Pr Roast Triage

PR Roast Triage

Review every open PR, close what has no independent impact, roast the rest with fix recommendations. The human merges manually in the emitted order — never auto-merge, never merge from this skill.

Step 1 — Triage baseline (run first, before any opinion)

./scripts/pr-triage.sh   # conflict status + CI status + suggested order
git status --short       # scope out unrelated local changes first
git diff HEAD --stat

Step 2 — CI truth (annotations, not badges)

A red badge is not a verdict. Pull the failing step's annotation:

gh api repos/<owner>/<repo>/check-runs/<job_id>/annotations \
  --jq '.[].message | .[0:200]'

Known infra flakes (re-run, do not "fix code"):

• Unable to authenticate to FlakeHub → Cargo Deny nix setup broken, pre-existing.
• hosted runner lost communication → CodeQL/runner loss, pre-existing.
• External-only red (Codacy/Sonar) with all GitHub checks green → code issue, real.

Step 3 — Duplicate detection (same base blob + same hunk)

Bot swarms file near-identical PRs. Compare, don't assume:

gh pr diff <a> --name-only
gh pr diff <b> 2>/dev/null | sed -n '1,80p'   # same base hash? same function?

Same file + same base blob + same effect = duplicates. Keep exactly one:

1. Green CI beats mergeable-but-red.
2. Fewest unrelated files (no export.json timestamp, no lockfile churn).
3. Backward-compat wrapper kept beats private-fn signature break.
4. Newest branch (rebases cleanest onto latest main).

Close the losers with reason superseded by #<keeper>.

Step 4 — Roast rubric (every keeper PR)

• Atomicity: one logical change. Mega-PRs (>5 files across concerns) must split.
• Title honesty: title must describe the diff (re-export hiding a BFS cap = reject/retitle).
• commitlint scope: must exist in commitlint.config.cjs scope-enum (perf(hyperdim) fails; only listed scopes pass).
• Rationale comments: never deleted to "shorten" (Never delete rationale comments). Under 500 LOC gate pressure, bot PRs may attempt to strip docstrings/comments to make room for new code. Reject/roast this behavior; require extracting child submodules (e.g. hyperdim_binary_serde.rs) instead.
• deny.toml: must only ADD ignores with advisory ID + reason. Deleting or commenting out existing ignores re-breaks cargo deny — reject.
• Perf claims: no criterion output or flamegraph = not review-ready.
• export.json / Cargo.lock noise: timestamp-only or resolver-churn hunks must be dropped before merge.
• Bot comments (Jules hello, Sonar/Codacy pass notes) are noise, not reviews. Zero human comments = no hidden requirements, but also no approval.
• GitHub CLI GraphQL deprecation: gh pr edit and gh issue view fail if querying deprecated projectCards. Query specific --json fields for reads, or use gh api -X PATCH repos/<owner>/<repo>/pulls/<id> -f title="..." for editing PR titles.
• Multi-issue parent linkage: When a parent PR implements multiple child issues, ensure all child issues (Fixes #A, Fixes #B) are declared in the PR body so all issues close cleanly upon merge.

Step 5 — Manual merge order (emit, do not execute)

1. Trivial green first (dependabot 1-liners).
2. Security/clamp fixes (small, high value).
3. Foundation before dependents (commitlint scope additions, owner/facade migrations before perf touches on the same files).
4. One keeper per duplicate cluster; closes reference the keeper.
5. Mega-PRs and SIMD/unsafe last (need evidence + rebase after everything).
6. After each merge: rebase next, re-run CI, re-check mergeable.

# per PR, in order:
git fetch origin main
gh pr checks <n>                       # all green?
git diff --stat origin/main...<branch> # no surprise reverts?
gh pr merge <n> --squash --delete-branch

Step 6 — Record in codebase

• Full roast → plans/PR_ROAST_<YYYY_MM_DD>.md (per-PR verdict table + order + commands).
• Pre-existing flakes/learnings → append progress/LEARNINGS.md.
• plans/GOAP_STATE.md: update counters in place, set action_last_completed (exactly once, last key) to the triage action.