[automation] Repository drift report

[github-actions]

Canonical delivery-state reconciliation

Generated: 2026-08-03T14:05:42.502Z

• Open PRs: 67
• Total remote branches: 319
• Ready PRs without exactly one canonical issue: 0
• Issues with competing implementation PRs: 1
• Unattached branches older than 14 days: 164

PRs requiring canonical issue

• None

Competing PRs

• Issue brace-expansion override floors permit OOM-vulnerable versions (^1.1.16 / ^2.1.2 are below the actual backport) #1115: fix(security): raise brace-expansion override floors above the OOM backport #1226, fix(deps): raise brace-expansion override floors to the actual backport (1.1.17 / 2.1.3) #1117

Stale unattached branches

• bolt/fix-batch-query-n1-bottleneck-16664381751066617032 — 065475a, last commit 2026-07-12T17:57:46Z
• bolt-fix-database-optimizer-n1-queries-14843617805558882255 — 265ab89, last commit 2026-07-14T14:41:06Z
• bolt-memoize-segment-row-14076716723355865204 — 3755b95, last commit 2026-07-12T00:10:21Z
• claude/dazzling-edison-95u1m4 — f6f083a, last commit 2026-07-17T09:47:19Z
• claude/dazzling-edison-ymkr67 — f024fd7, last commit 2026-07-17T09:13:23Z
• claude/determined-maxwell-1y1im5 — 5d491a0, last commit 2026-07-18T02:48:59Z
• claude/determined-maxwell-3vgs17 — 9d52047, last commit 2026-07-09T05:59:01Z
• claude/determined-maxwell-3z56dm — 13ce226, last commit 2026-07-07T03:27:39Z
• claude/determined-maxwell-4dxv0e — a395c9c, last commit 2026-07-07T03:21:31Z
• claude/determined-maxwell-4ky0tz — c42776d, last commit 2026-07-14T14:25:40Z
• claude/determined-maxwell-4l1ghl — 1d5ace8, last commit 2026-07-17T02:22:03Z
• claude/determined-maxwell-4nb64a — dddb4db, last commit 2026-07-02T07:02:52Z
• claude/determined-maxwell-4thx4d — 1d5aa56, last commit 2026-07-10T18:46:42Z
• claude/determined-maxwell-5c11cx — 1e32f53, last commit 2026-07-07T04:48:27Z
• claude/determined-maxwell-06popt — 70f6db6, last commit 2026-06-29T22:25:32Z
• claude/determined-maxwell-6rg53d — 78331cd, last commit 2026-07-09T06:45:21Z
• claude/determined-maxwell-7vrh8v — 850a796, last commit 2026-06-28T23:07:41Z
• claude/determined-maxwell-7wu0ng — cfe7cd8, last commit 2026-07-17T01:40:51Z
• claude/determined-maxwell-8p5n7u — d9d12cd, last commit 2026-07-17T02:36:20Z
• claude/determined-maxwell-9erquc — e3c0238, last commit 2026-07-18T03:37:17Z
• claude/determined-maxwell-9ld7rc — d39c63d, last commit 2026-07-15T03:45:23Z
• claude/determined-maxwell-9qcq2r — bb98a67, last commit 2026-07-06T23:43:16Z
• claude/determined-maxwell-9qx6ue — 4f5e59f, last commit 2026-07-07T03:22:01Z
• claude/determined-maxwell-9wecpk — 16be441, last commit 2026-07-12T02:10:48Z
• claude/determined-maxwell-11jydn — 471a705, last commit 2026-07-12T02:19:35Z
• claude/determined-maxwell-14jgh5 — b3bad4c, last commit 2026-07-14T13:49:01Z
• claude/determined-maxwell-61nzcv — 5e90d0a, last commit 2026-07-12T04:09:08Z
