LM Studio + Qwen: prompt prefix is not byte-stable across turns, breaks prefix cache #26750

Description

LM Studio prefix caching requires the tokenized prompt prefix to remain identical across turns. With Qwen / QwQ-family chat templates served via LM Studio's OpenAI-compatible endpoint, two things make the prefix unstable across turns when OpenCode replays the conversation history as-is:

1. Assistant reasoning content is replayed. Qwen chat templates render historical assistant reasoning (think) blocks differently on later turns once a new user message is appended, so the same historical assistant turn tokenizes differently the next time it appears in the prompt.
2. Raw role: "tool" messages are tokenized non-deterministically. Qwen-style chat templates render past tool results inconsistently when they sit as standalone role: "tool" messages, but render them stably when they appear as <tool_response>...</tool_response> blocks inside the preceding user/assistant text.

Either behavior breaks LM Studio prefix caching even though the underlying conversation history has not meaningfully changed, so a long context can fully re-tokenize on every turn.

Expected: when OpenCode replays the same conversation history through LM Studio's OpenAI-compatible endpoint with a Qwen-family model, the prompt prefix is byte-stable, so the prefix cache hits on every turn after the first.

Actual: the prefix changes turn-over-turn, causing a large or full prompt-cache miss against LM Studio.

This is provider/model compatibility behavior specific to LM Studio + Qwen-style templates and should not affect Anthropic, OpenAI, or non-Qwen OpenAI-compatible backends.

Plugins

None

OpenCode version

1.14.46

Steps to reproduce

1. Run a Qwen / QwQ-family model under LM Studio (any model whose chat template uses the Qwen <tool_response> convention).
2. Configure OpenCode to use LM Studio's OpenAI-compatible endpoint as the provider.
3. Have at least one tool-using turn in the session, then send another user message so the previous tool result becomes history.
4. Observe LM Studio's prefix-cache stats on the second model request (and any later turn): cache hit on the prefix is much smaller than the actual unchanged history would imply, often near zero, because the rendered prompt prefix is not byte-stable.
5. (Optional) Compare against a non-Qwen OpenAI-compatible LM Studio model — prefix cache behaves normally.

Screenshot and/or share link

No response

Operating System

macOS

Terminal

iTerm2