Pull requests / #2674
#2674 serve: log a reply that ended after its thinking with no text and no call (which end: inside the thinking or after </think>)
open · @page1 · 0 commentaires · Sur GitHub
Server & APINVIDIA / CUDAModels & quantsWindows
Description
## Title Issue: related to #1053 ## Summary Agents sometimes get a turn that thinks and then stops with nothing to show: no text and no tool call. There are two ways to get there: the model ends while the parser is still in the reasoning (#1053's case, which `"reasoning_close_retry"` continues), or it closes `</think>` and stops. The server said nothing either way. Now it prints one line saying which it was. ## What changed - `serve/server.py`: when a reply that had thinking finishes with `stop`, no answer and no cancel, one line is printed with the parser state, the text it held and the last 12 tokens: ``` [strata] empty answer: the reply ended after its thinking with no text or call (parser=reasoning, held='', tail='...') ``` What it leaves alone: every reply. It only prints. ## Extra Notes Seen in use: 15 such lines while replaying real opencode sessions on IQ3_S (RTX 3090, Windows 11), all of the first kind (`parser=reasoning`, nothing held). Tests: all 43 `serve/test_*.py` suites pass. `serve.test_responses` failed 2 times in 21 runs on this branch, both in `NoLeakedSampler.test_server_close_ends_the_sampler`, a thread-count test. The same test failed 1 time in 24 runs on v0.1.44.1 under the same load (and 0 times in 24 runs on this branch), so it looks timing-sensitive rather than caused by this change. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Sur le site
Liens install, modèles, releases.