Commit review
Use /Users/ale/.local/bin/agent-review and the native subscription CLIs. Default
independent review uses Opus or Sol; the runner selects the other provider from the
author when possible, then tries a fresh reviewer from the other default provider.
Astra or Fable review requires an explicit --reviewer selection following a user
request. GLM is for chores and easy tasks and is never selected as a reviewer or
used to clear this gate.
Only an actual agent-commit attempt triggers this workflow. Human commits in
Lazygit or a terminal use normal Git hooks without AI review. Do not review on save,
on stop, after each turn, or while the user is iterating. The review command itself
does not commit or push.
- Stage only intended files and inspect
git diff --cached --stat. Attempt the
authorized commit with /Users/ale/.local/bin/agent-commit -m "message". Routine
documentation, lockfiles, known manifest/config bookkeeping and generated files may
skip review. Generated means: build_runner-style suffixes (.g.dart, .freezed.dart,
.gr.dart, .pb.dart, .mocks.dart, ...), files whose header carries a generated-code
marker, and paths matching globs listed in a repo-root .agent-review-exempt file
(one gitignore-style glob per line, e.g. curriculum.json, */*/data.json).
Any source/test file or substantive/unfamiliar configuration triggers the gate,
including when mixed with exempt files. Agents must use agent-commit, not raw Git.
- When blocked, run
agent-review start --author <model>, using the actual author
identity: opus, sonnet, fable, sol, terra, astra, or glm. Optionally
pass --requirements /absolute/path: include the user's requirements, not the
implementer's reasoning. The runner freezes the staged tree and obtains one
independent review covering correctness and simplicity. It defaults to Opus or
Sol and retries once with the other fresh default reviewer on quota/rate limits.
Pass --reviewer astra or --reviewer fable only for an explicitly requested
state-of-the-art-model review.
- Read the report and its actual reviewer/attempt history. Adjudicate every finding
in the printed
decisions.json: use fixed or rejected, with evidence. Do not
manufacture findings or reject them merely to clear the gate.
- Have a separate fix worker apply accepted findings with an explicit scope and
one writer per worktree. Record its model and session/task ID in
fixer. Use an
appropriate worker; GLM is suitable only for simple mechanical chores. Stage
accepted fixes explicitly and run relevant validation. Skip this step if no
findings are accepted.
- Run
agent-review finish /absolute/path/to/review-directory. Changed content
receives at most one confirmation pass restricted to the accepted fixes, using
the initial reviewer selection and its fallback. If confirmation fails, report
the issue rather than starting another full review or confirmation loop.
- Once cleared, run
agent-commit -m "message" separately when authorized. Existing
repository hooks still run; the exact staged tree and HEAD are checked after
hooks. No -a, pathspecs, chained staging, --no-verify, or push without authorization.
A provider's quota/rate error switches to the next reviewer once, using the same
frozen files, prompt and requirements. A completed valid report is required. Both
providers unavailable, authentication failures, malformed reports, unresolved
findings, changed HEAD or stale staged content leave review pending. Fallback is
not triggered by ordinary findings or content refusals. Never substitute GLM.
If both providers are limited during confirmation, it remains pending without
clearance. After limits reset, finish may resume for the exact same staged fixes;
changed fixes are rejected. This does not rerun the initial review.
The runner remembers reviewer limits under /Users/ale/.agents/state/review-provider-limits.
It uses a machine-readable reset time when available; otherwise the backoff is one
hour for quota exhaustion and one minute for transient rate limits. These defaults
are retry delays, not claims about the subscription's actual reset. There are no
background retries or automatic purchases/resets. Logs record every attempted or
skipped provider. No paid API-key fallback is introduced.
Snapshots, reports and ledgers live under the worktree's private Git directory
(git rev-parse --git-path agent-reviews); never stage them. agent-review check
verifies clearance without committing. An unchanged snapshot may reuse clearance;
any staged-tree or HEAD change invalidates it. The start command requires a gate
request for the current staged state.
All agents must use agent-commit. Raw Git remains available for human commits;
this is a workflow requirement, not a security boundary against an agent that
ignores instructions. Reviewers and confirmation workers are read-only, do not
delegate reviews, and do not invoke this workflow recursively.
