded web-dist, configured AWS and printed its identity, then stayed in sync::assets, the first S3 sync, until GitHub stopped the job: the run’s annotation reads “The job has exceeded the maximum execution time of 10m0s”, and the job ran 22:48:04–23:03:04 UTC. Its log is not kept (gh run view --job 118930653473 --repo altrium-io/forge-et-fe --log → log not found: 118930653473, EXIT=1, twice). sync::html never ran, so the bucket kept the previous index.html and the site served the earlier build; the next run, 39521584931 (b2dc3ee7, 00:29 UTC), deployed main including 5b4a0138’s change. Result: not met for that run — recorded as BUG-254 (.forge/bugs.md).

The CRM’s deploys after wave 2. forge-et-fe#298’s merge started no run (AC-3 above); the first CRM deploy from a main push is run 39287294321 on 8c0bc93a (forge-et-fe#301, F11-h), whose build job ran check::config before pnpm::build and whose deploy job ran artifact::download, configure::aws, identity::aws (AC-1 above), sync::assets, sync::static, sync::html, sync::prune and cache::invalidate (aws cloudfront create-invalidation --distribution-id "E2RVC3893O10TJ" --paths "/*"), each success.

Still WI-6.1’s and WI-7.1’s: their own runs’ rows, and the PR comments and records AC-12 asks of each workflow PR (WI-7.1).

Recorded 2026-10-26 by F38-d-WI-6.1. The five runs this work item started, counted the same way (a skipped job is not billed). Estimates: the WI-6.1 plan’s revised D-14 (2–3 for a forge-et-be run, about 1 for W-9, which stops at the guard; about 13–15 in all).

2026-10-26 15:05 UTC — Claude (F38-d-WI-6.1), gh as tharinda-altrium

$ gh run view <id> --repo altrium-io/<repo> --json jobs --jq '[.jobs[] | select(.conclusion!="skipped") | …]'      # the formula above, once per row
EXIT=0 for each of the five
Run Repository Trigger Head Conclusion Jobs Billed minutes Estimate
40331078887 (W-1) forge-et-be workflow_dispatch e88fef50 success build 37 s (1), deploy 28 s (1) 2 2–3 (WI-6.1 D-14, revised)
40331100359 (W-2) forge-et-be workflow_dispatch e88fef50 success build 57 s (1), deploy 12 s (1) 2 2–3 (WI-6.1 D-14, revised)
40334043628 (W-4) forge-et-be workflow_dispatch 864e2f34 success (the deployment then failed and rolled back — AC-2) build 39 s (1), deploy 42 s (1) 2 2–3 (WI-6.1 D-14, revised)
40337010507 (W-7) forge-et-be workflow_dispatch 3c9e13f0 success (the deployment then failed and rolled back — AC-8) build 43 s (1), deploy 40 s (1) 2 2–3 (WI-6.1 D-14, revised)
40340419206 (W-9) forge-et-fe workflow_dispatch 326c481a failure at check::config (AC-5) build 5 s (1); deploy skipped 1 about 1 (WI-6.1 D-14)

Against the estimates. 9 billed minutes in all, under the revised estimate of 13–15, because W-11 was not needed (AC-5 above) and the backend runs billed 2 each, as forge-et-be#286’s 2–3 expected. The deploy jobs W-4 and W-7, which built a new image, took 42 s and 40 s, against 12–28 s for the skipped builds W-1 and W-2.

Recorded 2026-10-26 by F38-d-WI-7.1. The capture app’s runs, counted the same way, against WI-5.1 D-13’s estimate (2 a run; about 1 for a run that stops at the guard); then the minutes posted on each workflow PR, and each F38-d app-repo PR’s records.

2026-10-26 15:40–15:43 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium

$ gh run view <id> --repo altrium-io/forge-et-pwa --json jobs --jq '[.jobs[] | select(.conclusion!="skipped") | …]'      # the formula above, once per row
EXIT=0 for each of the four
Run Repository Trigger Head Conclusion Jobs Billed minutes Estimate
40224118047 forge-et-pwa workflow_dispatch 777fdc4b success build 24 s (1), deploy 16 s (1) 2 2 (WI-5.1 D-13)
40228839882 forge-et-pwa push c4b5db1e success build 15 s (1), deploy 11 s (1) 2 2 (WI-5.1 D-13)
40279386333 forge-et-pwa push 0ecd1322 success build 18 s (1), deploy 16 s (1) 2 2 (WI-5.1 D-13)
40351896602 (W-3) forge-et-pwa workflow_dispatch 0ecd1322 failure at check::config (AC-5) build 5 s (1); deploy skipped 1 about 1

Against the estimates. Every capture run billed what WI-5.1 estimated. The PR-body edits below each started a PR validation run on the merged PR’s old head (forge-et-be 40352469741, forge-et-fe 40352542448, forge-et-pwa 40352553328), because the workflow runs on edited; Claude cancelled each within a minute, since a run on a merged PR checks nothing.

The minutes on each workflow PR (W-5), 2026-10-26 15:47 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium; approved by the lead at 15:47 UTC

Approval: AskUserQuestion “W-5a: may I post this comment on forge-et-be#275? …” → “Yes, post it”; “W-5b: may I post on forge-et-fe#298 …” → “Yes, post it”; “W-5c: may I post on forge-et-be#286 …” → “Yes, post it”; “W-5d: may I post on forge-et-pwa#91 …” → “Yes, post it”. Each comment, headed “Billed minutes, measured (F38-d WI-7.1, 2026-10-26)”, gives one deploy run’s job times and billed minutes beside the PR’s own estimate and points here.

