[cirolini]

Fixes found reviewing v3.0.0 before announcing it. Meant to ship as v3.0.1, together with the max_tokens 8192 default that is on main but not in any tag (@v3 still runs 2048).

Fixed

1. Unreviewed files dropped out of the summary on the next push. last_reviewed_sha advanced even when files were skipped, the diff was truncated, or the model output was unusable. It now advances only after a complete run whose comments were delivered, and the summary says the next push reviews that range again.
2. Findings counted as posted when the review failed to post. They are now listed in the summary and retried on the next push.
3. Incremental review after a merge from main or a rebase → 422. The compare is checked first (status != ahead, or a merge commit in the range, falls back to a full review with the reason in the summary). As a second guard, findings are placed against the PR's own diff and files outside it are not sent to the model.
4. One-click suggestions on relocated findings. A finding snapped up to 3 lines, or with its range shortened, gets relocated=True and its suggestion is shown as a plain code block, not as ```suggestion.

Security

5. .genai-review.yml is read from the base commit over the API, not from disk. Before, with a checkout the PR author controlled it (ignore_paths: ["**"], max_comments: 0, base_url pointing at their own server so it receives the key). Without a checkout, which the README recommends, the file was never read. base_url is now an action input only.

Docs

• action.yml: "ChatGPT GitHub Actions" → "GenAI Code Review" (that Marketplace slug is free). The listing itself changes only when a release is published to the Marketplace.
• The migration guide and release notes said @v2 was untouched. It points at v2.1; they now say so.

Tests

284 pass (was 267), lint clean. New regression tests for each item. The compare fields (status, total_commits, commits[].parents) were checked against this repository's real API: 4dafa5d...a3372f9 is ahead with three 2-parent commits, which is exactly case 3.

🤖 Generated with Claude Code

[github-actions]

Code review did not run

The openai provider failed and no review was produced: the API key was rejected (401). Check that the secret you pass to api_key is set and has not been revoked.

_{model: gpt-5.6-luna}