Intermittent empty/truncated responses from qwen3.5

[GiuseppeR84]

Intermittent empty/truncated responses from qwen3.5 — likely think:false placement bug (see ollama/ollama#14793)

Symptom
Document ingest fails intermittently on nearly every PDF with:
FILE block "wiki/sources/<name>.md" was not closed before end of stream — likely truncation (model hit max_tokens, timeout, or connection dropped). Block dropped.

This happens regardless of document size (reproduced on both a 3-page and an 8-page source), and persists after:

• increasing repeat_penalty on the model (ruled out a token-repetition loop as the sole cause)
• raising the Context Window slider from 66K to 131K/205K characters (ruled out max_tokens as the sole cause)

Isolated repro via Settings → Provider tests
With both "Reasoning / thinking" and "Reasoning during ingest" set to Off, repeatedly pressing "Test connection" and "Test function" fails ~50% of the time with:
Model produced N characters of reasoning/chain-of-thought, but no actual response content...

Likely root cause
This matches ollama/ollama#14793 exactly: qwen3.5 silently ignores think:false when placed inside options on /api/generate — it only works as a top-level parameter on /api/chat. I confirmed directly against Ollama (bypassing LLM-Wiki):

• think:false top-level on /api/chat: 12/12 success
• reasoning_effort:"none" on /v1/chat/completions: 8/8 success
• No disable flag at all: intermittent — 4 fast successes then one call hung >2 min with no output

Question
Could you confirm which endpoint/parameter shape LLM-Wiki uses to disable reasoning for Ollama-native providers (both for chat and for "Reasoning during ingest" specifically), and whether it's placed as a top-level think field rather than inside options? Given how closely this matches #14793, this seems like a very plausible explanation for widespread ingest truncation with qwen3.5 on Ollama.

Environment

• LLM-Wiki version: v0.6.11
• Ollama version: 0.24.0
• Model: qwen3.5:latest (custom derivative with repeat_penalty 1.3)
• Endpoint: http://localhost:11434 (native, no /v1)

Helped by ClaudeCode in writing this