PR Run cited Billed minutes Estimate in the PR Comment
forge-et-be#275 39228765669 (its merge) 7 about 15–16 issuecomment-6420387267
forge-et-fe#298 39287294321 (the first CRM deploy after it) 2 2 issuecomment-6420388243
forge-et-be#286 39484112323 (the first backend deploy after it) 3 2–3 issuecomment-6420389021
forge-et-pwa#91 40228839882 (the first merge-started capture deploy) 2 2 issuecomment-6420389953

Each F38-d app-repo PR’s records, 2026-10-26 15:40 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium

$ gh pr view <n> --repo altrium-io/<repo> --json body,mergedAt,headRefOid      # once per row
$ gh api repos/altrium-io/<repo>/commits/<headRefOid>/status --jq '[.statuses[] | select(.context=="Forge review" or .context=="Test record") | [.context,.state]]'
EXIT=0 for each of the five
PR (merged, head) CI-change approval in the body Forge review on the merged head Browser gate: line “Shyamin Ayesh was told”
forge-et-be#275 (2026-10-19, 5ee0564a) yes — the WI-2.1 plan review, 2026-10-19 success n/a (backend) added 2026-10-26 — see below
forge-et-fe#298 (2026-10-19, 70cfb875) yes — the WI-2.2 plan review, 2026-10-19 success yes — “Browser gate: not needed — …” added 2026-10-26 — the body said “Pending”
forge-et-be#286 (2026-10-20, 2e449e20) yes — the lead in session, 2026-10-20 success n/a (backend) yes — told by Tharinda, 2026-10-20
forge-et-be#306 (2026-10-25, d28b5ff9) yes — the WI-4.1 plan review, 2026-10-19, and Rev 1, 2026-10-23 success n/a (backend) yes — confirmed by the lead, 2026-10-23
forge-et-pwa#91 (2026-10-25, 699d9b28) yes — the WI-5.1 plan review, 2026-10-19 success added 2026-10-26 — the body said only that no gate input changed yes — confirmed 2026-10-26

2026-10-26 15:45 UTC — Claude (F38-d-WI-7.1), gh as tharinda-altrium; approved by the lead at 15:44 UTC

Approval: AskUserQuestion “AC-12 needs each F38-d app-repo PR body to say Shyamin Ayesh was told what it changes. forge-et-be#275 has no such line and forge-et-fe#298 still says “Pending” … Was Shyamin told about #275 and #298?” → “Told; edit bodies”; “AC-12 also needs a Browser gate: line on each frontend PR. … How should I handle #91?” → “Add the line to #91”.

$ gh pr edit 275 --repo altrium-io/forge-et-be --body-file <the body plus a "Note to Shyamin Ayesh (AC-12)" section>      # 15:45:26
EXIT=0
$ gh pr edit 298 --repo altrium-io/forge-et-fe --body-file <the body with its "Pending: …" line replaced>      # 15:45:54
EXIT=0
$ gh pr edit 91 --repo altrium-io/forge-et-pwa --body-file <the body with a "Browser gate: not needed — …" line after the gate sentence>
EXIT=0

The line added to #275 and put in place of #298’s “Pending” line: “Shyamin Ayesh was told what this PR changes in the deploy files — confirmed by the lead (Tharinda) on 2026-10-26, recorded by F38-d WI-7.1 (the date he was told was not recorded).” The line added to #91: “Browser gate: not needed — this PR changes no gate input (…; node scripts/browser-gate.mjs touches-gate-input --base origin/main --head HEAD → none). Added after merge by F38-d WI-7.1 on the lead’s approval, 2026-10-26.” Nothing else in the three bodies changed.

Result: met — every workflow PR carries the measured minutes of one deploy run beside its estimate (the ARM64 build, the artifact and ECR storage are set against sow-cost-annexe.md §3 in forge-et-be#275’s “AWS cost” lines and WI-2.1 D-12; the frontends’ PRs add no AWS resource); and every F38-d app-repo PR carries the lead’s CI-change approval, a success Forge review record on its merged head, a Browser gate: line on each frontend PR, and the sentence that Shyamin Ayesh was told. The last two were completed after merge, on the lead’s word, for #275, #298 and #91. With WI-4.2’s and WI-6.1’s rows above, AC-12 is met.

F38-e — one-off resources on dev, checked

The short-lived resources spec NFR-2 allows outside Terraform, each created and removed in one session by AWS CLI calls the lead approved: FR-6’s test alarm (F38-e-WI-3.1) and FR-8’s SES probe (F38-e-WI-3.2). The full records are the work items’ record files. On 2026-11-21 Claude ran them in the lead’s session; the lead approved the writes at 20:16 UTC (“F38-e AWS writes”, “Throwaway BE deploys”, “iaac manual dev.yaml”), ran the two IAM additions himself (~/.F38-e-iam-add.sh, 20:20 UTC), and allowed Claude to run the removals.

