Claude Code review cap honored exactly; all-findings flag adds none

The test was deliberately built so the cap would have room to bite: a 41-line cart.py with six planted defects presented as a 34-line-addition, 2-line-deletion diff. Every run used the same model, claude-sonnet-5-5, on version 2.1.296, with one process per call and each condition run in a fresh copy of the test folder — except the check on whether the cap persists, which reused one folder.

The most interesting signal is not the exact count but where it comes from. Because the same defect led every run, and because a 20-finding cap still returned only the 8 findings that existed, the cap looks like a ceiling applied to the response rather than an instruction the model follows. Timing supports the same picture: the slowest capped runs topped out at 8,208 ms while the fastest uncapped runs started at 11,769 ms, with no overlap, and output length differed by an order of magnitude (366–446 characters versus 2,746–3,245).

The author also flags what the setup could not settle. The claim that the feature arrived in version 2.1.288 was not verified, since only 2.1.296 was tested. Whether the cap survives inside an interactive session, why line numbers drift by one to three lines after the first finding, and why the uncapped count wobbles between 7 and 8 all remain open. An attempt to measure a third-party plugin was dropped after a 403 on retrieval and an environment guard blocking external code.