source: .forge/discovery/meeting-notes/2026-10-01-aws-setup-shyamin.md · generated — edit the markdown, not this file

Empire Teas CRM — AWS setup call with Shyamin Ayesh

Date: 1 October 2026 Meeting: Working call on the AWS setup — the account structure, the region, access, the dev estate, the deploy pipelines and the AI models Attendees: Shyamin Ayesh, Tharinda Rajapaksha, Dinil Gamage (Altrium) Agenda used: the pre-meeting checklist “AWS setup — what to agree with Shyamin”, prepared on 30 September 2026. It is not in the repository. Its items are numbered 1–19 and are cited below as “checklist #N”. Its sources were .forge/design/environments-and-identity.md §4, .forge/design/ai/ai-provider-strategy.md §4b, the F38 specs and .forge/cloud-gated-validation.md. Source: Dinil’s account of the call, written up on 1 October 2026 from memory and from answers to follow-up questions, and checked against the checklist and the harness documents. Wording is approximate. The DNS section also records a message exchange after the call (§8).

Why this file exists: it is the dated record of what was agreed with Shyamin about the AWS estate in Empire Teas’ account. The decisions here feed the F38 specs (F38-a account baseline, F38-b dev estate, F38-d pipelines), environments-and-identity.md §1 and §3–§5, and the cloud register. Those documents are updated from this note; this note does not replace them. Three of the agreements differ from approved specs or architecture decisions; they are listed in § Differences from the approved design and must be raised, not silently adopted (.claude/rules/source-trust.md). Update 2 Oct 2026: checked with Shyamin afterwards — D-1 is not a difference, D-2 is, and D-3 is still open (§ Confirmed with Shyamin after the call).


Summary

  • Region: the estate moves from us-west-2 to us-east-1 this week, because Bedrock AgentCore Web Search is not offered in us-west-2.
  • Accounts: Empire Teas’ main account is the management account; a dev member account exists under it; production comes later.
  • Terraform: one set of shared modules, with each environment adding only its own settings (“overlays”).
  • Access: Tharinda and Dinil already have write access to forge-et-iaac and may write Terraform, with Shyamin reviewing before merge. A read-only IAM user (console and CLI) on the dev account is coming for Tharinda, Dinil and Claude.
  • GuardDuty: Malware Protection for S3 (the file scan) goes ahead. Account-wide threat detection was not decided.
  • Database: db.t3.micro for dev. Production sizing is left for later.
  • Pipelines: Tharinda and Dinil write dev deploy workflows for all three app repos, merge them to main, and Shyamin changes what the infrastructure needs.
  • AI models: Shyamin has requested Bedrock access to Claude Sonnet 5.5 and Claude Haiku 4.5; AWS has not granted it yet.
  • DNS: after the call, Empire IT reported a problem delegating crm.empire.lk; it is not resolved (§8).

Decided on the call

1. Region: us-east-1 (checklist #1)

  • forge-et-iaac was built in US West (Oregon), us-west-2. It was one of the default choices, not picked for cost.
  • Decision: move to us-east-1. The reason is the lead-research tool: Bedrock AgentCore Web Search, the recommended provider behind the SearchProvider seam (SP-ARCH-004, .forge/design/ai/lead-research-source-options.md), is not offered in us-west-2.
  • Cost did not come up as a reason either way. Shyamin noted that Singapore would be expensive; it was not a candidate.
  • Shyamin makes the change this week. He described it as a one-line change in forge-et-iaac, cheap now because only a network and an empty ECS cluster exist.
  • Still to check: that the Terraform state bucket (et-ops) also ends up in us-east-1 (checklist #1), and that the Bedrock model request (§7) is valid in us-east-1 (§7).
  • No harness change is needed for the region itself: AD-O-9 and the harness documents already say us-east-1.

2. Account structure (checklist #3)

  • Empire Teas’ main AWS account (the one Kavindu Anthony holds) is the management account of an AWS Organization.
  • A dev member account has already been created under it.
  • A production member account will be created later.
  • This matches F38-a FR-1: dedicated project accounts in one Organization, with no project workload in the management account.
  • Not discussed: a separate log-archive account (F38-a FR-3 sends CloudTrail to the management account’s organization-trail bucket when there is none), and which account the temporary uat instance runs in (F38 OQ-A).

