[provider.api_error] The API returned a response containing only thinking content without any text or tool calls.

[devxia]

Additional Diagnostic Information

Environment

• Kimi Code Version: 0.5.0
• Node.js: v24.15.0
• Platform: macOS Darwin 24.6.0 x86_64
• Model: kimi-code/kimi-for-coding (thinkingEffort=high)

Local Environment Ruled Out

• Proxy not involved: Confirmed via Clash Verge Rev traffic logs that api.kimi.com traffic is marked as DIRECT and does not go through any proxy node.
• Environment variables: HTTP_PROXY=http://127.0.0.1:7897 and HTTPS_PROXY=http://127.0.0.1:7897 are set, but proxy logs confirm the Kimi API connection is direct.
• Other models work fine: In the same session and network environment, switching to deepseek/deepseek-v4-pro produces no terminated or APIEmptyResponseError errors at comparable context lengths.

Error Types and Statistics

Error counts across historical sessions in the current working directory:

Session	Model	terminated	APIEmptyResponseError
session_04ec300e	kimi-for-coding	8	0
session_1bdc5439	kimi-for-coding	4	3
session_27f2da98	kimi-for-coding	4	0
session_29f5f704 (current)	kimi-for-coding	6+	1
session_d194ec6b	kimi-for-coding	0	0
session_f0e7f83a	kimi-for-coding	0	0

Key finding: Errors occur exclusively on the kimi-for-coding model. All turns using deepseek-v4-pro are error-free.

Current Session Error Timeline (session_29f5f704)

02:33:10  turnId=0 step=8  input=36,560 tokens → terminated (attempt 1/3)
02:33:31  turnId=1 step=2  input=37,480 tokens → terminated (attempt 1/3)
02:35:37  turnId=3 step=2  input=40,868 tokens → terminated (attempt 1/3)
02:49:34  turnId=6 step=1  input=43,704 tokens → APIEmptyResponseError (attempts 1, 2, 3 all failed)
02:xx:xx  User reports another terminated error

Root Cause Analysis of Both Error Types

A. Error: terminated

• Location: KimiStreamedMessage._convertStreamResponse → convertOpenAIError
• Nature: The SSE stream is abnormally interrupted during transmission.
• Pattern: Intermittent; retrying the same request may succeed. Frequency increases as context grows beyond ~36K tokens.

B. APIEmptyResponseError: The API returned a response containing only thinking content without any text or tool calls...

• Location: Agent.generate (main.cjs:53690)
• Nature: The SSE stream ends normally, but the response contains only thinking fragments with no actual text or tool_call content.
• Pattern: At 43K+ input tokens, all 3 retry attempts fail consecutively, ruling out a transient network issue.

Preliminary Conclusions

1. Not a local network/proxy issue — DeepSeek works fine on the same network, and Clash logs confirm direct connection to the Kimi API.
2. Not a context window overflow — 43K tokens is far below the 256K limit.
3. Suspected Kimi API server-side issue:
   • SSE long connections become unstable when generation time increases.
   • The kimi-for-coding thinking mode may consume an excessive portion of the output token budget, leaving no room for the final response.

Suggested Investigation Directions

1. Check whether the kimi-for-coding server-side SSE long-connection keep-alive / timeout mechanism is defective.
2. Verify if the output token budget allocation for thinking mode is reasonable (43K input causing exhaustion seems abnormal).
3. Investigate whether HTTP/2 stream control is triggering connection resets (Node.js v24 + undici enables HTTP/2 by default).

[wbxl2000]

Thanks for the detailed report — there are two separate issues here.

Error: terminated is a transport-level drop: the stream gets cut off
mid-flight and we were treating it as non-retryable, so it failed instead of
retrying. Fixed in #201 — these now retry transparently.

APIEmptyResponseError (only thinking content) is a different problem, and we
can't reproduce it without more data. To look into it we'd need either the
complete session ID or the kimi export for a session that
hit it — the ones in your report look truncated (e.g. session_29f5f704). If you
can grab either, please open a new issue for the APIEmptyResponseError case with
it and we'll investigate there.