Action Contract & Harness Repo Maintainer

Build and maintain repo-local action contracts and harnesses where agents execute and humans steer. The historical mcp- name remains because many adopters arrive through MCP work, but this skill is not MCP-only. For general app, library, tool, plugin, or meta-repo stewardship baselines, use repo-quality-system-lifecycle first; use this skill only when typed actions, probes, benchmarks, or CLI/MCP parity are in scope.

Core principle (action-contract and harness repos)

MCP and CLI are thin interfaces—APIs for agents and CI. Core contains the real logic, schemas, and registries. Adapters parse wire format (argv, MCP JSON-RPC); they delegate immediately.

Agents / CI  →  CLI ──┐
                      ├──► Core (logic, contracts, tests)
Agents / chat →  MCP ──┘

Full layering: core-and-interfaces.md. Parity: every MCP tool must call the same core entrypoint as its CLI twin.

Progressive Automation (Agent-Driven Workflows): Harnesses should let agents turn repeated friction into reviewed, durable capability. If an agent discovers a complex fix or command sequence, it should capture an unknown case or typed action candidate with owner, risk class, inputs, outputs, effects, provenance, and verification. Permanent steward.yaml changes must go through reviewable diffs and validation; do not teach agents to save raw bash permanently from MCP.

Goal-first adoption: The original user goal remains the acceptance check. Tool repair, install work, wrappers, action candidates, evals, and refactors are detours unless they directly solve that goal or preserve a reusable lesson. After two failed repair/setup attempts, stop tool restoration, use a type-native command or portable fallback when possible, record the friction, return to the task, and do not promote from that same detour.