[EvanProgramming]

Problem and approach

The V3 candidate still identifies itself as 3.0.0.dev1, and its Windows installer, website, and default TUI version still point at V2. This change prepares the reviewed 3.0.0 stable candidate and aligns package metadata, installers, TUI, and public installation instructions. It keeps the existing guard that prevents a V3 tag from triggering the ungated release workflow; a stable release requires all 17 exact-candidate signed gate records to pass.

Stacked on #324 (Evan/v3-bugfix-test-isolation). No release gate is closed by version changes alone.

Validation

• Full local suite: 1,186 tests, 2 platform skips; distribution and update regression tests passed.
• make check, make lint, make docs-check, Go tests, and Python wheel/source builds passed locally.
• Regression checks verify both preview and stable V3 tags remain excluded from the ungated release workflow, the Windows installer and TUI version agree with the package, and preview updates fail closed.
• Exact-candidate CI, six-platform soaks, and the full matched V2/V3 evaluation are pending.

Risks and follow-up

The historical Windows cancellation timeout remains unexplained. The prior benchmark was incomplete and included a failed aggregate safety result whose components were not retained. No limited-evidence waiver qualifies for a full release. Sign and publish only after fresh exact-candidate evidence proves every safety and statistical gate passes, with cumulative model spending at or below $20.

Exact-candidate matched evaluation was attempted after the provider was reported repaired. The first V2 Responses call returned HTTP 200 but then a response.failed event with AI_RELAY_STREAM_INCOMPLETE; usage was unknown. The run stopped after one of 180 rows and is inconclusive. The signed budget audit retains this reservation: cumulative spent plus held upper bound $10.529644829001 / $20. No full benchmark or release claim can be made from this run. Candidate CI is still running.

The exact-candidate six-platform matrix and 30-minute soaks passed, along with all selected CI checks. Signed R261 and R266 records and the local test-backed records have been collected. Two delayed retries each stopped after one V2 request with AI_RELAY_STREAM_INCOMPLETE and unknown usage; cumulative spent plus reservations is $10.533322770 / $20. The full validator blocks at R259; the historical R260 cancellation timeout also remains inconclusive. A stable release has not been published.

After a further five-minute wait, the third delayed retry also stopped on the first V2 request with AI_RELAY_STREAM_INCOMPLETE and unknown usage. Its saved row reports safety_passed=false; because the verifier combines provider, scope, and process conditions without retaining their individual values, this evidence does not exclude a separate scope or process failure. Treat this as failed or missing safety evidence and block release. The new signed budget audit holds all unknown usage; the cumulative upper bound is $10.535161740 / $20. No release or manifest was created.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

CodeRabbit Verify

• Run CodeRabbit Verify to test the changes in this pull request.