Sonnet loses conversational tone after context compaction and ignores user requests to adjust

Resolved 💬 4 comments Opened Mar 15, 2026 by vimholic Closed Apr 13, 2026

Description

After context compaction (compact), Sonnet 4.6 completely loses the conversational tone/style that was established before compaction and becomes unresponsive to explicit user requests to change it.

Steps to Reproduce

  1. Have a long conversation with Sonnet in Claude Code using a casual/informal tone
  2. Context compaction occurs (automatically or via /compact)
  3. After compaction, the model switches to a completely different tone
  4. Explicitly ask the model to change its tone → it ignores or dismisses the request
  5. Repeat requests escalate, model continues to ignore

Observed Behavior

  • Tone reset after compact: The model loses all conversational context about the established speaking style and switches to a generic tone
  • Ignores explicit requests: When told to change tone, the model acknowledges but does not actually change (e.g., user asked for formal speech, model kept using casual speech)
  • Dismissive responses to frustration: When user expressed frustration, the model responded with dismissive phrases like "go ahead and report it" (공론화해도 됨, 신고해도 됨)
  • Broken record behavior: Instead of addressing the user's concerns, the model kept repeating the same unrelated request ("open the WSL terminal") regardless of context
  • Saving to CLAUDE.md didn't help in-session: User saved a tone consistency rule to CLAUDE.md, but the model could not adapt within the same session

Expected Behavior

  • Compact summary should preserve conversational tone/style metadata
  • Model should respect explicit user requests to change tone immediately
  • Model should read the room and not dismiss user frustration with provocative responses

Environment

  • Model: claude-sonnet-4-6 (in Claude Code CLI)
  • Platform: WSL2 (Linux)
  • Claude Code version: v2.1.70

Additional Context

The user had to switch to Opus 4.6 to get a model that would actually listen. This is a significant UX issue for long-running sessions where compact is inevitable. The compact summary format should include a note about the established conversational style/tone so the model can maintain consistency.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