trace-audit

trace-audit path [--json] [--fail-on low|med|high]

trace-audit reads a Claude Code session log and runs five deterministic checks. It does not call a model, does not use the network, and does not change the session. The report starts at 100 and subtracts 20 for each high finding, 10 for each medium finding, and 3 for each low finding (floor 0). Trust is HIGH at 90 or above, MEDIUM at 70 or above, and LOW below that.Each finding names an id, a short name, a severity, a message, and the turn it attaches to.TA001 CLAIM-PASS (high)

The agent says tests or the build succeeded, but no verify command ran, or the last verify command before that claim failed

Code files were edited and no test, build, lint, or typecheck command ran after the last edit

The session claims a fix while editing an existing test file. Removing or weakening an assertion is high. Other edits to existing tests are medium. Creating a new test file does not fire

A tool result contains an error and the agent's next message never acknowledges it

Edited code files far exceed what the user messages mentioned (more than three times the mentioned set, and at least eight files). Skipped when the session has more than three real user turns

path

One Claude Code session log (a `.jsonl` file), or a directory. A directory is walked recursively and every `.jsonl` file is audited. Missing paths and directories with no session logs exit 3

One JSON object per session (score, trust, findings) instead of the text report. Sessions are separated by a newline

Exit 1 when any finding is at or above that severity. high is the strictest of the three thresholds (only high findings fail). Omit it to always exit 0 after a successful read, findings or not

Input is Claude Code session JSONL (the files under `~/.claude/projects/`). Other agent transcripts are not parsed. A directory audit includes every `.jsonl` file below that path, not only Claude sessions.Exit 0 means the files were read and, if --fail-on was set, nothing met the threshold. Exit 1 means the threshold was met. Exit 3 means the path was missing or the directory had no session logs. Findings are a report only; the tool does not edit files or re-run tests.Rules are data in YAML next to the checker. They look for English and some Vietnamese claim phrases. A session written in another language can miss TA001 and TA003.

trace-audit is Apache-2.0, by Vũ Văn Tâm, the same author as Yana AI. The five checks are the v0.1 set. TA003 and TA005 were calibrated against real sessions on 4 July 2026.

yana-ai(1), yana-rt(1), yana-ai-rt(1), claude(1), python(1), pip(1)

Source code · Documentation