• claude/determined-maxwell-81rne1 — 17958ac, last commit 2026-07-07T03:07:30Z
• claude/determined-maxwell-5593h9 — baf93b2, last commit 2026-07-02T21:38:19Z
• claude/determined-maxwell-92049h — e6ad8b6, last commit 2026-07-07T00:20:43Z
• claude/determined-maxwell-a6apnm — 18cb319, last commit 2026-07-03T20:50:31Z
• claude/determined-maxwell-aacju9 — c4cd028, last commit 2026-07-04T05:08:30Z
• claude/determined-maxwell-ayuxew — 365f94b, last commit 2026-07-14T17:18:08Z
• claude/determined-maxwell-bfoiqr — 58b1284, last commit 2026-07-12T02:47:55Z
• claude/determined-maxwell-bloicm — aee31b4, last commit 2026-07-17T04:03:52Z
• claude/determined-maxwell-bmfsgm — 32d7ef1, last commit 2026-07-02T07:02:48Z
• claude/determined-maxwell-bowgqi — 98d16fb, last commit 2026-07-15T02:43:52Z
• claude/determined-maxwell-brk70z — c9adf78, last commit 2026-07-12T02:48:28Z
• claude/determined-maxwell-cmkdxq — ddf6ac0, last commit 2026-07-12T04:07:30Z
• claude/determined-maxwell-dczolv — 42e9ff9, last commit 2026-07-17T03:59:38Z
• claude/determined-maxwell-e0iuif — df2daa2, last commit 2026-07-07T00:16:44Z
• claude/determined-maxwell-e1ur8w — b9bb83a, last commit 2026-07-12T04:05:59Z
• claude/determined-maxwell-e8uv5v — 01d1e4f, last commit 2026-07-07T00:41:03Z
• claude/determined-maxwell-e9hha1 — 2d027d8, last commit 2026-07-17T04:08:45Z
• claude/determined-maxwell-en9kru — 62cdc86, last commit 2026-07-06T23:38:34Z
• claude/determined-maxwell-ey01vy — c290856, last commit 2026-07-07T03:23:10Z
• claude/determined-maxwell-f02xls — 0d8d387, last commit 2026-07-12T04:07:11Z
• claude/determined-maxwell-fuumwi — e18d1e4, last commit 2026-07-06T23:38:56Z
• claude/determined-maxwell-g31z0u — f3226e7, last commit 2026-07-18T04:30:07Z
• claude/determined-maxwell-gfxhok — 25315d9, last commit 2026-07-07T03:21:53Z
• claude/determined-maxwell-gvupbt — cb56b19, last commit 2026-07-12T04:02:43Z
• claude/determined-maxwell-h7gz7l — dc3b09f, last commit 2026-07-07T04:56:41Z
• claude/determined-maxwell-h904ct — 660d3dc, last commit 2026-07-07T03:26:13Z
• claude/determined-maxwell-hli561 — afa02ac, last commit 2026-07-03T23:31:06Z
• claude/determined-maxwell-hu0rea — 0e22d42, last commit 2026-07-12T02:25:36Z
• claude/determined-maxwell-iox2ok — 08c15da, last commit 2026-07-17T02:43:42Z
• claude/determined-maxwell-ip1f59 — cbe919a, last commit 2026-07-06T23:46:27Z
• claude/determined-maxwell-jm8plv — 2caf995, last commit 2026-07-07T03:22:43Z
• claude/determined-maxwell-kaypem — cc4df87, last commit 2026-07-08T22:25:02Z
• claude/determined-maxwell-kdbbzs — bd1c5eb, last commit 2026-07-03T20:48:49Z
• claude/determined-maxwell-kpfd1c — de73686, last commit 2026-07-12T02:15:32Z
• claude/determined-maxwell-l4kzow — 318a59d, last commit 2026-07-12T05:17:22Z
• claude/determined-maxwell-ldvlx6 — 8b39f2a, last commit 2026-07-03T01:49:28Z
• claude/determined-maxwell-ldye72 — 8fbee60, last commit 2026-07-15T03:03:22Z
