[Bug] Anthropic models: tool calls fail with "Input tag 'input_json_delta' ... does not match any of the expected tags" (streamed tool_use never assembled)

[sepivip]

MaxKB Version

v2.10.4-lts (also reproduced on v2.10.2-lts)

Problem Description

When an application has a custom tool attached and uses an Anthropic model (model_anthropic_provider), every message that triggers a tool call fails. The Anthropic API rejects MaxKB's follow-up request:

BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error',
'message': "messages.1.content.1: Input tag 'input_json_delta' found using 'type' does not match any of the expected tags"}}

and, in the other variant:

BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error',
'message': 'messages: text content blocks must be non-empty'}}

MaxKB returns this to the caller as HTTP 200 with the error text as the answer body, so it reaches end users as the assistant's reply.

Cause

input_json_delta is an Anthropic streaming event type, not a message content type. Anthropic streams tool arguments as a series of content_block_delta events of type input_json_delta, each carrying a partial_json fragment. These must be accumulated and assembled into a completed tool_use content block before the assistant turn is replayed in messages.

MaxKB appears to replay the raw streamed deltas instead. A search of the v2.10.4-lts tree finds no handling for either field:

grep -rn "input_json_delta\|partial_json" apps/     # zero hits

_yield_mcp_response() in apps/application/flow/tools.py handles only OpenAI-shaped tool deltas (tool_call_chunks / tool_calls / invalid_tool_calls). There is no branch for Anthropic content blocks, which is consistent with both error messages: the unassembled delta is sent verbatim, and the accompanying text block is empty.

Steps to Reproduce

1. Register an Anthropic model (reproduced with claude-sonnet-5, model_anthropic_provider).
2. Create a SIMPLE agent using that model.
3. Create a custom Python tool and attach it (tool_enable: true), e.g.

def live_price(query):
    return 'DJI Neo: 4242.42 GEL, in stock.'

4. Set the system prompt so the tool is actually used, e.g. "For any question about price or stock you MUST call the live_price tool."
5. Call the OpenAI-compatible endpoint with stream: true (required — with tools attached, stream: false is unsupported per [Bug] Exception:'generator' object has no attribute 'content' #6300 / [Bug] agent添加mcp,调用对话API 报错 #4846 / [Bug] Using non-streaming parameters when using MCP will return an error #4363):

curl -N "$MAXKB/chat/api/$APP_ID/chat/completions" \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"model":"gpt-3.5-turbo","stream":true,"messages":[{"role":"user","content":"How much is the DJI Neo?"}]}'

Expected: the answer quotes the tool's value (4242.42 GEL).
Actual: the answer is the BadRequestError text above.

The same setup works on an OpenAI-shaped provider

Identical agent, identical tool, identical prompt, only the model swapped to model_openai_provider (gpt-4o-mini): no error, and the answer contains 4242.42 GEL, confirming the tool executed. This isolates the defect to the Anthropic provider's streamed tool-call handling rather than to tools, sandboxing, or the endpoint.

Impact

Custom tools and MCP cannot be used at all with Anthropic models. Because tools require stream: true, and streaming is exactly where the assembly fails, there is no working combination for Anthropic users.

Tested on v2.10.2-lts and v2.10.4-lts (CE), Docker, single-node.

[Shenguobin0102]

感谢反馈，我们目前没有Anthropic的账号，暂时无法复现并解决这个问题。