The incremental </think> search missed tags split across tokens, inverting
reasoning/content classification. Detect the first </think> on the full
buffer and track the content boundary as a byte offset. If the model never
closes </think>, flush the buffered text as content so clients always
receive the response. Chat UI now renders reasoning_content too.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- Unified synchronous generate() core with callback; both streaming and
non-streaming paths run it via spawn_blocking (context: std Mutex).
- build_prompt now passes tool definitions to the template (was dead) and
embeds assistant tool-call history as XML matching the parser format;
fixes double <tool_response> wrap and template set-scoping bug.
- Tokenize with AddBos::Never (template owns <s>) to remove double BOS.
- Streaming: preserve inter-word spaces (per-chunk trim removed), add
[DONE] + usage chunk, emit error events, single-shot tool_calls delta.
- Strict model validation (400 on unknown model); health/UI/README aligned
to minicpm5-1b-fable5-v2-thinking; auth returns JSON errors; n_ctx/
n_batch/n_threads env-configurable.
- Added 18 unit tests; cargo check/clippy/fmt clean.
- scripts/smoke-test.sh for post-deploy verification on the VPS.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
