extend-harness

Edit manifests/libraries, not generated configs. REFERENCE.md: severities, options, parser contracts, debugging.

Claude mods: MODS.md. Cross-runtime enforcement stays in settings hooks.

Evidence bar

Compose existing skills before adding. Justify rules with a real failure or unmet contract; prove clean -> violation -> clean.

Add rule

1. Prefer Biome/Ultracite; hooks cover cross-element/file, workflow, agent behavior.
2. Copy .claude/hooks/checks/*.lib.sh: one run_* function and thin .claude/hooks/*.sh wrapper.
3. Register in skill-manifest.json, usually PostToolUse.Edit|Write.
4. Add focused evals/ fixture; prove RED -> GREEN.
5. Regenerate and test:

bash scripts/generate-hook-configs.sh --apply
echo '{"hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/tmp/x.ts"}}' | bash .claude/hooks/my-check.sh

Use hook_warn for style, hook_block correctness, hook_block_strict security, hook_info observation. Prefer skill-scoped hooks for one vertical.

Implementation

• Manifest objects carry permission/async filters; retain stdin guards: Codex drops Claude-only filters.
• Structure: Biome/AST; judgment: review.
• Sync toolchain bans with hooks/frontend-skills.rules.

Audit/debug

/hook-audit --all: latency/firing/zero-fire; HOOK_DEBUG=1 HOOKS_FAIL_CLOSED=1 claude: missing hooks; claude --safe-mode: isolate customization. /doctor latency is a P95 budget failure.

Done

Manifest owns matcher; executable script sources _hook-lib.sh, parses stdin, filters paths, documents escape; fixture proves RED -> GREEN; generator --check and bash evals/run.sh pass.