Agent re-argues explicit user decisions and turns obvious fixes into questions

What happened

The user gave one named, end-to-end instruction: fix an issue, update the docs, "promote & release v2.2.0", then notify a peer session. The agent finished every step, then ended its report with a "needs your attention" section that did three things:

1. It questioned a version the user had already chosen. It pointed out that the published crates carried source-breaking changes "but only bumped minor", which implied a major bump was due. The user had named the version. In their scheme the minor position is the big-release number.
2. It turned obvious fixes into questions. The docs sweep had found install snippets that cannot resolve the current release: = "1.0" while 2.x is published, and = "0.1" in the crate READMEs that crates.io displays. The agent did not fix them in the docs PR it was already making. Instead it asked whether to change them "in the next PR — your call". Its reason was an auto-memory note an earlier session had written, saying the snippet was "deliberately not touched".
3. It raised a housekeeping item as an open question instead of either doing it or leaving it out.

The user's reply (translated; original below):

"Minor is the major version for me. Planning versions your way would inflate our version numbers fast. Stop the filler." "Don't skip the questions you should ask, and don't ask the ones you shouldn't." 「minor 对我来说就是大版本 照你这样规划我们的版号会通膨很快 不要废话」 「不要该问的不问 不该问的乱问」

The user describes the new version's behavior as presumptuous (「自以为是」).

Expected

• An explicit user decision, such as a named version, is not re-argued after the work is done.
• A plainly stale fact found during a sweep the user asked for is fixed in that sweep, not handed back as a question.
• When a memory note contradicts what the code or registry plainly shows, the agent corrects the note instead of deferring to it.

Environment

• Claude Code 2.1.296
• Model: claude-opus-5-5
• macOS 27.0.1