COORD-Harness AGENTS.md

CoordHarness agent entrypoint

COORDHARNESS_AGENT_INSTRUCTION_SENTINEL=v1

CoordHarness coordinates multi-agent work through one local SQLite authority. coord.db owns work, claims, leases, handoffs, and lifecycle events. Markdown provides stable orientation only; dashboards, job sidecars, knowledge.db, facts, KFTS, and accepted memory are evidence or recall, never lifecycle state.

Start and finish

1. Use $operating-coordharness (repo skill: .agents/skills/operating-coordharness/).
2. Run .venv/bin/coord onboard --register-clients once, then MCP preflight. If the board is empty, use .venv/bin/coord create with a concrete repository-relative done_signal.
3. Expand with bounded board, next_work, and work_context reads. Use the context lenses only when the active row needs broader history or source evidence.
4. Claim before substantive work. Renew only near lease expiry or a material milestone.
5. Complete only after the declared proof passes. Otherwise park, block, release, or use the typed handoff command with exact row fences.

Do not edit SQLite or lifecycle projections by hand. Preserve unrelated dirty work and never weaken identity, proof, review, or authority checks to make a transition pass.

Context hierarchy

• Clean-room installation and verification: docs/agent-onboarding.md
• Claims, reviews, events, and handoffs: docs/agent-protocol.md
• Bounded board and federated retrieval: docs/context-architecture.md
• Facts, knowledge, and memory boundaries: docs/context-and-memory.md
• Durable CPU/GPU processes: docs/jobs-and-runs.md
• Detailed skill references: .agents/skills/operating-coordharness/references/

Read the focused document that answers the current gap; do not load the documentation set or a full board dump by default. Run focused tests for changed surfaces and report exact files, commands, uncertainty, and remaining risk. Do not commit or push unless the user explicitly asks.