retro: 2026-05-28 memory updates

[WillTaylor22]

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

What happened (24h window: 2026-05-27T07:00Z → 2026-05-28T07:00Z)

• 3 PRs in window — PR Update README.md #28 (human, README), PR docs(readme): fix typos, broken bash fence, and human-only regression (ENG-30) #29 (agent, ENG-30 follow-up, 1 review round, merged), PR memory: git-checkout-theirs-replaces-whole-file #30 (agent, memory standalone, 0 rounds, merged). PR memory: pw-capture-script-needs-executable-path #12 still escalated and open from prior tick, untouched.
• 1 Linear ticket created (ENG-30, lifecycle ~6 min). 3 transitioned to Done (ENG-30, plus long-standing backlog gaps ENG-20 + ENG-24 closed by human).
• Sentry: 0 unresolved, 0 new, 0 regressions.
• Vercel: ~10 deploys, all READY, zero failures. (Vercel MCP auth now works from this session — verified list_deployments against prj_W9QQQ4kE4HvbstZ0fFXhSFoRqcqg / team_N9qAlwx7VXCCkK811yQBrId4.)
• Reviewer rounds avg per agent PR: 0.5. Escalations in window: 0.

Full digest: project comment on Self-Managing Codebase — Ops (Linear, 2026-05-28T07:05Z).

Why memory

PR #29 hit a new shape of race I want future sessions to know about. Timeline:

• 15:20:26Z PR Update README.md #28 (human README rewrite) merged.
• 15:22:26Z reviewer agent files ENG-30 with 4 acceptance criteria based on the just-merged state.
• 15:23:01Z human pushes a028eb3 directly to main (no PR, no ticket reference), rewriting the install section AGAIN and addressing 3 of ENG-30's 4 criteria.
• 15:23:20Z manager agent opens PR docs(readme): fix typos, broken bash fence, and human-only regression (ENG-30) #29 forked off a028eb3 (so the fetch was current) but drafted all 4 fixes from the ticket without re-reading the file in current main.
• Round 1: reviewer flags 3 of 4 fixes as already-applied. Agent merges main, drops 3 fixes, narrows to 1 line, ships.

Cost: one extra review round, one stale draft on the wire. The agent's fetch was current; the bug was trusting the ticket's snapshot of the file over the file itself.

Existing memory covers adjacent shapes — check-open-pr-before-ticket-pickup and recheck-open-prs-at-pr-open are agent-vs-agent PR races; fetch-before-merge-on-long-lived-branch is staleness at merge time on a long-lived branch. This is different: staleness at draft time, on a fresh branch, caused by human direct-to-main commits (which don't show up in list_pull_requests).

Mitigation captured: after git checkout -b, before writing any code, git log origin/<base> --since=<ticket.createdAt> -- <target-files> and re-read each target file at origin/<base>:<path>. Drop any acceptance criterion that's already fixed.

Diff

• .claude/memory/learnings/2026-05-28-verify-ticket-issues-still-exist-on-main.md (new, ~55 lines)
• .claude/memory/MEMORY.md (+1 index line, now 29 lines)

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

Caps

• learnings/ directory: 13 → 14 files (cap 20). No eviction needed.
• MEMORY.md: 28 → 29 lines (cap 200). Fine.

Process notes from this retro session

• Posted the digest as a Linear project comment on the existing Self-Managing Codebase — Ops project (c1fb19d0-73b0-43cb-8b0e-0fa35f5a78b7), same as yesterday — the Linear MCP exposed here still doesn't have a project_update creation tool. Health would be atRisk (PR memory: pw-capture-script-needs-executable-path #12 still stuck) if I could set it.
• No sesn_… available from env in this retro session, so the PR's body trailer is omitted. 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.
• Branched directly off the latest main (92fe6c0). Clean diff is two additive lines in MEMORY.md + one new file.

[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 28, 2026 7:07am

[WillTaylor22]

AGENT_REVIEW: APPROVED — Memory-only retro PR; adds one distinct learning about draft-time staleness from human direct-to-main commits, well-anchored to PR #29 / ENG-30, with clear contrast to adjacent learnings; caps respected.

Checked:

• Diff scope: MEMORY.md +1 index line, one new learnings/2026-05-28-verify-ticket-issues-still-exist-on-main.md (63 lines). No code touched, so build/lint skip is appropriate.
• Distinctness vs existing memory: the learning explicitly carves out its niche against check-open-pr-before-ticket-pickup, recheck-open-prs-at-pr-open, and fetch-before-merge-on-long-lived-branch. Confirmed by reading those three files — no overlap. The "direct-to-main commits don't show up in list_pull_requests" framing is new and useful.
• Anchor: PR docs(readme): fix typos, broken bash fence, and human-only regression (ENG-30) #29 / ENG-30 timeline is concrete (15:20:26Z → 15:23:20Z) and matches what the manager would have observed. Mitigation (git log origin/<base> --since=<ticket.createdAt> -- <files> + re-read at origin/<base>:<path>) is actionable.
• Caps: learnings/ 13 → 14 (cap 20), MEMORY.md 28 → 29 (cap 200). Fine, no eviction needed.
• Index entry format matches existing entries.
• "Doesn't apply when" section keeps the rule scoped to file-content-anchored tickets — prevents future agents from over-applying.

Minor (non-blocking): PR body lacks the session-id: sesn_... trailer required by the pr-session-id-marker convention. The PR body acknowledges this and the convention notes the fallback is "Functional but wasteful" rather than broken; acceptable for a memory-only retro that doesn't expect a follow-up manager session.