• claude/determined-maxwell-lxth4b — 06fa2ad, last commit 2026-07-17T04:14:03Z
• claude/determined-maxwell-mij7u9 — 89e4a6d, last commit 2026-07-07T08:15:24Z
• claude/determined-maxwell-msnh15 — 4855637, last commit 2026-07-04T04:31:11Z
• claude/determined-maxwell-n1dpmx — b48b6ce, last commit 2026-07-15T02:32:01Z
• claude/determined-maxwell-nb1zz3 — 609985b, last commit 2026-06-29T14:43:08Z
• claude/determined-maxwell-nbnstz — 410c9be, last commit 2026-07-07T04:42:39Z
• claude/determined-maxwell-ostftd — b41cff0, last commit 2026-07-17T02:07:00Z
• claude/determined-maxwell-p4s1gj — 7503c72, last commit 2026-07-11T02:38:40Z
• claude/determined-maxwell-p6cak6 — 39d0dfa, last commit 2026-07-11T02:58:36Z
• claude/determined-maxwell-p9v5qd — 8569484, last commit 2026-07-17T09:24:19Z
• claude/determined-maxwell-pcaeli — d0a6ac6, last commit 2026-07-06T23:38:16Z
• claude/determined-maxwell-pduakf — f6da8ac, last commit 2026-07-15T04:06:16Z
• claude/determined-maxwell-pzoqq8 — fb9a5ee, last commit 2026-07-07T03:24:08Z
• claude/determined-maxwell-q4p6rk — 2ff61c3, last commit 2026-07-04T05:08:03Z
• claude/determined-maxwell-r838un — 802d531, last commit 2026-07-06T23:37:55Z
• claude/determined-maxwell-rn9d53 — b9efe61, last commit 2026-07-15T03:40:50Z
• claude/determined-maxwell-s49460 — f6e0653, last commit 2026-07-15T03:11:37Z
• claude/determined-maxwell-shjn00 — c86fdaa, last commit 2026-07-06T23:44:06Z
• claude/determined-maxwell-t6hfkk — 826d75c, last commit 2026-07-12T03:52:11Z
• claude/determined-maxwell-tbufwf — 3b3cbbe, last commit 2026-07-12T03:57:35Z
• claude/determined-maxwell-u7kfre — 1dcbc16, last commit 2026-07-14T17:11:30Z
• claude/determined-maxwell-ubh3hf — a693c78, last commit 2026-07-06T23:37:23Z
• claude/determined-maxwell-ubv6b6 — 8c47297, last commit 2026-07-07T03:37:24Z
• claude/determined-maxwell-udy1ne — 53fa3fc, last commit 2026-07-17T02:44:01Z
• claude/determined-maxwell-uybuwc — 53fa3fc, last commit 2026-07-17T02:44:01Z
• claude/determined-maxwell-vkacey — 9f8af98, last commit 2026-07-07T01:23:47Z
• claude/determined-maxwell-vupf5j — 52ea9dc, last commit 2026-07-07T03:22:17Z
• claude/determined-maxwell-w20w27 — b804a94, last commit 2026-07-11T02:43:03Z
• claude/determined-maxwell-xyhnam — 6ccc622, last commit 2026-07-07T05:06:59Z
• claude/evaluate-unused-folders — c4b64f4, last commit 2026-07-03T01:37:08Z
• claude/explore-codebase-implementation-plan — 8be0d91, last commit 2026-06-19T03:16:33Z
• claude/help-github-docs-page — 73a637f, last commit 2026-06-19T02:45:50Z
• claude/prompt-start-subagents-for-action — 34ae55f, last commit 2026-07-07T02:00:40Z
• claude/repo-architecture-review-Aewa3 — f330e9f, last commit 2026-06-29T07:44:07Z
• claude/review-session-history-tips — 1f3c082, last commit 2026-06-21T05:52:44Z
• coderabbitai/utg/c89b9b7 — 285ee5e, last commit 2026-07-03T22:25:55Z

