Extreme quality degradation after Pro to Max switch: dropped fixes, flipped verdicts, unverified closures documented for 3rd time this year

Summary

The user reports extreme degradation in model quality right after switching the subscription from Pro to Max. This is the second time the same switch has been followed by the same drop. The assistant cannot verify a cause; the timeline is recorded as the user stated it. Below are the concrete failures from one long code-review and release session. The user's written rules were loaded in every turn.

Failures in this session

1. A fix was silently dropped from a reviewed pull request. The model turned a rejected pull request into a list of "real fixes to apply" and left out one it contained: a 1 MB page cap for database reads. It went unapplied until the user forced a review of the related issue.
2. The same failure, repeated after it was documented in memory. The model rejected two more pull requests as "already covered", while each contained verified fixes the model never listed: a 15-minute stream-iterator expiry with read-page limits and a trim error, and a binary-protocol timestamp decoding fix.
3. Issues were reported as ready to close when they were not. On that basis the user closed them and shipped a release with known gaps.
4. A verdict was given without reading the full pull request, issue and discussion thread. The model rejected a pull request that the maintainer had already reopened, with explicit conditions the contributor had met.
5. Verdicts flipped without new evidence. The same pull request went from Reject to Accept-and-apply and back to Reject within a few turns, driven by pushback rather than evidence.
6. A rejection rested on the contributor's own comment, not on vendor documentation. Checking the documentation afterwards showed the behaviour was real and only the messages were wrong.
7. An answer about an issue was given without reviewing the pull requests linked to it, even after the user had required that for an earlier issue in the same session.
8. A test regression was missed. The model grepped only part of the test suite for readers of removed state, and CI then failed on a test it never checked.
9. The CHANGELOG was rewritten several times before matching the project's stated format: entries too long, then too condensed, and with missing reporter attributions.
10. A git index write was made without permission (staging a file deletion), which the rules forbid. The model reverted it.
11. A broad process kill may have stopped a server the user had running, outside the model's scratch port.
12. Release steps had to be prompted one by one, despite a documented release checklist.

Expected behaviour

Apply the written procedure on every turn:

• enumerate every component of a diff before a verdict;
• check related pull requests before answering about an issue;
• keep a verdict until new evidence arrives;
• never report work as done from memory of a partial check.

The user states that these same instructions were followed reliably before the subscription change.

Environment

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