adversarial-review

Adversarial review of local changes

Codex tries to break confidence in a git change: it looks for the strongest reasons the change must not ship. The review is read-only. This skill edits no file; fixing a finding is a separate request.

The diff and the untracked files it lists go to OpenAI. Stop and ask when an untracked file looks like a secret or a dependency tree.

For a proposal that lives in the conversation, not on disk, use critique.

Workflow

1. Parse the arguments. --base <ref> selects the branch diff against <ref>; without it, the target is the uncommitted work (staged, unstaged, untracked). -m <model> selects the Codex model. Everything else is the user's focus text, kept verbatim. <ref> and <model> go into commands inside single quotes, as written below.

2. Check that the target is not empty. A Codex run on nothing wastes minutes, so stop here with the reason when a check fails.

   • Uncommitted: git status --short --untracked-files=all must print at least one line. When it prints nothing, suggest --base <ref>.
   • Branch: git rev-parse --verify --end-of-options '<ref>' must succeed, and git diff --shortstat '<ref>...HEAD' must print a line. Also run git status --short: when it prints lines, tell the user that this uncommitted work is outside the review.

3. Create the run directory. Use the printed path literally in every later step: shell variables do not survive between Bash calls.

   "${CLAUDE_PLUGIN_ROOT}/scripts/codex-run.ts" new review
   

4. Write the prompt with the Write tool to <dir>/prompt.md. User text goes in this file, never inline in the shell command: quotes and backticks break it.

   Run an adversarial code review. Read ${CLAUDE_PLUGIN_ROOT}/skills/adversarial-review/prompt.md
   and follow it exactly.
   
   ## Target
   
   uncommitted            <- or: base <ref>
   
   ## Focus from the user
   
   <focus text verbatim, or: none>