CI, MIT license, and jitagent 0.0.1 on PyPI

[Birfy]

Adds the plumbing to ship this: GitHub Actions CI, a license, and a release path to PyPI as jitagent.

The name

agent-jit was rejected by PyPI — its similarity check strips -, _ and . before comparing, so it normalises to agentjit, which exists. Looking at what holds that name settled the question: 5 releases, the most recent one day before this, and the same concept ("Just-In-Time Compiler for AI Agent Trajectories"). Its wheel also installs a top-level agentjit package, so sharing the import name would mean the two distributions silently overwrite each other's files in site-packages.

So everything is jitagent — distribution, import, CLI, JITAGENT_HOME, ~/.jitagent, and the repository.

pip install jitagent
jitagent selftest

Two workflows

ci.yml — on every push to main and every PR:

• unit tests and the corpus selftest on Python 3.11 / 3.12 / 3.13
• audit_traps.py against the committed audit results, so the numbers the README quotes cannot drift silently
• builds the distribution, runs twine check --strict, then installs the wheel into a clean venv and runs the CLI out of it

That last step matters: everything else tests the source tree, so a broken console script or missing package data would sail through and only surface after upload. It caught a real bug immediately (below).

release.yml — publishes over Trusted Publishing (OIDC, no token stored anywhere) when a GitHub Release is published, not on a tag push. A tag is cheap to create by accident and a PyPI version can never be reused, even after deletion. It re-runs the tests on the released commit and refuses to publish if the tag and the version in pyproject.toml disagree. workflow_dispatch publishes to TestPyPI for a rehearsal.

Three bugs found while preparing the package

• selftest was broken for anyone who pip-installed it. It resolved the corpus relative to the source tree, so it only worked from a git checkout — while the README tells you to run it immediately after installing. The corpus now lives in src/jitagent/corpus/ and ships as package data. It is the one check worth running on a new machine anyway: how the sandbox contains a memory bomb is platform-dependent (RLIMIT_AS on Linux, a watchdog on macOS).
• coverage was a mandatory dependency nothing could reach. The branch-coverage gate was cut earlier, but its plumbing survived in sandbox.py and _child.py with no caller able to switch it on. Removed along with the dependency, so jsonschema is the only runtime dependency left.
• capture_demo.sh swallowed its own failures. The CLI's output goes to the transcript, so when a synthesis call failed transiently the operator got a bare exit code — and the EXIT trap then deleted the log that would have explained it. It now prints the tail of the log on failure. (Found by hitting it.)

The demo was re-captured, not renamed

docs/demo.txt claims to be real captured output, so a transcript edited by search-and-replace would have made that claim false. The run behind the animation is fresh under the new name — different generated cases, different handle.

Also

LICENSE (MIT), CHANGELOG.md, and CONTRIBUTING.md — which records what a new verification gate has to prove, given that five were removed from this repo for never having caught a real model's mistake: a real failure it would have caught, a corpus case, and evidence it does not fire on correct code.

Verified locally

pytest                    102 passed
jitagent selftest         6/6
tools/audit_traps.py      18/18
python -m build           jitagent-0.0.1{.tar.gz,-py3-none-any.whl}
twine check --strict      PASSED, PASSED
pip install dist/*.whl    then `jitagent selftest` outside the source tree → 6/6
                          and `agentjit` is not importable, so no collision

Wheel metadata: Name: jitagent, Version: 0.0.1, License-Expression: MIT,
Requires-Python: >=3.11, one runtime dependency (jsonschema>=4.21).

Releasing 0.0.1

The PyPI trusted publisher is configured. After this merges, publishing a GitHub Release tagged v0.0.1 runs release.yml, which re-runs the tests, checks the tag against the version, builds, and uploads over OIDC.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FeJE3eqihzPMnwr1TDJGvp

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.