Big PR? CodeMouse Now Reads It in Chapters

Every AI model has a reading limit — a maximum amount of text it can hold in its head at once. The industry calls it a context window. For code review, it means one simple thing: if your pull request is bigger than the window, the model literally cannot see all of it.

Until this week, CodeMouse handled that the way most tools quietly do: it reviewed what fit and moved on. A customer with a genuinely huge PR — ten million added lines, most of it a recorded dataset — showed us how badly that can go. The models saw a thin slice, found nothing wrong in it, and approved the whole thing. A clean bill of health for a PR nobody had actually read. That's worse than no review at all.

We shipped two rounds of fixes. Here's what changed, in plain terms.

First: no more pretending

A review that only saw part of your PR now says so, right on the PR. You'll get a clear note — "this PR changes 9,975 files; only 25 fit in the review window" — and CodeMouse will never auto-approve a PR it couldn't fully read. An approval from CodeMouse now means every reviewed line was actually reviewed.

Your source code also goes to the front of the line. If a PR mixes real code with generated files, lockfiles, or big data dumps, the code is what fills the review window — not ten thousand JSON files crowding out the one C++ file that mattered.

Second: reading in chapters

If the code in your PR is bigger than one reading window, CodeMouse now splits it into chapters and has the model read them in separate passes — up to four per model. Each pass reviews its slice; the findings are merged, deduplicated, and posted as one review, exactly as if a single very patient reviewer had read the whole thing. If any model manages to cover the entire change set this way, auto-approve switches back on.

Third: right-sized windows for every model

Different models have very different reading limits, and treating them all the same wastes the good ones. CodeMouse now looks up each OpenRouter model's actual context size and gives it as much of your diff as it can comfortably hold — bigger window for a 200k-token model, a safe minimum for smaller ones. Claude, Codex, and Gemini reviews were already special: they browse a full clone of your repository during the review, so they can go look at anything the diff summary didn't include.

What this means for you

Nothing to configure. Reviews of normal-sized PRs are unchanged. Large PRs get more coverage automatically, and impossibly large ones get honesty instead of a rubber stamp.

One habit still pays off: keep giant data drops (datasets, snapshots, generated artifacts) in separate PRs from your code changes. Models can't meaningfully review recorded data anyway, and a code-only PR gets you full coverage in a single fast pass.

FAQ

Does chunked review cost more?

Only when it actually runs. A PR that fits one window costs the same as before. A PR that needs three passes costs roughly three passes' worth of tokens — on your own API key, at your provider's normal (or batch) rates. CodeMouse adds no markup.

Which models get chunked reviews?

Models you run through OpenRouter — they're API-based, so the diff we send is all they see. Claude, Codex, and Gemini run with full repository access and browse the code directly, so they've always been able to look beyond the diff.

What happens if my PR is still too big after four passes?

CodeMouse reviews as much as fits, posts its findings, and marks the review as partial — visibly, on the PR, with exact file counts. Auto-approve stays off for partial reviews.

Why don't you just review everything, no matter the size?

Cost and signal. A ten-million-line dataset would burn hundreds of dollars of your tokens per review round, and models produce noise, not insight, when asked to "review" machine-generated data. We'd rather tell you honestly what was covered.