Codex review round four: workflows execute helper scripts from the
workflow ref, so comparing only the .github/workflows tree missed
script-only harness changes. Harness equivalence now means the git tree
diff between the candidate run's head SHA and the current workflow SHA
is itself release-metadata-only, using the same canonical classifier as
the target delta. The classifier's worktree overlay no longer applies
when --head is an explicit SHA, keeping SHA-exact comparisons exact.
Codex review round three: a prior green run may have executed with
different lane definitions; the resolver now compares the candidate
run head SHA's .github/workflows tree against the current workflow
ref before accepting its manifest.
Codex review round two: the reusable advisory input is global, so beta
would also have softened hermetic repo/OpenShell E2E — add a scoped
live_advisory input that only covers the live-provider suite jobs and
narrow the umbrella fail-fast exclusion to those job names. Evidence
reuse now also runs the macOS source-version consistency check against
the target (release-preflight --macos-versions-only) so version-stamp
deltas cannot reuse evidence while version surfaces disagree.
Codex review findings: evidence reuse must not let a default-input run
stand in for a focused provider/mode/filter/package-spec run, and
recorded child runs can be re-run to failure while the parent stays
green. Manifests now record validationInputs; the resolver requires an
exact input match and healthy child runs, and the summary re-verifies
the chain root plus its recorded children. Reuse-mode evidence dispatch
notes now disclose the chain-root relationship.
- Full Release Validation now checks for a prior green validation whose
target differs only by release metadata (changelog/version stamps) and
reuses that evidence instead of re-running every lane; disable with
reuse_evidence=false. Evidence manifests chain through reuse runs via
evidenceReuse.runId and the summary re-verifies the chain root.
- The Docker runtime-assets preflight no longer serializes the CI,
plugin-prerelease, release-checks, and performance lanes; it runs in
parallel and stays enforced by the umbrella verifier.
- Umbrella runs for release/* refs now supersede in-progress runs with
the same ref and rerun group instead of requiring manual cancels.
- release_profile=beta treats the live-provider E2E suites as advisory
(third-party model deployments move underneath releases); stable and
full profiles keep them blocking, and beta live failures no longer
fail-fast-cancel the remaining release-check matrix.
