Review features

The review dashboardCopy Anchor Link

The review page is a side-by-side 30/70 layout: calibrated 0–100 gauges on the left; on the right, the streaming Markdown review, an architecture and blast-radius Mermaid diagram (rendered offline, with a raw-source toggle), cohort file walkthroughs and severity-ranked findings. Re-run scores or the written review independently with one click.

Context the models seeCopy Anchor Link

• Smart context (outside-diff impact): CodeOtter extracts the functions, methods and classes a diff modifies and scans files outside the diff for their callers and importers, so a change cannot silently break an unmodified call site.
• Dynamic context: each code hunk is extended upward to its enclosing function, method or class, read from the PR head (JS/TS, Python, Go, Rust, Java/C#, Kotlin, Swift, Ruby, PHP), and trailing context is cut to one line. Setting: Repository settings → Review → Context above a change (Off, 8, 16 default, 32 lines).
• Large pull requests: a diff over the budget is not cut at a character count. Files in the main language go first, then other code, then docs and config, then lockfiles and generated output. A file too big to fit shows its leading hunks (at most half the budget), hunks that only delete lines are dropped, and every file left out is still named with its line counts.

Self-checkCopy Anchor Link

After a language model review, a second pass of the same model scores each finding 0–10 against the diff (wrong or unsupported, plausible, certain defect; a wrong suggested fix lowers the score). Findings below the minimum are dropped and the rest are ordered by score, which is shown on each finding. Setting: Repository settings → Review → Self-check (Off, 3, 5 default, 7).

PR toolsCopy Anchor Link

The Tools card under the review runs these with the same context, diff budget and guidelines as reviews:

PR comment commandsCopy Anchor Link

Off by default (Repository settings → Review → PR comment commands). A comment whose first line is /review, /describe, /improve, /ask <question>, /docs or /changelog runs that action within a minute and replies on the pull request. Only owners, members and collaborators can trigger one (GitHub author_association; Forgejo and Gitea write permission), because anyone can comment on a public repository and spend its model budget.

Suggested fixesCopy Anchor Link

Findings carry line-level :L<line> anchors and exact suggestion replacements, shown as Suggested Fix cards with Copy fix and Post inline suggestion to GitHub (POST /api/review-suggestions). Authors can then click Commit suggestion on GitHub. Forgejo and Gitea review comments cover one line, so they show the replacement as a code block.

Team learnings and accepted fixesCopy Anchor Link

• Dismiss & remember rule: dismissing a finding stores a repository learning that is added to future language model and System 1 prompts.
• Accepted fixes: when a re-review finds that the author applied an earlier CodeOtter suggestion, the fix is remembered, and later reviews look for the same kinds of problems. The 10 most recent are included in reviews; the list (up to 30) is editable.

Both are managed under Settings → Manage repositories → ⚙️.

Incremental delta reviewsCopy Anchor Link

Re-reviewing a pull request runs an incremental, commit-by-commit review (prevSha → headSha) of the new commits and tracks which earlier findings were resolved.

Posting to the pull requestCopy Anchor Link

Enable Post scores on PR, Add PR review as a comment and Post inline suggestions to publish scorecards, review summaries and committable fixes. Comments are updated in place on each run instead of piling up.