Writing Hooks

Guide for adding hooks to harness/hooks/hooks.json and writing the scripts they invoke.

Hook Types

Type	When to use	Cost
command	Deterministic checks, fast guards, file ops	Low
prompt	Single LLM judgment with {"ok": true/false} response	Medium
agent	Multi-step LLM reasoning with file tools	High
http	POST to external service	Varies

Prefer command whenever the logic can be expressed in any executable script or the Go binary. Reserve agent/prompt for cases that genuinely need judgment — they are slower and cost tokens on every invocation.

Note: agent hooks do NOT work on PreCompact. Use command instead.

Adding a command Hook

1. Write the script

Place scripts in harness/scripts/. Use whichever language fits the task — bash, Python, and Node.js are all fine. The only requirement is that the script is executable and produces the correct exit code and stdout.

Exit codes control the hook outcome:

Exit code	Meaning
0	Allow / pass
2 + JSON to stdout	Block (PreToolUse / Stop)
0 + JSON to stdout	Inject context (PostToolUse, PreCompact)

Bash example — blocking Stop hook:

#!/usr/bin/env bash
if [[ some_condition ]]; then
  echo '{"decision":"block","reason":"Explain why"}'
  exit 2
fi
exit 0

Python example — blocking Stop hook:

#!/usr/bin/env python3
import sys, json
if some_condition:
    print(json.dumps({"decision": "block", "reason": "Explain why"}))
    sys.exit(2)
sys.exit(0)

Node.js example — warning injection (PreCompact):

#!/usr/bin/env node
if (someCondition) {
  console.log(JSON.stringify({ systemMessage: "Warning: something needs attention" }));
}
process.exit(0);  // never block compaction

2. Set execution permission

Scripts without chmod +x fail silently or with permission denied, regardless of language:

chmod +x harness/scripts/your-script.py   # or .sh, .js, etc.

Alternatively, invoke the interpreter explicitly in the hook command (step 3) — then chmod +x is not required.

3. Register in hooks.json

Edit harness/hooks/hooks.json (SSOT). Reference the script via ${CLAUDE_PLUGIN_ROOT} and invoke with the appropriate interpreter:

{ "type": "command", "command": "bash \"${CLAUDE_PLUGIN_ROOT}/scripts/your-script.sh\"", "timeout": 5 }
{ "type": "command", "command": "python3 \"${CLAUDE_PLUGIN_ROOT}/scripts/your-script.py\"", "timeout": 5 }
{ "type": "command", "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/your-script.js\"", "timeout": 5 }

For bin/harness subcommands, use timeout: 3. For scripts, use timeout: 5.

4. Validate

make validate   # runs tests/validate-plugin.sh

---

Checklist for New Hooks

• Script has the correct shebang for its language (#!/usr/bin/env bash, #!/usr/bin/env python3, #!/usr/bin/env node)
• Script is executable (chmod +x) or interpreter is explicit in the hook command (python3 "...", node "...")
• Script path uses ${CLAUDE_PLUGIN_ROOT}/scripts/ — never absolute paths
• Timeout matches type: 3 for bin/harness, 5 for scripts, 30 for agents
• Tested locally by running the script directly with its interpreter
• make validate passes after editing hooks.json

---

Common Issues

Hook not running (silent failure)

Three likely causes:

1. Not executable — check with ls -la harness/scripts/your-script.*. Either chmod +x the file or prefix the command with the interpreter (python3 "...", node "...", bash "...")
2. Wrong path — verify the path resolves correctly; run the script directly to confirm
3. Syntax error — run the script manually (python3 script.py, node script.js, bash -n script.sh) to surface errors before the hook fires

Permission denied

chmod +x harness/scripts/your-script.py   # or .sh, .js
# OR use explicit interpreter in hooks.json — no chmod needed

Hook fires but doesn't block

For blocking hooks (PreToolUse / Stop), you must both print JSON to stdout and exit 2. Printing alone or exiting alone won't block.

agent hook not firing on PreCompact

Agent hooks silently do nothing on PreCompact. Convert to a command script — see harness/scripts/check-wip-precompact.sh as a reference.

---

Event Reference

Key events and their typical use:

Event	Use for
PreToolUse	Guard writes, validate inputs before execution
PostToolUse	Cleanup, logging, tracking after tool runs
Stop	Block session end if work is incomplete
PreCompact	Warn (don't block) before context compaction
UserPromptSubmit	Inject policy context, track commands
SessionStart	One-time session initialization (once: true)

Full event list and response schemas: .claude/rules/hooks-editing.md

---

Related

• harness/hooks/hooks.json — SSOT for all hook configuration
• harness/scripts/ — All command hook scripts
• .claude/rules/hooks-editing.md — Complete event reference and timeout guidelines
• update-config skill — For adding automated behaviors via settings.json