.github/company/agents/savannah-savings/MEMORY.md

Tacit Knowledge — Savannah Savings (CTO)

How I operate and patterns I've learned.

Organization

• Manager: Coupon Carl (CEO, f2395b62-cb26-4595-b026-d506fde1c2c1)
• Direct reports: Barcode Betty (71f37521-8e62-4d27-bd9c-cfd52b5b3a07), Stockboy Steve (01dfbf79-c93d-4224-a7d9-05b2779e425e), Checkout Charlie (b8b294e3-a12d-4bff-b321-6f020792b21c)
• Deal Dottie (ff0b8079-5823-4c4f-ad40-6a5147246594) — QA role, UAT tester
• Rollback Rhonda — TERMINATED as of 2026-04-01. Cannot assign tasks.
• Handoff chain: Engineer → QA (Checkout Charlie) → CTO (me). UAT run by Deal Dottie.
• Markdown Martha (9becc57b) — CMO, idle. Not a direct report.

Memory System Notes

• Layer 1 (PARA): $AGENT_HOME/life/ — entity knowledge graph
• Layer 2 (Daily Notes): $AGENT_HOME/memory/YYYY-MM-DD.md
• Layer 3 (Tacit): this file ($AGENT_HOME/MEMORY.md)
• Memory bootstrapped 2026-03-28 by CEO (CAR-64)

UAT Ownership Model (CEO Directive, 2026-03-30; updated 2026-04-02)

• CTO owns ALL UAT knowledge in a playbook at $AGENT_HOME/playbooks/uat-playbook.md
• Rollback Rhonda terminated 2026-04-01. Deal Dottie (ff0b8079-5823-4c4f-ad40-6a5147246594) is the active UAT tester.
• HEARTBEAT.md still references Rhonda for UAT assignment — route UAT tasks to Deal Dottie instead.
• When UAT work arrives, CTO decomposes into atomic tasks — one URL, one action, one verification per task
• Playbook matures continuously after every UAT cycle
• Plan: CAR-198

GitHub App Review Pattern

• CTO GitHub App (cartsnitch-cto) CAN submit formal APPROVE reviews via API: gh api repos/{owner}/{repo}/pulls/{n}/reviews -X POST -f event=APPROVE -f body="..."
• Do NOT use gh pr review --comment — that creates COMMENTED state, not APPROVED
• Branch protection requires 2 APPROVED reviews before merge — comment reviews don't count
• QA has a separate app (cartsnitch-qa) that can also submit approvals
• Always submit formal approvals, not just comment-based reviews
• Self-approval blocker: If CTO app pushed commits to a branch, GitHub blocks APPROVE review from that same app ("Can not approve your own pull request"). In this case, submit a COMMENT review noting CTO approval and hand off to CEO with a note about the constraint.

GitHub CLI Multi-Account Auth

• Multiple gh accounts exist in /paperclip/.config/gh/hosts.yml (CEO, CTO, engineer, QA bots + groombook bots)
• gh auth login --with-token adds/updates account but does NOT switch active account
• After login, MUST run gh auth switch --user cartsnitch-cto[bot] to make CTO account active
• Check with gh auth status before any GitHub operations
• Active account determines which bot identity is used for API calls

Task Decomposition Lessons

• Never delegate "investigate and fix" — investigate first, then delegate the specific fix
• MiniMax M2.7 agents cannot follow 500-line instruction docs — keep under 120 lines
• Each delegated task: 1 file change, exact repo/branch/path, step-by-step, no ambiguity
• Include full context in every task — don't assume delegate has context from previous tasks

QA Handoff Lessons (2026-04-14)

• Charlie (QA) uses Playwright browser tools — cannot run unit tests or CLI commands
• Engineering task descriptions are NOT QA briefs. Always write a separate QA section when creating tasks that will need QA review.
• QA brief MUST include: (1) PR URL, (2) numbered test steps with explicit PASS criteria per step, (3) On FAIL → reassign to Betty, On PASS → hand off to CTO
• Frame QA steps as code review checks (open file in diff, verify X is present), NOT test execution
• Three tasks bounced (CAR-557, CAR-576, CAR-579) because original descriptions were engineer-oriented with no QA-ready steps
• Best pattern: engineer opens PR and comments with URL → CTO writes QA brief with the PR URL and reassigns to Charlie

Playwright MCP

• Correct Playwright URL for CartSnitch: http://playwright-cartsnitch:8931/sse (NOT http://playwright:8931/mcp)
• Board confirmed 2026-03-31: Playwright was never broken, just misconfigured URL
• Each app gets its own Playwright sidecar with naming convention playwright-<app>

Infra & CI Lessons

• Never use :latest image tags in production k8s manifests — containerd caches can serve stale digests. Always pin to CalVer or SHA tags.
• The k3s cluster has nodes: dot (control-plane), mindy (worker), wakko (control-plane), yakko (control-plane), buttons (gpu/worker).
• When debugging pod CrashLoops, always exec into the container and check the actual config files.
• GitHub App token script needs chmod +x before each session — permissions don't persist.
• Merge-conflict silent drops: When GitHub reports a PR as "merged", always verify with git merge-base --is-ancestor <sha> main. Merge conflicts can silently drop changes. PR #76's auth session fix was lost this way during PR #61 merge — the merge resolved auth.ts without the session mapping. After any merge, verify target file content on main matches expectations.
• Foreign commit in infra PR #105 (2026-04-02): "Flea Flicker" (flea-flicker@groombook.io) committed to cartsnitch/infra. Not a CartSnitch identity — groombook is a different org. Check all PRs for non-CartSnitch commit authors going forward. Investigate how cross-org push access occurred.
• Infra repo IS Flux GitOps: cartsnitch/infra is referenced by the cartsnitch GitRepository in the cluster. It IS the Flux GitOps source of truth for CartSnitch. Do NOT escalate Flux CRD/RBAC changes to the bootstrap repo (cpfarhood/kubernetes). Add Flux resources (Kustomization CRDs, RBAC) directly to the infra repo. Board was clear (2026-04-03): "the infra repo is flux gitops."
• Flux architecture: Bootstrap (flux-system/cartsnitch) creates GitRepository + Kustomizations in cartsnitch namespace. Existing Kustomizations: cartsnitch-dev (path ./apps/overlays/dev), cartsnitch-prod (path ./apps/overlays/prod). All use cartsnitch-flux SA. New environment overlays need: a Flux Kustomization CRD + RBAC in the infra repo.