Blueprint Engine

Blueprint Engine — Hybrid DAG Primitive

Inspired by Stripe Minions: deterministic scaffolding + bounded agentic intelligence. Hard caps are non-advisory — they are enforced by iteration-counter.py.

Node Types

Deterministic Nodes (No LLM — Always Idempotent)

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 HEAD
• create-pr — gh pr create --title "minion: {task}" --body "{template}" --label "minion-generated"
• lint — pnpm turbo run lint
• type-check — pnpm turbo run type-check
• test — pnpm turbo run test
• circular-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)

Agentic Nodes (LLM Reasoning — Hard Caps Apply)

• implement — Implement the requested feature. Cap: 1 pass only
• reproduce-bug — Write a failing test that reproduces the bug. Cap: 1 pass only
• implement-fix — Fix the reproduced bug. Cap: 1 pass only
• implement-migration — Execute a database or code migration. Cap: 1 pass only
• implement-refactor — Refactor per system supervisor audit findings. Cap: 1 pass only
• fix-ci — Fix CI/test failures. Cap: 2 rounds maximum
• fix-lint — Fix non-auto-fixable lint errors. Cap: 1 round maximum

Hard Iteration Caps (Non-Advisory)

When any cap is reached → output BLUEPRINT_ESCALATION block, halt, write to .claude/memory/current-state.md, stop.

Auto-Fixes (Deterministic — Zero Iteration Cost)

These are applied automatically before any agentic node. They do NOT count against caps:

Escalation Format

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}

Execution Rules

1. One-shot mandate: Execute the full blueprint DAG without asking clarifying questions.
2. Context scoping: Read ONLY files in the pre-hydration manifest. No additional file reads.
3. Iteration counting: Every agentic node execution increments the counter in .claude/data/minion-state.json.
4. Human review gate: NEVER auto-merge PRs. The blueprint always ends at create-pr.
5. en-AU locale: All PR bodies, commit messages, and output use Australian English.
6. Deterministic first: Always attempt auto-fixes before invoking any agentic node.

Blueprint DAG Notation

node-type → node-type → [optional-node] → node-type

Brackets [...] indicate conditional nodes (only executed if the preceding deterministic check fails).