Skip to content

Releases

Harness Workbench v0.2.0

Choose a tag to compare

@explorefailure explorefailure released this 08 Nov 04:28
v0.2.0

What's Changed

Full Changelog: v0.1.0...v0.2.0

Verification

The signed tag, reproducible-build evidence, workflow URLs, published hashes, downloaded-asset verification, and residual limitations are recorded in the release conformance record.

Harness Workbench v0.1.0

Choose a tag to compare

@explorefailure explorefailure released this 25 Aug 21:04
v0.1.0

Harness Workbench v0.1.0 is the first final GitHub release of the deterministic harness-testing workbench. It promotes the verified v0.1.0-rc.2 contents through a new commit, build, signed tag, hosted matrix, and clean artifact-install gate.

Highlights:

  • Run command workloads with append-only attempts, preserved specs and feature source, integrity closure, replay, and zero runtime dependencies.
  • Measure confinement, effects, blast radius, interruption recovery, steadiness, efficacy, sensitivity, feature interaction, fidelity, and replay behavior.
  • Use the public vendor-neutral harness_workbench.capture and harness_workbench.canon primitives for bounded subprocess evidence and canonical digests.
  • Materialize the experimental five-subject adapter tree for Claude Code, Codex CLI, DeepSeek Harness, Hermes Agent 0.20.5, and Pi. The shared adapter envelope remains experiment-local rather than a stable public API.
  • Build reproducible wheel and normalized source-distribution artifacts with strict metadata, archive, executable-mode, checksum, Gitleaks, and separate clean-install verification.

Compatibility and boundaries:

  • Supported target: CPython 3.11–3.14 on Linux and macOS. Windows is unsupported.
  • Harness Workbench executes trusted commands and Python feature modules with the current user's permissions; it is not a security sandbox.
  • The release is not published to PyPI. Install from these GitHub assets or a tagged clone.
  • Live subject runs can require third-party clients, credentials, network access, and paid quota. The core tests and artifact verification are network-free.

The attached SHA256SUMS authenticates the wheel and normalized source archive. A release-final conformance record will be attached after the GitHub-hosted assets are downloaded and independently rechecked against those digests.

Harness Workbench v0.1.0-rc.2

Pre-release

Choose a tag to compare

@explorefailure explorefailure released this 25 Aug 20:31
v0.1.0-rc.2

Harness Workbench 0.1.0rc2 is the second public release candidate.

Highlights:

  • Adds the public harness_workbench.capture module: bounded subprocess capture, per-stream limits, deadlines, process-group cleanup, redaction, file/tree evidence, containment helpers, minimal environments, and JSONL decoding.
  • Declares and mechanically checks the public harness_workbench.canon surface.
  • Updates the experimental Hermes subject to stable 0.20.5 / v2026.8.19, retains the strict no-allowance steadiness result, and carries fresh five-subject repair comparison evidence for Claude Code, Codex CLI, DeepSeek Harness, Hermes Agent, and Pi.
  • Keeps the five-subject evidence envelope experiment-local after an explicit API/schema promotion review; capture and canon remain the supported vendor-neutral primitives.
  • Hardens release reproducibility with a pinned build backend, normalized source-distribution modes and metadata, archive/source executable-set agreement, history and archive secret scans, separate clean wheel/sdist installs, and checksum verification.

Release identity and verification:

  • Signed immutable tag: v0.1.0-rc.2
  • Source commit: 66ce4ed853ac4a60f975bddf7e84cfccd4505cc6
  • Supported target: CPython 3.11–3.14 on Linux and macOS
  • Package forms: wheel and ownership-neutral normalized source distribution
  • SHA256SUMS authenticates the downloadable package bytes
  • The release conformance asset records the final checks and limitations

This remains a release candidate. Harness Workbench executes trusted commands and modules; it is not an operating-system sandbox. Windows is unsupported. The subject adapters and their exact-five-subject envelope remain experimental rather than a supported integration API. No independent outside artifact check is claimed.

This GitHub prerelease does not publish the package to PyPI.

Full changelog: v0.1.0-rc.1...v0.1.0-rc.2

Harness Workbench v0.1.0-rc.1

Pre-release

Choose a tag to compare

@explorefailure explorefailure released this 12 Aug 19:12
v0.1.0-rc.1
5e13e33

Harness Workbench 0.1.0rc1 is the first public release candidate.

It provides a small, dependency-free Python harness for running trusted JSON-described workloads, preserving inspectable run records, replaying and comparing evidence, and exercising opt-in harness-measurement features.

Release identity and verification:

  • Signed immutable tag: v0.1.0-rc.1
  • Source commit: 5e13e33a391eda1ed67729bbc4a0afd3497210a0
  • Supported target: CPython 3.11–3.14 on Linux and macOS
  • Package forms: wheel and ownership-neutral normalized source distribution
  • SHA256SUMS authenticates the downloadable package bytes
  • The release conformance asset records the final checks and limitations

This remains a release candidate. It executes trusted commands and Python modules; it is not an operating-system sandbox. Windows is not a supported target. No independent outside artifact check is claimed for this candidate.

This GitHub prerelease does not publish the package to PyPI.