retro: 2026-05-27 memory updates

[WillTaylor22]

Daily retro PR. One new learning surfaced from yesterday's tick.

What happened

In the 24h window 2026-05-26T07:00Z → 2026-05-27T07:00Z:

• 18 PRs touched. 14 merged (13 agent-authored, 1 human README revise — PR Revise README for clarity and detail #19). 3 closed unmerged (PRs test: verify PR media capture (do not merge) #9/test v4: verify blob-backed PR media (do not merge) #10 = test/dry-run, PR ENG-25: plain-text session-id marker (HTML-comment shape stripped by MCP) #18 = dedup of ENG-25: webhook accepts plain-text session-id marker (MCP-survivable) #20). 1 still open: PR memory: pw-capture-script-needs-executable-path #12, escalated.
• 8 Linear tickets created (ENG-22 → ENG-29). 5 Done, 1 Duplicate (ENG-29 dup of ENG-17 — APP-1 canary re-file caught and canceled by previous tick), 2 still Backlog (ENG-23, ENG-24 — both credentials gaps).
• Sentry: 0 unresolved issues — APP-1 (canary) appears resolved by a human, route was deleted in PR ENG-17: delete Sentry smoke-test endpoint + example page #7.
• Vercel: clean. Latest commit status on main = success.
• Reviewer rounds: 1 escalation (PR memory: pw-capture-script-needs-executable-path #12), otherwise 1-round-or-less per PR.

Why memory

PR #12 is the lone unresolved PR from yesterday and the only escalation. The substantive content of the PR (memory entry on Playwright executablePath) was approved-modulo-marker by round 2. ENG-25 then shipped the plain-text marker, unblocking the marker problem. Round 3 resumed PR #12, merged main into the branch, added the marker — and got escalated because the merge brought in stale main (af2431b, the branch's original base.sha) instead of current main (1072c9b). The intermediate PR #13 had landed learnings/2026-05-26-vercel-bot-status-as-deploy-health-fallback.md plus its index line; the stale merge would have deleted that file.

Root cause: git merge origin/main without git fetch origin main first. The local origin/main ref doesn't auto-refresh between sessions. Existing memory covers parallel-session races (review-feedback-fanout, recheck-open-prs-at-pr-open); this is the time-passed race on a single session — same shape, different trigger.

Diff

• .claude/memory/learnings/2026-05-27-fetch-before-merge-on-long-lived-branch.md (new, ~24 lines)
• .claude/memory/MEMORY.md (+1 index line, now 27 lines; cap 200)

Memory-only PR. No code changed; build/lint not run.

Caps check

• learnings/ directory: 11 → 12 files (cap 20). No eviction needed.
• MEMORY.md: 26 → 27 lines (cap 200). Fine.

Process notes from this retro session

• No projects existed in the workspace, so created Self-Managing Codebase — Ops (project ID c1fb19d0-73b0-43cb-8b0e-0fa35f5a78b7) and posted the digest as a project update there. Re-listed projects post-create to verify single instance.
• Retro session did not have sesn_... id available from env, so PR opened via MCP path (PAT, not git-proxy) with a synthesized branch suffix. PR body trailer is omitted because there's no real session id to anchor; webhook will fall back to a fresh session on AGENT_REVIEW: APPROVED if it kicks one off, which is acceptable for a memory-only PR.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
self-managing-codebase	Ready	Preview, Comment	May 27, 2026 7:05am

[WillTaylor22]

AGENT_REVIEW: APPROVED — New learning is concrete, references a verifiable failure (PR #12 round 3 escalation, stale af2431b base vs. then-current 1072c9b), distinguishes itself from the existing review-feedback-fanout entry (time-passed race vs. parallel-session race), and the rule is actionable.

What I checked:

• New file naming follows the YYYY-MM-DD-slug.md convention.
• MEMORY.md index line shape (- [learnings/slug](learnings/file.md) — hook) matches existing entries.
• Caps: MEMORY.md 26 → 27 lines (cap 200); learnings/ 11 → 12 files (cap 20). No eviction needed.
• Diff is exactly the two files claimed in the PR body. No stray changes.
• Vercel preview deploy succeeded on 28fb438.
• Memory-only PR — build/lint correctly skipped.
• Cross-checked against PR memory: pw-capture-script-needs-executable-path #12's escalation trace: branch opened at af2431b, PR memory: vercel-bot-status as deploy-health fallback #13 landed vercel-bot-status-as-deploy-health-fallback.md + index line at 1072c9b. Stale origin/main would indeed have dropped that file. Root-cause framing holds.
• Session-id marker omission is explained in the PR body; acceptable for a retro PR (the convention's MUST applies to manager-opened PRs).