🤖 Fix thinking slider memory and auto-resume for slow models

[ammar-agent]

Summary

Two bugs fixed:

1. Thinking slider didn't remember for Ctrl+Shift+T

Root Cause:

• Slider only updated thinkingLevel:{workspaceId} storage
• Ctrl+Shift+T toggle reads/writes lastThinkingByModel:{modelName}
• Slider changes weren't visible to the toggle

Fix:

• Added handleThinkingLevelChange that updates both storage locations
• Only saves active levels ("low"/"medium"/"high"), not "off"
• Matches the logic in useAIViewKeybinds.ts

2. Auto-resume didn't work after app restart during slow models

Root Cause:

• Models can take 30-60 seconds before first token
• If app restarts during this time:
  • History has user message but no assistant response
  • hasInterruptedStream() returned false (no partial/error messages)
  • Auto-resume didn't trigger

Fix:

• Updated hasInterruptedStream() to return true when last message is a user message
• Rationale: History never ends with user message in normal flow
  • Either assistant is responding (active stream, checked separately)
  • Or assistant has responded (completed message in history)
• If history ends with user message → incomplete conversation → needs resume
• Respects user preference: autoRetry flag is checked separately and persists across restarts

Changes

src/components/ThinkingSlider.tsx           |  16 ++- (thinking slider fix)
src/utils/messages/retryEligibility.test.ts | 176 +++ (comprehensive tests)
src/utils/messages/retryEligibility.ts      |   8 +++ (auto-resume fix)
3 files changed, 198 insertions(+), 2 deletions(-)

Testing

• ✅ 669 unit tests pass
• ✅ 8 new tests for retry eligibility logic
• ✅ Type checking clean
• ✅ Integration tests pass

Generated with cmux