README
01Agent Claude Code
Time de 7 subagents especializados pra desenvolver o Harness MCP. Cada agente é dono de um bloco da SPEC.md.
Agent Claude Code
Time de 7 subagents especializados pra desenvolver o Harness MCP. Cada agente é dono de um bloco da SPEC.md.
Agent Claude Code
Use proactively when starting new components, making cross-cutting tech decisions, resolving conflicts between subsystems, or whenever a change touches more than one module of anyharness. Owns coherence with SPEC.md, the positioning vs @dotcontext/cli, and the design principles in §5.
Agent Claude Code
Use when designing or implementing Contracts — Specs, Tasks, contract.spec.validate, contract.task.next, contract.task.complete, the resource handler for harness://contracts/specs/{id}, or anything in src/contracts/ and .harness/contracts/. Owns the behaviour harness layer defined in SPEC §10.
Agent Claude Code
Use when implementing judge.list, judge.review, the Anthropic-backed inferential runner, rubric prompt management, output JSON validation, fixture regression, or anything in src/judges/ and .harness/judges/. Owns the inferential feedback side defined in SPEC §3, §4, §7, §9.
Agent Claude Code
Use when implementing or modifying the MCP server core — stdio transport, capabilities declaration, tool/resource/prompt registry, request handlers, file watcher wiring, error envelopes, or anything in src/server/ and the cross-cutting plumbing. Owns the surface defined in SPEC §3 and §4.
Agent Claude Code
Use when implementing sensor.list, sensor.run, sensor.register, sensor adapters (eslint, tsc, semgrep, ArchUnit), output normalization, or anything in src/sensors/ and .harness/sensors/. Owns the computational feedback side defined in SPEC §3, §4, §9.
Agent Claude Code
Use when implementing session.start, session.append, session.get, the steering log consumer, harness.steer.suggest, or anything in src/sessions/ src/steering/ and .harness/sessions/ .harness/.local/steering/. Owns persistence of execution state and the human-in-the-loop steering loop defined in SPEC §8 and §11.
Agent Claude Code
Use when setting up test infrastructure, writing integration tests with a real MCP client (Claude Code, Cursor), e2e tests of the server, mock LLM runner, fixture management, or anything in tests/. Owns dogfood-of-Harness MCP testing.