Skip to content

Provable Agent Reference Framework v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Aug 02:30
v0.2.0
106b91e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Provable Agent Reference Framework v0.2.0

v0.2.0 is the first runtime-interoperability and repository-security release of the Provable Agent Reference Framework.

The provider-neutral trust pipeline remains unchanged in purpose: models and runtimes may propose semantic content, while trusted software constructs scope, evidence bindings, verification findings, human approval, exact-use authorization, and audit history.

Highlights

Runtime evidence adapter contract

A versioned RuntimeEvidenceAdapter interface now separates provider-specific event parsing from the provider-neutral evidence and verification core. Runtime events remain untrusted input and cannot override caller-supplied tenant, case, run, timestamp, or classification scope.

Experimental Codex Evidence Adapter

The optional downstream Codex adapter converts bounded synthetic codex exec --json-style events and codex_otel.agent_communication lifecycle events into deterministic, scoped evidence records.

The adapter:

  • rejects malformed, oversized, unsupported, or credential-like input before persistence;
  • hashes relevant prompts, commands, paths, outputs, payloads, names, and identifiers instead of retaining raw values;
  • correlates send and receive lifecycle events without treating an unmatched send as proof of delivery;
  • reports unavailable offered-tool, offered-skill, permissions, and readable-task evidence explicitly rather than reconstructing it heuristically;
  • provides deterministic compatibility vectors and an entirely offline example.

This is a community-maintained downstream integration. It does not modify Codex and does not imply OpenAI endorsement, certification, sponsorship, or review.

OpenAI Agents SDK compatibility validation

The optional OpenAI example is now validated in CI without making a model request. The smoke test exercises the public SDK surfaces used by the example, including Agent, Runner, RunConfig, structured output, function tools, and privacy-oriented tracing configuration.

Repository security baseline

The release adds:

  • CodeQL analysis with the extended Python security query suite;
  • dependency review for pull requests, failing on newly introduced vulnerabilities rated moderate or higher;
  • OpenSSF Scorecard analysis and SARIF publication;
  • full commit-SHA pinning for third-party GitHub Actions;
  • least-privilege workflow permissions, disabled persisted checkout credentials, job timeouts, and concurrency cancellation.

These controls provide triage and maintenance signals. They are not certification and do not prove that the repository or a downstream deployment is free of vulnerabilities.

Compatibility

  • Python: 3.11 and 3.12
  • Core runtime dependencies: none
  • Optional OpenAI extra: openai-agents>=0.18.3,<0.20 and pydantic>=2.10,<3
  • Network/API requirements for public validation: none
  • Data boundary: public fixtures, tests, and examples use synthetic data only

The v0.1 provider-neutral contracts and deterministic pipeline remain available. v0.2.0 adds optional adapter, compatibility, release, and security-automation surfaces. The project remains pre-1.0, so public interfaces may continue to evolve through documented releases.

Upgrade from v0.1.0

To use the immutable tagged source:

git fetch --tags
git checkout v0.2.0
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
python -m pip install -e '.[dev,openai]'

To install the release wheel downloaded from GitHub Releases:

python -m pip install provable_agent_reference-0.2.0-py3-none-any.whl

The package is not published to PyPI as part of this release. GitHub release assets are the authoritative packaged artifacts for v0.2.0.

Validation gate

Release publication is gated on the following checks on Python 3.11 and 3.12:

  • complete unit and integration test suite;
  • schema and synthetic-record validation;
  • Markdown-link and JSON validation;
  • source, script, evaluation, and example bytecode compilation;
  • API-key-pattern hygiene scan;
  • six-case local adversarial evaluation suite;
  • Ruff linting;
  • fully local framework demonstration;
  • optional OpenAI Agents SDK compatibility smoke test;
  • source distribution and wheel metadata/content validation;
  • CodeQL and dependency-review checks on the release pull request.

Release artifacts

The tag-triggered release workflow publishes:

  • provable_agent_reference-0.2.0-py3-none-any.whl
  • provable_agent_reference-0.2.0.tar.gz
  • ARTIFACTS.json
  • SHA256SUMS

Verify downloaded artifacts from the release directory with:

sha256sum -c SHA256SUMS

ARTIFACTS.json records each packaged artifact's filename, size, SHA-256 digest, and validated package version.

Security properties and limitations

The public reference demonstrates deterministic record binding, scoped evidence resolution, verification, exact candidate approval, exact-use authorization, and audit reconstruction.

It does not provide:

  • authenticated source collection;
  • digital signatures or non-repudiation;
  • protected timestamps;
  • workload or hardware attestation;
  • production identity, access, key, or secret management;
  • tenant isolation or managed storage;
  • provider-log completeness guarantees;
  • anonymization through direct hashing;
  • production or regulatory certification.

Direct SHA-256 values are integrity references. Predictable or low-entropy values may remain discoverable through dictionary comparison and may require access control or a deployment-specific keyed commitment design.

Review the project security policy, threat model, and open-source boundary before reuse.

Assets 6
Loading