docs(retrospective): 2026-05-21..25 multi-day sprint retrospective#898
Conversation
Capture the 5-day sprint (35 PR, 6 release, 1 production CVE close) as durable record before starting the next session. Codex consultation was the foundation; the primary takeaway is: > 最大の問題は速度ではなく、停止条件の弱さ。次は「作る量」 > より「どこから保留に戻すか」を先に設計すべき。 Sections - Numerical summary - What worked (5 items: multi-perspective review, silent-skip thread, memory persistence, gh-api recovery, dogfood detection) - What didn't work (6 items: overproduction past Codex stop, eval-less skill increase, manual release-please kick x4, AGENTS.md pollution, reviewer mis-suggestions, single-shot multi-perspective) - Surprises (npm pack audit findings, CLI boundary, fs-loss recovery) - 5 pivot points (timeline of where direction should have changed) - Codex 100-char summary on the sprint - Improvement priorities split into Skills / Agents / Workflow with P0/P1/P2 ranking - Stop-on-stop-judge operational rule with Continue / Consult / Park buckets, plus post-hoc classification of this sprint's 35 PRs (Continue: A2-fix and CVE-class items; Consult: skill registry decisions; Park: ~13 PRs of docs-about-docs / marginal polish) - Next sprint opening checklist No code change. Foundation for two follow-up PRs (S3 registry section, W1+W5 release-please kick automation). Refs: Codex consultations during the 5-day sprint (PR #860 78→91, PR #864 dogfood, PR #877 riskAssessment, post-stop 13 PRs). Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
No changes to review compared to main. |
There was a problem hiding this comment.
Code Review
This pull request introduces a retrospective document for the multi-day sprint from 2026-05-21 to 2026-05-25, detailing metrics, successes, failures, and future improvement priorities. The review feedback identifies inconsistencies in the PR reclassification table, specifically regarding overlapping PR ranges and inconsistent punctuation usage.
| | Continue (やるべきだった) | #860 (concept), #861 (B'), #864 (A2-1), #865 (A2-fix-1), #869 (A2-fix-2), #871 (A2-fix-3), #877 (A2-fix-4), #888 (CVE), #896 (risk-map example), #897 (smoke runbook) | | ||
| | Consult (相談を挟むべきだった) | #862-#874 (release-please の手動 kick が繰り返された)、#873 / #875 / #881 / #884 (community skill 増産) | | ||
| | Park (本来後送りで良かった) | #866 (docs D2+D3 — 統合可能だった), #872 (docs cleanup), #887 (agent-learnings entry), #889 (README drift)、#883 (eng-practices mapping) | |
There was a problem hiding this comment.
「このスプリントの再分類」テーブルにおいて、PR番号の指定方法と区切り文字の整合性に改善の余地があります。\n\n1. 分類の重複と範囲指定: Consult 行の #862-#874 という範囲には、Continue 行にある #864, #865, #869, #871 や Park 行にある #866, #872 が含まれており、分類が重複しています。また、#873 が範囲内と個別の両方でリストされています。事後評価の正確性を期すため、範囲指定を避け、各バケットに属する PR を明確にリストアップすることをお勧めします。\n2. 区切り文字の統一: カンマ(,)と読点(、)が混在しています(例:103行目の #889 (README drift)、#883)。文書内の他のリスト(101行目や110行目など)に合わせて、カンマに統一することをお勧めします。
PlanGate Review
PlanGate review decision: pass
ポリシー: critical=fail / major=fail-if-required (warn) / minor=comment-only / info=skipped — spec: |
#902) Follow-up to multi-perspective review (Self/Codex/Gemini) of PR #898-#900. 3 convergent defects fixed in one PR per Codex's process-hardening grouping. A. release-please-kick.yml: add concurrency guard (group: release-please-kick) - prevents double-click 422 race on PATCH refs B. branch input default: blank → auto-detect via gh pr list - eliminates silent 404 if release-please component changes - applied to both workflow and scripts/release-please-kick.sh C. retrospectives/2026-05-21-25.md: add reopen conditions - 4 explicit triggers (new info / explicit scope / risk down / time+observation) - "対応を自律的に進めて" is re-judgment trigger, not auto-release - allowed-in-Stop list + 3-step release procedure
Direct follow-up to retrospectives/2026-05-21-25.md. Captures 27 PRs / 10 releases / 3 issues closed (#906 / #911 / #868) / 4 issues opened (#906 / #911 / #921 / #929) / 8+ Stop-on-Stop-Judge invocations. Key findings: - Codify-then-validate caught its own gap in same session (#898 → #902) - Multi-perspective review found convergent defects (#902) - Stop-on-Stop-Judge stress-tested and held under 8+ autonomous prompts - GITHUB_TOKEN no-recursion was a same-session defect (#906) that 3-perspective review missed - "Posture, not progress" guard rejected hand-written goldens multiple times - Issue carve-out (#921, #929) enabled graceful epic closure New improvement priorities: - A8: gh-actions-privileges line in Code-gen self-review (catches #906 class) - A9: cached-Codex-judgment heuristic (cap consults on repeated situations) - A10: pre-merge dist rebuild for src/cli.mjs (prevents #917 class) - W7: gh release-ratchet helper (10/27 PRs were releases) Stop-on-Stop-Judge upheld; ~3 PRs of "stop weakness" surface vs ~13 in prior sprint. Rule helped.
Summary
Persist the 5-day sprint retrospective as a durable doc. The sprint shipped 35 PRs / 6 releases / 1 production CVE close, and the most important lesson — per Codex's analysis — is not about velocity, but about stop conditions:
This PR is PR 1 of 3 in the retrospective follow-up:
docs/development/retrospectives/2026-05-21-25.md.skills/registry.yamlcommunity section /recommended: falseflag (decide which community skills are experimental).workflow_dispatch+ supporting script).Doc structure
Key proposals carried by this doc
registry.yamlcommunity section /recommended: falseflagTest plan
No code change. Pure documentation.
🤖 Generated with Claude Code