harness-maintenance - SKILL.md Agent Skill

Harness Maintenance

You maintain Harness framework guardrails without turning ordinary research code debugging into framework-policy work. 1. Read `AGENTS.md`, `CLAUDE.md`, `tooling/codex_hooks/README.md`, `tooling/codex_hooks/Lightweight_Hook_Policy_Guide.md`, `schemas/skill_contracts.json`, `schemas/skill_contracts.schema.json`, and `tooling/.tests/test_codex_hooks_contracts.py` before changing hooks or contracts. 2. Use this skill for Codex hook runtime, hook trust/status scripts, skill contracts, evidence tooling guardrails, schema validation, trigger/routing behavior, `.agents/skills/**`, `.agents/references/**`, `.claude/Workflow_Guide.md`, `.claude/skills/**`, `.claude/rules/**`, `.claude/shared/**`, bootstrap `templates/**`, permission boundaries, and guardrail tests/docs. 3. Do not use this skill for ordinary implementation changes under `src/`, `scripts/`, or `configs/`; route those to `/code-debug` or `/code-expert`. 4. Keep `.agents/` and `.claude/` behavior semantically aligned when routing or skill rules change. 5. Preserve fail-fast behavior. Do not hide missing state with broad fallback behavior. 6. Use `.claude/shared/ubiquitous-language.md` before changing workflow terms, operator guidance, or application-codebase vocabulary rules. Distinguish Conclusion Evidence from Gate Evidence. 7. Treat hooks as runtime guardrails, not proof that a workflow gate passed. 8. After Python hook/test edits, run `python -m py_compile` and `ruff check --select=E,F,I` on modified Python files. 9. After contract, hook, generator, or detection changes, run `python tooling/codex_hooks/check_contracts.py --workspace-root .` and `pytest tooling/.tests/test_codex_hooks_contracts.py`. 10. When validating a live Codex install, run `python tooling/codex_hooks/hook_status.py --workspace-root . --trust-status` and report `NOT_RUN` if unavailable. 11. Report a Gate ledger for hook, skill contract, routing, permission, generator, or trust behavior changes. Summarize the changed guardrail boundary, tests run, remaining risks, and the Gate ledger.