This report is intentionally non-destructive. Branch deletion requires a merged PR or an explicit retention decision.

Canonical governance: #898

[linear-code]

GRV-117

[coderabbitai]

🔗 Related PRs

#222 - docs: branch-cleanup assessment + reproducible fail-test harness [merged]
#277 - Vercel remediation complete: waitUntil adoption + active middleware + @vercel/functions package (resolves three user-defined outcomes; 16-agent verification-gate consolidated PASS on full matrix) [closed]
#404 - fix(web): sync package-lock with @vercel/functions (unblocks #277 auto-merge) [merged]
#1061 - revert: remove unauthorized broad #1057 merge [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[google-labs-jules]

Jules is on it. When finished, you will see another comment and be able to review a PR.

[coderabbitai]

Coding Plan

Summary

• Treat the drift report as a triage trigger, not a delete order: re-derive fresh evidence (with a current PR snapshot) rather than trusting the two-week-old branch list — that stale snapshot is a trap waiting to bite whoever runs this blindly.
• Reuse the existing 6-gate harness and its history-rewrite-safe logic verbatim — two-dot diffs + G1/G5, never three-dot/merge-tree as sole justification — to avoid mis-classifying orphaned branches as empty. Get this wrong and you'll silently nuke branches that only look empty.
• Produce the standard governance artifacts (dated audit, matrix, followups, recovery refs) under canonical governance #898 so every disposition is evidence-backed and reversible.
• Execute only through the archive-tag-then-delete gated workflow with a live open-PR guard and dry-run validation, keeping REVIEW branches for explicit human retention decisions — no shortcuts, no raw deletes.

Design Choices

Design Choice 1: Should this ticket actually delete branches, or stop at a staged, human-approved classification?

Options Considered:

1. Run the harness and immediately execute deletions of everything not KEEP.
2. Run the harness, produce the evidence matrix + audit docs, regenerate the delete script, and leave execution as a separately-gated, human-triggered step (dry-run validated first).

Chosen Option: 2

Rationale: Option 2 is the only defensible call here — it matches the repo's non-destructive governance, the #227 collateral-deletion incident lesson, and the 'explicit retention decision' requirement. Auto-deleting 100+ branches off a two-week-stale snapshot is exactly the kind of production landmine this pipeline was built to defuse, and I'm not signing off on that risk.

💡 User Tips

Regenerate the plan with different choices with @coderabbitai <feedback>.

Implementation Steps

Phase 1: Evidence Gathering and Classification

Reconstruct authoritative evidence for the 114 flagged branches and classify each into KEEP / CLOSE-SAFE / CLOSE-STALE / REVIEW using the existing 6-gate harness. The report snapshot is dated 2026-07-27 but branch state and PR state drift daily — evidence must be re-derived against current refs, not the ticket list taken on faith.

Task 1: Capture a fresh PR-state snapshot for gate G1

Produce the authoritative open/closed PR map so the 'never touch open-PR branches' gate is trustworthy. Generate a current head.ref|state|number PR map (all states) via the GitHub MCP list_pull_requests tooling, as documented in .claude/skills/branch-cleanup/SKILL.md step 1, saved to a --pr-json input file. This is a real safety requirement, not a formality: the report is ~2 weeks old and any of the 114 branches may have gained an open PR since. G1 is the only authoritative gate — do not run the harness without it.

Task 2: Run the 6-gate harness to regenerate the evidence matrix

Score all flagged branches against current main and PR state. Run scripts/maintenance/branch-fail-test.sh --pr-json <snapshot> after git fetch --all --prune, regenerating docs/branch-cleanup-matrix.md and docs/branch-cleanup-matrix.csv. Respect the history-rewrite caveat baked into the harness: rely on two-dot git diff --shortstat main branch (G4) plus G1/G5, and treat any G3 merges_clean=yes as a known false positive for orphaned branches. Do NOT reintroduce three-dot diffs or git merge-tree as sole justification — orphaned branches from the secret-purge force-push falsely report as empty/clean and would be wrongly classified CLOSE-SAFE. Optionally pass --build (G6 CI fail-test) only for REVIEW-bucket branches that carry unique diff, to confirm dead-vs-salvageable.

