WorkTrust for Python tool runners

This package launches the same JavaScript measurement engine shipped in the WorkTrust npm package. It does not implement another measurement engine. Version 0.10.14+ also includes the explicit observation SDK documented below; 0.10.13 does not include it. See the official PyPI project for published versions. Python 3.10+ is required. Supported platform wheels include Node.js 22.23.3; no separate Node or npm installation is needed. The launcher uses that bundled runtime, never downloads a runtime when invoked, and does not silently switch to another CLI version.

Wheels target macOS 11+ (Intel and Apple Silicon), Linux with glibc 2.28+ (x64 and arm64), and Windows 10+ (x64 and arm64). Linux also requires the Node-supported system C++ runtime (GLIBCXX_3.4.25 or newer). Alpine/musl, 32-bit systems and other platforms are not supported. A source distribution cannot fetch Node automatically: a source build requires the explicit maintainer preparation step below. Supported users should install the matching wheel.

Run the CLI through either Python tool runner:

uvx worktrust preserve
pipx run worktrust preserve

Omit preserve to connect your computer. Installing through uvx or pipx does not instrument Python applications automatically. To test a reviewed local candidate before publication:

uvx --from /absolute/path/worktrust-VERSION-py3-none-PLATFORM.whl worktrust --help
pipx run --spec /absolute/path/worktrust-VERSION-py3-none-PLATFORM.whl worktrust --help

Use your package manager to choose and update the installed version. The package contains the CLI modules and Node runtime, so preserve itself needs no network after installation. It reads supported local session files and can change retention settings or create local metadata archives; only the person runs it. Existing WorkTrust agent safeguards and consent prompts remain. Other commands can contact WorkTrust. Package installation itself can contact registries.

Explicitly request the latest published package when needed:

uvx worktrust@latest preserve
pipx run --no-cache worktrust preserve

For persistent installations, use uv tool upgrade worktrust or pipx upgrade worktrust. An existing cache or version constraint can otherwise keep an older release. worktrust update refreshes the computer coupling using the selected package; it does not upgrade the Python package.

Runtime bundle hashes detect missing or changed files; they are not hardware attestation or independent proof of authenticity. Request and archive signatures retain the existing CLI's software-key threat model. See the npm package SECURITY.md and README.md for the data contract.

The 0.10.21 release also retains metadata before submission in ~/.worktrust/measurements/ and advances progress only on explicit receipts. This is the same implementation used by npm, uvx and pipx. The journal contains available timelines, opaque references and counts, not conversation text, paths or credentials. It remains on the computer after disconnecting; keep your own backup. Its contents remain user-controlled observations, not independent proof of human effort. Changed historical replacements wait for server review without overwriting earlier measurements. Updating does not restore support-withdrawn lines.

Maintainers

Running node tools/worktrust-cli/sync.mjs from the monorepo also regenerates the Python JavaScript bundle and version. node scripts/build-python-cli.mjs --check refuses drift. The native runtime is a separate build artifact, excluded from Git and sdists. Explicitly prepare it before building:

python tools/worktrust-python/build_runtime.py
python -m build --wheel tools/worktrust-python
python -m build --sdist tools/worktrust-python

build_runtime.py downloads one official Node archive from nodejs.org at build time, verifies its committed SHA256 from runtime-lock.json, and extracts only the executable and complete LICENSE. It accepts --target for a reviewed target and --cache for a build cache. Platform wheel tags follow the selected runtime. Never label these wheels any. Building a wheel from the sdist fails closed until this explicit preparation step has been performed; no install hook downloads code.

Run python tools/worktrust-python/tests/test_runtime_build.py, then node scripts/test-python-cli.mjs with a Python containing the build requirements (select it with WORKTRUST_TEST_PYTHON). Tests build and exercise a real wheel using its own runtime with no Node on PATH; they use synthetic homes and metadata only. Native execution must pass independently on all six claimed OS/architecture combinations before release. Cross-building alone is not that proof. Set WORKTRUST_REQUIRE_TOOL_RUNNERS=1 to require actual uv and pipx wheel smoke tests, and WORKTRUST_TEST_ARTIFACT_DIR to a fresh directory to retain the exact tested wheel and sdist. Existing artifacts are never overwritten. Canonical module bytes match across channels; bit-for-bit reproducibility of whole wheel or sdist archives is not asserted.

The public fhomey/worktrust-cli repository carries this package under python/, alongside the npm package at its root. There, install python/requirements-release.txt, run python python/build_runtime.py, then python python/tests/test_distribution.py with WORKTRUST_TEST_CLI_SOURCE set to the repository's absolute path and WORKTRUST_REQUIRE_TOOL_RUNNERS=1. The test derives Node from the built wheel; it does not require a system Node installation. Keep build-only workflow results distinct from successful publication.

