NPU route: reasoning_content, as llama-server returns it

[bong-water-water-bong]

Checklist item 6: the NPU route returned a thinking model's reasoning inside content. Lemonade's clients expect reasoning_content, as llama-server returns it.

Change: app/think_split.h splits the output:

• it handles tags split across tokens;
• it handles a think block the chat template opened (Qwen3.6);
• a literal <think> after the answer stays in the content;
• whitespace around the block is dropped.

It works for streamed and non-streamed responses. "reasoning_format": "none" keeps the raw text, as in llama-server.

Checks:

• tests/think_split_test.cpp: 6 cases, run in CI.
• Qwen3.6-35B-A3B on the NPU (strixhalo): content Paris with reasoning_content of 471 characters, the same streamed and non-streamed, across 3 non-streamed repeats. reasoning_format: none returns the raw thinking in content.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 21d37fa

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 21d37fa)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Potential Buffer Overflow The ThinkSplit::feed function may potentially lead to a buffer overflow or undefined behavior if the tag string is empty or if buf.size() is less than tag.size(). This could happen if the tag variable is constructed incorrectly or if the buf is manipulated in an unexpected way. The code should validate that tag.size() is not zero and that buf.size() is greater than or equal to tag.size() before performing string operations. const std::string tag = in_think ? "</think>" : "<think>"; const size_t p = buf.find(tag); const bool opens = p != std::string::npos && !in_think && !content_started && buf.find_first_not_of(" \t\r\n") >= p; // only whitespace before it if (p != std::string::npos && (in_think || opens)) { out(buf.substr(0, p), emit); buf.erase(0, p + tag.size()); in_think = !in_think; continue; } // hold back the longest suffix that could start the tag size_t keep = 0; for (size_t k = std::min(tag.size() - 1, buf.size()); k > 0; --k) if (buf.compare(buf.size() - k, k, tag, 0, k) == 0) { keep = k; break; } out(buf.substr(0, buf.size() - keep), emit); buf.erase(0, buf.size() - keep); return; } Logic Error in Tag Detection The logic for detecting whether a tag opens a thinking block (opens) may not correctly handle all edge cases. Specifically, the condition buf.find_first_not_of(" \t\r\n") >= p might not correctly identify when the tag is preceded only by whitespace, especially if the buffer contains only whitespace or if the tag is at the beginning of the buffer. This could lead to incorrect splitting of content. const bool opens = p != std::string::npos && !in_think && !content_started && buf.find_first_not_of(" \t\r\n") >= p; // only whitespace before it if (p != std::string::npos && (in_think || opens)) {

[github-actions]

Persistent review updated to latest commit 21d37fa