Version 0.1.0 is a local alpha under the MIT license. Building artifacts is separate from publishing them; the verification tool does not tag, push, create a GitHub release, or upload to a package index.
From a trusted checkout, install the development tools and run:
python -m pip install -r requirements-dev.txt -r requirements-prototype.txt
python -m unittest discover -s tests -v
python -m tools.verify_release --work-dir state/release-check-001The work directory must be new and ignored. The tool builds both a source archive and a wheel, matches distributed source to public Git inputs, and checks generated metadata for local paths and runtime identifiers. It rejects private guidance, state, logs, unexpected members, and archive links. Gitignore alone is not the artifact review.
It then creates two fresh virtual environments outside the source package. The
base install has no optional dependencies. From an unrelated project it exports
the profile and skill, uses the installed skill helper, runs one harmless job,
checks its receipt, refuses relaunch, and revokes a fixture before dispatch. No
worker is contacted. A second environment installs the codex extra from a local
wheelhouse and checks that the WebSocket client can load. Downloading that optional
dependency requires network access; no model or Codex service is started.
report.json records artifact hashes, sizes, timings, and outcomes. Logs and
virtual environments remain in the ignored work directory for local review.
An alternate existing interpreter can be selected with --python PATH. Only the
versions actually exercised by a report should be described as tested.
Use --artifacts PATH to check the same wheel/source pair with another interpreter
without rebuilding. Both archives are still audited against the current source.
| Measure | Current evidence |
|---|---|
| Normal completion | Two short local worker trials each completed one job and one continuation |
| Worker context | Both trials retained context and returned to idle after two turns |
| Dispatch latency | About 2.6 and 3.4 seconds through final verification; isolated observations |
| Duplicate/missed events | None observed in those normal-path trials; failure cases covered offline |
| Restart behavior | Saved receipts, counters, uncertain sends, and controller reopen tested offline |
| Permissions and stop | Snapshot drift, approval handoff, revocation, and exhausted bounds tested offline |
| Installation | The release tool exercises fresh base and optional-extra environments |
| Manual polls avoided | Not measured against a manual baseline |
| Model calls/tokens saved | Not established; two turns per live trial is not a token-use measurement |
| Maintenance cost | Not yet measured over ongoing use |
The live trials used macOS, Python 3.14.3, Codex CLI 0.164.0, and websockets 17.3. The offline suite passes on Python 3.11 and 3.14; each release report records its clean-install interpreter. Machine reboot, extended idle periods, native child agents, implicit skill activation, and competing interactive writers remain unverified. The Codex app-server protocol is experimental.