Signing off manual checks

Some criteria only a person can judge: a design review, a deploy, a page that has to look right. Mark them [manual]. Zurdo can't check them, so it never lets a run pass them on its own. A person signs each one off, and the sign-off is logged.

1. Run. zurdo run verifies every automated hint. A task whose remaining criteria are [manual] stops at passed-pending-review.
2. List. zurdo review --no-prompt prints what's waiting, with the automated evidence beside it.
3. Sign. zurdo review --sign-off task:n --by name signs one criterion and appends it to the review log.
4. Pass. Signing a task's last [manual] criterion flips it to passed.

[manual] criteria and passed-pending-review

- [ ] the stack deploys and the site answers [manual]
- [ ] the contact form sends mail [shell: npm run test:e2e] [manual]

• A criterion with only [manual] carries no machine signal. Its task can reach passed-pending-review, never passed.
• Beside automated hints, the automated hints still gate. The [manual] part adds a sign-off on top.
• A task whose criteria are all [manual] reaches passed-pending-review at pre-flight, without calling the agent.
• passed-pending-review satisfies Depends-on, so later tasks don't wait for your sign-off.

Listing what's pending

--no-prompt prints one JSON object instead of opening the review screen. It writes nothing and takes no lock, so it's safe beside an open review session or a running agent.

$ zurdo review prds/launch.md --no-prompt
{"prd":"prds/launch.md","slug":"launch-4c1e","log_valid":true,
 "run_diff":".zurdo/launch-4c1e/run-diff.patch",
 "pending":[{"address":"task-deploy:2","task_id":"task-deploy",
   "task_title":"Deploy the stack","criterion":"the site answers",
   "attempts":1,"automated":[{"criterion":"the stack synthesizes",
   "hints":["shell: npx cdk synth"],"passed":true}]}],
 "coverage_notes":[]}

It needs a prior zurdo run of the PRD; without one it exits 3.

Signing one off

Check the work yourself, then sign:

zurdo review prds/launch.md --sign-off task-deploy:2 --by alice --note "checked www and apex"

{"address":"task-deploy:2","task_id":"task-deploy","criterion_index":1,
 "by":"alice","via":"headless","task_status":"passed"}

The gate is the same as the interactive screen's: the review log validates, the task is passed-pending-review, and the criterion carries [manual] and is still unsigned. A refusal names the address and exits 2, writing nothing. Another process holding the run lock exits 3.

The review log

Every sign-off, interactive or headless, appends a record to .zurdo/<slug>/review-log.jsonl. The records form a tamper-evident chain, and log_valid in the listing says whether it checks out. Headless sign-offs also record "via": "headless" and by; interactive ones carry neither.

Interactively

zurdo review <prd> opens a terminal screen that walks every task's evidence and diff, and signs [manual] criteria with an explicit confirmation naming the task. It holds the run lock for the whole session. Without a terminal and without flags, it prints a pointer to --no-prompt and exits 2.

From an agent (MCP)

zurdo mcp serve publishes the same two operations as tools, answering in exactly the JSON above:

pending_signoffs is read-only. sign_off writes, and the write is just as irreversible. An agent can sign on your behalf only if its MCP client lets it call the tool, so decide who may approve that call.