✅ OpenCodeReview: No comments generated. Looks good to me.

Thanks a lot for digging into the review-filter behavior, @ScarletCarpet — it's a genuinely subtle area and the change itself is clean (nice nil-guard and test coverage). 🙏

After looking closely, though, I don't think this is the right layer for the fix, so I'm going to close it. Here's the reasoning:

The review-filter output is a machine-parsed control signal, not user-facing text. The prompt asks the model to return a bare JSON array of comment IDs, and the consumer (executeReviewFilter → parseFilterResponse) does StripMarkdownFences + json.Unmarshal and then just calls RemoveByPathAndIndices. Nothing this task emits is ever shown to the user — the comments that survive filtering were already produced by MAIN_TASK, which already carries the language directive.

Because of that, injecting Always respond in <language>. here has no user-visible field to act on, and it can actually work against us:

• It nudges the model to emit prose/explanation, which directly contradicts the prompt's own "without any explanation" rule.
• Any surrounding prose makes json.Unmarshal fail, so parseFilterResponse returns nil and the filter silently becomes a no-op.

Contrast with DedupTask: it may look like precedent since ScanTemplate.ApplyLanguage applies the directive to dedup too — but dedup emits merged_content, a genuinely user-facing field, so localizing it there is correct. The review filter has no such field, which is exactly why it's different.

Overlap with #295: that PR moves the filter to function calling specifically to stop models from emitting natural language. This change pushes in the opposite direction, so if #295 lands the language directive becomes moot anyway.

If the real symptom is chatty models breaking the parse, the more robust fixes are the function-calling approach in #295 or more lenient JSON extraction inside parseFilterResponse. If you've hit a concrete case where filter output actually reaches the user, please do reopen and point me at that path — I couldn't find one, and it would change the assessment.

Thanks again for the contribution and for the careful tests — really appreciate the effort here!