Polygraph — Development skill for Claude Code

polygraph — A lie detector for coding agents

**Your agent says “done.” Check the evidence.**

[](https://github.com/guarrr816/polygraph/actions/workflows/ci.yml)

polygraph checks the records behind coding agents' claims. Its current Claude Code adapter checks tool calls and transcripts for three familiar situations:

• “Tests passed.” Checks code-changing turns for a completed verification command reported as successful. A claim in the reply does not count.
• “You asked me to stop.” Flags known user-turn patterns inside assistant replies. The said command checks whether the transcript contains matching human input.
• “I'll disable the hook.” Blocks recognizable commands targeting hook settings, plugin files, and other protected configuration.

Python 3.11+ · No third-party runtime dependencies · MIT

Platform	Current status
Claude Code	Adapter implemented; live session acceptance pending
Codex	Hook interfaces confirmed in official documentation; adapter planned

[Install](#install) · [The story](#why-i-built-it) · [How it works](#what-it-does) · [Limits](#known-residuals) · [Chinese](README.zh-CN.md)

See the delivery gate in action

A turn changes `foo.py` and replies **“Done. All tests passed.”** Its transcript contains no test command. The Stop hook exits **2** and prints:

[polygraph] this turn changed foo.py;
no verification command succeeded this turn (subagents included);
...
block 1/2; after the cap the turn is released and recorded as an unverified delivery.

Actual hook output from a **local test replay with a synthetic transcript**, excerpted and wrapped for readability. Live Claude Code session acceptance is still pending. Reproduce the blocking check from a source checkout with the test dependency installed:

python -m pytest tests/test_gate.py::test_baseline_write_without_evidence_blocks -q