Reasoning & effort

Liminal separates three independent dials that are easy to confuse:

Dial	What it controls	Keys
Reasoning budget	How deeply the model deliberates internally per turn	AGENT_REASONING_*
Output effort	How thorough the produced deliverable is	AGENT_EFFORT
Model routing	Which model runs the turn	AGENT_INTENT_*, AGENT_FAST_MODEL

Reasoning budget & surface

Per-turn reasoning effort and think-depth are inferred from the intent classifier (AGENT_REASONING_BUDGET, on). The surface decides how the model reasons:

• external (default) — the model reasons via the observable think() + reason() tools, so you can see its deliberation in the UI.
• native — use the provider's native reasoning stream.
• auto — pick per turn.

When the classifier is off or low-confidence, AGENT_REASONING_DEFAULT_EFFORT (high) applies. AGENT_EFFORT_LEARN (on) records per-intent reasoning outcomes and reuses the best as a prior.

Variable	Default	Purpose
AGENT_REASONING_BUDGET	on	Infer per-turn reasoning effort / think-depth
AGENT_REASONING_SURFACE	external	native | external | auto
AGENT_REASONING_DEFAULT_EFFORT	high	Fallback when classifier is off/low-confidence
AGENT_EFFORT_LEARN	on	Learn best reasoning effort per intent

Output effort

AGENT_EFFORT (low · medium · high · xhigh, default medium) is not reasoning — it's how complete, edge-case-covered, and polished the deliverable is. It's injected each turn as a system directive and scales completion max_tokens (≈0.75× low … 1.5× xhigh). The named rule R-EFFORT ties model behavior to it; high/xhigh turns suspend the default brevity rule unless you asked for something short.

Set it in Settings or .env. Raise it for thorough research/reports; lower it for quick lookups.

Model routing

A fast-model intent classifier (AGENT_INTENT_INFERENCE, on) labels each turn and can route knowledge/introspection turns to the fast model (AGENT_INTENT_ROUTING) when confidence clears AGENT_INTENT_FAST_THRESHOLD. The main model runs the full ReAct loop; the fast model (AGENT_FAST_MODEL) handles classification, distillation, rewrites, critics, and other background JSON tasks. See Configuration for the model keys.

Rule of thumb

• Quick factual or navigation turn → low effort, fast model is fine.
• Hard multi-step implementation → high reasoning, main model.
• Comprehensive report/deliverable → AGENT_EFFORT=high/xhigh (output), independent of reasoning depth.

These dials are also distinct from workflows (parallel fan-out).