vLLM glm47 parser duplicates tool call arguments in streaming mode causing HTTP 400

[Million-mo]

Problem

When using vLLM with the glm47 tool-call parser (used by GLM-4.7/5.0/5.1/5.2 models), streaming tool calls intermittently produce duplicated arguments — two complete JSON objects concatenated together:

{"path": "scratchpad:///fta-workspace/excavator-engine-overheat-workspace"}{"path": "scratchpad:///fta-workspace/excavator-engine-overheat-workspace"}

This causes HTTP 400 errors on the next model request:

Extra data: line 1 column 76 (char 75)

Root Cause

vLLM's glm47 parser (vllm/parser/glm47_moe.py) has a streaming bug where _create_remaining_args_delta() re-emits already-streamed arguments in the final chunk. This is documented in:

• [Bug]: GLM tool_call required + streaming causes JSON repetition vllm-project/vllm#47504 (OPEN) — GLM tool_call required + streaming causes JSON repetition
• [Bug]: GLM tool-call streaming final chunks repeat metadata and combine arguments with finish_reason vllm-project/vllm#44098 (CLOSED, upstream unfixed) — GLM tool-call streaming final chunks repeat metadata and combine arguments with finish_reason
• [Bug]: GLM-5-FP8 malformed tool calls vllm-project/vllm#34449 — GLM-5-FP8 malformed tool calls with MTP

The same class of bug exists in SGLang:

• [Bug] MiniMax-M2 streaming tool_calls: malformed response with name=null and duplicated arguments sgl-project/sglang#23071 — MiniMax-M2 streaming tool_calls: duplicated arguments
• [Bug] GLM-4.7 Streaming Response Returns Invalid JSON Format for Multiple Tool Calls sgl-project/sglang#16371 — GLM-4.7 multiple tool calls produce concatenated JSON without separators
• Tool-call parsers lose or corrupt data at streaming chunk boundaries (multiple detectors) sgl-project/sglang#31915 (OPEN) — Tool-call parsers lose/corrupt data at streaming chunk boundaries

Environment

• Model: GLM-5.2 (deployed as svc/kimi-k2)
• Backend: vLLM (confirmed by chatcmpl-tool-* tool_call_id format)
• Proxy: LiteLLM
• Framework: AgentPool with pydantic-ai
• Context: >100k tokens, step_index=2

Workaround

PR #335 adds a defensive fix in AgentPool that detects and repairs concatenated JSON in ToolCallPart.args before the message history is sent to the next model request. This prevents the 400 error but does not fix the root cause in vLLM.

Upstream Fix Needed

The root cause is in vLLM's glm47 parser streaming path. The _create_remaining_args_delta() function should not re-emit arguments that have already been streamed. A fix exists only in the vllm-ascend plugin, not in upstream vLLM.