Highlights
This release rewrites Gentle AI's 4R review framework around one principle: a review should find defects that matter and then STOP. v2 replaces recall-at-any-cost loops with deterministic routing, adversarial verification, and structural budgets that bound reviewers, refuters, and fix rounds.
Reviews that stop at the right time (#1083)
The previous contract could keep discovering smaller findings, reopening already reviewed work, and spending more effort without improving merge safety. Precision-gated 4R makes the cost and stopping conditions explicit:
- Trivial diffs containing only documentation, comments, formatting, or string typos run no review lens.
- Standard diffs run exactly one lens selected from the dominant risk.
- Hot-path or large diffs run the full risk, readability, reliability, and resilience set.
- Sweep budgets stop the first pass after one exhaustive standard sweep or at most two sweeps per lens for full-4R.
- Severity floors allow only verified
BLOCKERandCRITICALfindings into fix and re-review loops.WARNINGandSUGGESTIONfindings remain non-blocking information. - Convergence budgets cap remediation at two fix rounds, and each re-review receives only the persisted ledger and fix diff.
Findings must survive a challenge
High-severity findings now face an adversarial review-refuter before they can drive code changes. This protects maintainers from expensive false positives without weakening real blockers.
- Standard reviews launch exactly one refuter total with the complete candidate list.
- Full-4R reviews launch exactly three refuters total, one for correctness, one for impact, and one for reproducibility.
- Every refuter returns a verdict per finding. Full-4R applies an independent two-of-three vote to each finding.
- Missing, malformed, or inconclusive verdicts fail closed to
stands. - The task ceiling is fixed per review. It no longer grows as
3 × findings.
What you should notice
- Small changes finish faster because they no longer trigger unnecessary review fan-out.
- Everyday code changes receive one focused reviewer instead of four generic opinions.
- High-risk changes still receive full 4R coverage, but with bounded sweeps and bounded refuters.
- False-positive blockers are challenged before they consume a fix cycle.
- Re-review stays focused on confirmed findings and changed lines instead of reopening the original diff.
Runtime parity and safe migration
- Propagated the same v2 contract across Claude, Cursor, Kimi, Kiro, OpenCode, Codex, Gemini, Windsurf, and the remaining supported adapters.
- OpenCode and Kilocode now register the refuter explicitly, enforce structurally read-only tools, and require deterministic ledger output from every review agent.
- Inline-only adapters execute equivalent review and refutation passes sequentially instead of attempting unavailable subagent fan-out.
- Preserved OpenCode profiles migrate idempotently from the v1 loop-until-dry contract while retaining unrelated user-authored prompt content.
- Judgment Day continues to use two-judge convergence and never adds refuter fan-out.
Dogfooded under real load
The framework was reviewed with its own rules before release. That run exposed the original per-finding panel design, which spawned roughly 18 refuter tasks for six candidates and overwhelmed OpenCode. The final protocol batches every candidate by lens, fixing the ceiling at one standard refuter or three full-4R refuters regardless of finding count. The same review also caught runtime permission, migration, routing, inline-adapter, and ledger-contract defects that were fixed and scoped-reverified before merge.
Validation
- Unit Tests and E2E passed on Ubuntu, Arch, and Fedora in CI run 29056868492.
- GoReleaser completed successfully in release run 29057085108.
- Release archives and checksums were published for macOS, Linux, and Windows on AMD64 and ARM64.
- Homebrew Formula and Scoop manifests were updated to
1.46.0.
Upgrade
brew upgrade gentle-ai && gentle-ai sync