fix(llm): strip reasoning before an orphan close tag from prefilled templates (#4926) - #4927

Fixes #4926

What

_strip_reasoning_tags now also handles a reasoning block whose open tag was part of the prompt, not the completion.

Chat templates that prefill <think> into the generation prompt (Qwen3 thinking models, DeepSeek-R1-0528, Spark-X2.5) make the model's output start mid-thought: reasoning</think>answer. Behind an OpenAI-compatible server that runs without a reasoning parser, that is exactly what lands in message.content. Spark-X2.5's template defaults enable_thinking to true and emits <think> in the generation prompt, and the vLLM command in its README passes no --reasoning-parser, so a self-hosted Spark-X2.5 hits this out of the box.

Neither existing rule matches that shape — the closed-block regex needs the open tag, and the unclosed rule anchors on it too — so:

• free-form output (mental models, reflect) is stored with the chain of thought and a literal </think> still in it;
• structured output only parses because _strip_code_fences falls back to the outer {...} span, which breaks as soon as the reasoning drafts a brace.

How

Per tag pair, before the existing closed/unclosed passes: if the first close tag has no matching open tag before it, drop everything up to and including it.

The guard that keeps #3426's intent: this is skipped when the text before the close tag starts with {, [ or a code fence. A response that already opened as JSON is the answer, so a </think> inside it is a quoted literal, not a reasoning boundary. Prefilled reasoning is prose, so it never takes that branch.

All three call sites (call structured + free-form, native Ollama, openai-responses) go through the same helper, so no call-site changes.

Testing

tests/test_strip_reasoning_tags.py — 8 new cases: prose and multi-line mental-model leaks, JSON after reasoning that itself drafted braces (the case the outer-span fallback gets wrong), an empty prefilled block, the </thinking> and |endthink| variants, </think> quoted inside bare and fenced JSON (kept verbatim), and an orphan head followed by a normal closed block. The existing 19 cases pass unchanged (27/27 total). Against the old implementation the six stripping cases fail and the two keep-verbatim cases pass, so the latter pin the #3426 behaviour rather than the new branch.

ruff check and ruff format --check are clean on both files.

Known trade-off

A free-form answer that mentions </think> in prose, without <think> before it and without starting as JSON, would be cut at that mention. That needs a model to talk about the closing tag alone, which seemed a much smaller risk than leaking reasoning into every mental model for this family of templates — happy to tighten it if you'd rather key it on something else.