Skip to content

Reasoning Kernel v0.6.2

Choose a tag to compare

@gianlucamazza gianlucamazza released this 04 Oct 19:27
· 17 commits to main since this release
v0.6.2
55ba79b

Corrective release for executable host conformance and provider dependency security.

  • Reject observed effects after capability/egress denial or pre-dispatch audit failure.
  • Correlate outcomes with unique authorized starts and matching completions; reject missing, duplicated, reordered or inconsistent evidence, including empty traces paired with completed effects.
  • Validate root terminals while preserving legitimate partial-audit outcomes and delegated runs.
  • Require AnyIO >=4.14.2 in provider/dev extras; lock 4.15.1 to address the TLS certificate-spoofing and process-worker advisories.
  • Require urllib3 >=2.8.0 for development/publishing tools to address all three open urllib3 advisories.
  • Add a pinned, strict dependency audit over all locked extras/platforms as a CI/release prerequisite.
  • Include packaging in the local aggregate and use the same aggregate in CI.

API signatures, JSON report schema v1, profile names and CLI exit codes are unchanged. Diagnostic check identifiers are additive. Previously accepted contradictory observations now fail.

Validation: 237 kernel tests, Python 3.12–3.14 CI, required live DeepSeek qualification, isolated wheel/reference CLI checks, and immutable-artifact promotion through TestPyPI and PyPI. Vibemonitor at 806304fe231b31afc6784b90d1ee25bff8c33e75 passes 173 tests and all four gate-v1 cases on Python 3.12 and 3.13 with a disposable candidate dependency pin. This is compatibility evidence, not production adapter UAT; the deployed consumer was not changed.

The immutable v0.6.1 tag was not published: its workflow was cancelled before build/TestPyPI/PyPI after the refreshed dependency graph exposed urllib3 advisories. Version 0.6.2 includes all corrections.

Post-release OpenAI qualification (2026-10-04)

The unchanged published 0.6.2 wheel passed both OpenAI live contracts with requested model gpt-5.5 and SDK 3.19.2 in an isolated environment. The combined OpenAI/DeepSeek CI run passed all six required live cases; only Anthropic was excluded. No runtime adapter change or package republication was needed.

Qualification evidence · Combined live CI · Required release gates (PR #58)

Future releases require both OpenAI and DeepSeek. This qualification covers the existing Plan and operational Q-LLM contracts, not other OpenAI models, Anthropic, or production host-adapter acceptance. The original tag, distributions and publication evidence remain unchanged.