[divyangchauhan]

Summary

DiffVouch previously reviewed only the supplied patch and had no native ChatGPT subscription transport or reproducible evaluation runner. The native reviewer can now inspect repository context and run local checks, authenticate directly with a ChatGPT subscription, and resume frozen evaluations without launching Codex or Claude Code.

review
  collect the selected Git diff
  run the provider review loop
    inspect files and run Bash checks
    return structured findings on changed lines

eval
  prepare pinned benchmark and public-PR cases
  check subscription allowance before model calls
  save reviews, judgments, and resume checkpoints
  report coverage and compare completed cases

• Add native device sign-in, token refresh, and streamed subscription responses. Subscription exhaustion stops the run; there is no fallback to billable API calls. Earned resets require an explicit cutoff option.

• Add repository and Bash access to native API and CLI reviews, preserving provider reasoning and tool history with execution, output, and request limits.

• Freeze Martian, SWE-PRBench, and public personal PRs; retain development/held-out separation, grade archived competitor outputs independently, and report failures and uncertainty. Comparisons do not enforce a regression gate.

• Record the completed ten-PR main-versus-tools experiment in docs/eval-main-vs-tools-10.md and its paired data file. Downloaded corpora, sessions, binaries, and runtime checkpoints remain outside the commit.

• Add docs/review-quality-improvements.md, a sourced plan for deterministic context, staged discovery and verification, cross-file coverage, and concrete maintainability findings. These are research proposals, not implemented changes.

Evidence

• Before: Main 119d547 restricted the native reviewer to supplied patch chunks. On the ten matched PRs, it matched 13/32 benchmark references, scored 41.4% F2, and produced 28 findings judged valid and one judged invalid.
  After: The saved repository-enabled reviews matched 15/32 references, scored 47.2% F2, and produced 31 findings judged valid and one judged invalid. All ten pairs completed. Main used an isolated subscription compatibility adapter; the tool-enabled results were reused. The +5.8-point F2 interval is [-5.6, +15.2], so this is not conclusive evidence of general improvement and does not isolate file reading from the prompt change.
• GitHub CI passed the test job and all six Linux/macOS/Windows amd64/arm64 cross-builds on implementation commit afaa290. go test -race ./... and go vet ./... also passed locally. Transport tests cover native tool execution without provider CLIs, opaque history, token refresh, stream completion, allowance failures, and no API fallback. The main experiment adapter also passed these checks and retained the original main prompt.
• Paired counts, precision, recall, and F2 independently match the native Go reports. No prompt tuning or new full-campaign run was performed; the full campaign remains paused.

Merge Danger

Door: two-way

Code changes are reversible. Session files and evaluation checkpoints persist locally; reverting code does not remove them. The subscription transport depends on ChatGPT authentication and response protocols, which may change.

Blast Radius: review

File and Bash tools use the invoking user's filesystem and network access. Review-only behavior is instructed, not enforced by a filesystem sandbox. Existing API and CLI review behavior also changes because repository inspection is enabled. The ten-case result is a limited quality check, not a regression gate or an official competitor ranking.