Reasoning and Inference

Logician exposes three separate controls. Keeping them distinct makes configuration predictable.

Control	Purpose	Default
Inference mode	Sampling parameters sent to the provider	none (Provider)
Thinking level	Provider-specific reasoning effort	off
Structured reasoner	Optional advisory pre-reasoning strategy	none

Inference modes

none leaves provider sampling parameters untouched. Other presets include auto, thinking-general, thinking-coding, instruct-general, instruct-reasoning, instruct-coding, deterministic, creative, and analytical.

Open the selector with /mode or Ctrl+M/Alt+M. /inference-mode-cycle advances through presets.

Thinking level

Supported values are off, minimal, low, medium, high, and xhigh. Availability depends on the selected model; Logician clamps unsupported levels when models change.

Use /thinking <level> or the settings overlay. Thinking level is independent of whether reasoning text is collapsed, summarized, or expanded in the transcript.

Structured reasoners

Reasoners run before the ordinary tool-capable agent loop and provide advisory analysis. Available IDs are ssr, tot, got, reflexion, self_consistency, best_of_n, auto_cot, in_context_cot, and cover.

{
  "reasoner": "reflexion",
  "reasonerConfig": {
    "maxTrials": 2
  }
}

Use /reasoner <mode> to select and persist a reasoner, or /reasoner none to disable it. Reasoners do not replace verification: the main loop still owns tools, edits, and completion.

A practical starting point

Keep Provider mode, thinking off, and no structured reasoner until a task benefits from tighter sampling or additional deliberation. For an ambiguous design decision, try tot; for revision after a failed attempt, try reflexion; for output diversity with selection, try best_of_n.