FR-6 — the test alarm (F38-e WI-3.1)

  • 20:20:36 UTC — the lead: aws iam put-user-policy --user-name svc-harness --policy-name temp-f38e-test-alarm (cloudwatch:PutMetricAlarm, SetAlarmState and DeleteAlarms on arn:aws:cloudwatch:us-east-1:570417736903:alarm:et-dev-f38e-test-alarm only).
  • 20:22:39 UTC — aws cloudwatch put-metric-alarm --alarm-name et-dev-f38e-test-alarm --namespace ET/F38e --metric-name TestAlarmTrigger --statistic Sum --period 300 --evaluation-periods 1 --threshold 1 --comparison-operator GreaterThanOrEqualToThreshold --treat-missing-data notBreaching --alarm-actions arn:aws:sns:us-east-1:570417736903:et-dev-alerts — refused (AccessDenied) while the new policy took effect; 20:23:09 UTC attempt 2 — EXIT=0, created.
  • 20:23:23 UTC — aws cloudwatch set-alarm-state --alarm-name et-dev-f38e-test-alarm --state-value ALARM; alarm history at 20:23:24 UTC: “Successfully executed action arn:aws:sns:us-east-1:570417736903:et-dev-alerts”.
  • 20:24 UTC — aws cloudwatch delete-alarms --alarm-names et-dev-f38e-test-alarm; describe-alarms → 0. The alarm existed from 20:23:09 to 20:24 UTC, about one minute.
  • 20:24:30 UTC — aws iam delete-user-policy --user-name svc-harness --policy-name temp-f38e-test-alarm; list-user-policies → spike-invoke-gateway only.

FR-8 — the SES probe (F38-e WI-3.2)

  • 20:20:37 UTC — the lead: aws iam put-user-policy --user-name svc-harness --policy-name temp-f38e-ses-probe (ecs:RegisterTaskDefinition, DeregisterTaskDefinition, DeleteTaskDefinitions on *; ecs:RunTask on task-definition/et-dev-ses-probe:* in cluster et-dev; no PassRole statement, since IAMFullAccess already holds it).
  • 20:22:42 UTC — aws ecs register-task-definition → et-dev-ses-probe:1 (Fargate, ARM64, 0.25 vCPU / 0.5 GB, public.ecr.aws/aws-cli/aws-cli:2.37.12, execution role et-dev-backend-execution, no task role, awslogs to /ecs/dev/backend prefix ses-probe).
  • 20:22:56 UTC — aws ecs run-task --cluster et-dev --task-definition et-dev-ses-probe:1 --launch-type FARGATE --network-configuration 'awsvpcConfiguration={subnets=[subnet-02fb26b2351e546d1],securityGroups=[sg-0097bf895557b7716],assignPublicIp=ENABLED}' --overrides '{"taskRoleArn":"arn:aws:iam::570417736903:role/et-dev-backend-task"}' → task 2487094ed18a4d17933dd550ffdeb803, ran 20:23:15–20:23:42 UTC, exit 0.
  • 20:24:11 UTC — aws ecs deregister-task-definition --task-definition et-dev-ses-probe:1 → INACTIVE; 20:24:12 UTC — aws ecs delete-task-definitions --task-definitions et-dev-ses-probe:1 → DELETE_IN_PROGRESS; list-task-definitions --family-prefix et-dev-ses-probe --status ACTIVE → none.
  • 20:24:31 UTC — aws iam delete-user-policy --user-name svc-harness --policy-name temp-f38e-ses-probe; list-user-policies → spike-invoke-gateway only.
  • Left in the account: the log stream ses-probe/ses-probe/2487094ed18a4d17933dd550ffdeb803 in /ecs/dev/backend, removed by the 30-day retention after 2026-12-21; ECS finishes deleting the DELETE_IN_PROGRESS revision itself.

FR-2 — the trace check (F38-e WI-3.3)

  • 20:27:50 UTC — aws logs start-query --log-group-name /ecs/dev/backend filtering on traceId = "6b01fea55f8d844684a413d9d1b2b42c" → queryId 147a307a-3520-45e8-888f-c5fee18092d1 (931 bytes scanned). The request was a note added to the dev test account 1 as CRM01 at 20:27:17 UTC (note 55), deleted at 20:28:33 UTC.
  • Left in the account: nothing other than the query’s results, which CloudWatch Logs keeps for a limited time.

F38-f — changes made on dev, and put back

Every AWS change F38-f made on dev (spec AC-10), each approved by the lead in the session that made it, with the time it was made and the time it was put back. The full records are the work items’ record files.

F38-f-WI-2.2 — the manual path

Recorded 2026-11-21 by F38-f-WI-2.2; the full record is .forge/plans/F38-f/F38-f-WI-2.2-record.md. Run by Claude in Tharinda Rajapaksha’s session at his request. The lead approved the five writes below as one group at 20:16 UTC on 2026-11-21. The window was not announced, by his decision. Every other call was a read with the profile et-dev (svc-harness).

Write Approval Time (UTC) Result
Commit on the throwaway forge-et-be branch chore/F38-f-manual-path-check, cut from main abc9f368: six ET_AI_FEATURES_*_ENABLED=false lines in .aws/td.dev.json (INQUIRYEXTRACT, EMAILDRAFT, LEADTRIAGE, ACCOUNTBRIEFING, LEADRESEARCH, LEADRANKING) Tharinda Rajapaksha, 20:16 22:07 head b679b6e60b1956fc821f2380fa870aaaf6772dc5, diff 6 0
The branch pushed (no PR opened) Tharinda Rajapaksha, 20:16 22:08:01 pushed
gh workflow run build.yaml --ref chore/F38-f-manual-path-check — run 44955151808 Tharinda Rajapaksha, 20:16 dispatched 22:08:15; the branch deployed at 22:13:40 success; revision backend:179 (the six switches off) serving, one task
gh workflow run build.yaml --ref main — run 44956014556 Tharinda Rajapaksha, 20:16 dispatched 22:17:56; main deployed again at 22:22:22 success; revision backend:180 (main abc9f368, no _ENABLED line) serving, one task. Put back.
The branch deleted on the remote, and its local branch and worktree removed Tharinda Rajapaksha, 20:16 22:26:05 ls-remote prints nothing; no PR was ever opened; b679b6e6 is not on origin/main

