feat(blueprint-plugin): autonomy level 3 — scheduled out-of-band pipeline with approved work-order execution

[laurigates]

Context

PR #2003 (merged) shipped ADR-0020's autonomy levels 0–2: the manifest 3.4.0 automation block, the level-1 deterministic runner (scripts/blueprint-autorun.sh + SessionStart probe), the on-change feature-tracker sync, quiet interaction_mode, and the level-2 /blueprint:autopilot skill with work-order-draft proposal issues.

Level 3 — the scheduled out-of-band pipeline — was deliberately deferred. It is designed in docs/adrs/0020-blueprint-autonomy-levels.md but not implemented. This issue tracks it.

Scope (from ADR-0020)

Two workflows, following existing repo conventions:

1. blueprint-autorun.yml — daily scheduled run, patterned on the blueprint-health job in .github/workflows/scheduled-audits.yml:

   • deterministic tasks via blueprint-autorun.sh (commit or PR the writeback)
   • due agent-judgment tasks via anthropics/claude-code-action with --model opus --effort pinned per .claude/rules/workflow-model-effort.md
   • dedupe-check → pre-compute → issue/PR pattern; label-based dedupe

2. blueprint-wo-execute.yml — triggered when a human relabels a draft work-order-draft → work-order-approved:

   • reconstruct the WO packet from the issue body (local WO files are gitignored and absent in CI — the issue is the WO carrier)
   • execute TDD-style on a branch, open a PR (Fixes #N); human PR review is the final gate

Preconditions (hard requirements before shipping)

• .claude/rules/loop-integrity.md machinery: an independent verifier judges "done" (CI suite + a fresh reviewer agent — never the executing agent), and each iteration writes a compact state packet as issue comments (objective / ref / files in scope / exit condition / verifier result / changed-since)
• Budgets: per-run and per-day WO execution caps; --max-cycles-style ceiling with "same failure N× = stuck → surface to human"
• Gating: automation.autonomy_level >= 3 AND work_orders.auto_execute: true in the target repo's manifest; both default off
• Security baseline: .claude/rules/github-actions-security.md (least-privilege GITHUB_TOKEN, env-var indirection for issue-body content — the WO packet is untrusted run context)
• A consumer repo that actually wants it — this repo dogfoods at level 1 (constrained mode) and cannot exercise level 3 itself; ship only when a downstream blueprint repo opts in

References

• ADR-0020: docs/adrs/0020-blueprint-autonomy-levels.md (Options Considered Add configure command for README generation #3, Phase 5 in the design)
• PR feat(blueprint-plugin): ambient blueprint operations via autonomy levels (ADR-0020) #2003 — levels 0–2 implementation
• .claude/rules/loop-integrity.md, .claude/rules/workflow-model-effort.md, .claude/rules/github-actions-security.md, .claude/rules/workflow-naming.md

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y98JSyiJLti5ckFnpBwstw