llms.txt

# Orchestration Guide by DojoGenesis

> A progressive curriculum for Claude Code orchestration mastery — from single-repo builds to multi-repo infrastructure.

Extracted from 208 sessions, 432 hours, and 469 agent dispatches across 6 repositories over 18 days of production use (March-April 2026).

## What this teaches

Five-level progressive curriculum for learning to orchestrate infrastructure builds and multi-repo GitHub work using Claude Code:

- **Level 0 — Foundation**: CLAUDE.md as project constitution, TodoWrite for tracking, root-cause debugging
- **Level 1 — Single-Repo Mastery**: Compilation gates, granular visibility, seed extraction
- **Level 2 — Multi-Agent Dispatch**: Parallel dispatch lifecycle, file manifests, agent verification
- **Level 3 — Multi-Repo Orchestration**: DAG sequencing, interface contracts, convergence gates
- **Level 4 — Ecosystem Governance**: Seeds as memory, tiered authority, scheduled maintenance

## Key concepts

- **Directed Acyclic Graph (DAG) of deliverables**: Sequence work by dependency depth, not urgency
- **Compilation Gate**: Mechanical trust boundary — `go build ./...` is the only agent whose word you trust
- **Zero-Overlap File Manifests**: Each agent owns explicit files, no two agents touch the same file
- **Convergence Gate**: Every 3-4 sessions, stop building and verify (drift detection, ledger reset)
- **Friction-to-Seed Pipeline**: Every debugging session should extract at least one reusable pattern
- **Sonnet/Opus Split**: 80% Sonnet (execution) / 20% Opus (architecture + synthesis)

## 12 orchestration patterns

Pipeline, Map-Reduce, Hierarchical, Swarm, Generator-Critic, Adversarial, Jury, Blackboard, Chain-of-Responsibility, Circuit-Breaker, Event-Driven, Meta-Pattern Selector

## First session script

8 steps for a new user's first orchestration session:
1. Read/create CLAUDE.md
2. Run build + test suite
3. Pick a 3-5 file task
4. Write TodoWrite plan
5. Complete it yourself first
6. Reset and redo with 2-3 agents
7. Write one seed
8. Commit with conventional prefix

## Mechanical enforcement

Hooks infrastructure (PreToolUse blocks, PostToolUse informs):
- destructive-guard, force-push-guard, pre-commit-build-gate (blocking)
- go-build-check, go-vet-check, go-test-check, svelte5-lint, rust-check, ts-check, skill-validate (informational)
- milestone-gate: blocks `git tag` unless full test suite passes
- agent-model-enforce: warns on agent dispatch without explicit model parameter
- drift-detector: YELLOW/RED signals based on uncommitted files and sessions since convergence

## Related resources

- [DojoGenesis/plugins](https://github.com/DojoGenesis/plugins) — 83 first-party skills across 8 plugins
- [DojoGenesis/cli](https://github.com/DojoGenesis/cli) — Dojo CLI with skill management and CAS bridge
- [DojoGenesis/gateway](https://github.com/DojoGenesis/gateway) — Agentic Gateway with DAG workflow engine