Task 3: Sanity-check classification against the report's branch families

Confirm the verdict distribution is coherent before writing anything durable. Cross-check the ~60 claude/determined-maxwell-* agent branches (likely CLOSE-STALE/CLOSE-SAFE), the copilot/*, codex/*, jules-*, palette*, bolt* families, and the human-named feat/* / fix/* branches which are more likely to hold unique unmerged work and land in REVIEW. Flag any branch whose verdict is CLOSE-SAFE solely because of uniq_files == 0 for a manual glance — that is exactly the orphan false-positive signature and must be verified against G5 (behind/age) before trusting it.

🤖 Prompt for AI agents

Implement Phase 1: re-derive fresh, trustworthy evidence for all 114 flagged
branches using the repo's existing 6-gate harness — don't trust the two-week-old
report blindly.

- First, generate a fresh PR-state snapshot: use the GitHub MCP
`list_pull_requests` tooling (per `.claude/skills/branch-cleanup/SKILL.md` step
1) to produce a current `head.ref|state|number` map covering all PR states, and
save it to a file for use as `--pr-json` input. This is a hard safety
requirement — G1 is the only authoritative gate against accidentally touching a
branch that gained an open PR since the report was generated.
- Run `git fetch --all --prune`, then run
`scripts/maintenance/branch-fail-test.sh --pr-json <snapshot>` to regenerate
`docs/branch-cleanup-matrix.md` and `docs/branch-cleanup-matrix.csv` with fresh
scores across all 6 gates (G1 PR-state, G2 redundancy, G3 merge-clean probe, G4
unique-diff, G5 staleness, G6 optional CI).
- Critical correctness rule: use two-dot `git diff --shortstat main branch` for
G4 combined with G1/G5. Treat any G3 `merges_clean=yes` as a known false
positive for orphaned branches from the secret-purge force-push. Do NOT use
three-dot diffs or `git merge-tree` as sole justification for a verdict — this
WILL misclassify orphaned branches as empty/CLOSE-SAFE when they aren't. This is
a real bug pattern to guard against, not a style preference.
- Optionally pass `--build` (G6) only for REVIEW-bucket branches carrying unique
diff, to distinguish dead code from salvageable work.
- Once the matrix is regenerated, sanity-check the verdict distribution against
known branch families: ~60 `claude/determined-maxwell-*` agent branches (expect
CLOSE-STALE/CLOSE-SAFE), `copilot/*`, `codex/*`, `jules-*`, `palette*`, `bolt*`
families, and human-named `feat/*`/`fix/*` branches (expect more REVIEW verdicts
due to unique unmerged work).
- Manually flag any branch classified CLOSE-SAFE purely because `uniq_files ==
0` — cross-verify against G5 (behind/age) before trusting it, since this is the
exact signature of the orphan false-positive bug described above.

Phase 2: Governance Documentation

Turn the matrix into the durable, reviewable governance artifacts the #898 process expects, so any deletion later has a documented, reversible justification.

Task 1: Write the dated narrative audit document

Produce the human-readable audit that accompanies the matrix. Create docs/branch-audit-2026-07-27.md following the established structure: opening verdict-tally table (counts + meaning per bucket), then the full matrix, then the decision rule and the history-rewrite caveat note. State intent plainly: this audit exists to convert the reconciliation report's 114-branch backlog into evidence-backed dispositions under canonical governance #898.

Task 2: Record findings and pending execution tracking

Capture anything noteworthy and the execution state. Update docs/branch-cleanup-followups.md with a new dated section: any findings surfaced during triage (e.g., branches carrying unique unmerged security/perf work that should NOT be deleted), plus a pending-execution table (delete CLOSE-SAFE batch / delete CLOSE-STALE batch / triage REVIEW) with State and Plan columns. Reference the #227 incident lesson explicitly so the live open-PR guard requirement travels with this audit.

