This project exists to improve factual-verification behavior in Claude Code through host-native hooks.
- Improve the prompt classifier so it catches more stale-fact scenarios without over-triggering.
- Add more tests around
last_assistant_messageand transcript fallbacks. - Expand install docs for macOS and Linux.
- Tune trusted MCP defaults for common public servers.
- Add optional telemetry or debug logging that stays off by default.
- Keep the project standalone. No private framework dependencies.
- Prefer positive behavior rules over giant negative prompt blocks.
- Do not add proprietary connectors, credentials, or private MCP names.
- Favor small, reviewable PRs with a clear before/after behavior description.
Run the smoke test:
python scripts/smoke_test.pySyntax-check the hooks:
python -m py_compile hooks/common.py hooks/fact_prompt_gate.py hooks/track_verification.py hooks/verification_stop_gate.py