3. Terraform layout: shared modules plus per-environment settings (checklist #4)

  • Shyamin described the Terraform as one set of shared modules, with each environment (dev, later UAT and production) “overlaid” on top — only the environment’s own settings differ, so adding an environment is cheap.
  • He did not say how the overlays are laid out in the repository (for example a folder per environment, or a variables file per environment).
  • This matches F38 FR-8 (one set of Terraform modules, parameterised by environment) and the temporary-UAT commitment in AD #23.
  • ECR (the container registry): Dinil does not recall how ECR fits into this — whether each environment has its own registry or one registry is shared. The harness design (architecture.md § CI/CD Pipelines) builds the backend image once and deploys the same image to every environment, which works with either. To ask Shyamin.

4. Access (checklist #12, #13, #14)

The forge-et-iaac repository

  • Tharinda and Dinil already have write access.
  • They may write Terraform themselves, not only review it.
  • Shyamin reviews every change before it merges.
  • Not discussed: access for Biseni Perera and Nanndish Satgunarajah, and the repository suggestions in checklist #14 (CODEOWNERS, terraform plan on each PR with apply after merge, branch protection and the branch name, a .gitignore for state and plan files, the committed .terraform.lock.hcl, GitHub OIDC, actions/checkout@v4, a README, a team Owner tag).

The AWS account

  • A read-only IAM user is being set up for Tharinda, Dinil and Claude, so they can look at resources and debug without being able to change anything. Changes go through Terraform PRs in forge-et-iaac (checklist #13).
  • It covers the dev account only.
  • It is an IAM user with a console login and access keys for the AWS CLI. Claude would use the CLI keys.
  • No date was given.
  • IAM Identity Center (single sign-on with MFA) did not come up. This differs from F38-a FR-2 — see § Differences.
  • The access keys are a credential: they stay in each person’s local AWS CLI configuration, never in a repository, a chat or an email (environments-and-identity.md § Secrets boundary).

5. GuardDuty (checklist #8)

  • Decision: turn on Malware Protection for S3 (the scan of uploaded files) on the attachments bucket.
  • Price as discussed: about USD 1 per GB of uploaded files scanned. The CRM is expected to upload only a few GB a month, so this is a few dollars a month at most. The checklist had estimated under USD 1 a month. Check the exact per-GB and per-object rates for us-east-1 on the AWS pricing page before quoting a figure to Empire, and compare with .forge/client/sow-cost-annexe.md, which carries GuardDuty as one line.
  • This keeps the built scanner: the backend’s cloud profile refuses to start without a real FileScanProvider (register A-1), and the one built is GuardDuty’s (F38-c).
  • GuardDuty’s account-wide threat detection was not decided. Only the file scan was agreed. F38-a FR-4 requires threat detection — see § Differences.

6. Database size (checklist #5, #17)

  • Dev: db.t3.micro, as Shyamin recommended for this project’s load.
  • The cost annexe assumes db.t4g.micro for dev. The two classes are close in price and size; the annexe line is updated if the dev bill differs noticeably.
  • Production: left for later. A larger instance class and storage autoscaling were discussed; both cost more. The cost annexe assumes db.t4g.small, single-AZ, for production, and a Multi-AZ standby is a priced add-on Empire has to choose (checklist #17, architecture.md § Deployment Topology › Data). Multi-AZ did not come up.

7. Bedrock models (checklist #10)

  • Shyamin has requested access to Claude Sonnet 5.5 and Claude Haiku 4.5 on Bedrock. This matches ai-provider-strategy.md §4b (revised 2026-09-30).
  • Status: waiting on AWS.
  • The request was made in the dev account in us-west-2, before the region move. Check whether it carries over to us-east-1 or must be made again there. Bedrock quotas are set per region.
  • Still open from checklist #10: Anthropic’s one-time first-use form (F38-a says it is submitted once from the management account); the backend task role’s permission for both inference profiles and the Marketplace subscription permissions; the quota for each model; no policy that blocks other AWS regions, because Sonnet 5.5 uses the global inference profile; and the delivery team’s address as an account contact so that model-retirement notices reach us.

8. DNS for crm.empire.lk (checklist #2)

On the call. Shyamin explained how the DNS was set up: a hosted zone for crm.empire.lk was created in Route 53, and its name-server (NS) records were sent to Empire IT, so that Empire can hand the crm subdomain to AWS.

After the call, by message.

