review

name	review
description	Run a local multi-agent code review with the commands-com OSS CLI
argument-hint	[objective]
disable-model-invocation	true
allowed-tools	Bash(commands-com *), Bash(git *)

Review

Run a local, single-process review cycle using the commands-com OSS CLI. No desktop app, no account, no daemon — the CLI fans out to available provider CLIs (codex, claude, gemini, or explicit lists), synthesizes findings, and writes a Markdown report under .commands-com/runs/<run-id>/.

Default objective: Review the current working directory for correctness, regressions, and missing tests.

Arguments

The user invoked this with: $ARGUMENTS

Instructions

1. Decide what to review:

   • If the user gave a specific directory, pass it via --cwd <path>.
   • Otherwise use the current directory (the CLI's default).
   • If no objective was provided, look at recent git history first to craft a specific one. Run:

     git log --oneline -5
     git diff HEAD~1 --stat
     

     and base the objective on what changed.

2. Confirm commands-com is on PATH and a provider is available:

   commands-com doctor
   

   To verify auth/response parsing instead of PATH only, run:

   commands-com doctor --ping --json
   

   If no provider CLI is installed, fall back to --provider mock so the round-trip still works (it returns deterministic placeholder findings).

3. Run the review with --json so the result is machine-readable. Default to --changed when the working tree has uncommitted edits worth scoping to; omit it for a broader pass. Use --providers all unless the user asked for a specific provider. This is report-only — one cycle, no edits, no re-review pass.

   commands-com review "<objective>" \
     --cwd "<dir>" \
     --providers <all|codex,claude,gemini|mock> \
     --reviewers correctness,tests,maintainability \
     --json
   

4. If the user asked the agent to fix things (not just report), add --fix. For untrusted edits or large diffs, also add --worktree so the changes land in an isolated git worktree and can be inspected before merging. With --fix the loop becomes review → implement → re-review (and repeats up to --max-cycles):

   commands-com review "<objective>" --cwd "<dir>" --providers <ids> \
     --fix --worktree --test "<validation command>" --max-cycles 2 --json
   

   Heads-up before launching: a --providers all --fix --max-cycles 2 run typically makes ~24 LLM calls and takes 5–10 minutes. Drop providers, reviewers, or cycles to trade depth for speed. If you only have a single provider authenticated, prefer --provider <id> to skip fan-out.

5. Summarize the JSON result for the user:

   • runId and reportPath (the Markdown report).
   • providers, synthesizerProvider, and implementerProvider.
   • The final cycle's issueCount and (if --test ran) unresolvedTestFailure.
   • For --worktree runs: workspace.prune.ok (true means the worktree was auto-removed because no changes were made) or workspace.cwd / workspace.branch (the worktree to inspect manually otherwise).

6. Prior runs can be listed and inspected without re-running the provider:

   commands-com runs list
   commands-com runs show <run-id>
   

Notes

• This skill is local-only and does not require the Commands.com desktop app.
• Provider auth is whatever the underlying CLI already uses (codex login, claude login, gemini auth, etc.) — commands-com only forwards prompts.
• All run artifacts live under .commands-com/runs/<run-id>/ in the target directory and are gitignored by default.