fix: mistral tool parser drops calls with leading whitespace

[nhodge01]

Summary

The Mistral tool-call parser bails on detection when the model output has leading whitespace before [TOOL_CALLS] or [. This is the common case for Mistral on TRT-LLM — the chat template's BOS handling produces a leading space token, so generated text starts with " [{...}]" rather than "[{...}]". With the strict startswith check, neither branch matches and the parser falls through to the content-only path.

Symptom: tool_choice: "auto" returns the tool call as a raw JSON string in content rather than a structured tool_calls array. Most visible on parallel tool calls, which almost always come back with the leading space.

Reproduction

Against a Mistral Small 3.x model served via Triton with --tool-call-parser=mistral:

curl -s -X POST http://localhost:9000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model":"ensemble",
    "messages":[{"role":"user","content":"Weather in SF AND book a hotel for next Friday"}],
    "tool_choice":"auto",
    "tools":[
      {"type":"function","function":{"name":"get_weather","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}},
      {"type":"function","function":{"name":"search_hotels","parameters":{"type":"object","properties":{"city":{"type":"string"},"checkin_date":{"type":"string"}},"required":["city","checkin_date"]}}}
    ]
  }'

Before: content: " [{\"name\": \"get_weather\", ...}, {...}]", tool_calls: null, finish_reason: stop
After: content: null, tool_calls: [{...}, {...}], finish_reason: tool_calls

Fix

-        if not (full_text.startswith(self.bot_token) or full_text.startswith("[")):
+        stripped = full_text.lstrip()
+        if not (stripped.startswith(self.bot_token) or stripped.startswith("[")):

stripped is used only for the detection. The response still uses the original full_text for content, so non-tool responses where the model legitimately starts with whitespace pass through unchanged.

Testing

Manual end-to-end testing with Mistral Small 24B Instruct 2501 on Triton 26.03-trtllm-python-py3 (NVFP4, TP=2):

• Single tool call with tool_choice: "required" — passes before and after
• Single tool call with tool_choice: "auto" — fails before, passes after
• Parallel tool calls with tool_choice: "auto" — fails before, passes after
• Decline path (model chooses prose over tool) — passes before and after

I have not exercised the L0_* test suite locally — happy to add or run any specific tests if helpful.

CLA

CLA submission in flight to triton-cla@nvidia.com.