blueprint-autopilot

The level-2 executor of the ADR-0020 autonomy model: runs due agent-judgment maintenance tasks and (optionally) drafts work-order proposals — quietly, with a one-line receipt. The inverse of /blueprint:execute: no menus, no questions, act then report.

When to Use This Skill

Context

• Automation config: !bash ${CLAUDE_SKILL_DIR}/../../scripts/get-automation-config.sh
• Due tasks (report mode, no writes): !bash ${CLAUDE_SKILL_DIR}/../../scripts/blueprint-autorun.sh --report

Execution

Execute this autopilot pass:

Step 1: Gate on autonomy level

Read AUTONOMY_LEVEL from the Context automation config. If it is below 2, stop immediately with the single line Autopilot inactive (autonomy_level N) — raise automation.autonomy_level in docs/blueprint/manifest.json to enable. Nothing else. Also stop (silently, one line) when BLUEPRINT_AUTORUN_DISABLE=1 is set in the environment.

Step 2: Run due agent-judgment tasks (budget: 2 per pass)

From the Context --report output, collect the DUE_AGENT_TASKS list. Take at most the first 2 (session token budget — the rest surface again next session). For each task, dispatch ONE subagent (Task tool) with a focused brief:

• adr-validate → run the /blueprint:adr-validate --report-only procedure; return the findings summary only
• feature-tracker-sync → run the full-sync reconciliation in report-then-apply form, applying only unambiguous updates (drain evidence-backed completed WOs, recompute stats); ambiguous discrepancies are reported, not resolved
• story-audit → run the audit and write only its dated artifact under docs/blueprint/audits/
• any other task → run its skill's --report-only/dry-run form; never its mutating form

Each task's brief MUST state: the run is unattended, so never AskUserQuestion and never stop at "I'll now run X" — make the call and carry the task through to its report; and return a one-line result stating what the tool calls actually returned (counts, paths, exit status). Tasks whose skills would require interactive confirmation for writes stay report-only here — autopilot never bypasses a write gate.

After a task's subagent returns successfully, record completion in the manifest:

jq --arg task "<task>" --arg ts "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
   '.task_registry[$task].last_completed_at = $ts |
    .task_registry[$task].last_result = "ok (autopilot)" |
    .task_registry[$task].stats.runs = ((.task_registry[$task].stats.runs // 0) + 1)' \
   docs/blueprint/manifest.json > docs/blueprint/manifest.json.tmp \
   && mv docs/blueprint/manifest.json.tmp docs/blueprint/manifest.json

On a failed subagent, record last_result = "error: <one-line reason>" but do NOT update last_completed_at (the task stays due).

Step 3: Work-order auto-draft (only when WO_AUTO_DRAFT=true)

Skip this step entirely unless the Context config shows WO_AUTO_DRAFT=true. Work-order creation stays human-only (/blueprint:work-order keeps disable-model-invocation: true — never invoke it from autopilot). Autopilot may only file proposals:

1. Find ready PRPs: scan docs/prps/*.md frontmatter for confidence: 9 or higher (the confidence-scoring bar for delegation).

2. Exclude PRPs that already have a work order or draft: a WO exists when the manifest id_registry maps the PRP to a WO-* document, when a tracker tasks.pending entry carries source: <PRP-id>, or when an open GitHub issue labeled work-order-draft or work-order names the PRP id:

   gh issue list --state open --label work-order-draft --limit 100 --json number,title
   gh issue list --state open --label work-order --limit 100 --json number,title
   

3. Cap: if ≥ 5 open work-order-draft issues exist, draft nothing and note the cap in the receipt.

4. File the draft for each remaining ready PRP (at most 2 per pass): a GitHub issue titled [work-order-draft] <PRP-id>: <PRP title>, labeled work-order-draft, whose body is the full work-order packet built from the PRP (objective, minimal context, TDD requirements copied verbatim, success criteria, file list) plus the promotion instruction: Promote with /blueprint:work-order --from-issue <N>. Create the label first if missing (gh label create work-order-draft --description "Auto-drafted work-order proposal (ADR-0020)" --color FBCA04).

5. No GitHub remote / gh unauthenticated: write local draft files to docs/blueprint/work-orders/drafts/NNN-<slug>.md instead (same packet, status: draft frontmatter). Never touch tasks.pending or the id_registry — those mutate at promotion time only.

Step 4: One-line receipt

End with exactly one summary line, e.g.: Autopilot: ran adr-validate (2 findings), feature-tracker-sync (1 WO drained); drafted 1 work-order proposal (#214). If nothing was due and nothing drafted: Autopilot: nothing due.

Guardrails

• Never AskUserQuestion; never end on a question. Autopilot is fire-and-report.
• Never invoke /blueprint:work-order or /blueprint:prp-execute (human-only by design; the draft-issue side channel is the only WO surface here).
• Bounded per pass: ≤ 2 agent tasks, ≤ 2 drafts, draft cap 5 open issues — no loops, no self-continuation (see .claude/rules/loop-integrity.md).
• enabled: false tasks never run regardless of level.
• Drafts are additive and closable; closing a work-order-draft issue is a valid human veto and autopilot re-drafts only if the PRP still qualifies AND no open draft exists (a closed draft for the same PRP id counts as a veto — check --state all before re-drafting and skip PRPs with a closed, unpromoted draft).

Agentic Optimizations