API Error 400: invalid_request_error when context accumulates large tool results

[HiepPP]

Issue Description

The Kimi API returns a 400 invalid_request_error when the conversation context accumulates large tool results across multiple turns. The error provides no specific details about what validation failed.

Error Details

Timestamp: 2026-02-26T04:58:43.109Z
Error Response:

{
  "error": {
    "type": "invalid_request_error",
    "message": "Invalid request Error"
  },
  "type": "error"
}

HTTP Status: 400 (non-retryable)

Token Progression Pattern

The session accumulated tokens steadily before the crash:

Turn	Input Tokens	Cache Read	Output Tokens	Note
1	20,748	0	~26	Initial request
2	32,031	0	~26	After file read #1
3	36,374	0	~322	After file read #2
4	4,374	32,000	~81	ToolSearch call - CRASH

Total context at error: ~44,000+ tokens

Reproduction Steps

1. Start a conversation with kimi-for-coding
2. Perform multiple file read operations that return large text content (~180-200 lines each)
3. Each file content gets added to conversation history as tool results
4. Make a ToolSearch call that returns deferred tool references
5. API returns 400 error immediately after

Key Technical Details

Message Structure Before Error

The conversation contained:

• thinking blocks with signatures (encrypted thinking content)
• tool_use messages with large input parameters
• tool_result messages containing full file contents
• tool_result with deferred tool references (71 total deferred tools returned)

Cache Configuration

• Using 32k token cache (cache_read_input_tokens: 32000)
• Cache miss on final request

Client Behavior

• Claude Code CLI v2.1.59
• Implements 11-attempt retry logic
• First attempt fails with 400, no retries attempted (correct behavior for client errors)

Expected Behavior

API should either:

1. Accept the request and process normally (if within limits)
2. Return a specific error indicating:
   • Context length exceeded (with actual vs limit)
   • Message structure validation failed
   • Tool result size limit exceeded
   • Cache token limit exceeded

Actual Behavior

Generic invalid_request_error with message "Invalid request Error" provides zero diagnostic information.

Environment

• Model: kimi-for-coding
• Client: Claude Code CLI v2.1.59
• Context management: 128k max output tokens (capped to 64k)
• Tool use: Yes, multiple tool results in history
• Cache: 32k ephemeral cache enabled

Questions for Kimi Team

1. What is the actual context length limit for kimi-for-coding?
2. Are there per-message size limits for tool results?
3. Are there limits on deferred tool references count?
4. Can error messages include diagnostic details about which validation failed?
5. Is there a way to preview context usage before sending requests?

Suggested Improvements

1. Descriptive errors: Instead of "Invalid request Error", return:

   {
     "error": {
       "type": "context_length_exceeded",
       "message": "Context length exceeded limit",
       "details": {
         "current_tokens": 44374,
         "limit": 40000,
         "exceeded_by": 4374
       }
     }
   }

2. Headers: Return usage headers (x-request-tokens, x-context-limit) for debugging

3. Documentation: Publish clear limits for:

   • Total context length
   • Individual message size
   • Tool result size
   • Number of tool definitions
   • Cache token limits