🤖 Prompt for AI agents

Implement Phase 2: convert the regenerated matrix into the durable, reviewable
governance artifacts that canonical governance `#898` requires, so any future
deletion has documented, reversible justification.

- Create `docs/branch-audit-2026-07-27.md`, following the established audit
structure: start with a verdict-tally table (counts per bucket plus what each
bucket means), followed by the full matrix, then a decision rule section and an
explicit note on the history-rewrite caveat (two-dot diff requirement, G3
false-positive risk). State plainly in the doc that its purpose is to convert
the reconciliation report's 114-branch backlog into evidence-backed dispositions
under `#898`.
- Update `docs/branch-cleanup-followups.md` with a new dated section covering:
any findings from triage — especially flagging any branch carrying unique
unmerged security/perf work that must NOT be deleted — and a pending-execution
table (rows for delete CLOSE-SAFE batch / delete CLOSE-STALE batch / triage
REVIEW) with State and Plan columns.
- Explicitly reference the `#227` collateral-deletion incident in the followups
doc so the live open-PR guard requirement is clearly tied to and travels with
this audit.

Phase 3: Reversible, Gated Cleanup Execution

Stage the actual cleanup so it is archive-tagged, dry-run validated, and executed only through the gated workflow — never a direct, irreversible delete.

Task 1: Regenerate the deletion script from the new matrix

Rebuild the batch arrays so they reflect this audit, not a stale hardcoded list. Regenerate scripts/maintenance/branch-cleanup-delete.sh SAFE_BRANCHES / REVIEW_BRANCHES (and CLOSE-STALE) arrays from docs/branch-cleanup-matrix.csv. Preserve the mandatory safety mechanics: git tag -f archive/<branch> origin/<branch> + push tag before git push origin --delete, DRY_RUN support, and the live open-PR re-check immediately before each delete. A hardcoded list without the live re-check is precisely what caused the #227 collateral deletion — do not ship it without that guard.

Task 2: Dry-run validate and record recovery refs

Prove reversibility before any destructive action. Run the delete script with DRY_RUN=1 and confirm the planned actions match the audit's CLOSE buckets and skip every open-PR branch. Append the <SHA> <branch> pairs for the batch to docs/branch-cleanup-recovery-refs.txt with a timestamped batch header and the restore comment (git push origin <sha>:refs/heads/<branch>), so recovery is possible even if a tag is lost.

Task 3: Execute via the gated workflow (human-triggered)

Perform deletion only through the sanctioned, permissioned path. Execute cleanup through .github/workflows/branch-cleanup.yml — either workflow_dispatch (choosing batch: safe|stale|review, dry_run first) or a [run-cleanup:<batch>] sentinel commit on a claude/branch-cleanup-* branch — since the sandbox token lacks branch-delete permission. Confirm the workflow uploads the branch-cleanup-recovery-refs artifact. Leave REVIEW-bucket branches untouched pending explicit human retention decisions, satisfying the report's 'explicit retention decision' requirement.

🤖 Prompt for AI agents

Implement Phase 3: stage the cleanup so it is archive-tagged, dry-run validated,
and executed exclusively through the gated CI workflow — a raw, irreversible
delete is not acceptable here.

