Agent governance in practice

Required startup reading (two rule files, one command contract, the relevant specification), a mission statement that forbids expanding the product into a platform, and an instruction to follow links only when they answer a material question. AGENTS.md · CLAUDE.md

docs/specification/ is the sole authoritative design source, with declared precedence inside it. A spec-scoped AGENTS.md governs work within the specification tree itself. Agents may not maintain a second complete specification — which is why the backlog explicitly labels itself reviewable planning material. Spec agent guide

Planned, implemented, partial, mocked, and measured are distinct states (NFR-010), and Measured is an evidence qualifier that never substitutes for implementation. The acceptance audit enforces the distinction mechanically: a Planned capability's named file being present on disk does not make its rows pass — “the status decides, and presence alone never passes.” Implementation status

The .agents/skills/ (Codex) and .claude/skills/ (Claude Code) directories carry fpat-lite and atomic-commit skills that route into the shared command contracts. Platform files exist for discovery and invocation only; procedures live in commands/, invariants in rules/, artifact shapes in templates/. Agent-system graph

Handoffs write a checkpoint (.fpat/handoff.md); resume flow compares branch, diff, tests, and approvals against it, marks stale claims, re-runs evidence when needed, and continues a single recorded next action — without inheriting old approval when the target changed. The tracked .fpat/plan.md shows the artifact in real use.

The merged-PR trail (visible in the Evolution Timeline) runs epic by epic — data contract, baseline, transformer, evaluation, serving, validation gate, delivery — each landing with its validation evidence. The governance model above is what made that trail auditable after the fact.