1---2name: commit-review3description: Review staged code at agent-commit time with independent Opus or Sol sessions and a fresh cross-provider fallback on usage limits. Covers correctness and simplicity, adjudication, a separate fix worker, one confirmation pass, and staged-content clearance. Human commits and routine bookkeeping are exempt.4metadata:5 source: https://skillmd.com/skills/nank1ro/commit-review6---78# Commit review910Use `/Users/ale/.local/bin/agent-review` and the native subscription CLIs. Default11independent review uses Opus or Sol; the runner selects the other provider from the12author when possible, then tries a fresh reviewer from the other default provider.13Astra or Fable review requires an explicit `--reviewer` selection following a user14request. GLM is for chores and easy tasks and is never selected as a reviewer or15used to clear this gate.1617Only an actual `agent-commit` attempt triggers this workflow. Human commits in18Lazygit or a terminal use normal Git hooks without AI review. Do not review on save,19on stop, after each turn, or while the user is iterating. The review command itself20does not commit or push.21221. Stage only intended files and inspect `git diff --cached --stat`. Attempt the23 authorized commit with `/Users/ale/.local/bin/agent-commit -m "message"`. Routine24 documentation, lockfiles, known manifest/config bookkeeping and generated files may25 skip review. Generated means: build_runner-style suffixes (`.g.dart`, `.freezed.dart`,26 `.gr.dart`, `.pb.dart`, `.mocks.dart`, ...), files whose header carries a generated-code27 marker, and paths matching globs listed in a repo-root `.agent-review-exempt` file28 (one gitignore-style glob per line, e.g. `curriculum.json`, `*/*/data.json`).29 Any source/test file or substantive/unfamiliar configuration triggers the gate,30 including when mixed with exempt files. Agents must use agent-commit, not raw Git.312. When blocked, run `agent-review start --author <model>`, using the actual author32 identity: `opus`, `sonnet`, `fable`, `sol`, `terra`, `astra`, or `glm`. Optionally33 pass `--requirements /absolute/path`: include the user's requirements, not the34 implementer's reasoning. The runner freezes the staged tree and obtains one35 independent review covering correctness and simplicity. It defaults to Opus or36 Sol and retries once with the other fresh default reviewer on quota/rate limits.37 Pass `--reviewer astra` or `--reviewer fable` only for an explicitly requested38 state-of-the-art-model review.393. Read the report and its actual reviewer/attempt history. Adjudicate every finding40 in the printed `decisions.json`: use `fixed` or `rejected`, with evidence. Do not41 manufacture findings or reject them merely to clear the gate.424. Have a separate fix worker apply accepted findings with an explicit scope and43 one writer per worktree. Record its model and session/task ID in `fixer`. Use an44 appropriate worker; GLM is suitable only for simple mechanical chores. Stage45 accepted fixes explicitly and run relevant validation. Skip this step if no46 findings are accepted.475. Run `agent-review finish /absolute/path/to/review-directory`. Changed content48 receives at most one confirmation pass restricted to the accepted fixes, using49 the initial reviewer selection and its fallback. If confirmation fails, report50 the issue rather than starting another full review or confirmation loop.516. Once cleared, run `agent-commit -m "message"` separately when authorized. Existing52 repository hooks still run; the exact staged tree and HEAD are checked after53 hooks. No -a, pathspecs, chained staging, --no-verify, or push without authorization.5455A provider's quota/rate error switches to the next reviewer once, using the same56frozen files, prompt and requirements. A completed valid report is required. Both57providers unavailable, authentication failures, malformed reports, unresolved58findings, changed HEAD or stale staged content leave review pending. Fallback is59not triggered by ordinary findings or content refusals. Never substitute GLM.60If both providers are limited during confirmation, it remains pending without61clearance. After limits reset, finish may resume for the exact same staged fixes;62changed fixes are rejected. This does not rerun the initial review.6364The runner remembers reviewer limits under `/Users/ale/.agents/state/review-provider-limits`.65It uses a machine-readable reset time when available; otherwise the backoff is one66hour for quota exhaustion and one minute for transient rate limits. These defaults67are retry delays, not claims about the subscription's actual reset. There are no68background retries or automatic purchases/resets. Logs record every attempted or69skipped provider. No paid API-key fallback is introduced.7071Snapshots, reports and ledgers live under the worktree's private Git directory72(`git rev-parse --git-path agent-reviews`); never stage them. `agent-review check`73verifies clearance without committing. An unchanged snapshot may reuse clearance;74any staged-tree or HEAD change invalidates it. The start command requires a gate75request for the current staged state.7677All agents must use agent-commit. Raw Git remains available for human commits;78this is a workflow requirement, not a security boundary against an agent that79ignores instructions. Reviewers and confirmation workers are read-only, do not80delegate reviews, and do not invoke this workflow recursively.