- Regenerate `scripts/maintenance/branch-cleanup-delete.sh`'s `SAFE_BRANCHES` /
`REVIEW_BRANCHES` (and CLOSE-STALE) arrays directly from
`docs/branch-cleanup-matrix.csv` so the script reflects this audit, not a stale
hardcoded list.
- Preserve every mandatory safety mechanic in the script: `git tag -f
archive/<branch> origin/<branch>` plus pushing that tag BEFORE `git push origin
--delete` runs, full `DRY_RUN` support, and a live open-PR re-check performed
immediately before each individual delete. This live re-check is non-negotiable
— a hardcoded list without it is exactly what caused the `#227` collateral
deletion incident, and shipping without this guard risks repeating a real
production incident.
- Dry-run validate before anything destructive: run the script with `DRY_RUN=1`
and confirm planned actions match the audit's CLOSE buckets exactly and that
every open-PR branch is skipped.
- Append the batch's `<SHA> <branch>` pairs to
`docs/branch-cleanup-recovery-refs.txt`, with a timestamped batch header and a
restore comment (`git push origin <sha>:refs/heads/<branch>`) for each entry, so
recovery is possible even if an archive tag is somehow lost.
- Execute the actual cleanup only through
`.github/workflows/branch-cleanup.yml`, either via `workflow_dispatch`
(selecting `batch: safe|stale|review`, running `dry_run` first) or via a
`[run-cleanup:<batch>]` sentinel commit on a `claude/branch-cleanup-*` branch —
required since the sandbox token lacks branch-delete permission.
- After execution, confirm the workflow uploads the
`branch-cleanup-recovery-refs` artifact. Leave every REVIEW-bucket branch
untouched pending an explicit human retention decision — this satisfies the
report's explicit 'explicit retention decision' requirement and is not optional.

Research

This repo already owns a mature branch-hygiene pipeline governed by canonical issue #898. scripts/maintenance/branch-fail-test.sh scores every remote branch across six evidence gates (G1 PR-state, G2 redundancy, G3 merge-clean probe, G4 unique-diff, G5 staleness, G6 optional CI) and emits docs/branch-cleanup-matrix.{md,csv}. Narrative audits live in dated docs/branch-audit-<date>.md docs, with findings/execution tracking in docs/branch-cleanup-followups.md. scripts/maintenance/branch-cleanup-delete.sh is regenerated from the matrix CSV and archive-tags each branch (archive/<branch>) before deleting, with a live open-PR guard; it runs only via .github/workflows/branch-cleanup.yml (manual dispatch or a [run-cleanup:...] sentinel commit). docs/branch-cleanup-recovery-refs.txt logs SHA/branch pairs for belt-and-suspenders recovery.

---

🚀 Next Steps

🤖 All AI agent prompts combined

Task: 1

Implement Phase 1: re-derive fresh, trustworthy evidence for all 114 flagged
branches using the repo's existing 6-gate harness — don't trust the two-week-old
report blindly.

- First, generate a fresh PR-state snapshot: use the GitHub MCP
`list_pull_requests` tooling (per `.claude/skills/branch-cleanup/SKILL.md` step
1) to produce a current `head.ref|state|number` map covering all PR states, and
save it to a file for use as `--pr-json` input. This is a hard safety
requirement — G1 is the only authoritative gate against accidentally touching a
branch that gained an open PR since the report was generated.
- Run `git fetch --all --prune`, then run
`scripts/maintenance/branch-fail-test.sh --pr-json <snapshot>` to regenerate
`docs/branch-cleanup-matrix.md` and `docs/branch-cleanup-matrix.csv` with fresh
scores across all 6 gates (G1 PR-state, G2 redundancy, G3 merge-clean probe, G4
unique-diff, G5 staleness, G6 optional CI).
- Critical correctness rule: use two-dot `git diff --shortstat main branch` for
G4 combined with G1/G5. Treat any G3 `merges_clean=yes` as a known false
positive for orphaned branches from the secret-purge force-push. Do NOT use
three-dot diffs or `git merge-tree` as sole justification for a verdict — this
WILL misclassify orphaned branches as empty/CLOSE-SAFE when they aren't. This is
a real bug pattern to guard against, not a style preference.
- Optionally pass `--build` (G6) only for REVIEW-bucket branches carrying unique
diff, to distinguish dead code from salvageable work.
- Once the matrix is regenerated, sanity-check the verdict distribution against
known branch families: ~60 `claude/determined-maxwell-*` agent branches (expect
CLOSE-STALE/CLOSE-SAFE), `copilot/*`, `codex/*`, `jules-*`, `palette*`, `bolt*`
families, and human-named `feat/*`/`fix/*` branches (expect more REVIEW verdicts
due to unique unmerged work).
- Manually flag any branch classified CLOSE-SAFE purely because `uniq_files ==
0` — cross-verify against G5 (behind/age) before trusting it, since this is the
exact signature of the orphan false-positive bug described above.
===============================================================================

