What an AI agent actually did.
Read from its transcript, not its own report.
Agents say “done” and “tests pass”. osop reads the Claude Code session transcript instead and writes a one-page receipt: every command the agent ran, what came back, and whether it failed.
Rules only. No model, zero tokens, no cooperation from the agent.
pip install git+https://github.com/Archie0125/osopNot on PyPI yet. The PyPI package named “osop” is a different project.
Open source · Apache 2.0 · Claude Code transcripts only
How to use it
Run a Claude Code session. Then run osop log in the same directory.
- osop log
- Writes the receipt for the latest session in this directory.
- osop log --brief
- The same, plus a one-line summary. A Stop hook can show it at the end of every turn.
- osop log --stdout
- Prints the receipt instead of writing a file.
- osop init
- Adds an OSOP section to the project's CLAUDE.md.
Receipts go to ~/.osop/receipts/, never into the repo the agent worked on.
What the receipt shows
One page per session, sub-agents included. Facts first.
- ClaimsWhat the agent's final message says it did (tests pass, lint passes, build succeeds, committed, pushed, merged), checked against the transcript. Each claim gets a verdict and the command that decided it. English and Traditional Chinese messages are read.
- TestsEvery test run and its result, read from the exit status and the whole output. When a pipe or a semicolon hid the exit status, the receipt says so instead of guessing.
- Files changedFiles edited with Edit, MultiEdit or Write, plus shell commands that may have changed files (>, tee, sed -i, mv …).
- ErrorsEvery tool call that failed.
- No result recordedCalls that never got a result, so there is nothing to judge.
- GitEvery git commit, push and merge, and gh pr merge, with its result. This is what backs a “committed and pushed”.
- Sub-agents and commandsWho ran what. The full command table sits in a collapsed block at the end.
Claim verdicts
- ✅ backed
- The transcript shows it happened.
- ❌ contradicted
- The transcript shows the opposite.
- ⚠️ cannot tell
- There is evidence, but it does not settle the claim.
- ❓ no evidence
- Nothing in the transcript backs it.
Questions, instructions and negated sentences are not claims.
Does it work?
We checked 173 claims from real sessions by hand. In that data osop has not yet caught a real false claim, and some of its verdicts were wrong. The cases page shows where it holds and where it does not.
See the cases →