kb-kimi

To fix Kimi K2 Thinking tool call failures, apply explicit handling for tool-call reasoning events and implement targeted retry strategies. This documents model quirks and provides workarounds to stabilize unreliable tool calls during agent workflows.

Moonshot.kimi-k2-thinking throws AccessDenied errors when certain cache options are enabled. You must avoid these specific cache configurations during cross-region inference to prevent AccessDenied errors and ensure reliable model execution.

Differences between Claude and Kimi K2 Thinking include distinct cross-region inference, cache behavior, and tool call stability. Kimi K2 exposes think-process events requiring explicit think-tag cleanup and markdown extraction, unlike Claude environments.

To clean up think tags and extract markdown from Kimi K2 reasoning events, apply explicit handling logic within your agent flow. This addresses model-specific reasoning exposure quirks to ensure clean markdown extraction.

You cannot use cache options with moonshot.kimi-k2-thinking for cross-region inference. Applying these cache configurations causes AccessDenied errors, so you must avoid them to maintain stable cross-region operations.