review agent repeating the same nit comment on every file: loop detection fix

TL;DR

Deduplicate before posting. Collect all candidate comments first, group them by rule and normalized message, and post each unique nit once as a summary instead of once per file. Cap per-file nits so one style gripe can never produce thirty comments.

Verbatim error

review agent repeating the same nit comment on every file: loop detection fix

Steps

1. Confirm the pattern. List the agent's comments on the PR and group by message text; you are looking for one message appearing on many files with only the filename differing.

Expected: a clear cluster, like the same "consider renaming" note on 20 files.

1. Add a dedupe pass before posting. Normalize each candidate comment (lowercase, strip the filename) and keep a set of already-planned messages; drop any candidate whose normalized form is already in the set.

Expected: 20 identical nits collapse to 1 before anything is posted.

1. Convert repeats into a summary. When a nit would have fired on 3 or more files, post one top-level review comment naming the pattern and listing the affected files, instead of any inline copies.

Expected: the PR shows one summary comment like "naming nit found in 12 files" with the file list.

1. Cap nits per review. Set a maximum (say 5) for low-severity style comments per run; anything beyond the cap goes into the summary or is dropped.

Expected: a run can never spam the timeline no matter how many files match.

Use this when

• A review agent posts the same comment on many files.
• You need comment deduplication in an agent review pipeline.
• Nit-level comments are drowning out real findings.

Not for this skill when

• The repeated comments are genuinely different findings that happen to share a template (check the details before deduping).
• A human reviewer posts similar comments manually (that is a style choice, not an agent loop).
• Comments duplicate because of a retry bug posting twice (that is an idempotency problem, not a nit loop).

Variant phrasings

• review bot posting same comment on every file
• agent spamming duplicate nitpick comments PR
• dedupe review agent comments

Why it happens

The agent evaluates each file independently with no memory of what it already said. A style preference that fires on one file fires on all of them, and since findings are posted per file as they are found, there is no moment where the agent sees the whole set and realizes it is repeating itself. The repetition is an emergent property of file-at-a-time processing.

Edge cases

• Severity matters: never dedupe high-severity findings into oblivion. Dedup and summarize nits; post real bugs inline even if similar.
• Normalization that is too aggressive merges distinct findings ("null check missing" on different variables is arguably one pattern, arguably two). Keep the file and line in the summary list so nothing is truly lost.
• If the team disagrees with the nit itself, the fix is turning off that rule in the agent's config, not better deduping. Dedupe handles volume; config handles correctness.
• Track suppressed duplicates in the run log so tuning the rules later has data to work from.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_80-cGW66vAgWQCGH8G2FZw