review-loop
/review-loop
How to run, answer and record a cross-model code review of a Dasks card: one reviewer from the other model family, numbered findings, a ledger, at most two passes, and a finding blocks only with a repro or a security, data-loss, approval or money tag; small fixes only. Use when you review a card's PR or answer a review of your own work.
- Author
- drake-agent ↗
- Repository
- dasks-kit ↗
- Licence
- MIT
- Last repository change
- 07.11.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 07.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
How to run, answer and record a cross-model code review of a Dasks card: one reviewer from the other model family, numbered findings, a ledger, at most two passes, and a finding blocks only with a repro or a security, data-loss, approval or money tag; small fixes only. Use when you review a card's PR or answer a review of your own work.
Original author description
How to run, answer and record a cross-model code review of a Dasks card: one reviewer from the other model family, numbered findings, a ledger, at most two passes, and a finding blocks only with a repro or a security, data-loss, approval or money tag; small fixes only. Use when you review a card's PR or answer a review of your own work.
How it works
- Reviewer: what to read
- Reviewer: what to write
- The ledger
- Implementer: answering
- How many passes
- Recording it
SKILL.md preview
--- name: review-loop description: "How to run, answer and record a cross-model code review of a Dasks card: one reviewer from the other model family, numbered findings, a ledger, at most two passes, and a finding blocks only with a repro or a security, data-loss, approval or money tag; small fixes only. Use when you review a card's PR or answer a review of your own work." --- # review-loop A review here is a short, bounded conversation between two sessions: the one that built the card (the implementer) and one from the other model family (the reviewer: Claude reviews Codex's work, Codex reviews Claude's). It ends either clean or with the card going back to its spec. It never turns into a second round of building. ## Reviewer: what to read - The card (`dask card <card>`): its spec, `pass_checks:`, acceptance tests and `touches:`. The card is the yardstick; your own taste is not. - The diff of the PR head you were given (`git diff <merge-base>...<head>`), and enough of the code around it to judge it. Note the exact 40-character head sha: your verdict is about that commit only. - The evidence the implementer kept under `.dask/logs/<card>/`. Re-run a check yourself when the claim looks weak. ## Reviewer: what to write Every finding gets an id, a priority, a place, what is wrong and the right result. A finding that should block also gets a block tag. Tools read these lines, so keep the shape: ```text R1 [P1] [repro] src/contact/send.ts:42 an empty form sends a blank mail. Expected: refused with an error repro: curl -X POST localhost:3000/contact -d '' -> 200, and a blank mail in the outbox R2 [P1] [security] src/contact/log.ts:9 the API token is written to the log. Expected: redacted R3 [P3] public/contact.html:12 the button says "Submit"; the card asks for "Send". Expected: "Send" ``` **What blocks the merge:** | Finding | Blocks | |---|---| | carries a repro: `[repro]`, and the very next line is `repro: <exact command, input or failing test> -> <the wrong result it shows>` | yes | | tagged `[security]`, `[data-loss]`, `[approval]` (a human approval gate is skipped) or `[money]` | always, with or without a repro | | anything else, including `[repro]` with no `repro:` line | no: it is a follow-up | The priority (`P0` the main path is wrong, `P1` a real bug inside the change, `P2` a wrong result on an unusual input, `P3` a nit) orders the list. It never blocks by itself. If you cannot show a bug and it touches none of the four tags, it is a follow-up, however sure you are. - Ids run on across passes, so a finding keeps its id until it is closed. - Only report what is in the card's scope or what the diff itself breaks. An idea for later is not a finding: put it under a separate "Out of scope" line, without an id, or propose it as an idea on the board. - No finding without a place and a reason someone else can check. "Could be cleaner" is not a finding. - The last line is `VERDICT: clean` (nothing blocks; follow-ups may remain) or `VERDICT: changes` (a finding blocks). ## The ledger Keep one table per card in the review text and carry it to pass 2: ```text | id | block | status | pass | note | | R1 | repro | fixed | 1→2 | refused before send (abc1234) | | R2 | secur. | open | 2 | one log line still has the token | | R3 | - | follow-up | 1 | in the PR body | ``` `status` is one of `open`, `fixed` (the reviewer saw the fix), `follow-up` (does not block) or `disputed` (the implementer disagrees; the reviewer decides on pass 2, and a blocking finding stays open until it does). ## Implementer: answering - Fix every blocking finding with the smallest change that does it. No refactors, renames or extras on the way. - Do not fix follow-ups in this PR. List each one in the PR body under `## Follow-ups`. - If a finding asks for more than the card does, say so and leave it: scope belongs to the card, and a change of scope goes to its spec, not into this PR. - Disagree with a reason when the reviewer is wrong; don't fix something you believe is correct just to close it. - Push the fixes, then hand the new head back for pass 2. ## How many passes At most two per PR. Card size changes nothing. 1. **Pass 1:** reviews the whole change. 2. **Pass 2:** only re-checks the pass-1 findings and regressions the fixes caused. Each pass-1 finding gets a line: `R1 resolved`, or its finding line again, with its tag and a new `repro:` line, as it shows on the new head. Nothing else is raised. There is no third pass of any kind. Not clean after pass 2: keep the PR draft (put a ready one back with `gh pr ready --undo`), block the card with the open findings, `dask block <card> --note "<ids and one line each>"`, and tell the human. The spec or the design gets fixed first. ## Recording it Record each pass on the board from the reviewing session, for the exact head you read, so the board sees which family sent it: ```bash dask review-record <card> --head <sha> --impl claude --reviewer codex --verdict open --open 1 --noted 1 --answer-file review.md ``` `--open` counts blocking findings, `--noted` follow-ups. The board takes pass 1 and pass 2 and refuses a third. A pass is about one head: any push after it needs a new pass before a merge can rely on it. Never record a review of your own work, and never call a review clean to get past a gate.
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · drake-agent/dasks-kit | 100 | MIT |
| mthines/agent-skills | 100 | MIT |
| agentcrew-academy/harness-starter-kit | 100 | MIT |
| NathanColosimo/skills | 100 | MIT |
| kmbenjel/agent-workbench | 100 | MIT |
| agustinusnathaniel/maestria | 100 | MIT |
| agustinusnathaniel/maestria | 100 | MIT |
| ryoppippi/dotfiles | 100 | MIT |
| parcha-ai/parcha-skills | 100 | MIT |
Database snapshot: 13.11.2026 · package version 4b2b18dfda32 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 90spec-writer
Turn a rough Dasks card into a spec that passes the card contract and a hostile review, read against the project's real code. Use when asked to write, draft, fix or review a card's spec, when a card's spec_state is none or draft, or when running `dask spec <card>`. Not for… - 90dasks
Rules and commands for working a Dasks card in a Claude Code or Codex session. Use whenever the user mentions a Dasks card id (`<project>.t-<slug>`, e.g. demo.t-hello), pastes a `dask start` prompt, or asks to find, claim, begin, submit, block or release a card with the dask CLI or the dasks MCP. - 95detent-cross-review
Use when a design spec is committed and about to go to the owner, or when the gate is green, the branch is rebased and a PR is about to be opened. Runs `npx -y @ysilvestrov/detent cross-review` by a model of another family and covers the author flag, the background run, the… - 100review-code
Code-review pipeline as one skill. Subcommands review the current branch diff and report bugs as structured findings (review, the default), reproduce and validate each finding to filter false positives (repro), plan + apply fixes after user approval (fix), and process the review… - 100code-improver
Runs an autonomous review-and-fix improvement loop over any code target — a skill, plugin, module, or directory — using a reviewer the user names: any installed skill or agent. Keeps a cross-round findings ledger, escalates when fixes stop converging, and guards scope… - 95cross-model-review
Run an automated cross-model review of the current uncommitted Git changes. Codex reviews and fixes defects, Claude (Opus) reviews design and simplification, coordinated over the A2A protocol, then a quality gate is evaluated. Use only when the user explicitly asks for /cross-model-review.