SpecPrefill bug in Claude Code

[HLGRL]

When using SpecPrefill feature in Claude Code it works up to sometimes 3 to 4 messages or tool usage or file edits, then it abruptly stops displaying message like "Baked for..., Cogitated for.." without any work done.
I tried many draft models configurations, changing from Auto-detect to LLM/VLM for main model and draft models mainly for Qwen-35B family, but after couple of requests in Claude Code it just stops.

[fparrav]

I tested this more and can separate multiple failure modes:

1. With SpecPrefill enabled, Claude Code conversation can stop unexpectedly mid-task (matches this issue).
2. TurboQuant crash is separate (tracked in Continuous batching fails with broadcast shape error when Turbo Quant is enabled (Qwen3.5-27B 4-bit) #559): Cannot broadcast array ....

Important update:
Even with both specprefill_enabled=false and turboquant_kv_enabled=false, I still see interruptions in long-context agent runs.

What I see in logs (specprefill OFF + turboquant OFF):

• repeated [vlm_stream_generate] GeneratorExit for request ...
• and in one long prefill request:
  Aborting request ...
Prefill interrupted at 67584/110879 tokens: 1 request(s) aborted
Rescheduled 1 requests for re-prefill

So there may be an additional client-disconnect/timeout path in long-context agent sessions, independent of SpecPrefill/TurboQuant crashes.