extend-harness

name	extend-harness
description	Extend frontend-skills hook harness. Add new rules via skill-manifest.json, pick severity tier (block/warn/nudge/info/block-strict/diagnostic), view analytics via /frontend-skills-stats. Use when adding enforcement, tuning rule severity, or debugging hook that isn't firing.

Extend the Harness

1. Never hand-edit generated configs

.claude/settings.json and hooks/hooks.json generated. Edit skill-manifest.json, regenerate:

bash scripts/generate-hook-configs.sh --apply
bash scripts/generate-hook-configs.sh --check   # drift check (lefthook runs this pre-push)

2. Add a new rule (grep-expressible)

1. Write my-check.sh in .claude/hooks/ -- start from any existing *-check.sh as template.
2. Add filename to matcher block in skill-manifest.json (usually PostToolUse.Edit|Write).
3. Regenerate: bash scripts/generate-hook-configs.sh --apply.
4. Test: feed synthetic edit event on stdin:

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

3. Pick the right severity tier

Function	Exit	Claude sees	Logged	Use when
hook_block	2	systemMessage	JSONL	Must fix before continuing
hook_block_strict	2	[STRICT] prefix	JSONL	Security-critical, no escape hatch
hook_warn	0	systemMessage	JSONL	Should fix, but proceed
hook_nudge	0	[nudge] prefix	JSONL	Pattern suboptimal most cases
hook_info	0	--	JSONL	Telemetry only, no UI
hook_emit_diagnostic	0 or 2	LSP JSON with range + fix	JSONL	Machine-parseable with auto-fix

Default hook_warn for style, hook_block for correctness, hook_info for observation.

4. When grep isn't enough

Grep can't express nested interactives, exhaustive switches, or useState-object-ref leaks reliably. AST-level patterns: handle in code review for now, or file issue for future Biome custom-rule integration. Don't fake with fragile multi-line regex -- too many false positives.

5. View analytics

/frontend-skills-stats

Reports per-hook P50/P95 latency, blocks/warns/nudges/infos per rule, zero-fire candidates (prune), over-aggressive hooks (soften). Needs >=5 session summaries in ~/.claude/hook-metrics/.

6. Debug a hook that isn't firing

HOOK_DEBUG=1 HOOKS_FAIL_CLOSED=1 claude

Then: tail -f /tmp/hook-session-*/debug.log. Fail-closed turns crash-into-silent-exit into crash-into-visible-block.

7. Verification checklist

• skill-manifest.json lists script under correct event+matcher
• bash scripts/generate-hook-configs.sh --check exits 0
• Script executable (chmod +x)
• Sources _hook-lib.sh, parses input, filters extension, handles escape hatch
• Test fixture added to evals/ if non-trivial
• bash evals/run.sh passes