Dynamic per-stage `think` control for Ollama reasoning models (`?think=true|false`)

[acato]

Summary

The current codebase generates <think>...</think> blocks with any Ollama reasoning model (Qwen 3.x, gpt-oss, DeepSeek-R1, etc.), then post-processes them out via RemoveThinkTagFromAssistantMessages in Wrapper.py. This is correct for cleanup but expensive at generation time — the framework pays the latency cost of thinking even when the call's purpose doesn't benefit from it.

A small dynamic affordance — per-stage ?think=true|false in the model URL — would let users (and the framework) get the best of both behaviors with no breaking change.

Background

Reasoning models default-on to thinking mode unless the chat API explicitly passes think: false. With the framework's URL parameter system (ollama://model?param=value) this should be a natural fit, except for one constraint: the existing URL parameter parser hard-casts every value through float():

AIStoryWriter/Writer/Interface/Wrapper.py

Lines 506 to 508 in 161b712

	QueryParams = parse_qs(parsed.query)
	# Flatten QueryParams

# Flatten QueryParams
for key in QueryParams:
    QueryParams[key] = float(QueryParams[key][0])

So ollama://qwen3.5:122b-a10b?think=false raises ValueError: could not convert string to float: 'false' before the value ever reaches the chat call. Booleans (and strings) need a small parser upgrade before this feature is possible.

Motivation — why dynamic rather than a single Config flag

Different pipeline stages benefit from different think settings:

Stage	Want thinking?	Why
Outline generation, story elements, per-chapter outlines	off	Wastes tokens on internal reasoning the framework cannot see; output is the artifact.
Scene-by-scene chapter writing (Stage 1)	off	Prose generation. Tokens should produce prose, not silent thinking.
Character / dialogue passes (Stage 2/3)	off	Same — refinement, prose-quality.
LLMSummaryCheck (chapter ↔ outline alignment verdict)	on	Structured judgment, true/false outcome. Thinking improves calibration.
GetFeedbackOnChapter, GetChapterRating (Stage 5 critique loop)	on	Critique is the canonical reasoning task; thinking lifts quality.
Eval / Info / Scrub structured-output passes	on	Same — better verdicts when the model is allowed to reason internally.

A single global flag misses this. A per-stage flag in Config.py (one per *_MODEL) bloats the config surface. A per-URL ?think=true|false rides on the existing URL-as-config mechanism, defaults to model-native behavior when omitted, and lets users mix-and-match across the 13 model knobs already in Config.py.

This also dovetails with a related architecture choice many users land on for v1 of dual-model setups: a fast generation model with thinking off (Qwen 3.x 122B, Mistral Medium 3.5, etc.) for the writing knobs, and a thinking-capable critic model (gpt-oss 120b, DeepSeek-R1 distills) for EVAL_MODEL / REVISION_MODEL / CHECKER_MODEL. The framework already supports per-stage model selection; per-stage ?think completes the picture.

Proposed implementation

Two small edits in Writer/Interface/Wrapper.py:

1. Smarter URL query parameter parser

Replace the float-only cast with bool → float → raw-string coercion so ?think=true, ?think=false, ?temperature=0.8, and (future) string parameters all work:

for key in QueryParams:
    raw = QueryParams[key][0]
    if raw.lower() == 'true':
        QueryParams[key] = True
    elif raw.lower() == 'false':
        QueryParams[key] = False
    else:
        try:
            QueryParams[key] = float(raw)
        except ValueError:
            QueryParams[key] = raw

2. Pop think from ModelOptions and pass it to the Ollama chat call

think is a chat-level kwarg in the ollama Python client, not a model option, so it has to live outside the options= dict and the ValidParameters allowlist:

# Extract chat-level kwargs (separate from Ollama model options)
ChatExtras = {}
if "think" in ModelOptions:
    ChatExtras["think"] = bool(ModelOptions.pop("think"))

# ... existing ValidParameters check / num_ctx default / JSON-mode handling ...

Stream = self.Clients[_Model].chat(
    model=ProviderModel,
    messages=_Messages,
    stream=True,
    options=ModelOptions,
    **ChatExtras,
)

Behavioral contract

• No ?think= in URL: behavior unchanged. Default Ollama / model behavior preserved. No breaking change for anyone.
• ?think=false: thinking suppressed; <think> blocks not generated; post-hoc strip becomes a no-op. Faster generation, no quality loss for non-reasoning tasks.
• ?think=true: thinking forced on (useful when overriding a model that defaults off, e.g., gpt-oss when used as a critic).

Compatibility

• Requires ollama Python client ≥ 0.4 (introduced the think kwarg). Current requirements.txt is unpinned at ollama; latest stable already supports this.
• No change to any non-Ollama provider path. The smarter parameter parser is provider-agnostic (it runs before the provider branch) but is strictly more permissive than the current parser — any URL that worked before still works.
• No change to Config.py defaults. Users opt in by adding ?think=true|false to whichever model URLs they want.

PR to follow.