SKILL.md is the adaptive workflow protocol. Role profiles remain in agents/roles/ and are passed with delegated tasks or applied to inline execution. The installed commands are:

This directory is the complete Coordinator installation artifact. Copy it unchanged to ~/.agents/skills/coordinator. The command adapters load their bundled standard-library-only runtime from scripts/lib/coordinator, so installation needs no repository checkout, assembly step, or PYTHONPATH configuration.

• scripts/coordinator_state.py for durable workflow and controller state. The normal path is plan-apply → next → node-route-auto → node-claim → node-start → node-complete → workflow-complete. Runtime adaptation adds node-observe, graph-reconcile, graph-expand-auto, judge-gate-add, and judge-complete; low-level mutation commands remain available for reconciliation and recovery. Role packets load from agents/roles/<role>.toml. Closeout requires integrated review coverage for completed artifact work or an explicit review_waiver reason, which is persisted as a review_waived event. Schema-v8 nodes also require descriptive evidence plus positive and negative proof commands; closeout reruns every non-exempt node's pair and succeeds only when positive exits zero and negative exits nonzero.
• scripts/model_router.py choose for direct evidence-based role and runtime-candidate selection when testing or integrating the selector independently.

Project source and documentation: https://github.com/alanhoff/agent-coordinator

Runtime graph payloads, bounds, gate semantics, nested shapes, and logical feedback iterations are documented in references/dynamic-runtime-graphs.md.