Left in AWS (WI-2.2 plan D-7, the lead’s answer to its Q-1): the revision backend:179 stays registered and ACTIVE, superseded by backend:180; the branch’s image crm/backend:build-b679b6e60b1956fc821f2380fa870aaaf6772dc5 (digest sha256:04d82859351eb2ef442203485f196cd59bcc2851936b7476b94a308acfa485d0) stays in ECR. Dev ran with the six CRM AI features off from 22:13:40 to 22:22:22 UTC.

Billed minutes (each job rounded up, as #### AC-12 — billed minutes and PR records counts them): run 44955151808 — build 39 s (1), deploy 46 s (1), 2; run 44956014556 — build 46 s (1), deploy 12 s (1; the image of main was already in ECR), 2.

F38-f-WI-2.4 — the mailbox key grant

Run 2026-11-21 by Claude in the lead’s session; each change approved by Tharinda Rajapaksha (20:16 UTC, “F38-f risky writes”; the temporary policy under his permission of 20:24 UTC). Full record: .forge/plans/F38-f/F38-f-WI-2.4-record.md.

  • 21:40:08 UTC — aws iam put-user-policy --user-name svc-harness --policy-name temp-f38f-wi24-redeploy (ecs:UpdateService on service/et-dev/backend only); removed 21:51:19 UTC (delete-user-policy), inline policies back to spike-invoke-gateway.
  • 21:40:12 UTC — aws iam put-role-policy --role-name et-dev-backend-task --policy-name kms without the DirectKeys statement; put back 21:51:09 UTC with the document read first, read back identical (cmp).
  • 21:41:24 UTC — aws ecs update-service --cluster et-dev --service backend --force-new-deployment (same revision backend:177), completed 21:45:33 UTC, task 384b6f575dd44f1396474f13e83b6a38.
  • Left in the account: nothing; the dead-letter queue held 2 messages before and after.

F38-f-WI-2.6 — the audit log at DEBUG

Run 2026-11-21 by Claude in the lead’s session, as CRM01 through /actuator/loggers, approved by Tharinda Rajapaksha with the dev windows. Full record: .forge/plans/F38-f/F38-f-WI-2.6-record.md.

  • 22:02:05 UTC — com.empireteas.platform.audit set to DEBUG; put back to null (effective INFO) at 22:04:25 UTC, read back.
  • 22:02:16 UTC — one 1×1 PNG uploaded (attachment 21), discarded at 22:04:23 UTC.
  • Left in the account: nothing; the upload’s S3 object is deleted with the attachment.

Agenda for the meeting with Shyamin Ayesh (drafted 30 Sep 2026)

The record of the 30 Sep agenda; what is still open is the list above. Our position is given where we have one; the F38 siblings are the acceptance criteria his work is checked against.

A. Region, DNS and the account

  1. Region — us-west-2 (his finding: cheaper) or us-east-1 (AD-O-9). Decided 1 Oct 2026: us-east-1 (§3). Both Claude models we use are offered the same way from either Region (ai-provider-strategy.md §4b). What us-west-2 costs us:

    • AgentCore Web Search — SP-ARCH-004’s recommended lead-research provider — runs only in us-east-1, eu-west-1 and ap-northeast-1 (AWS’s Web Search Tool page, read 2026-09-30; .forge/design/ai/lead-research-source-options.md). From us-west-2 its gateway runs in us-east-1 (same account, one cross-Region call), or the fallback, Tavily, is used.
    • Two Regions in the Terraform instead of one. CloudFront’s ACM certificates, a WAF in front of CloudFront, and CloudWatch’s billing metrics exist only in us-east-1, so a us-west-2 estate needs a second provider for those.
    • New AWS features usually reach us-east-1 first (AgentCore Web Search is an example).
    • Documents and the client: AD-O-9, the proposal’s 21 August post-issue note, dependencies.md, the cost annexe and the F38 specs all say us-east-1, and Empire is told that their data sits in Oregon rather than Virginia (both in the US).

    What it gains: the lower cost he found, to be checked against sow-cost-annexe.md; and us-east-1, the largest and oldest Region, is where most widely felt AWS outages have happened. Latency from Sri Lanka is to be measured, not assumed. The team’s leaning (2026-09-30): us-east-1, unless the saving Shyamin found is real and noticeable — for the main services the two Regions are usually priced the same, so the question for him is which costs were lower. Whichever is chosen, move now, while only a VPC and a cluster exist.

  2. DNS. Import the crm.empire.lk zone into Terraform; Terraform writes every record (§1.1, F38 Rev 2); which address CRM mail is sent from (F38 OQ-E) decides whether Empire must add SES records.

  3. Account structure. One account, or dev and production as separate accounts under an Organization, as AD #4 and F38-a FR-1 expect? Most items below depend on the answer.

  4. Other environments. Production after Phase 1 (F38-i); UAT as a temporary terraform apply / destroy per acceptance round (AD #23), the first needed inside the Phase 1 UAT window that ends Thu 10 Dec 2026; the production administrator’s email in production’s Secrets Manager (2026-10-28, AD #40: as a plain value in production’s task definition) (§5).

