review
Dispatch Canon's independent reviewer subagents for this branch and read their verdicts back from canon_review. Use once a change is built and verified, before shipping. This skill only dispatches and reads back -- the review itself runs inside the subagent, in a fresh context, and never in this session.
- License 许可
- BSD-3-Clause
- 所属仓库
- urban233/Canon
- 最近更新
- 2026-10-01
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- review
- description
- Dispatch Canon's independent reviewer subagents for this branch and read their verdicts back from canon_review. Use once a change is built and verified, before shipping. This skill only dispatches and reads back -- the review itself runs inside the subagent, in a fresh context, and never in this session.
03 SKILL.md 正文
Review
Invariant III is that nothing ships on the agent's own word, and a review you perform yourself is your own word. So this skill's whole job is to hand the change to something that did not write it, and to read back what that something actually said.
Dispatch
-
Call
canon_review. It reportsreviewers_called_for-- alwaysreviewer, plusrisk-reviewerwhen the changed paths touch a risk surface. You do not choose them and there is no waiver to write; the diff decides. -
Dispatch each reviewer it names, giving it the
base..HEADrange, the saved plan and its## Non-goals, and the evidence that checks ran. Never your own reasoning about the change -- a reviewer's value comes precisely from not having seen it.Antigravity bundles both reviewers with this plugin, under
agents/, soinvoke_subagentcan reach them by name once the plugin is enabled. If neither resolves there is nothing to dispatch, and the honest answer is to say so -- not to review the change yourself, which is the one thing this skill exists to prevent. -
Read the verdict out of the reviewer's own final message.
On Claude Code and Codex a
SubagentStophook captures that message andcanon_reviewreads it back, so the verdict reachingshipis the reviewer's own words rather than this session's retelling. Antigravity fires no such event, so that capture does not happen here andcanon_reviewwill report no stored verdict. Quote the reviewer's closing verdict line verbatim instead of paraphrasing it, and say that it was read in-session. Do not present a paraphrase as though it came back throughcanon_review: a weakened guarantee stated plainly is worth more than a strong one that isn't true. See this plugin's README for the standing limitation.
When a reviewer has seen this branch before
canon_review's verdicts names, per reviewer, the head that reviewer
last looked at and whether that is stale. When a reviewer's own verdict
is stale, tell it which commit it last reviewed and ask it to review
<head>..HEAD as a delta in addition to the full base..HEAD range,
never instead of it.
This matters because the most expensive failure in a repair loop is a fix that closes the reported finding and quietly breaks something the reviewer already passed. Re-reading the whole diff makes that regression look like part of the change; reading the delta against a commit already approved is what makes it visible.
After a READY verdict, confirm the delta only. When a reviewer's own
last verdict is READY FOR HUMAN APPROVAL and its head is an ancestor of
HEAD -- you only closed its notes, or touched docs -- it has already
passed everything up to that commit. Ask it to review <head>..HEAD
alone and return a fresh verdict. The full range is for repair after
CHANGES REQUIRED, where a fix can break what was passed; after READY
there is nothing unpassed outside the delta, and re-reading the whole
range costs a full round for no new coverage. The verdict is still fresh
at HEAD, so canon_ship is satisfied the same way.
Read stale per reviewer, not from the combined verdict. The
combined stale is true when any reviewer's verdict is stale, while the
combined head belongs to whichever verdict ranked worst. With two
reviewers in play those two fields can describe different agents, and a
delta computed from the combined pair can span the wrong commits entirely.
verdicts[<name>].head and verdicts[<name>].stale always describe the
same reviewer.
Stop after two rounds on the same change
A third dispatch for the same change is the signal that the problem is not
the code. When you have already addressed two CHANGES REQUIRED verdicts
on this branch, stop and ask the developer rather than dispatching
again. Grinding is the most expensive failure mode there is, and the
point at which to spend a human's attention is before the third attempt,
not after the fifth.
This is imposed on you, not on the developer: it is not a gate, it blocks
nothing, and canon_ship never reads it. Nothing counts the rounds for
you either, deliberately -- a counter in the repository is
round-state.json, and a counter in session state is both unreadable by
canon_ship and lost on the next session anyway. So this is yours to
hold, from the rounds in this conversation. The honest consequence: after
a compaction the count is gone, and the rule degrades to judgement. On
Claude Code and Codex, canon_review counts the rounds from captured
verdicts; Antigravity captures none (step 3 above), so here the count
really is only in this conversation.
04 同仓库更多 Skills
testing-craft
Provides Google-derived testing strategy, test-writing craft, and test-suite health reference, distilled from Software Engineering at Google (the Flamingo book), the Google Testing Blog, and Google's published testing-infrastructure research. Canon's `reviewer` subagent reads the writing-craft and strategy references as review criteria. Invoke directly to design a test strategy, audit an existing suite's health, or triage a flaky or brittle test.
ship
Decide whether a change is ready to hand to a human, and open the pull request when it is. Use once the work is done and verified, before opening a PR or telling the developer the change is ready.
review
Dispatch Canon's independent reviewer subagents for this branch and read their verdicts back from canon_review. Use once a change is built and verified, before shipping. This skill only dispatches and reads back -- the review itself runs inside the subagent, in a fresh context, and never in this session.
review-change
暂无简介
plan
Shape what plan mode drafts, so the plan Canon saves needs no follow-up. Use whenever drafting a plan for approval in plan mode, before writing it to .canon/plans/.
frame
Surface the open questions before planning a multi-branch feature, and shape the feature plan mode drafts. Use before entering plan mode for work that obviously won't fit in one branch and one pull request -- never for a change that already does.