Paul Huliganga
|
5db61dd321
|
feat: agent harness v2.0 — wave-based agentic project methodology
A complete system for running autonomous AI coding agents on complex projects.
Proven in practice: Fintrove project — 4 waves, 11 streams, 44 tasks, 1,254→1,597 tests, 0 regressions.
Core templates:
- AGENT-INSTRUCTIONS.md — agent system prompt, core loop, commit attribution
- PROJECT-SPEC.md — project definition template
- DECISIONS.md — Architecture Decision Records
- EXECUTION-BOARD-TEMPLATE.md — stream planning artifact (write before coding)
- VALIDATION-TEMPLATE.md — per-packet evidence
- PROCESS-EVAL-TEMPLATE.md — stream retrospective
Process guides:
- WAVE-BASED-MANAGEMENT.md — plan-then-implement discipline, wave gates, known-answer tests
- SPEC-CREATION-GUIDE.md — interview protocol for building specs
- PLAN-MANAGEMENT.md — living IMPLEMENTATION_PLAN.md
- REVIEW-AND-QA.md — evaluating agent output
- PARALLEL-AGENTS.md — running multiple agents simultaneously
- COST-OPTIMIZATION.md — getting more work per dollar
- OPENCLAW-INTEGRATION.md — sessions_spawn, cron, automation
- TROUBLESHOOTING.md — five failure modes + recovery
- TUTORIAL.md — 30-min hands-on walkthrough
- EXAMPLES.md — real project examples
Tooling:
- ralph-loop.sh — bash loop for Claude/Codex/Aider/Gemini
- model-report.ts — per-model quality reporting from git trailers
|
2026-04-01 21:20:26 -04:00 |