B. The dev estate (F38-b, F38-e) 5. ECR, Secrets Manager, RDS, ECS, SQS, ALB, IAM, the task definition. Share the F38-b spec. Fixed points: no NAT gateway, tasks in a public subnet, the database in private subnets in one Availability Zone (architecture.md); SQS queues with dead-letter queues; EventBridge Scheduler and its role; KMS keys, including the one that wraps mailbox tokens (F20-a); backup retention of 3 days on dev and 14 on production; register §D lists every task-definition value. 6. CloudFront and S3 for both frontends — dev.crm.empire.lk and capture.dev.crm.empire.lk; a static build of each can be tested before any backend exists. 7. The Entra certificates and private keys in Secrets Manager — the non-prod pair into dev (the vault copy stays while laptops and UAT read it); the production pair only at F38-i; where UAT reads its key from; rotation (§2.3). 8. (Superseded 9 Oct 2026: GuardDuty off — both Malware Protection for S3 and account-wide threat detection. Thilina, with Malinda, decided to go without it and to add it later only if it is needed or Empire Teas asks; Tharinda told Shyamin Ayesh the same day. Every environment runs F38-j’s setting ET_FILES_SCAN_PROVIDER=none with ET_FILES_SCAN_OFF_ACKNOWLEDGED=true, as dev already does; F38-c’s scanner code stays on main unused, so adding GuardDuty later is Terraform plus two settings, no code. The proposal (§A.2.2, §A.4) and the client-facing PRD (uploads “virus-scanned”) promise both; on 2026-10-11 the lead decided Empire Teas is not told separately, since the client-facing PRD (revised 9 Oct) already states it. The text below is the 2 Oct decision, kept for the record.) GuardDuty — decided 2 Oct 2026: on. Malware Protection for S3 goes on the attachments bucket, built to forge-et-be/docs/messaging/scan-result-cloud-contract.md (the protection plan with no object-prefix scoping, the EventBridge rule, the scan-result queue and its dead-letter queue, the plan’s role and the KMS and queue-policy grants). It costs under $1 a month at our volume (AD-O-4). The file scanner built for the cloud (F38-c) is GuardDuty’s. Removing the protection plan on its own leaves every new file PENDING and undownloadable. Switching Malware Protection off follows forge-et-be/docs/messaging/scan-result-cloud-contract.md § “Switching scanning off and back on (F38-j)”, once F38-j is done (both its waves on main and deployed): the setting first, the protection plan after, and production only after Empire is told. Account-wide threat detection is enabled by F38-a FR-4. 9. Observability and mail — CloudWatch logs and alarms, SNS to the delivery team’s address, SES production sending access (it needs AWS’s approval, so ask early), AWS Budgets and who at Empire receives their alerts.

C. AI 10. (Update 9 Oct 2026: the Marketplace permissions were granted to svc-harness; Haiku 4.5 is subscribed and answers; Sonnet 5.5 and Haiku 5.5 are refused as “not available for this account” — §3. The open part is getting AWS to enable those two models for the account, which only the account holder can ask AWS for.) Bedrock — who submits Anthropic’s first-time-use form; the Marketplace subscription permissions; model access to Claude Sonnet 5.5 (global. profile) and Haiku 4.5 (us. profile), with Haiku 5.5 added when it arrives; the task role’s IAM policy allowing both profiles; the quotas for both models (F38-a FR-7). No policy refusing other Regions while Sonnet 5.5 is called through the Global profile (F38-a Q-2). The delivery team’s address is added as an alternate contact on the account, so AWS’s model legacy and end-of-life notices reach the team (ai-provider-strategy.md §4c rule 7; Haiku 4.5 may be retired from 16 Oct 2026 onwards, with at least six months’ legacy period).

D. Pipelines 11. Deployment pipelines in the app repos (F38-d; architecture.md § CI/CD Pipelines). From him: the ECR repository and ECS service names, the S3 bucket names and CloudFront distribution IDs, CI’s AWS credentials — an IAM service user per account, its key stored as a GitHub Environment secret (F38 Rev 4, which replaced the OIDC role) — GitHub environment variables per environment. Actions minutes are scarce across the organisation, so deploy jobs stay small. Both halves of the startup check are on main (register A-1).

E. Access 12. (Superseded for dev on 14 Oct: the team uses svc-harness’s access key, item 8 of the list above; production keeps personal logins.) Console access for each developer — IAM Identity Center with MFA, one login per person, then the shared administrator credentials retired (F38-a FR-2); permission sets per role. 13. (Update 13 Oct 2026: a read-only AWS role cannot check data before a migration, such as BUG-162’s V41 pre-check on dev, which Shyamin Ayesh ran by hand for forge-et-be#242. That needs a SELECT-only Postgres role on the dev database with its password in Secrets Manager, secretsmanager:GetSecretValue on that one secret, and a network path to the database, such as SSM port forwarding through a bastion or an ECS task. Asked for with the command-line access in item 7, as the lower priority.) (Update 14 Oct 2026: Claude uses svc-harness in dev for reads only and asks the lead before any other call (.claude/CLAUDE.md § ASK FIRST); database access is the list’s item 15 — a free tunnel through the backend with a dev_readonly login, set up when first needed, a check like BUG-162’s being such a need.) Access for Claude — our position: Claude changes AWS only through Terraform PRs in forge-et-iaac, and holds a read-only role for inspection and debugging; no direct write access, and none on production. 14. forge-et-iaac — write access for the four developers; CODEOWNERS with Shyamin as reviewer; plan on each PR and apply only after merge; protection on dev, and whether the trunk should be main as in the other repos; the hygiene points listed above. (Update 2026-11-19: Tharinda, Dinil and Nanndish have write access; Biseni is still read-only — gh api --paginate repos/altrium-io/forge-et-iaac/collaborators --jq '.[]|.login+" "+.role_name'.)

