Generate multi-agent Prophet Understand skill manifests

[mdheller]

Purpose

Make agent-registry the source of truth for Prophet Understand / Repo Intelligence agent skill manifests.

Parent epic: SocioProphet/prophet-platform#407

Required capability

Generate portable agent-facing instructions and manifests for codebase understanding across supported surfaces: Copilot, Codex, OpenClaw, Claude-style tools, Cursor-style tools, Gemini-style tools, OpenCode-style tools, and VS Code surfaces where relevant.

Initial implementation path

1. Add docs/PROPHET_UNDERSTAND_AGENT_SKILLS.md.
2. Define a single registry contract for the prophet-understand skill.
3. Generate or document target manifests for:
   • repo scan
   • graph validation
   • guided tour
   • PR impact analysis
   • graph-aware fix proposal
   • evidence-preserving review
4. Each generated skill must require:
   • bounded scope
   • no secret exfiltration
   • source anchors
   • validation evidence
   • policy status preservation
   • no invented graph facts
5. Add fixtures for at least two agent surfaces.

Acceptance criteria

• There is one canonical skill definition that can be rendered into multiple agent tool surfaces.
• The manifests point to the Prophet Understand artifact contract rather than inventing local schemas.
• Graph facts must include evidence/source anchors or be marked as inferred with confidence.
• PR review skills must emit affected nodes/edges/tests/docs/policies when available.

Non-goals

• Do not create agent-specific schema forks.
• Do not grant agents mutation authority based only on repo graph output.
• Do not vendor third-party agent plugin systems as trust roots.