Task: 2

Implement Phase 2: convert the regenerated matrix into the durable, reviewable
governance artifacts that canonical governance `#898` requires, so any future
deletion has documented, reversible justification.

- Create `docs/branch-audit-2026-07-27.md`, following the established audit
structure: start with a verdict-tally table (counts per bucket plus what each
bucket means), followed by the full matrix, then a decision rule section and an
explicit note on the history-rewrite caveat (two-dot diff requirement, G3
false-positive risk). State plainly in the doc that its purpose is to convert
the reconciliation report's 114-branch backlog into evidence-backed dispositions
under `#898`.
- Update `docs/branch-cleanup-followups.md` with a new dated section covering:
any findings from triage — especially flagging any branch carrying unique
unmerged security/perf work that must NOT be deleted — and a pending-execution
table (rows for delete CLOSE-SAFE batch / delete CLOSE-STALE batch / triage
REVIEW) with State and Plan columns.
- Explicitly reference the `#227` collateral-deletion incident in the followups
doc so the live open-PR guard requirement is clearly tied to and travels with
this audit.
===============================================================================

Task: 3

Implement Phase 3: stage the cleanup so it is archive-tagged, dry-run validated,
and executed exclusively through the gated CI workflow — a raw, irreversible
delete is not acceptable here.

- Regenerate `scripts/maintenance/branch-cleanup-delete.sh`'s `SAFE_BRANCHES` /
`REVIEW_BRANCHES` (and CLOSE-STALE) arrays directly from
`docs/branch-cleanup-matrix.csv` so the script reflects this audit, not a stale
hardcoded list.
- Preserve every mandatory safety mechanic in the script: `git tag -f
archive/<branch> origin/<branch>` plus pushing that tag BEFORE `git push origin
--delete` runs, full `DRY_RUN` support, and a live open-PR re-check performed
immediately before each individual delete. This live re-check is non-negotiable
— a hardcoded list without it is exactly what caused the `#227` collateral
deletion incident, and shipping without this guard risks repeating a real
production incident.
- Dry-run validate before anything destructive: run the script with `DRY_RUN=1`
and confirm planned actions match the audit's CLOSE buckets exactly and that
every open-PR branch is skipped.
- Append the batch's `<SHA> <branch>` pairs to
`docs/branch-cleanup-recovery-refs.txt`, with a timestamped batch header and a
restore comment (`git push origin <sha>:refs/heads/<branch>`) for each entry, so
recovery is possible even if an archive tag is somehow lost.
- Execute the actual cleanup only through
`.github/workflows/branch-cleanup.yml`, either via `workflow_dispatch`
(selecting `batch: safe|stale|review`, running `dry_run` first) or via a
`[run-cleanup:<batch>]` sentinel commit on a `claude/branch-cleanup-*` branch —
required since the sandbox token lacks branch-delete permission.
- After execution, confirm the workflow uploads the
`branch-cleanup-recovery-refs` artifact. Leave every REVIEW-bucket branch
untouched pending an explicit human retention decision — this satisfies the
report's explicit 'explicit retention decision' requirement and is not optional.

💡 Iterate on the plan with: `@coderabbitai` <feedback>

Example Feedback
- `@coderabbitai` You can skip phase 3. Add a simple unit test case for phase 2.
- `@coderabbitai` For design choice 1 go ahead with option 3 and replan.

---

💬 Have feedback or questions? Drop into our discord!

[google-labs-jules]

I apologize, but I encountered an unexpected error and wasn't able to complete the task.