  • Kavindu Anthony (Empire IT) reported a problem creating the subdomain at the .lk registry (NIC.lk), and asked whether empireteas.com could be used instead of empire.lk.
  • The team passed this to Shyamin. His answer, in summary:
    • empire.lk is registered with NIC.lk, but its DNS is served by Cloudflare. So the NS records for crm must be added in Cloudflare’s DNS, not at NIC.lk.
    • empireteas.com is in the same position; only the registrar is different.
    • If NIC.lk pointed the domain’s name servers back to its own DNS to create the subdomain, empire.lk would lose its Cloudflare setup, which would cause a larger problem.
    • He pointed to the public DNS report for the domain, https://intodns.com/empire.lk, as evidence.
  • Tharinda has passed Shyamin’s answer to Kavindu.
  • Status: not resolved. Waiting for Kavindu’s reply.

Why the domain matters. Changing from crm.empire.lk to an empireteas.com address is not a DNS-only change. The eight Entra reply URLs Kavindu registered on 24 September are on api.*.crm.empire.lk hosts (environments-and-identity.md §2.2), and the XSRF-TOKEN cookie domain per environment (AD #33) and every environment address in §1 assume crm.empire.lk. A domain change would mean registering the reply URLs again and changing those settings. Adding the NS records in Cloudflare keeps everything as it is.

9. Deploy pipelines (checklist #11)

  • Who: Tharinda and Dinil write the GitHub Actions deploy workflows.
  • Which repos: all three app repos — forge-et-be, forge-et-fe and forge-et-pwa.
  • Which environment: dev only for now. UAT and production come later.
  • How Shyamin is involved: the workflows are written and merged to main, and Shyamin then changes whatever the infrastructure needs.
  • Backend steps Shyamin described, in order:
    1. build the backend’s Docker image;
    2. push the image to ECR;
    3. register a new revision of the ECS task definition that points at the new image, and update the ECS service so that the running tasks are replaced with it.
  • Frontend steps (from architecture.md § CI/CD Pipelines, not discussed in detail): Vite build with the dev settings, copy the bundle to the S3 bucket, then clear the CloudFront cache.
  • How GitHub Actions signs in to AWS: AWS access keys stored as GitHub secrets. GitHub OIDC did not come up. This differs from the architecture — see § Differences.
  • No date was given.
  • This is F38-d (.forge/specs/F38-d-pipelines-first-deploy-spec.md, still a draft stub). The F38 parent spec allows the workflow files to be written before the AWS estate exists, as a lead-approved CI change for each workflow PR (.claude/CLAUDE.md § Boundaries).
  • Not discussed: what the workflows do before the dev estate exists. A workflow that deploys on every merge to main fails on every merge until ECR, ECS, the buckets and the distributions are there. One option is a manual trigger only (workflow_dispatch) until dev is ready, then switch on deploy-on-merge. To agree with Tharinda and Shyamin before the first workflow merges.
  • Needed from Shyamin to finish them: the ECR repository name or names; the ECS cluster and service names; the S3 bucket names and CloudFront distribution IDs for both frontends; the credentials GitHub Actions uses; and the per-environment values to hold as GitHub environment variables.

Differences from the approved design

These were agreed on the call, or are the effect of it, and differ from an approved spec or architecture decision. Each needs a decision — keep the design and tell Shyamin, or accept the change and revise the spec or decision with a revision entry. Owner: Tharinda (lead).

# Agreed on the call Approved design Where
D-1 People get IAM users with access keys (read-only, dev) People sign in through IAM Identity Center with MFA and time-limited roles; no long-lived access keys for any person F38 FR-2, F38-a FR-2; architecture.md security baseline (IAM Identity Center)
D-2 GitHub Actions uses AWS access keys stored as GitHub secrets GitHub Actions assumes an IAM role through GitHub OIDC; no static keys architecture.md § CI/CD Pipelines cross-cutting rules (“AWS access by GitHub OIDC → IAM role, no static keys”)
D-3 Only GuardDuty Malware Protection for S3 is turned on GuardDuty threat detection is enabled in each account F38 FR-2, F38-a FR-4; the proposal’s security baseline

Smaller points, recorded but not needing a decision now: dev runs db.t3.micro where the cost annexe assumed db.t4g.micro (§6); the deploy workflows are merged before the estate exists (§9).

Confirmed with Shyamin after the call (2 October 2026)

Because this record was written from memory, Dinil checked D-1 to D-3 with Shyamin by message on 2 October 2026. The table above is kept as first written; this is what the check found.