F. Also to settle 15. Who builds what — which parts of F38-a, F38-b, F38-d, F38-e, F38-f, F38-g and F38-h Shyamin builds and which the team builds, so nothing is built twice or missed. 16. Dates — the Foundation milestone is Fri 16 Oct 2026, assuming the environment and the team’s access are ready by Fri 2 Oct (.forge/client/project-timeline.md); every milestone moves day for day after that. 17. Production decisions for Empire (register §C-8): single- or multi-AZ database, cross-Region backup copy, WAF at go-live. 18. Cost — a monthly estimate for dev against the annexe, and resource tags that name a team or role rather than a person.

5. Owners & status

Item Owner Status
Entra app registration (IDs, activation) Altrium / Empire IT Done (registration created; IDs assigned)
M365 test accounts (CRM01–05@empire.lk, Business Basic) Empire IT Done — 5 accounts provisioned (passwords held out-of-band, not in the repo). crm01@empire.lk is the seeded IT Admin on local, dev and UAT — the value of the F5-k Flyway placeholder outside production (AD #34), so the first sign-in on each of those environments is as CRM01. CRM01, CRM02, CRM03 and CRM05 have a working Outlook mailbox; CRM04 has none (corrected 2026-11-19, F38-h WI-5.2 record R-10: Outlook on the web as CRM04 answered OwaUserHasNoMailboxAndNoLicenseAssignedException, and its CRM mailbox connection on dev failed with MAILBOX_NOT_ENABLED; .forge/owed.md). MFA is off on all five (the lead, 2026-10-19, answering F38 OQ-H): whoever holds an account’s password can sign in as it, so the passwords stay out-of-band only. CRM05 had no Outlook or calendar access on 2026-09-29, which is how BUG-048 was found; as of 2026-10-04 its Outlook mailbox opens and holds mail, so it can be used for mailbox tests like CRM01–CRM03 (a CRM mailbox connect as CRM05 has not been repeated since). The no-mailbox path (MAILBOX_NOT_ENABLED, BUG-048) is proved by the backend tests against a stubbed Graph error, and live by CRM04 on dev on 2026-11-19 (F38-h WI-5.2 record R-10)
Smoke account (F38-g) — CRM05 (crm05@empire.lk), holding the role Smoke test (view only) with exactly account.all, account.view, buying_standard.view, lead.all and lead.view; reserved for pnpm e2e:smoke in forge-et-fe (nothing else signs in as it, the retest included; the role is never changed or assigned to anyone else; the account is never deactivated). The SMOKE name prefix is reserved for the records the smoke run reads, which nothing edits, archives or withdraws (.forge/plans/F38-g/F38-g-WI-1.1-plan.md, contract C-3) Tharinda Set up on dev 2026-10-21 (F38-g-WI-2.1). Role key smoke_test_view_only; first green runs on 2026-10-22, after a forge-et-fe deploy (R-3, run 39714930703) and a forge-et-be deploy (R-4, run 39727038992); record .forge/plans/F38-g/F38-g-WI-2.1-record.md. The after-deploy smoke run (.forge/team-guide.md § Environments, “After a dev deploy — the smoke run”) applies from this date. On dev, the IT Admin role also holds account.manage, account.all, lead.manage, lead.all, buying_standard.manage and price.publish, added on 2026-10-21 for this setup and kept for the retest (F38-g-WI-2.1 D-1); it does not hold email.manage
Production administrator’s email (the F5-k placeholder value in production — AD #34) Kavindu Anthony to confirm the address; Tharinda to ask and to store it To confirm — not yet asked. The address itself is never written in this file. Held in the production account’s Secrets Manager once AWS exists, never in git (secrets table above). (2026-10-28, AD #40: a plain value in production’s task definition once confirmed; asked of Kavindu in the message drafted in .forge/owed.md, “Empire IT asks”.) Without it the production backend cannot start: Flyway refuses the migration while the placeholder has no value (.forge/cloud-gated-validation.md §B F5-k)
Entra reply URLs (Web platform) Kavindu Anthony (Empire IT) — the only person with Azure portal access Done — all eight in §2.2 are registered. The three sign-in reply URLs were configured on the 10 Sep session and finalised afterwards; the three mailbox and two laptop URLs were added on 24 Sep 2026 (confirmed by Kavindu’s portal screenshot). No further reply-URL ask is expected (§2.2). The prod/uat/dev hosts only resolve once those environments + DNS exist
Entra login certificates (two public certs: prod, and non-prod for dev, UAT and laptops) Shyamin Ayesh generated; Kavindu Anthony uploaded Done — both uploaded 24 Sep 2026 (thumbprints in §2.3). Private keys in a restricted vault until each environment’s Secrets Manager exists (F38)
Certificate expiry dates Tharinda Recorded: prod — 23 Sep 2028 / non-prod — 23 Sep 2028. Owed: a calendar reminder about a month before (late Aug 2028), and the replacement pairs uploaded before then (§2.3, validity and rotation)
Graph delegated permissions + admin consent Altrium / Empire IT Granted 10 Sep 2026 on the session with Kavindu Anthony (delegated set + offline_access; admin consent given) — see 2026-09-10-azure-pipedrive-configuration.md
AWS account Kavindu Anthony (Empire IT) Done — Empire Teas’ own account, on Empire Teas’ card; administrator IAM credentials given to Shyamin Ayesh (confirmed 29 Sep 2026, §3). The schedule assumes the environment, with the team’s access, is ready by Fri 2 Oct 2026 (.forge/client/project-timeline.md)
AWS environments (dev/prod, us-east-1) and the developers’ logins Shyamin Ayesh, with Senuthmi Thimansa In progress from 29 Sep 2026 — Terraform and the environments; in dev the team uses svc-harness’s access key instead of a login per developer (decided 14 Oct, §4 item 8). In us-east-1 since 5 Oct 2026. The backend has run on dev since 9 Oct 2026 at https://api.dev.crm.empire.lk (liveness UP), deployed by forge-et-be’s build workflow on every merge to main; since 13 Oct the CRM frontend runs at https://dev.crm.empire.lk, deployed by forge-et-fe’s build workflow, and sign-in as CRM01 works. Since 20 Oct the attachments bucket, the KMS keys, the ingest queue and EventBridge Scheduler exist (forge-et-iaac#3), and since 22 Oct their values are in the backend’s task definition (forge-et-be#300), so file uploads and the mailbox wait only on being walked on dev; the schedules run (read 25 Oct). The capture app’s hosting exists since 22 Oct (forge-et-iaac#5), its dev GitHub Environment since 23 Oct, and its deploy since 25 Oct (forge-et-pwa build.yaml, F38-d wave 5); https://capture.dev.crm.empire.lk serves