AFK automation and loop governance

Goal

Queue and run optional sandboxed work with explicit readiness, autonomy level, budget, and failure controls.

Prerequisites

• a Change eligible for automation
• Docker daemon for container execution
• the sandcastle:local image or an explicitly selected image
• credentials for the selected Codex/Claude runner
• an explicit loop binding when loop policy is required

Interactive workflows do not require Docker. Missing optional AFK prerequisites should appear as visible degradation, not invalidate the core installation.

Check readiness

tenon doctor --json
tenon afk status --json
tenon loops status --json

Doctor never prints credential values. It reports whether each supported runner has a usable source.

Enqueue a Change

tenon afk enqueue <change-name>

Bind it explicitly to a loop:

tenon afk enqueue <change-name> --loop <loop-id>

PM's built-in policy can enqueue after Spec completes. Enqueueing does not automatically start Docker or a runner.

Run safely

L1 is the default report-only level:

tenon afk run <change-name> --level L1

Use another image only when it is trusted and prepared:

tenon afk run <change-name> --level L1 --image sandcastle:local

Cancel an active job:

tenon afk cancel <change-name>

Cancellation records intent and terminates the owned container. Inspect final status and preserved artifacts before retrying.

Loop lifecycle

Create a paused draft:

tenon loops init \
  --id nightly-fix \
  --goal "Repair flaky tests" \
  --runner codex \
  --yes

Or use a starter template:

tenon loops init \
  --id ci-loop \
  --template ci-sweeper \
  --skill-bundle backend \
  --yes

Inspect policy:

tenon loops list --json
tenon loops status --json
tenon loops enforce --loop <loop-id>
tenon loops budget <loop-id>

Autonomy levels:

Level	Meaning
L1	report-only safe default
L2	assisted execution within admitted policy
L3	explicitly graduated unattended execution

Changing level requires admission and explicit confirmation:

tenon loops level <loop-id> set L2 --confirm

Directed loop execution:

tenon loops run <loop-id> --dry-run --level L1 --json

Do not use --commit or L3 until sandbox, allowlists, verification/review gates, budget, credentials, and recovery have been reviewed.

Enforcement

Loop governance evaluates R1–R11 policy, including budget, maximum in-flight work, failure and dry-run streaks, admission, and kill/warn decisions. Level labels do not bypass normal Workflow gates or Skill/document evidence.

Expected result

Queued, running, waiting, cancelled, and failed states remain distinct; every worker action is bounded by the selected loop/autonomy policy.

Verification

tenon afk status <change-name> --json
tenon loops enforce --loop <loop-id>
tenon loops status --json

Common failures

Queued but not running

Check Docker, image, runner credential, loop admission, budget, and concurrency. Queue state alone does not launch a worker.

L2/L3 graduation rejected

Read the enforcement verdict. Do not edit the registry to bypass failed admission.

Worker failed

Preserve logs and artifacts, inspect the terminal reason, then retry only after the underlying condition changes.

Next action

Read advanced tools or troubleshooting.