Revue de code commit par commit

---

name: code-review description: "Review a PR or feature branch commit by commit so sync-merged changes are not misattributed. Skip feedback on uncommitted edits (verify) and quality questions with no PR or branch context." argument-hint: "[base-branch]" allowed-tools: Bash, Read, Glob, Grep, Agent effort: max disable-model-invocation: false

Review the current branch's PR changes per-commit.

Base branch: $ARGUMENTS (if empty, infer from branch naming or ask)

Current Context

• Branch: !git branch --show-current 2>/dev/null || echo "N/A"
• Commits on branch: !git log --oneline main..HEAD 2>/dev/null || git log --oneline -10 2>/dev/null || echo "N/A"

When to dispatch to reviewer agent

For PRs with many commits or a large diff, spawn the reviewer subagent via the Agent tool to keep per-commit analysis out of the main context. Run the steps below inline only for small PRs.

Use inline when:

• 1–2 unique commits
• Diff under ~200 lines
• Review scope is narrow (single feature or fix)

Dispatch to reviewer when:

• 3+ unique commits
• Diff exceeds ~400 lines
• Multiple modules touched

Dispatch to a separate context to keep the main thread clean:

• Claude Code: Agent(subagent_type: "reviewer", description: "<short>", prompt: "Base branch: <base>. Focus: <optional>. Branch: <current>. Unique commit/surface inventory: <...>. Requirement/convention source inventory: <...>.")
• Codex CLI: codex exec with the same focused prompt and provenance packet. Codex's multi_agent_v1.spawn_agent is explicit-request-only (not for auto-dispatch), so spawn a codex exec subprocess instead.

The parent completes Steps 1–2 before dispatch. Do NOT re-run delegated per-commit analysis, synthesize missing preflight content, or retry an invalid return into success. Before surfacing it, mechanically validate:

1. Unique-commit inventory appears before Context Preflight, which appears before findings.
2. Requirements, Conventions, Impact, Tests, and Unknowns are all present and non-empty or explicitly none.
3. Every Impact/Tests [seed:...] SHA/path/symbol-or-@file exists in the dispatched unique-commit inventory; a sync-merged seed is invalid.
4. Every returned requirement, convention, Unknown, and finding-context source ID exists in the dispatched inventories; all available requirements and conventions are covered; all unavailable/conflicting requirements appear under Unknowns. Requirements: none or Conventions: none is valid only when its inventory is empty.
5. Every finding whose [context:<source-id>] is unavailable/conflicting is [status:Unverified].
6. Every finding's [commit:<sha>] exists in the dispatched unique-commit inventory.

A valid return reports Review status: COMPLETE. An invalid return reports Review status: INCOMPLETE, includes a fenced validator result naming every failed check, and keeps delegated findings reference-only with context-dependent claims unverified.

Steps

1. Identify unique commits

<base>..HEAD already excludes commits reachable from <base>, so no extra cross-reference is needed.

Prefer --first-parent — it collapses each sync-merge from base into a single commit on the feature branch, making sync-merged code easy to identify and skip:

git log --first-parent --oneline <base>..HEAD

Fallback: if the branch has unusual merge topology and raw merge commits still appear in the output, use --no-merges to drop them entirely:

git log --no-merges --oneline <base>..HEAD

These are the ONLY commits to review. When git show <merge-commit> produces a massive diff, that diff is the sync-merged code from base — SKIP it (per Step 5 Cross-verify).

2. Context Preflight

After identifying unique commits and before inline analysis or delegation, build this bounded read-only provenance packet from only their changed surfaces:

Unique commit inventory
- <sha>: <changed paths and changed symbols>

Requirement source inventory
- <source-id>: status=<available|unavailable|conflicts-with:<source-id>> locator=<path|URL|user-scope> content=<compact requirement excerpt or access error>

Convention source inventory
- <source-id>: locator=<repository path> content=<compact relevant convention excerpt>

Read available user scope, PR/Jira requirements, active-plan criteria, repository-local AGENTS/rules, and relevant project patterns. Add no network authority; record unavailable remote context instead. If .agents/review-lessons.md exists, read it as past review evidence, NOT as convention: its Adopted entries rank below every written project rule and never override one, Candidates are context only, and Rejected points must not be re-raised as findings. Its absence is recorded as absent — never a reason to collect, which is /review-lessons, the only network-touching step in that loop. Use code-graph support when available, falling back to repository search and language-native tools. Keep discovery relevant to the unique changed surfaces rather than auditing the whole repository.

Present the result before findings:

Context Preflight
- Requirements: [source:<source-id>] <available requirement>, or none
- Conventions: [source:<source-id>] <relevant convention>, or none
- Impact: [seed:<unique-sha>:<changed-path>:<changed-symbol|@file>] <impacted target>, or none
- Tests: [seed:<unique-sha>:<changed-path>:<changed-symbol|@file>] <test path or observable>, or none
- Unknowns: [source:<source-id> status:unavailable|conflict] <detail>, or none
- Past lessons: [lesson:<id> state:adopted|candidate type:correctness|convention|preference] <point>, or none

Use canonical @file for config, documentation, assets, deletions, or any file-level change without a language symbol. Impacted targets and related tests may be outside the diff, but every seed must be in the unique-commit inventory. Never seed discovery from a branch-wide diff or sync-merged surface. Surface conflicting requirements without resolving them silently.

Every finding must carry [commit:<unique-sha>], [context:<source-id>|none], and [status:Verified|Unverified]. A finding dependent on unavailable or conflicting context is Unverified, not asserted.

3. Review per-commit

For each unique commit:

git show <commit> --stat   # scope check
git show <commit>          # full diff

Do NOT use git diff <base>..HEAD for line-level review — it includes sync-merged changes.

4. Branch diff for context only

git diff <base>..HEAD --stat is acceptable for overall scope, but flag issues only if they trace back to a unique commit.

5. Cross-verify before flagging

Before reporting any issue:

1. Identify which commit introduced it
2. If the commit is NOT unique (i.e., from a sync merge), do NOT flag it

6. Next step (optional)

After fixes are committed, suggest running /pr-body to generate the PR description.

Rules

• Only review changes from unique commits (Step 1)
• Run Context Preflight before findings and preserve its provenance tags
• Match the project's language and conventions
• Keep feedback actionable and concise