Repository navigation
0.1.42: stop match in assistant prefill ends generation with finish=stop #1950
Description
Environment: Strata 0.1.42 release binary with serve/server.py at the same tag, Windows 11, RTX 5070 12 GB, single GPU. Reproduced with two models and their stock configs (Swift IQ2_XS, IQ3_XXS). Did not occur on 0.1.39 with the same stop list and full prompt shape.
Repro (<served-model> = the loaded model name):
curl -s http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"<served-model>","messages":[{"role":"user","content":"Write the body."},{"role":"assistant","content":"Here:\n```python\nfrom typing import List\n\n\ndef f(x):\n pass\n"}],"max_tokens":2048,"temperature":0,"reasoning_budget_tokens":1024,"stream":true,"stop":["\ndef"]}'Expected: the model continues the body. Actual: finish_reason: stop after ~35 served chars (fence and imports echo). The served text contains no newline-def.
Notes, each run repeatedly with identical results:
"stop": []or an impossible stop string: full correct body.["\ndef"]: 35-char stub. Same on the 337-token variant of this prompt.["\nclass"]on the same prompt: full 550-char body."\ndef"in the user message only, or"\n#"in a prefill: full answers.stream: falseand omitting the budget key give the same result on the full prompt.
Effect: any client that prefills the assistant turn and passes code stops gets truncated completions with a stop finish reason.
Activity
[-]0.1.42: stop string present in the assistant prefill halts generation after the first chunk (finish=stop), though never generated[/-][+]0.1.42: stop match in assistant prefill ends generation with finish=stop[/+]Niko1221 commentedon Oct 19, 2026
Thanks for taking the time to report this.
We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to read properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.
To reopen it takes about two minutes:
/reopen. A check runs and either reopens the issue or lists what is still missing.The rules are in CONTRIBUTING.md.