ABOUTME: Commit-aware PR review that reads each commit's intent before judging code.

ABOUTME: Composes the existing review tiers and reclassifies findings by commit narrative, never re-implementing them.

PR Review - Commit-by-Commit

Review Tiers - pick the right one

pr-review is the PR-level tier. It COMPOSES the tiers below; it never re-implements them.

pr-review's unique value is commit-narrative awareness: reading each commit's intent to tell a deliberate deferral from a bug, then reclassifying findings. Everything else it delegates to the tiers above.

Trigger

Activate when user says: "review PR", "review pull request", "analyze PR", "pr review", or /pr-review.

Arguments

/pr-review <number|url>           Review a specific PR
/pr-review <number> --quick       Skip the gated second-opinion (faster, less thorough)
/pr-review <number> --no-gemini   Skip Gemini batches (offline mode)

Why Commit-by-Commit

A monolithic diff loses context. Each commit carries intent via its message:

• A "fix(security):" commit means the author knew about the vuln.
• A "deferred to phase N" note means incomplete code is intentional.
• A fix commit after a feature commit means the issue was caught and addressed.

This lets the review reduce false positives (intent vs oversight), reveal the development narrative (feature to review to fix cycles), spot incomplete fixes, and catch regressions where a later commit breaks an earlier one.

Execution Flow

Phase 0: Gather PR metadata

gh pr view <N> --json title,body,state,baseRefName,headRefName,additions,deletions,changedFiles,commits,labels,author
gh pr diff <N> --name-only
gh pr diff <N> > /tmp/pr<N>-diff.patch                                    # full diff for delegated reviewers
gh pr view <N> --json commits --jq '.commits[] | "\(.oid) \(.messageHeadline)"'  # chronological commits

Scope assessment:

• < 300 lines, < 5 files: simple, standard review.
• 300-1000 lines, 5-15 files: moderate, commit-by-commit adds value.
• > 1000 lines or > 15 files: large, commit-by-commit essential + flag scope concern.
• > 5000 lines or > 50 files: excessive, recommend reject-and-split.

Phase 0b: Prior review record

Read what the implementing loop already recorded before judging anything (see the orchestrator skill, Review Artifacts):

• Committed: quality_reports/approvals/<YYYY-MM-DD_slug>.md on the PR branch. Redacted by design: branch, commit, rounds run, severity counts, CWE ids, final SCORE, residual risks, and the path of the local findings directory.
• Local, if present: that findings directory, quality_reports/reviews/<slug>/NNN-findings.md. It is gitignored, so it exists only when the PR was built on this machine; its absence says nothing about the PR.

How it changes the review:

Phase 1: Isolated build verification

Never check out the PR on the active repo: review always happens in a throwaway clone so uncommitted work is never contaminated. Clone, checkout the PR, run the project gate (make check or the CLAUDE.md equivalent), and export $PR_REVIEW_DIR for every later phase. Full mechanics, partial-clone budget, and base-branch comparison: references/isolated-clone.md.

Phase 2: Commit narrative analysis

Read each commit oldest-first. For each, extract intent (message subject + body), scope (--stat), type (conventional prefix), and whether it responds to a prior review ("address findings", "fix Gemini review").

Build a commit graph tracking introduced_in[finding], fixed_in[finding], and still_open[finding]. Signals to watch:

• fix after feat: author caught something (verify the fix is complete).
• "deferred to" / "phase N" / "TODO": intentional incompleteness (note, don't penalize as a bug).
• duplicate commit messages: rebase or amend residue (process smell).
• "address review findings": cross-reference the review it responds to.
• nolint / shellcheck disable: conscious suppression (verify justification).

Phase 3: Delegate specialized review

Route reviewers by file pattern using the Review Routing table in the orchestrator skill (Step 3). Do not duplicate that table here; apply it as written. In parallel, run /gemini-review on the diff segmented by package (target < 3000 lines per segment) using prompts/gemini-segment.md.

