fix: enable tool call parsing in streaming when reasoning parser is active

[Thump604]

Problem

When both --reasoning-parser and --tool-call-parser are configured, streaming tool calls are completely broken. The reasoning parser branch in stream_chat_completion() yields chunks directly without ever passing content through the tool parser.

Symptoms:

• tool_calls is null in every SSE chunk
• Tool call XML (<tool_call>, <function=...>) appears as raw text in content
• finish_reason is "stop" instead of "tool_calls"
• Non-streaming works correctly

Impact: Every CLI agent using streaming (Claude Code, OpenCode, Roo, etc.) cannot execute tool calls when a reasoning parser is active. This affects all models that use both reasoning and tool calling: Qwen3, Qwen3.5, DeepSeek-R1 with tools, etc.

Root Cause

In stream_chat_completion(), lines 1926-1952:

if _reasoning_parser and delta_text:
    # ... reasoning extraction ...
    yield chunk  # <-- emits directly, NEVER runs tool parser
else:
    # ... tool parsing only runs HERE ...

The reasoning parser branch and tool parser branch are in opposite sides of an if/else — mutually exclusive. When _reasoning_parser is set, the else branch (containing all tool parsing logic) is never reached.

Fix

After reasoning extraction produces delta_msg.content, pass that content through the tool parser before emitting the chunk. This creates a sequential pipeline matching upstream vLLM's architecture:

1. Phase 1: Reasoning extraction (separate thinking from content)
2. Phase 2: Tool call parsing on the content portion
3. Phase 3: Emit chunk with reasoning + structured tool_calls

Handles three tool parser states:

• None (inside XML markup): suppress content, still emit reasoning
• tool_calls detected: emit structured tool_calls with reasoning
• Normal content: pass through as before

The else branch (no reasoning parser) is unchanged.

Testing

Tested on Apple Silicon (M2 Ultra, 128GB) with:

• Model: Qwen3.5-122B-A10B-5bit
• Parsers: --tool-call-parser qwen3_xml --reasoning-parser qwen3
• Flags: --enable-auto-tool-choice --mllm

Before (broken)

tool_calls: null (all chunks)
content: "\n\n<tool_call>\n<function=write_file>..."
finish_reason: "stop"

After (fixed)

tool_calls: [{"name": "write_file", "arguments": "{\"path\": \"hello.txt\", ...}"}]
content: null (tool XML consumed by parser)
finish_reason: "tool_calls"

Non-streaming regression test: still works correctly.

Related

• Fixes Streaming mode returns tool calls as raw content text instead of structured tool_calls delta #107
• Related: Fix streaming tool calls when reasoning parser is active #93, feat(server): implement sequential reasoning and tool-call parsing bridge #118, Add Qwen3.5 model support (text-only) and fix reasoning+tool streaming #127, fix: integrate tool call parsing with reasoning parser in streaming mode #148 (all attempt to fix the same issue)
• Upstream vLLM uses the same sequential pipeline approach (reasoning + tool calls docs)

[Thump604]

Closing — this fix was incorporated upstream in v0.2.6. Tool call parsing in the streaming reasoning path is now working in main.