Opus 5.5: /code-review auto-spawns 10 subagents, hits limit in 8 minutes

The author is a non-programmer who runs a small business and uses Claude Code's desktop app to delegate work, starting in July with a rule that any change to calculation logic for client documents must go through /code-review before pushing. From 7/30 to 9/17, those reviews ran with at most 1 subagent 23 times. After Opus 5.5 launched on 9/22 and the model, effort, and app version all changed on 9/23, the review launched 10 subagents at once for the first time, and the pattern repeated the next day.

The 10 subagents split into correctness and cleanup roles, but only 2 of 13 findings were real bugs, and the author adopted none of the 9 cleanup suggestions. The 10-agent runs hit the Pro plan's 5-hour limit in 8 and 4 minutes, with 16 of 26 subagents stopping mid-run over two days. The author's fix was to use one subagent with no context about the code's history, limited to correctness checks, following official guidance that a reviewer should not be biased toward code it just wrote. That run took 19 minutes and found 5 issues, 2 of which were real bugs that had not appeared in the client-facing document numbers.

The outcome hinges on whether the parallel behavior stems from the model's improved delegation, the max effort setting, or the app version, since the author did not isolate the trigger. For others who use /code-review habitually, the author suggests checking the app's usage breakdown when limits seem to drain faster, as the cause may be written there. The official settings to cap concurrent subagents or make the review user-invocable only have not been tested by the author.