The active .github/workflows/release-pypi.yml uses the protected pypi environment and PyPI Trusted Publishing. In the 0.10.22 candidate, the successful signed npm release workflow triggers PyPI, which consumes its exact artifacts without rebuilding. Failed jobs can be rerun through GitHub. The retained release-pypi.yml.example is historical, not the active workflow. All six native artifact checks must pass before either registry publishes. Both channels use the same version and module bytes. Approvals and registry failures remain separate; confirm both publications before announcing a release as available on both channels. Neither workflow invokes a Vercel deployment.

Do not edit generated bundle/, VERSION or LICENSE. Publication requires confirmed PyPI ownership and a configured Trusted Publisher; source readiness does not grant that.

Official runtime platform requirements: https://github.com/nodejs/node/blob/v22.23.3/BUILDING.md Pinned upstream archive checksums: https://nodejs.org/dist/v22.23.3/SHASUMS256.txt

Explicit agent observations (requires 0.10.14+)

The optional worktrust_cli.telemetry SDK records events supplied by your own Python orchestrator. It does not discover agents or inspect prompts, outputs, files, credentials or provider SDKs. Recording is local, requires consent=True, and makes no network requests. The shared JavaScript engine calculates the diagnostics for both the CLI and WorkTrust app; Python contains no separate efficiency or scoring calculations.

Install the SDK in your agent application's own virtual environment:

python -m pip install "worktrust>=0.10.14"

This requires a published 0.10.14 or newer wheel for your platform. A GitHub release waiting for approval is not a completed PyPI publication. A separate uvx/pipx tool environment does not make worktrust_cli.telemetry importable in your application.

from pathlib import Path
from worktrust_cli.telemetry import Recorder

# Keep opaque, randomly generated UUID4 identities stable across comparable runs.
# Generate a new version UUID when changing the agent configuration.
agent_id = "416973a6-fad6-4f20-a59c-f66b01946dc0"
agent_version = "1809b614-3212-44b6-b668-c46cfca07cb8"

with Recorder(Path.cwd().resolve() / "agent-events.jsonl", consent=True) as recording:
    agent = recording.run(agent_id=agent_id, agent_version=agent_version,
                          role="worker", task_kind="build")
    agent.budget(1000)  # Optional; before usage, once per run.
    # One call's mutually exclusive counters, taken from your provider's actual usage.
    agent.usage(input_uncached_tokens=100, cache_read_tokens=40,
                cache_write_tokens=0, output_tokens=80)
    agent.verification("passed")  # A reported check, not independent corroboration.
    agent.finish("succeeded")    # Only after your application knows its outcome.

Use a fresh file for each recording: existing files, symlink paths and overwrite attempts are refused. Parents must already exist; create a child with recording.run(..., parent=agent) and finish children before their parent. The recorder serializes threads in one process; multiple processes need separate recordings. All elapsed timestamps share the recorder's monotonic clock. Exiting a context or encountering an exception never invents completion. An unclosed run remains incomplete. Header-only and partially written files cannot be imported.

Usage fields are optional: omit unavailable counts instead of substituting zero. Never pass inclusive provider input totals into input_uncached_tokens, cumulative usage snapshots, or child usage again on its parent. This SDK supplies no automatic provider normalization. security(code, source=...) accepts only the fixed event categories documented by the API; use it for explicit client_report or boundary_report events, never classifications inferred from arbitrary prompt text. intervention(code) records an explicit challenge, reframe, alternative or verification request. These observations do not establish causation, human presence, capability or independently verified outcomes.

Limits are 64 runs, 2,048 events, 256 KiB per recording, 4 KiB per line and seven days on the monotonic timeline. UUIDs, times, enums and counts only: no names, model names, task descriptions or exception text. Files request mode 0600; Windows ACLs remain the responsibility of the containing folder. The SDK does not change existing directory permissions. Stop writing with close() and delete the explicitly chosen local file to erase it; no remote copy is sent by this SDK. File metadata remains visible to other software running as the same user.

Run python tools/worktrust-python/tests/test_telemetry.py for synthetic writer and shared-engine checks. Distribution tests also run this suite against the actual wheel and bundled runtime.

Inspect a recording locally with worktrust orchestration report --file agent-events.jsonl. This command uses the shared engine, reads only the selected file and sends nothing. The separate private Sources → Agent teams browser import requires its platform deployment and your explicit file selection and consent; CLI publication does not deploy that page. Application reports never become measured human hours, capability scores or employer evidence. For the wider service, see privacy, terms and support.

Signed startup (0.10.22+)

The Python launcher uses pycryptodomex>=3.24.1,<3.25 to verify the signed file manifest before starting the bundled Node runtime. The runtime, descriptors, shared engine and Python modules are covered. Missing or changed files and invalid signatures stop execution. Package managers install this verification dependency; npm is never invoked or installed by the Python launcher. The initial package-manager installation, Python environment and bootstrap remain trusted; a computer owner can change or bypass them. This is not execution attestation.

This is release integrity, not proof that a local transcript is authentic. Local Claude records are not release files, and cli_digest is diagnostic client metadata. Hashes are not encryption. Computer-key revocation stops submissions; passport withdrawal invalidates a publication; neither silently erases historical account data. See the shared CLI SECURITY.md for the limits.