[EvanProgramming]

Tracks #268 and remains stacked on #311. It does not close #268 until all release evidence passes.

Root cause: release gates lacked a machine-enforced evidence chain bound to the exact clean candidate SHA, and the six-platform results had no reproducible collector. Partial or stale evidence could otherwise be published as a release record.

Approach: validate signed gate records against the exact candidate, configured GPG signer, fixture/command hashes, result-artifact hashes, classifications, and six-platform R266 matrix. Evidence and manifest paths inside the candidate checkout are rejected. Signature and manifest publication do not overwrite existing files, and partial publication removes the signature. CI uploads platform logs, environment and dependency metadata; the collector rejects stale, incomplete, escaping, or failed results. Linux and macOS run the full unit and gateway integration suites. Per the agreed Windows policy, Windows runs supported-feature agent/sub-agent acceptance and Go TUI tests/build; POSIX-dependent Python suites and interactive PTY acceptance are explicitly recorded as unavailable. Every platform still requires acceptance evidence and a 30-minute soak.

Validation: local make test passes (1,125 tests, 2 skipped); the macOS 3.13 delegation regression passes with a 15-second wait (10.2 seconds locally); make check, make lint, and make docs-check pass. Commits are GPG-signed and verified. Exact-candidate run 39971830170 passed all checks on 2104cd3, including six platform suites and 30-minute soaks. Its 50-artifact R266 record was inspected, signed with the pinned GPG key, and validated locally at /Users/evangong/Documents/ChatGPT/OpenKyrozenV3Evidence/r266-candidate-2104cd3c1d4f030b42ca9276c96d8eabc80c5f97. All six platform suites and 30-minute soaks passed, and the resulting R266 record was reviewed, detached-GPG-signed, and validated against the exact candidate with 50 hashed evidence artifacts.

Risks and follow-up: R266 is signed and verified locally at /Users/evangong/Documents/ChatGPT/OpenKyrozenV3Evidence/r266-candidate-5f5a612. The R250–R265 records are not available in the evidence directory, so a complete release manifest cannot yet be produced; the full release-gate validator was run against this exact SHA and correctly blocked at missing R250 without writing a manifest. The matched V2/V3 campaign is incomplete and inconclusive. Accounted provider exposure across the two partial attempts is conservatively held at US$10.20. The approved cumulative cap is now US$20, leaving at most US$9.80 for a fresh matched run. The runner reserves worst-case usage before each call and stops when the remaining budget cannot cover it; earlier partial reports remain inconclusive. #266, #267, and #268 remain open; keep this PR in draft until the missing evidence gates pass.

[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

_{Comment @coderabbitai help to get the list of available commands.}