fix: retry failed LLM requests at the bridge layer

[Chumor]

• Retry provider exceptions and assistant errors with one bridge policy
• Resume failed harness turns and prevent duplicate streaming output

Fixes #58

Summary by CodeRabbit

• Bug Fixes
  • Improved reliability when completions and ongoing tasks encounter temporary failures.
  • Added controlled retries with increasing delays, retry limits, and cancellation support.
  • Prevented duplicate or conflicting streamed output during retry attempts.
  • Preserved conversation and extension state when recovering from failed tasks.
  • Retry settings are refreshed for each task session.
  • Final failures are clearly reported instead of leaving incomplete results.

[Chumor]

@coderabbitai review

[coderabbitai]

Note

GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer.

[coderabbitai]

Note

GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer.

[Zhou-Shilin]

Hi @Chumor! Thank you for taking the time to open this pull request. However, I am going to close it as the proposed fix is applied in the wrong location and introduces critical issues. During our testing, sending a message caused an immediate crash on iOS, and the reconnect logic failed to take effect on Android. Since similar issues have come up previously, I kindly ask that you thoroughly test all changes on real devices before submitting future pull requests, especially when utilizing AI tools to generate code. We appreciate your efforts and understanding~