the app’s shell. Since 23 Oct the alert topic and the failed-deploy email rule exist (forge-et-iaac#7), and ECR’s tags are immutable (#6). The differences from F38 are listed in §4 (“Checked against F38”). SES in dev can send from dev.crm.empire.lk since 21 Oct 2026, when production access was granted (F38-a wave 3, recorded in §4 “F38-a — the development account, checked”, AC-5), and the lead’s test message from noreply@dev.crm.empire.lk reached the CRM01 Inbox the same day with SPF, DKIM and DMARC pass (AC-5); the backend sending through its own task role is still to be built and shown (F38-e). Lead triage (F11-d) runs on dev since 20 Oct (backend:44) and lead research (F11-c) since 23 Oct; the service runs backend:65 as of 26 Oct (registered 25 Oct, 20:04 UTC, with ET_DATASOURCE_SSL_MODE=verify-full from forge-et-be#306, F38-d wave 4). Bedrock serves Haiku 4.5 at 10 requests a minute; AWS refused the quota requests for Sonnet 5.5 and Haiku 5.5 again on 21 Oct, Shyamin Ayesh sent AWS the lead’s reply on 22 Oct, and on 26 Oct AWS answered with no increase: the account keeps new-account limits until it has a few billing cycles of Bedrock usage and spend, with no date (§4 item 4, the 26 Oct update). Production is not started. The dev account’s baseline was checked on 15 Oct 2026 (F38-a wave 1) and is recorded in §4 “F38-a — the development account, checked”; the resources forge-et-iaac#3 applied were checked on 21 Oct 2026 (F38-b wave 2) and are recorded in §4 “F38-b — the dev estate, checked”; on 26 Oct 2026 F38-b wave 4 recorded there the network, the database, both frontends’ hosting (the capture app’s included) and a sweep finding nothing outside Terraform, and a dev.yaml run on dev the same morning (40278225094) planned no changes, so dev shows no drift. Cloud-gated items that need uploads, the mailbox or the 5.5 models wait on the rest of it; A-5 closed on 13 Oct when sign-in worked (.forge/cloud-gated-validation.md). The bring-up is tracked as F38; its order of work is the register’s “Start here” checklist
Terraform repo Shyamin Ayesh, with Senuthmi Thimansa Decided 29 Sep 2026: forge-et-iaac (created empty that day; networking and the ECS cluster added 30 Sep; RDS and ECR 5 Oct; the backend service, load balancer, certificate and scheduler queue 9 Oct; the AgentCore Gateway, the Entra secrets and the CRM frontend’s hosting 13 Oct; the delivery team’s first PR, forge-et-iaac#3 (F38-b), opened 17 Oct, merged and applied 20 Oct, with the KMS keys, the attachments bucket, the ingest and lead-triage queues and EventBridge Scheduler — §4; its merge run, run 39475850266, is recorded in §4 “F38-b — the dev estate, checked”, AC-1). Review decided 2 Oct 2026: a developer’s change is a PR to dev, reviewed by the others with write access (Shyamin, Senuthmi, Tharinda, Dinil); the DevOps team may push to dev directly (the lead, 5 Oct 2026); its other conventions are on hold until the setup is finished — see §4
crm.empire.lk delegation — four NS records in the empire.lk zone (§1.1) Dinil Gamage asked; Kavindu Anthony added (in Cloudflare); Shyamin Ayesh copied Done — live 1 Oct 2026 (§1.1). No further DNS request is expected unless the CRM sends mail from @empire.lk itself (F38 OQ-E). Still owed on our side: importing the zone into Terraform (F38-i, before production’s records are written — F38 Rev 13)
GuardDuty — account threat detection and Malware Protection for S3 Altrium management (Thilina, with Malinda) decided Decided 9 Oct 2026: both off, until needed or Empire Teas asks (§4, agenda item 8); every environment runs scanning switched off (F38-j). The client-facing PRD (revised 9 Oct) states it, and the lead decided on 11 Oct that Empire Teas is not told separately. Before 9 Oct, and kept for the record: Decided 2 Oct 2026: both on (§4, agenda item 8). Malware Protection follows forge-et-be/docs/messaging/scan-result-cloud-contract.md (F38-b); threat detection is F38-a FR-4. Switching Malware Protection off follows forge-et-be/docs/messaging/scan-result-cloud-contract.md § “Switching scanning off and back on (F38-j)”, once F38-j is done (both its waves on main and deployed): the setting first, the protection plan after, and production only after Empire is told. Removing the protection plan without the setting leaves every file uploaded afterwards PENDING, and it cannot be downloaded. Account-wide threat detection needs no application code and can be turned on or off in AWS alone (the proposal lists it in §A.2.2’s cost table and §A.4’s security posture, so turning it off is a change to tell Empire about). Per AWS’s documentation (read 2 Oct 2026): suspending or disabling GuardDuty does not stop Malware Protection for S3 on a protected bucket; the one optional add-on that touches the running backend is Runtime Monitoring for ECS Fargate, which adds a GuardDuty sidecar container to each new task (it does not stop a task from starting if the sidecar fails), so turn it on only by decision. To verify (Tharinda): once forge-et-iaac has the GuardDuty resources and the team has read-only console access, compare what is built with this row, with F38-a FR-4 and with forge-et-be/docs/messaging/scan-result-cloud-contract.md, and correct the harness documents
Mailbox reply URLs (dev/UAT/prod …/authorized/microsoft-mailbox) Kavindu Anthony (Empire IT) Done — registered 24 Sep 2026 (§2.2)
Laptop reply URLs (§2.5) — laptops use the non-prod certificate, so no laptop certificate is needed Kavindu Anthony (Empire IT) Done — registered 24 Sep 2026 on the shared CRM registration; removal owed before go-live (row below)
Certificate plan (§2.3) Tharinda with Shyamin Ayesh Count, timing, generator and validity decided (two — prod and non-prod — generated by Shyamin Ayesh and uploaded 24 Sep 2026, valid two years). Where UAT reads the non-prod key: decided by the lead on 2026-10-23 — from dev’s own Entra secrets (§2.3), for Shyamin Ayesh to confirm. To decide: the rotation routine
Real-Entra laptop mode in forge-et-be and forge-et-fe (§2.5 item 3) Tracked as F5-i — Tharinda Done — built by F5-i Wave 1 (how-to forge-et-be/docs/auth/laptop-sign-in.md); the live run as CRM01 (F5-i-WI-2.1) was recorded in forge-et-harness#322, merged 2026-09-30
Remove the laptop reply URLs before production go-live, and consider replacing the non-prod pair (§2.5) Tharinda to ask; Kavindu Anthony to remove Owed at go-live — also a row in .forge/cloud-gated-validation.md §B FS-6

