claude-code - 💡(How to fix) Fix [BUG] Thinking stalls exceeding conversation cache TTL [1 participants]

I tried waiting it out but got hit with API Error: Claude's response exceeded the 32000 output token maximum. To configure this behaviour, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable

Error Messages/Logs

TL;DR

Set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable to a higher value to potentially mitigate the issue of Claude Code stalling at round token counts.

Guidance

• Investigate the relationship between the stalling token counts (33.0k, 65.0k, 97.0k) and the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable to determine if increasing this value can prevent stalling.
• Verify that the conversation cache expiration is indeed causing the token usage jumps by monitoring the cache TTL and token usage patterns.
• Test with a smaller task to see if the stalling issue still occurs, which could help determine if the problem is task-size related.
• Consider reaching out to the Anthropic API support to inquire about any known issues with the Sonnet model and token usage.

Example

No code snippet is provided as the issue seems to be related to configuration and API interaction rather than code.

Notes

The exact cause of the stalling issue is unclear, and it's uncertain whether increasing CLAUDE_CODE_MAX_OUTPUT_TOKENS will fully resolve the problem. The fact that this worked in a previous version suggests a potential regression, but without the last working version specified, it's difficult to pinpoint the change that introduced this issue.

Recommendation

Apply workaround: Increase the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable to a higher value to potentially mitigate the stalling issue, as this directly addresses the error message received and may help in preventing the conversation cache from expiring due to token limits.