Skip to content

AgentGuard v0.3.0 Release Verification

This versioned record documents the completed v0.3.0 production release. It preserves the earlier release-candidate-v0.3.0.md record unchanged as pre-release evidence and does not replace historical v0.2.2 evidence.

The annotated tag dereferences exactly to the source commit and has the exact annotation AgentGuard v0.3.0. The GitHub Release is published, non-draft, and non-prerelease, targets the same commit, and has no manually attached wheel or source distribution.

Published Identity

Identity Value
Product AgentGuard
Repository richinmrudul/agentguard
PyPI distribution agentguard-evals
Python import package agentguard
Console command agentguard
Version 0.3.0
Supported Python 3.9–3.12
Requires-Python >=3.9

Production PyPI exposes one wheel and one source distribution. The project metadata does not declare a Project-URL mapping, so the PyPI JSON project_urls value is absent rather than inconsistent. No legacy agentguard==0.3.0 distribution metadata exists.

Protected Publication

The release event triggered workflow run 31545391719 at refs/tags/v0.3.0 and source commit f19b54564bdd45fd438f7e48b055c102d2994a04. Its build job validated the release event, ref, version, metadata, archive members, isolated installed wheel, and package smoke before uploading the exact distributions as Actions artifact agentguard-evals-v0.3.0-validated-distributions.

The publication job waited for required reviewer approval in environment pypi. Approval was granted by the configured reviewer, after which the same job downloaded artifact ID 9122261829 with archive digest sha256:caa8c15e806eff6e25e80b16fb08bc7560051ecd660370e711d5dc692918423b, verified SHA256SUMS, removed the checksum manifest, and uploaded only the wheel and source distribution. GitHub deployment 5860189127 ended in success.

The publication used GitHub OIDC Trusted Publishing. No repository or publication secret was used, the pypi environment has no secrets, and no local artifact, API token, username/password credential, Twine command, or TestPyPI path was used.

Artifact Verification

The exact workflow files and the public PyPI downloads were compared by filename, SHA-256, and bytes and were byte-identical:

File Workflow SHA-256 PyPI SHA-256 Byte comparison
agentguard_evals-0.3.0-py3-none-any.whl 446ba25ef9f3eebb2d056606e6493b45ab8d0f4a6431e4d3ecabbfff859e8e26 446ba25ef9f3eebb2d056606e6493b45ab8d0f4a6431e4d3ecabbfff859e8e26 identical
agentguard_evals-0.3.0.tar.gz 2c156ff2817b38158dd7fbbf04122ade551b8bda9d4f1cc4a4d59a6ea6182fdf 2c156ff2817b38158dd7fbbf04122ade551b8bda9d4f1cc4a4d59a6ea6182fdf identical

Both public archives passed the repository release-artifact validator. They report Name: agentguard-evals, Version: 0.3.0, and Requires-Python: >=3.9; retain the agentguard import package, console entry point, MIT license, and packaged JSON Schema; and exclude repository-only tests, docs, examples, workflows, scripts, caches, local databases, generated runtime state, secrets, and absolute workspace paths.

Attestations And Provenance

PyPI exposes a digital publish attestation for each file through its Integrity API. Each statement names the corresponding artifact and exact SHA-256 digest. The publisher identity is:

Field Value
Publisher kind GitHub
Repository richinmrudul/agentguard
Workflow publish.yml
Environment pypi
Ref refs/tags/v0.3.0
Commit f19b54564bdd45fd438f7e48b055c102d2994a04

Both attestations include Sigstore certificates and Rekor transparency-log entries. The publication action generated and uploaded them together with the corresponding distributions.

Release Validation

  • Local full pytest: 1,401 passed, 15 skipped, one expected duplicate-archive warning. All skips were Docker-dependent because the local Docker daemon was not running.
  • Exact release commit GitHub CI: Python 3.9, 3.10, 3.11, and 3.12 jobs passed; Ruff, coverage, artifact validation, isolated install, package smoke, and the full Docker-backed Python 3.11 integration job passed.
  • Coverage gate: 91.61% statement, 80.88% branch, 89.03% combined against an 88.00% requirement.
  • Focused publishing, release-hardening, initializer, preset, baseline, and PR reporting tests: 204 passed.
  • Strict MkDocs, workflow YAML parsing, schema generation/parity, 40 maintained configuration examples, ordinary and strict release validation, Ruff, git diff --check, adversarial metrics, and showcase metrics passed.
  • Local candidate wheel and source distribution both installed successfully in isolated environments. Their hashes were diagnostic only; workflow artifacts remained authoritative for publication identity.

Fresh Production Installation

A new temporary Python 3.9 virtual environment outside the checkout installed agentguard-evals==0.3.0 from https://pypi.org/simple with pip caching disabled and PYTHONPATH absent.

  • Installed metadata resolved as agentguard-evals==0.3.0 with Requires-Python: >=3.9.
  • import agentguard resolved from the temporary environment's site-packages, and no legacy agentguard distribution metadata existed.
  • agentguard --version, agentguard --help, agentguard presets list, and agentguard presets show recommended --format json passed.
  • The packaged Draft 2020-12 configuration JSON Schema was accessible and valid.
  • Python, Node.js, and Go project initialization detected the intended project types and commands. All three generated workflows parsed as YAML, retained contents: read, pinned agentguard-evals==0.3.0, and produced configurations accepted by the production loader.
  • The native Node.js and Go fixture tests passed. The available hosts were Node.js 22.18.0 and Go 1.22.12; the Go smoke disabled network access and automatic toolchain switching.
  • A network-free deterministic mock-safe evaluation passed with score 100/100 and modified only src/auth_example/login.py.
  • Baseline-aware reporting classified a second run with an available baseline as three new and two existing findings while preserving the conservative all-blocking-findings gate.

Historical And Scope Notes

The v0.3.0 source-candidate record and v0.2.2 release evidence were not rewritten. TestPyPI was not contacted or changed. The v0.2.2 tag, GitHub Release, production files, and hashes remain unchanged.

Initialization performs bounded inspection and file generation without running repository tooling or code. AgentGuard CI and its minimal, recommended, and strict presets perform post-execution validation; they do not contain agent or test execution. This release does not include an untrusted-agent preset or claim protection against arbitrary hostile code.