  • D-1 — not a difference. Shyamin’s answer: each person gets an SSO user with MFA, and uses aws configure sso for the CLI. That is IAM Identity Center with MFA and short-lived credentials, as F38 FR-2 and F38-a FR-2 require. On the call he showed the AWS access portal (a /start address), which is Identity Center’s sign-in page; “IAM user” in §4 was a loose description, and no person gets a long-lived access key. Claude uses the developer’s own SSO session in the developer’s terminal, so it needs no credentials of its own. §4’s “IAM user with … access keys for the AWS CLI” is superseded by this answer.
  • D-2 — confirmed as a difference. Shyamin’s answer: GitHub Actions reaches AWS through an IAM service user in each workload account, with its access key held as a GitHub secret. Nobody on the call raised GitHub OIDC; what was seen was the forge-et-iaac workflow’s existing secrets. The delivery team (Dinil) proposes to accept Shyamin’s approach. It changes the approved F38 spec (FR-19, NFR-1) and architecture.md § CI/CD Pipelines, so it follows as a spec revision approved by Tharinda, the F38 owner. The proposal is not affected: it promises “no long-lived keys for people” (§A.4), and a service user is a machine identity.
  • D-3 — still open. Shyamin will check the cost of GuardDuty threat detection and give his view. The proposal includes it (§A.2.2 prices it at $5 a month per account; §A.2.3 and §A.4 list it in the baseline), so turning it off would need Empire to be told.

Not discussed

These checklist items did not come up and remain open:

  • #4 when UAT and production are set up (only the overlay approach was discussed — §3);
  • #6 the frontend hosting (S3 + CloudFront) as a first milestone;
  • #7 the Microsoft sign-in private keys in Secrets Manager, where UAT reads its key from, and key rotation;
  • #9 CloudWatch alarms and the SNS alert address, SES production access, and who at Empire receives the AWS Budgets alerts;
  • #15 which F38 parts Shyamin builds and which the delivery team builds;
  • #16 whether Fri 2 Oct (environment and logins ready) and Fri 16 Oct (environments and pipeline) are realistic;
  • #17 the production choices for Empire (a Multi-AZ database, cross-region backup copies, a web firewall) — only the database size was touched on (§6);
  • #18 a monthly cost estimate for dev against the cost annexe;
  • #19 anything else Shyamin needs from us.

Actions

Action Owner When
Move forge-et-iaac to us-east-1, including the et-ops state bucket if it has to move Shyamin Ayesh This week
Create the read-only IAM users (console and CLI) on the dev account for Tharinda, Dinil and Claude Shyamin Ayesh No date given
Follow up the Bedrock model request with AWS, and confirm it is valid in us-east-1 Shyamin Ayesh —
Turn on GuardDuty Malware Protection for S3 on the attachments bucket, to the contract in F38-c FR-10 Shyamin Ayesh With the dev estate (F38-b)
Create the dev database as db.t3.micro Shyamin Ayesh With the dev estate (F38-b)
Write the dev deploy workflows for forge-et-be, forge-et-fe and forge-et-pwa, and merge them to main for Shyamin to adjust Tharinda Rajapaksha, Dinil Gamage No date given
Decide D-1, D-2 and D-3 and tell Shyamin Tharinda Rajapaksha Before the read-only users and the workflows are built. 2 Oct: D-1 resolved, D-2 proposed for acceptance, D-3 waiting on Shyamin (§ Confirmed with Shyamin after the call)
Agree how the deploy workflows behave before the dev estate exists (§9) Tharinda Rajapaksha, Dinil Gamage, with Shyamin Ayesh Before the first workflow merges
Get Kavindu’s reply on adding the crm NS records in Cloudflare Tharinda Rajapaksha Waiting on Kavindu
Send Shyamin the open questions: ECR layout (§3), the values the workflows need (§9), and the not-discussed checklist items Dinil Gamage / Tharinda Rajapaksha —
Check the GuardDuty Malware Protection rates for us-east-1 and record them in the cost annexe Owner to set —

Harness updates that follow (not yet made)

  • .forge/design/environments-and-identity.md — §1 (the DNS delegation problem and the Cloudflare answer), §3 (the account structure), §4 (Tharinda’s and Dinil’s write access, Shyamin reviewing), §5 (the GuardDuty row: Malware Protection decided, threat detection open; the AWS environments row: the region move and the read-only users).
  • .forge/cloud-gated-validation.md — a dated ## Log line for this call; the A-1 row (the scanner stays).
  • The F38-a and F38-b specs — the account structure, the dev database class and, once decided, D-1 to D-3.
  • .claude/CLAUDE.md § Stack still names “Claude Sonnet 5”; the AI strategy moved to Sonnet 5.5 on 2026-09-30. A separate fix, for the lead to agree.

Generated from .forge/discovery/meeting-notes/2026-10-01-aws-setup-shyamin.md by .claude/lib/render-docs-html.js. Markdown is canonical.