blueprint-autonomy-level3

/blueprint:autonomy-level3

Scaffold the ADR-0020 autonomy level 3 pipeline — the highest rung of the blueprint automation model — into a consumer repo: two GitHub Actions workflows plus the deterministic scripts they call. Level 3 runs blueprint out of band (scheduled cron + a label-triggered executor), where levels 1–2 only run in-session. This repo (claude-plugins) dogfoods at level 1 and cannot exercise level 3 itself; a consumer repo opts in.

When to Use This Skill

Context

• Automation config: !bash ${CLAUDE_SKILL_DIR}/../../scripts/get-automation-config.sh
• Blueprint present: !find . -maxdepth 3 -path '*/docs/blueprint/manifest.json'
• Autorun workflow already installed: !find . -maxdepth 3 -path '*/.github/workflows/blueprint-autorun.yml'
• WO-execute workflow already installed: !find . -maxdepth 3 -path '*/.github/workflows/blueprint-wo-execute.yml'

Parameters

Parse $ARGUMENTS:

• --check: audit-only. Diff the installed workflows/scripts against the plugin templates and report PRESENT / DRIFT / ABSENT per file. Make no changes.
• (no args): scaffold/install (idempotent — overwrites the managed files with the current template versions, re-syncing any drift).

Execution

Execute this level-3 scaffold.

Step 1: Verify preconditions

If the Context shows no docs/blueprint/manifest.json, stop with: Blueprint not initialized — run /blueprint:init first. Nothing else.

Read AUTONOMY_LEVEL and WO_AUTO_EXECUTE from the Context automation config — these are the activation gate, reported in Step 4. Scaffolding does not flip them; the workflows stay dormant (the blueprint-wo-guard.sh gate HALTs) until the human sets them.

Step 2 (--check mode): audit drift, then stop

For each managed file, diff the installed copy against the plugin source and report one line each:

for pair in \
  ".github/workflows/blueprint-autorun.yml|${CLAUDE_SKILL_DIR}/../../templates/blueprint-autorun.workflow.yml" \
  ".github/workflows/blueprint-wo-execute.yml|${CLAUDE_SKILL_DIR}/../../templates/blueprint-wo-execute.workflow.yml" \
  ".github/blueprint/blueprint-wo-guard.sh|${CLAUDE_SKILL_DIR}/../../scripts/blueprint-wo-guard.sh" \
  ".github/blueprint/blueprint-wo-packet.sh|${CLAUDE_SKILL_DIR}/../../scripts/blueprint-wo-packet.sh" \
  ".github/blueprint/blueprint-autorun.sh|${CLAUDE_SKILL_DIR}/../../scripts/blueprint-autorun.sh" \
  ".github/blueprint/get-automation-config.sh|${CLAUDE_SKILL_DIR}/../../scripts/get-automation-config.sh"; do
  dst="${pair%%|*}"; src="${pair##*|}"
  if [ ! -f "$dst" ]; then echo "ABSENT  $dst"
  elif diff -q "$src" "$dst" >/dev/null 2>&1; then echo "PRESENT $dst"
  else echo "DRIFT   $dst"; fi
done

Report the table and stop. DRIFT/ABSENT → advise re-running without --check to re-sync (the templates are the source of truth; per code-quality-plugin:code-scaffold-backport, a fix to the plugin template re-lands here on re-run).

Step 3 (install mode): copy templates + scripts

mkdir -p .github/workflows .github/blueprint
cp "${CLAUDE_SKILL_DIR}/../../templates/blueprint-autorun.workflow.yml"   .github/workflows/blueprint-autorun.yml
cp "${CLAUDE_SKILL_DIR}/../../templates/blueprint-wo-execute.workflow.yml" .github/workflows/blueprint-wo-execute.yml
cp "${CLAUDE_SKILL_DIR}/../../scripts/blueprint-wo-guard.sh"    .github/blueprint/blueprint-wo-guard.sh
cp "${CLAUDE_SKILL_DIR}/../../scripts/blueprint-wo-packet.sh"   .github/blueprint/blueprint-wo-packet.sh
cp "${CLAUDE_SKILL_DIR}/../../scripts/blueprint-autorun.sh"     .github/blueprint/blueprint-autorun.sh
cp "${CLAUDE_SKILL_DIR}/../../scripts/get-automation-config.sh" .github/blueprint/get-automation-config.sh

The workflows call the deterministic scripts from .github/blueprint/ (they are blueprint-plugin scripts, not present on a bare CI runner otherwise). The agent-judgment halves invoke anthropics/claude-code-action with the claude-plugins marketplace + blueprint-plugin enabled.

Step 4: Report activation requirements (do not perform them)

Print a receipt listing exactly what the human must do to activate — scaffolding alone leaves everything dormant:

1. Manifest gate — set in docs/blueprint/manifest.json:
   • automation.autonomy_level: 3 (both workflows), and
   • automation.work_orders.auto_execute: true (executor only). Report the current values (from Context) and whether the gate is met. Optional budget caps default to max_per_run 1 / max_per_day 3 / max_cycles 3 — see REFERENCE.md.
2. Secret — CLAUDE_CODE_OAUTH_TOKEN in the repo's Actions secrets.
3. Labels — work-order-draft (proposals) and work-order-approved (the human relabel that triggers execution). Offer to create the approval label: gh label create work-order-approved --description "Blueprint level-3: execute this approved work order (ADR-0020)" --color 0E8A16.
4. Branch protection (REQUIRED — load-bearing). The executor job runs with broad Bash + contents: write (it must, to run the consumer's arbitrary test suite and push its work branch), and GitHub cannot ref-scope that token to blueprint/wo-*. So the "human PR review is the final gate" guarantee rests on the default branch being protected: require a pull-request review, restrict who can push, and disallow force-push/self-merge. State plainly that without this, a prompt-injected work order could push to or self-merge into the default branch. Report whether the default branch is protected (gh api repos/{owner}/{repo}/branches/{branch}/protection returns 200) and refuse to call the setup "done" until it is.
5. Enable the plugin in CI — the caller repo needs blueprint-plugin available to claude-code-action (marketplace + .claude/settings.json).

Step 5: Summarize the safety model

State the guarantees in the receipt so the human understands what they are enabling (full detail in REFERENCE.md):

• Gating — blueprint-wo-guard.sh HALTs every run unless autonomy_level >= 3 (autorun) and additionally auto_execute: true (executor). Both default off.
• Human approval + provenance — a work order executes only after a human relabels its proposal issue to work-order-approved, and the gate refuses any issue that didn't come through the pipeline (no work-order-draft/ work-order label). The protected default branch + human PR review are the final gate (see requirement 4).
• Budgets + stuck ceiling — per-day cap (RAN_TODAY counted fail-closed) and "same order attempted max_cycles× → stuck → surface to human" (loop-integrity).
• Independent verifier — the executing agent never certifies its own work; the PR's CI suite plus a fresh reviewer agent (the verify job) judge "done", and each iteration writes a state-packet issue comment.
• Untrusted-input safe — the issue body is the WO carrier and is untrusted; it is bound to an env var, written to a file, and parsed by blueprint-wo-packet.sh — never interpolated into a shell command.

Agentic Optimizations

For the full workflow templates, the manifest gate/budget schema, the state-packet fields, and the security rationale, see REFERENCE.md.