What happened
PR #742 renamed slash commands from /triage to /fs-triage across 16 files. The review agent found stale references incrementally: iteration 1 found skill files, iteration 2 found retro spec/plan docs, iteration 3 found installer spec, iteration 4 finally approved. After each review, the author pushed a fix that addressed the specific files mentioned but missed others. This drove 5 commits and 4 review rounds just for stale-reference chasing — a task that a single grep -r could have resolved in iteration 1. See the review comment history and related issue #890.
What could go better
The review agent should detect when a PR is performing a systematic rename or migration (replacing one string pattern with another across multiple files) and proactively grep the entire repo for remaining instances of the old pattern. This would surface all stale references in a single review iteration instead of discovering them one file at a time. Confidence: high — the pattern is clear and the fix is mechanical. Issue #890 already proposes this but lacks the concrete triggering heuristic.
Proposed change
In the review agent definition (internal/scaffold/fullsend-repo/agents/review.md) or the code-review skill, add an instruction: when a PR's changes show a consistent pattern of replacing string A with string B across multiple files, the reviewer should run grep -r (or equivalent) for remaining instances of string A in the repo and report all hits, not just those in changed files. The heuristic for triggering this is: if the PR diff shows 3+ files where the same literal string is removed and replaced with a different literal string, treat it as a rename PR and scan exhaustively. This could also be implemented as a dedicated rename-audit skill that the review agent invokes.
Validation criteria
On the next rename-type PR in this repo (or a test PR that renames a string across multiple files), the review agent should report ALL remaining instances of the old string in a single review iteration, rather than discovering them across multiple iterations. Measure: the number of review iterations driven purely by stale-reference findings should drop from 3-4 to 1.
Generated by retro agent from #742
What happened
PR #742 renamed slash commands from
/triageto/fs-triageacross 16 files. The review agent found stale references incrementally: iteration 1 found skill files, iteration 2 found retro spec/plan docs, iteration 3 found installer spec, iteration 4 finally approved. After each review, the author pushed a fix that addressed the specific files mentioned but missed others. This drove 5 commits and 4 review rounds just for stale-reference chasing — a task that a singlegrep -rcould have resolved in iteration 1. See the review comment history and related issue #890.What could go better
The review agent should detect when a PR is performing a systematic rename or migration (replacing one string pattern with another across multiple files) and proactively grep the entire repo for remaining instances of the old pattern. This would surface all stale references in a single review iteration instead of discovering them one file at a time. Confidence: high — the pattern is clear and the fix is mechanical. Issue #890 already proposes this but lacks the concrete triggering heuristic.
Proposed change
In the review agent definition (
internal/scaffold/fullsend-repo/agents/review.md) or thecode-reviewskill, add an instruction: when a PR's changes show a consistent pattern of replacing string A with string B across multiple files, the reviewer should rungrep -r(or equivalent) for remaining instances of string A in the repo and report all hits, not just those in changed files. The heuristic for triggering this is: if the PR diff shows 3+ files where the same literal string is removed and replaced with a different literal string, treat it as a rename PR and scan exhaustively. This could also be implemented as a dedicatedrename-auditskill that the review agent invokes.Validation criteria
On the next rename-type PR in this repo (or a test PR that renames a string across multiple files), the review agent should report ALL remaining instances of the old string in a single review iteration, rather than discovering them across multiple iterations. Measure: the number of review iterations driven purely by stale-reference findings should drop from 3-4 to 1.
Generated by retro agent from #742