review

Review a code change adversarially: assume it is broken and try to prove it. The
deliverable is a set of verified findings the user can act on
as-is — ready-to-post comments for an MR/PR, concrete fixes for local targets — each
explained in plain language. Report verified defects and explicit repository
requirement violations. Include optional style suggestions only when requested.

Arguments: --test and --no-test (Manual testing), --fix (steps 4–5), and
--agents <model>[:<effort>][,…] (Delegation). A request in words to fix,
approve, or merge, or that names reviewer models counts as the matching flag.

Automatic completion review

For automatic reviews of this session's work, read when-to-run.md.

Delegation

Every review runs through the reviewer sub-agent. After the reviewers return, read delegation.md.

Manual testing

Manually test the changed behavior by default; --no-test skips it. Read
manual-testing.md for skip rules, cases, and evidence.

Fix authorization

Before deciding whether review findings may be fixed, read Fix authorization in fix.md.

0–2. Target, review, verify

Before the spawn, the main session reads no review references except when-to-run.md, and it never reviews
the code itself. Start one cheap sub-agent with exactly this prompt:

Set up a review pass per <absolute path of references/setup.md>.
Target: <the user's target verbatim; a range or MR/PR the calling skill or loop names; or none>.
Scratch: <task scratch directory>. Earlier worktree: <path or none>.
Skip check: <on for a run's first setup; off when the user asks to review again, and for later passes>.

• Run setups one at a time, never in parallel.
• On empty, say there is nothing to review and stop. On checks: unclear or a
  blocker, rerun the prompt with the worker role, adding Judgement: allowed, any
  reported worktree, and started checks. If the worker is still blocked, clean up
  per output.md, report the blocker to the user, and stop.
• Start one reviewer with the returned prompt verbatim, or one per --agents
<model>[:<effort>] entry on that model and effort; with several targets, one set
  per target. Otherwise the sub-agents skill sets model and effort.
• On skip (own MR/PR only: agent-review status --head <sha> on its store target
  shows reviews_at_head of 2 or more and 0 open threads; store.md),
  report Skipped: <short-sha> already has <N> reviews with the run IDs and stop.
  Setup stops before checks; clean up any worktree it reports per
  output.md.
  With --approve, --auto-merge, or --merge, skip only the review and still run
  step 7.
  A request to review again overrides this. Never pass records to reviewers.
• With own: yes, the review posts and resolves MR/PR comments per Own MR/PR
  comments in remote-comments.md. With own: no,
  it posts nothing unless the user asks.
• Never add a summary of the change, focus areas, what already passed, or earlier
  findings. That framing steers reviewers away from the code. Only a fix check and
  the cycle 2 reveal add the inputs loop.md lists.
• Use the returned worktree; the manual pass uses it too.
• On mutations unsupported, run no --fix, --threads, or merge step.
• Give each reported linked MR/PR its own pass, read-only: apply --fix, --threads,
  approval, and merge only to MRs/PRs that the user named or authored.
• Use the reported CI facts in the step 3 report.

After the reviewers return, re-verify per Delegation, reading only the code each
finding depends on: the cited lines, their callers, and any guard that could refute it.
Record every pass, its verified and killed findings, and the run in the store per
store.md.
Read reviewer-steps.md only for an inline fallback.

3. Output

When reporting results, read output.md. It owns the
finding format, remote comment links, evidence, and cleanup. Post comments only on
an own MR/PR, or when the user asks, and only with agent-review publish
(publish --not-own --findings --only <ids> on another person's MR/PR, per
output.md;
store.md).

4. Optional: --fix

For authorized fixes, including your own work, read fix.md.
Fixing always continues into step 5.

5. Fix loop

Every review that fixes runs the loop in loop.md.

6. Optional: --threads

Only when the user asks (--threads, "are all comments resolved?", "fix and resolve the
comments"). Follow threads.md for the named MR/PR.

7. Optional: --approve, --auto-merge, --merge

For --approve, --auto-merge, --merge, a request to approve or merge, or a fix
pushed to an MR/PR, read approve-merge.md. Each
marks a draft MR/PR ready.