Promote tool_call blocks from reasoning to content

[Thump604]

Summary

• Teach BaseThinkingReasoningParser to detect <tool_call> blocks inside reasoning and reclassify them as content so the downstream tool parser can find them
• Fixes Failure Mode B (tool calls inside <think> blocks lost by parser pipeline), confirmed on Qwen 3.5 27B/35B/122B/397B and Qwen 3.6 30B/35B
• Both non-streaming and streaming paths
• Related: Fix Qwen3 reasoning tool calls embedded inside think vllm-project/vllm#39055 (non-streaming only, known bugs)

Non-streaming

_promote_tool_calls(reasoning, content) scans reasoning after _extract_complete_reasoning splits it from content:

• Closed blocks (<tool_call>...</tool_call>): appended to content
• Unclosed blocks (<tool_call>\s*[\{<]... spanning </think> boundary): prepended to content to reassemble with continuation
• Structural \s*[\{<] guard on unclosed regex prevents false positives on prose that mentions <tool_call>
• content = (content or "") handles the content=None case
• Logs warning when promotion fires (model misbehavior recovery path)

Streaming

Tool-call buffering state machine in the thinking phase:

• <tool_call> detected: start buffering, stop emitting as reasoning
• </tool_call> detected: emit buffer as content, resume reasoning
• </think> while buffering: flush buffer as content, transition to content phase
• Stream ends while buffering: finalize_stream() flushes as content
• _transition_to_content applies non-streaming promotion as catch-all for single-delta cases

Test plan

• 11 non-streaming tests (closed/unclosed/boundary/prose/JSON/multiple/warning)
• 7 streaming tests (full-text, large-chunk, think-ends-while-buffering, finalize, multiple)
• 2 composition tests (preserves trailing content, tool_choice=required)
• 1 composition test (reasoning parser + tool parser end-to-end, skipped without transformers)
• 121 existing parser tests: 0 regressions

[janhilgard]

Review: PR #433 — Tool call promotion from reasoning to content

Clean design. The two-path approach (regex for non-streaming, state machine for streaming) is the right call, and the \s*[\{<] structural guard on the unclosed regex is a smart way to avoid false positives on prose.

Should fix

1. Ambiguous ternary at line 514 — add explicit parentheses

final_content = promoted + (
    content_msg.content or "" if content_msg else ""
)

Python's grammar makes this (content_msg.content or "") if content_msg else "" (safe), but the expression reads as if it could be content_msg.content or ("" if content_msg else "") (AttributeError when content_msg is None). Explicit parentheses would remove all doubt:

final_content = promoted + (
    (content_msg.content or "") if content_msg else ""
)

Same pattern at line 552 but there the + makes operator precedence unambiguous.

2. Missing test: </tool_call></think> with no trailing content

The _thinking_tool_call path where </tool_call> and </think> are in the same delta AND after_think is empty has no test coverage. This exercises the content_msg = ... if after_think else None branch:

def test_stream_tool_call_closed_immediately_before_think_end(self, parser):
    text = (
        "<think>R\n"
        "<tool_call>\n"
        "<function=f><parameter=x>1</parameter></function>\n"
        "</tool_call></think>"
    )
    reasoning, content = self._stream(parser, text)
    assert content is not None
    assert "<tool_call>" in content

3. Missing test: _transition_to_content defensive promotion

The _transition_to_content path also calls _promote_tool_calls as a catch-all. No test exercises this path specifically (where a single large delta contains <think>...<tool_call>...</tool_call>...</think> and hits _transition_to_content directly without going through the streaming tool call state machine).

Minor / nice to have

• finalize_stream() is clean — if stream ends mid-tool-call, buffer is flushed as content. Good edge case handling.
• Closed blocks appended after existing content, unclosed prepended — correct for the reassembly scenario. The test_tool_call_spanning_think_boundary test validates this nicely.
• Warning logging on promotion is useful for operators debugging model misbehavior.
• Regex _TOOL_CALL_CLOSED_RE uses non-greedy .*? — correct for multiple tool calls.
• The test_prose_mention_not_promoted test verifies the structural guard works.

Overall

This is solid work — the state machine is well-structured, the regex guards prevent false positives, and 21 tests cover the important paths. CI 9/9 pass. The parentheses fix (#1) and the two test gap additions (#2, #3) are straightforward.

[Thump604]

All three items addressed in ff62be4:

1. Explicit parentheses on the ambiguous ternary at line 353
2. Added test_stream_tool_call_closed_immediately_before_think_end -- exercises the after_think empty branch
3. Added test_stream_single_delta_promotion_via_transition -- exercises the _transition_to_content catch-all path directly

142 passed, 1 skipped (transformers).

[janhilgard]

@Thump604 Looks good, all three addressed. Thanks for the quick turnaround.

[janhilgard]

All three items addressed in ff62be4. LGTM.

[Thump604]

@waybarrios This has been reviewed and approved by @janhilgard. When you have a chance, could you take a look and merge if it looks good?

[Thump604]

Approved by Jan, CI green. Wayner, review and merge when you have a moment?