This release candidate captures the first real-repository review baseline for the supported CLI, GitHub Action, and MCP package line. It remains prerelease-only: npm latest promotion, stable GitHub Action refs, and public desktop support are still out of scope.
- Added the
rc.3throughrc.8prerelease roadmap and evidence workflow for the path from real-repo testing to a later stable-candidate packet. - Captured five real CodeAgora PR review baselines with the OpenRouter low-cost diverse model set.
- Recorded
rc.3ascomplete with blockers, making rc.4 false-positive/noise reduction the next required workstream. - Promoted package metadata and generated GitHub Action examples to
0.1.0-rc.3/v0.1.0-rc.3. - Fixed the tag-triggered release workflow so release-candidate tags require the
rcevidence manifest tier rather than the weaker beta tier.
The five selected samples completed successfully:
| Sample | Result | Final findings |
|---|---|---|
| R3-01 PR #502 | ACCEPT |
0 |
| R3-02 PR #512 | NEEDS_HUMAN |
4 |
| R3-03 PR #514 quick | NEEDS_HUMAN |
0 |
| R3-04 PR #525 | NEEDS_HUMAN |
12 |
| R3-05 PR #503 | ACCEPT |
0 |
--output json --quietcan emit logger/progress lines before JSON, which breaks machine consumers until fixed.- Low-confidence findings can still surface as CRITICAL/HARSHLY_CRITICAL, especially on large deletion/refactor diffs.
- Invalid final locations such as
unknown:0can survive into final findings.
- This release does not claim stable quality, npm
latest, or stable public desktop support. - Desktop remains private-preview evidence only.
- Direct GitHub PR mode still requires separate token-backed smoke evidence; the rc.3 baseline used patch-file inputs fetched from historical PRs.
docs/rc-evidence/rc3-real-repo-qa.mdrecords the real-repo baseline and blocker table.- Deterministic docs regression checks passed after evidence capture.
- Release-candidate validation must still pass before tagging and publishing.