Model skips a mandatory per-component review step and silently drops a fix from a reviewed pull request

Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 7, 2026

Summary

During a long code-review session, the model skipped a mandatory step of the user's written review procedure, which was loaded in every turn: split each pull request into its independent concerns and verify each one. As a result it dropped a real bug fix from a pull request it was told to take over.

What happened

  1. The user asked the model to review a ~1,400-line pull request and list every real issue it fixed. The procedure, in the auto-loaded memory and CLAUDE.md, requires a per-component review, with each concern verified against the vendor documentation before the verdict.
  2. The model produced a 10-item fix list from that pull request. The list left out one component the diff clearly contained: a server-side 1 MB page cap for paginated database reads. The diff had a dedicated helper function and a CHANGELOG line naming that behaviour.
  3. The user rejected the pull request and told the model to apply every real fix itself. The model applied the 10 items, so the missing fix was never applied.
  4. Later, the issue that reported the bug came up. Asked "is it implemented?", the model answered "no" correctly but did not connect it to the pull request it had already reviewed. The omission surfaced only after the user insisted on reviewing the related pull requests.

Expected behaviour

When a written procedure says "split the diff into its independent concerns, verify each", the model should enumerate every concern in the diff (every new function, every CHANGELOG line) before writing a verdict or a fix list, and should not drop any silently.

Impact

A verified bug fix was lost between review and implementation. The user had to find it by pushing the model, which is exactly what the written procedure exists to prevent.

Environment

  • Claude Code CLI
  • Model: claude-opus-5-5 (1M context)

View original on GitHub ↗