Review with Crit

Run this interactive browser review cycle only when the user invokes /skill:crit or directly asks to use Crit. A generic request to review code, a plan, a diff, a PR, or a page does not count.

1. Choose the target

Pass the user's arguments through. Crit auto-detects files, directories, live URLs, and local HTML previews; do not ask which mode to use.

crit <arguments>
crit --pr <number-or-url>
crit --mr <number-or-url>
crit --range <base>..<head>
crit

Without arguments, review the plan file written earlier in this conversation, if there is one; otherwise run bare crit for the branch diff.

2. Start the review and await completion

Crit opens the browser and stays running until the human clicks Finish Review. Use omp's bash tool to launch the command chosen in Step 1 as an asynchronous job with no deadline:

{"command":"crit","async":true,"timeout":0}

Save the returned job ID. timeout: 0 disables the default five-minute command deadline; async: true alone does not. If asynchronous jobs are unavailable, run the same command in the foreground with timeout: 0. Do not use a named service: each review round is a finite command that exits when feedback is ready.

Relay the printed review URL verbatim: "Crit is open at <URL>. Leave inline comments, then click Finish Review." If the initial job result omits the URL, read proc://<job-id> once for startup output.

Use wait when no independent work remains. Completion also auto-delivers. If a wait returns a running snapshot or an unrelated message, keep waiting for this job's completion; neither means the human finished reviewing. Do not poll process output or the review file, hold an eval cell open, launch a second Crit command, or ask the user to type a reply to signal completion. An existing daemon is reused automatically.

3. Read feedback

After the command exits, read its output and follow the finish prompt. Check for approved: true or approved: false; omp merges stdout and stderr. Retrieve omitted output through proc://<job-id> or the output's artifact:// link. If approved, stop the review loop. A command failure is not approval; report the error rather than proceeding.

For mid-round re-entry or headless workflows, use crit comments --json (or crit comments --plan <slug> for plan reviews).

Review-level comments matter as well as file comments. Read existing replies. Treat missing resolved as unresolved. Use quote to narrow the requested change, anchor to locate content after edits, and drifted: true as a sign that line numbers are approximate.

4. Address comments

Revise the referenced plan or source files and apply suggestion blocks when present. Reply with what changed:

crit comment --reply-to <id> --author 'omp' '<what changed>'
crit comment --plan <slug> --reply-to <id> --author 'omp' '<what changed>'

Do not pass --resolve unless the user explicitly asks; resolution belongs to the reviewer. For multiple replies, write JSON with the write tool:

crit comment --json --file /tmp/crit-replies.json --author 'omp'

Each entry has reply_to and body. Use the crit-cli skill for the full schema and session disambiguation. Plan edits reload live in the browser.

5. Start the next round

Run the next-round command printed by Crit using the same asynchronous bash pattern with timeout: 0. It signals round-complete and then waits for the next Finish Review.

Tell the user only: "Replied in Crit. Review the changes in your browser and click Finish Review when ready." Do not repeat your replies or list the comments in chat. The reviewer reads them in Crit. Await command completion, then return to Step 3. Continue until the reviewer approves.

Sharing

Only share or unpublish when the user asks:

crit share <file> [file...]
crit unpublish [file...]

Relay the full printed URL. Use --qr only in real monospace terminals. For another device, keep Crit on loopback and use an SSH tunnel or reverse proxy. --public-url requires --allow-unauthenticated-network and changes the printed URL, not the listener. Crit has no network authentication; confirm the user wants that exposure before making it reachable remotely.