Dw Check

dw-check — a fast look, then fixes

This is the review step, and it is repeatable mid-build. Run it twice while the diff is small; dw-land's closing verdict stays a thin last look and never grows a reviewer of its own.

What it reads

The diff against the default branch — both the branch and the ref of it resolved the way dw-git does, never origin/ by reflex; ` below is that ref. Plus the change's CHANGE.md for the goal, found by branch the same way dw-next finds it. $ARGUMENTS is read two ways, and this is the one skill in the catalog that mixes them: the single word **codex`** forces the delegated pass past the triviality floor (step 2), and anything else — including whatever follows that word — narrows the focus to a path or a topic.

It writes no .ai/ artifact — approved fixes land as code commits, and a gate you re-run freely doesn't need a report file rotting between runs.

Workflow

1. Establish the diff

git diff ...HEAD plus git log --oneline ..HEAD, narrowed by the focus part of $ARGUMENTS when given — codex is a mode, never a path. Read the CHANGE.md goal so findings are judged against what the change is trying to do, not against taste.

2. Delegate by default, above a floor

Bare delegates. The second model is what this gate adds over reading your own diff again, and gating that on a word you have to remember is how it went unused — so hand the diff to codex:rescue whenever the codex plugin is installed, without asking first.

Two things send the diff to step 3 unassisted, and each costs one line saying so, never a silent skip:

• A trivial diff — 2 files or fewer and under 50 lines touched, from

git diff --shortstat ...HEAD (insertions plus deletions). A typo or a copy edit is not worth the round trip; anything above the floor is.

• No reviewer installed — self-review, plus the fix in that same line:

/plugin marketplace add openai/codex-plugin-cc, then /plugin install codex@openai-codex. Where the plugin is there but not ready, /codex:setup is the fix instead — it checks the CLI and auth. Any other installed reviewer that takes a diff and returns findings folds in the same way, and this skill never depends on one being there.

codex overrides the floor, and only the floor — it forces the pass on a two-line diff, and it cannot conjure a plugin that isn't installed.

Two paths into the plugin, and they are not interchangeable:

• codex:rescue — model-invocable, so this skill can run it itself: the codex:rescue

command forwards to the codex:codex-rescue subagent via the Agent tool. No round trip through the user. This is what bare uses.

• /codex:review --wait — the purpose-built reviewer, and the richer report. It is

disable-model-invocation: true, so no skill can run it — name it for the user to type when the diff deserves the better tool, and wait for the findings to land in the conversation. Its moment is a change that already has an open PR — dw-land opens one before the merge decision, and that window is what the richer pass wants.

Either way the findings fold into step 3's report. codex:rescue returns Codex's output verbatim by its own contract, which pulls against that: quote it verbatim, then verify each finding against the file before it counts. A delegated finding is still a finding that has to point at a real file:line — an outside reviewer earns no exemption from step 3, and its line numbers are the first thing to check.

3. Two axes, judged separately

Whether delegated or self-reviewed, weigh two things and report them side by side, never merged into one score — a fit problem must not hide behind "it's correct":

• Correct? Does the diff do what the goal says? Edge cases, error paths, the empty input, the

call that fails. Anything actually broken.

• Does it fit? Compare against the neighbouring code, not best practice in the abstract. A

pattern used once elsewhere in this repo beats a better pattern used nowhere in it.

Every finding points at a real file:line you opened — if you can't ground it, don't report it. "No findings on either axis" is a normal, useful answer; say it plainly and stop.

4. Present, wait, then fix

List the findings with a severity-ordered recommendation and stop — nothing is fixed without approval. On approval, fix the accepted findings in-session and commit the way dw-git does — related fixes together, unrelated ones apart. Then offer to run again: the second look is the point of a gate this cheap.

Next: dw-next for the next task, or dw-land when the boxes are all ticked.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dominikwozniak
• Source: dominikwozniak/dw-solo-skills
• License: MIT

Install and usage instructions live in the source repository linked above.