Workflow guide

Dispatch bounded work with a meaningful gate

Name the account, isolate implementation and distinguish an agent claim from accepted evidence.

Pre-release. Reviewed November 3, 2026. This review covers code behavior, not a packaged release or provider certification.

Before you start

Use a repository you control, a signed-in named account and a Loadout/policy appropriate for the task. Read the installed commandem autopilot --help. The example below starts real work and consumes provider quota; it is not an automatic website test.

Define the deliverable before dispatch

For this example, the deliverable is a nonempty docs/acceptance.md with an Acceptance heading. The gate checks those structural conditions, not the quality of every proposed criterion. Add your repository's actual tests and review before treating the work as release-ready.

commandem autopilot PROJECT \
  "Create docs/acceptance.md with an Acceptance heading and project-specific checks. Do not change existing code." \
  --agent codex --account demo-codex --worktree --max-turns 3 \
  --verify 'test -s docs/acceptance.md && rg -q "^# Acceptance" docs/acceptance.md'

Replace PROJECT and the example account. This command requires rg. It requests a reviewable worktree; isolation is explicit, not a promise about every Commandem launch path. Confirm the launch only after checking authority and quota.

Inspect the outcome

  1. Open Monitor or inspect commandem queue --json.
  2. Inspect commandem queue --receipts --json. Confirm provider/account, worktree, stop reason and verification evidence.
  3. Review the diff and the meaning of the gate. A generic passing build on an unchanged repository is not proof that a deliverable exists.
  4. Merge only after the requested acceptance and project checks pass. The website example does not automatically merge anything.

Read the foreground exit status

Autopilot now exits non-zero when a backstop stops an unmet goal: 3 for the cost cap, 4 for the turn cap, 5 for the turn cap after rejected verification, 6 for a blocked goal, and 130 for an observed human stop. Errors use 1; existing usage refusals use 2. A later successful completion after a rejected check still exits 0. Zero records the loop's declared completion condition, not independent proof that the requested work exists.

Receipt reasons and JSON fields are unchanged; a verifier-exhausted receipt still has reason=max-turns and its rejection count. Inspect the receipt and diff. A detached launch returning zero confirms admission only; it does not report the child's outcome. No approval or verifier guard is weakened.

Intent

Specify deliverable, account, authority and bounds.

Execution

Inspect the isolated changes and run state.

Acceptance

Check evidence and limitations, then decide.

Recover without inventing success

A failed gate, budget stop or infrastructure error stays a failure or pending decision. Fix the named cause and review a retry; do not erase the original receipt. Quota and model capacity are different conditions. See recovery guidance and how experiments interpret outcomes.