Menu
Hybrid DAG execution primitive combining deterministic and agentic nodes with hard iteration caps
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | blueprint-engine |
| version | 1.0.0 |
| description | Hybrid DAG execution primitive combining deterministic and agentic nodes with hard iteration caps |
Inspired by Stripe Minions: deterministic scaffolding + bounded agentic intelligence. Hard caps are non-advisory — they are enforced by iteration-counter.py.
Commands executed directly with no iteration cost:
git-branch — git checkout -b minion/{task-id}git-commit — git add -A && git commit -m "minion: {message}"git-push — git push -u origin HEADcreate-pr — gh pr create --title "minion: {task}" --body "{template}" --label "minion-generated"lint — pnpm turbo run linttype-check — pnpm turbo run type-checktest — pnpm turbo run testcircular-deps-check — pnpm turbo run lint (includes circular dep detection)system-supervisor-quick-audit — read .claude/memory/current-state.md + grep for violations (deterministic)execution-guardian-check — evaluate risk level from task keywords (deterministic mapping)implement — Implement the requested feature. Cap: 1 pass onlyreproduce-bug — Write a failing test that reproduces the bug. Cap: 1 pass onlyimplement-fix — Fix the reproduced bug. Cap: 1 pass onlyimplement-migration — Execute a database or code migration. Cap: 1 pass onlyimplement-refactor — Refactor per system supervisor audit findings. Cap: 1 pass onlyfix-ci — Fix CI/test failures. Cap: 2 rounds maximumfix-lint — Fix non-auto-fixable lint errors. Cap: 1 round maximum| Node | Cap |
|---|---|
implement | 1 pass |
reproduce-bug | 1 pass |
implement-fix | 1 pass |
implement-migration | 1 pass |
implement-refactor | 1 pass |
fix-ci | 2 rounds |
fix-lint | 1 round |
| Total agentic per blueprint | 3 max |
When any cap is reached → output BLUEPRINT_ESCALATION block, halt, write to .claude/memory/current-state.md, stop.
These are applied automatically before any agentic node. They do NOT count against caps:
| Error Pattern | Auto-Fix Command |
|---|---|
Cannot find module | pnpm install |
ModuleNotFoundError | uv sync |
| Auto-fixable ESLint errors | pnpm turbo run lint -- --fix |
| Auto-fixable Python lint | uv run ruff check src/ --fix |
| Missing type stubs | pnpm add -D @types/{package} |
When a cap is reached or a blocking error occurs, output this exact block and halt:
BLUEPRINT_ESCALATION
task_id: {id}
node: {failing-node}
iteration: {n}/{max}
reason: {what failed in one line}
evidence: {last 5 lines of error output}
next_action: Human review required — do not retry automatically
Then write to .claude/memory/current-state.md:
## Minion Escalation — {DD/MM/YYYY HH:MM AEST}
task_id: {id}
status: ESCALATED
node: {failing-node}
reason: {reason}
.claude/data/minion-state.json.create-pr.node-type → node-type → [optional-node] → node-type
Brackets [...] indicate conditional nodes (only executed if the preceding deterministic check fails).
Apply this skill for Unite-Hub Supabase migrations, PostgREST/Data API visibility, founder-scoped Playwright journeys, or errors such as PGRST205, access=denied, stale Supabase linked refs, or migration history drift. Prevents repeating the SQL/cache/auth loop by enforcing the exact verification sequence for core journeys.
The compass for Unite-Hub's road to /shipit. Defines the single NorthStar (a real, comprehensive, working founder CRM in production, every section GREEN), the binding definition of GREEN, and the No-Invaders Manifest that keeps the build honest and surgical. Consult BEFORE deciding what to build/skip/finish — it resolves "200 ≠ real" temptations and scope-creep pressure. P1, auto-loaded.
Apply this skill WHEN scaffolding a new cron "pull" route that syncs external/derived data into Supabase on a schedule (Vercel cron). Encodes the Unite-Hub cron invariants: CRON_SECRET auth, FOUNDER_USER_ID actor, overlap safety, idempotent upsert, last-sync timestamp, and failure surfacing. Generic `cron-scheduler` covers scheduling; this covers the PULL handler body. P3.
Apply this skill WHEN verifying that a route, page, or integration serves REAL data and not silent mock/placeholder data. Detects the "false-green" failure mode: an endpoint returns 200 (or a page renders) while the underlying data is fabricated because a provider is unconnected. Trigger WHENEVER classifying a section's readiness, reviewing integration wrappers, or before marking anything GREEN. P2 — load on audit/verify tasks.
Manifest-first context isolation — each subagent receives only its scope, never the full codebase
Apply this skill for ANY decision with non-obvious tradeoffs: architectural choices, debugging without a clear root cause, performance strategies, security decisions, feature design with competing constraints, refactoring scope decisions. Forces multi-perspective analysis before committing to a solution. P1 auto-load — always active on complex reasoning tasks.