TruthHook — Stop Your Claude Code Agent From Lying

TruthHook is a Claude Code Stop-event hook that blocks responses claiming external API/tool results without citing the artifact path that proves the call happened.

It catches three classes of fabrication:

1. Phantom tool results ("the API returned…", "8 of 10 models said…", "OpenRouter sent back…") without a saved-response path.
2. Phantom completions ("I've sent the email", "I've committed", "I've posted to X") without a matching tool invocation in the transcript.
3. Plausible synthesis — multi-source claims (board, calendar, gmail) where the model is summarizing results it never fetched.

False-positive guard: honest framing ("I attempted to call X but couldn't") is exempt — the hook stays silent.

This is the actual hook used in production by the operator who built it, after a multi-model advisory board call was fabricated by a subagent that couldn't fire 8 parallel API calls. Shipping it unmodified except for path genericization.

What's inside:

• truthhook.py (Python 3.8+, no dependencies, 215 lines)
• SKILL.md — capability spec + extension points
• INSTALL.md — 90-second wiring into ~/.claude/settings.json
• PATTERNS.md — design rationale, worked examples for every check
• TEST-CASES.md — 12 validation cases (8 block, 4 pass)
• README.md

Designed to fail open. Any malformed input, missing transcript, or unexpected exception → exit 0 silently. Will not break your agent loop.

Claude Code only. Pattern-matching, not semantic. English-only triggers ship by default. Read SKILL.md limitations before buying.