eea36b9f92
Skills are opt-in (I have to remember to invoke them). Hooks are
ambient — they fire on every matching event automatically. This PR
moves the careful-mode and learnings discipline from "doc I should
read" to "harness-enforced behavior I cannot bypass".
## 6 new hooks (.claude/hooks/)
- pre-bash-careful — REFUSES git push --force to main, rm -rf at root,
DROP TABLE against prod schema. WARNs on force-with-lease, gh pr/
issue close. Tested: blocks the destructive case, allows safe ones.
- pre-edit-freeze — implements /freeze. When .claude/freeze contains
a path glob, edits outside it are denied. Tested: edits to PLAN.md
blocked when scope locked to platform/internal/handlers/.
- session-start-context — auto-loads last 20 cron-learnings, freeze
status, open-PR/issue counts as additionalContext at session start.
Tested: emits valid SessionStart JSON.
- post-edit-audit — appends every Edit/Write to .claude/audit.jsonl
(gitignored). One-line records {ts, tool, file, ok}. Tested writes.
- user-prompt-tag — injects context warnings when prompt mentions
force-push, drop-table, "delete all", "push to main", etc. Tested:
emits warning for "force push the fix to main".
- subagent-stop-judge — off by default; touch .claude/judge-subagents
to enable. When on, prompts orchestrator to verify subagent's last
message addresses the original task. Cost-free MVP (no LLM call yet).
All hooks are Python (jq isn't on the hook PATH on macOS — Python is).
Shared helpers in _lib.py: read_input, deny_pretooluse, add_context,
warn_to_stderr.
## settings.json — wires all 6 hooks
Adds SessionStart, UserPromptSubmit, SubagentStop event handlers.
Existing PreToolUse:Bash + PostToolUse:Edit chains gain the new hooks
alongside the existing ones (check-inbox.sh, echo reminder).
Adds @modelcontextprotocol/server-sequential-thinking MCP server for
structured chain-of-thought scratchpad — useful when triaging multiple
PRs in parallel without losing context.
## .claude/commands/triage.md — slash command shortcut
Manual /triage runs the same flow as the c5074cd5 hourly cron, on
demand. Saves ~4KB of prompt every invocation by pulling the cron
prompt out of working memory.
## CLAUDE.md additions
New "Agent operating rules (auto-loaded — read first)" section right
after Ecosystem Context. Documents:
- Cron / triage discipline (read learnings, treat docs PRs touching
CLAUDE.md/PLAN.md as noteworthy, write per-tick reflections)
- Table of all 6 hooks active in this repo
- List of skills and how to invoke them
- Standing rules (inviolable) consolidated for the agent
This block auto-loads into every conversation context — free behavior
change without me remembering to opt in.
## .gitignore
audit.jsonl, freeze, judge-subagents, per-tick-reflections.md are all
local operational state, never committed.
## Verification
- echo '{"tool_input":{"command":"git push --force origin main"}}' |
bash pre-bash-careful.sh → emits deny JSON ✓
- Same for git status (safe command) → empty output, exit 0 ✓
- pre-edit-freeze with .claude/freeze=platform/handlers/ blocks
edits to PLAN.md, allows edits inside the locked path ✓
- post-edit-audit appends valid JSONL ✓
- session-start-context emits additionalContext with PR/issue counts ✓
- user-prompt-tag emits warning for "force push to main" prompt ✓
- python3 -c "json.load(open('.claude/settings.json'))" → valid ✓
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
47 lines
1.4 KiB
Python
Executable File
47 lines
1.4 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""SubagentStop — optional self-check prompt before accepting subagent output.
|
|
|
|
Disabled by default. Enable per-tick with: touch .claude/judge-subagents
|
|
|
|
When on, asks the orchestrator to verify the subagent's output addresses
|
|
the original task. Cost-free MVP — does NOT call an LLM. Future versions
|
|
can plug in an actual llm-judge call gated by a separate toggle.
|
|
"""
|
|
import json
|
|
import os
|
|
import sys
|
|
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
|
from _lib import read_input, emit, warn_to_stderr # noqa
|
|
|
|
REPO = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
|
TOGGLE = os.path.join(REPO, ".claude", "judge-subagents")
|
|
|
|
|
|
def main() -> None:
|
|
if not os.path.isfile(TOGGLE):
|
|
return
|
|
|
|
data = read_input()
|
|
last = data.get("last_assistant_message", "")
|
|
agent = data.get("agent_type", "unknown")
|
|
if not last or len(last) < 100:
|
|
return
|
|
|
|
snippet = last[:400].replace("\n", " ")
|
|
emit({
|
|
"decision": "block",
|
|
"reason": (
|
|
f"subagent-judge: {agent} returned. Before proceeding, re-read its last message "
|
|
f"(snippet: {snippet}...) and confirm: did it actually address the original task? "
|
|
f"If unsure, re-spawn with a tighter prompt."
|
|
),
|
|
})
|
|
|
|
|
|
if __name__ == "__main__":
|
|
try:
|
|
main()
|
|
except Exception as e:
|
|
warn_to_stderr(f"[subagent-stop hook error] {e}")
|
|
sys.exit(0)
|