Agenda for the next meeting with Empire IT (Kavindu Anthony)

  1. Register the three mailbox reply URLs — done 24 Sep 2026 (§2.2).
  2. Register the two laptop reply URLs — done 24 Sep 2026, on the shared CRM registration (§2.2, §2.5).
  3. Upload both certificates’ public halves — done 24 Sep 2026; they were sent by email as .pem files, public halves only (§2.3).
  4. Agree how renewals will be handled. The registration holding several certificates at once is confirmed (both sit on it). Still to agree: before each expiry (23 Sep 2028) Altrium sends the new public half, Kavindu uploads it alongside the old one, and removes the old one once the switch is done.
  5. Re-read the granted permission list on the registration: is User.ReadBasic.All granted (§2.4)? No feature needs it today: F22 syncs a user’s own contacts under Contacts.ReadWrite, which is granted, and needs it only if its spec also reads other people in the organisation (2026-11-06).
  6. Tell Kavindu that calendar, contacts and the capture app need no further reply URLs — done, in the 24 Sep 2026 request.
  7. Confirm the production administrator’s email — the one address that becomes the first CRM user in production, holding the IT Admin role (AD #34). Explain that it is stored in the system’s configuration (2026-10-28, AD #40: a plain task-definition value, not a secret), and that production cannot start without it. Mention that the dev and UAT environments start with CRM01@empire.lk as their administrator, and recommend that Empire keeps two people in an administering role once the CRM is live, so one person leaving cannot lock everyone out.
  8. After each item: mark it done in the client-facing .forge/client/dependencies.md §1.2 (the 24 Sep 2026 items are marked there).

Status as of 24 Sep 2026: items 1, 2, 3 and 6 are done; items 4, 5 and 7 are still to ask.

Who can change what. Only Kavindu Anthony (Empire IT) has access to the Azure portal. Every change to the CRM app registration — reply URLs, certificates, permissions, consent — is made by him, so each one is gathered into a meeting with him (the agenda above). Shyamin Ayesh (Altrium DevOps) has no Azure access: his part is the AWS side (§3, §4) and generating the certificates, whose public halves go to Kavindu.

Key contacts: Kavindu Anthony (Empire Teas), and Shyamin Ayesh and Senuthmi Thimansa (Altrium DevOps) — roles in team-guide.md § The Team.

Generated from .forge/design/environments-and-identity.md by .claude/lib/render-docs-html.js. Markdown is canonical.