Each delegated reviewer receives a domain slice, not the full diff: only the hunks of files its routing pattern matched, within a soft budget of ~20k token per brief (drop whole out-of-domain files first; if still over, keep the hunks and cut surrounding context). The brief declares the cut per the page-fault rule in the orchestrator skill (Step 1): excluded: <files>; read on demand from /tmp/pr<N>-diff.patch or $PR_REVIEW_DIR. Simple PRs (< 300 lines) pass whole, slicing overhead is not worth it there. Every reviewer also gets the project CLAUDE.md conventions and $PR_REVIEW_DIR as the working directory, with instructions to read real source (never hallucinate) and classify each finding as Critical/Major/Minor with exact file:line. The review already runs in a throwaway clone, so these launches deliberately omit isolation: "worktree"; each brief therefore opens with the line ISOLATION-EXEMPT: pr-review throwaway clone $PR_REVIEW_DIR, verbatim and as the FIRST line of the prompt, which is the only place reviewer-isolation-guard reads it (a marker further down is quoted content, not an authored exemption, and does not exempt). That satisfies the launch guard only: the brief carries no isolation assertion, so the reviewer stays read-only agent-side.

Gemini hallucination patterns to filter (cross-validate every Gemini Critical against source): language-feature availability by version, database-engine limits misattributed across engines, standard-library APIs that do not exist.

Phase 4: Commit-context reclassification (unique value)

Cross-reference Phase 3 findings against the Phase 2 narrative. For each finding: which commit introduced it, was there a fix commit, is the fix complete, was it intentional?

Phase 4b: Evidence gate

No Critical or Major finding reaches consolidation as a bare claim. Bug-class claims require an executable red-green test written and run inside $PR_REVIEW_DIR (fails on the PR branch = demonstrated; fails on the merge base too = pre-existing). Non-executable finding classes use the evidence taxonomy instead (CWE + line, Big-O derivation, grep-able convention, named principle). A test that does NOT fail means the claim is unproven: escalate it to Phase 6, never report it as-is. Full protocol, taxonomy, budget, and skip conditions: references/evidence.md.

Phase 5: Consolidation and scoring

Deduplicate: when several reviewers find the same issue, keep one entry with cross-references, e.g. [Security/Architecture]. Order the report by severity (Critical first), then by component within each band.

Score per rules/quality-gates.md (Critical = auto-fail, Major -10, Minor -3; commit >= 80, PR merge >= 90, excellence >= 95). Do not restate the rubric here. Only evidence-backed Critical/Major findings count toward the score; claims escalated as unproven (Phase 4b) are excluded until second-opinion resolves them.

Phase 6: Second opinion (gated)

Invoke /second-opinion only when reviewer verdicts conflict, a Critical finding is contested, or a Critical/Major claim came back unproven from the Phase 4b evidence gate (its red-green test did not fail). Otherwise skip it: it spins up isolated containers and is not free. When invoked, pass the disputed finding(s), the conflicting verdicts (for unproven claims: the test and its passing output), and ask which classification is correct and why. For unproven claims, keep only if second-opinion supplies new evidence; otherwise drop and count under "Hallucinations caught". Synthesize the answer into the final severities.

Phase 7: Present report

Use the report structure in references/output-format.md: header (branch, scope, score, reviewers, hallucinations caught), commit narrative, Critical/Major/Minor sections, a reclassifications table, dependencies, process observations, and a recommendation (APPROVE / FIX BEFORE MERGE / REJECT AND SPLIT).

Phase 8: Cleanup

Always remove the temp clone after the report, whatever the outcome, and even on interruption (build failure, agent timeout, user abort). Evidence tests live only in the clone: make sure surviving repro tests were copied into the report (Phase 4b deliverable) before deleting.

rm -rf "$PR_REVIEW_DIR"; unset PR_REVIEW_DIR

Stale pr-review-* directories under $TMPDIR are safe to delete at any time.

Quality Notes

• Never relay raw agent output: synthesize, deduplicate, verify.
• Every Critical must be source-verified: read the actual file:line before reporting.
• Every Critical/Major must carry evidence (Phase 4b): a claim without a failing test or a taxonomy-backed proof is not a finding.
• Commit context changes severity: a conscious deferral is not a bug.
• Pre-existing issues: still block (green pipeline is everyone's responsibility) but don't penalize the author's score.
• Large PRs (> 50 files): always recommend reject-and-split, even if quality is high.

See references/isolated-clone.md for clone mechanics and the full troubleshooting table.