Function looksVisionCapable

src/setup/model-detector.ts:213–225

(id: string)

 * always spinning up a separate vision sub-agent.
 */
export function looksVisionCapable(id: string): boolean {
  const lower = id.toLowerCase();
  if (/gemma-?4/.test(lower)) return true;            // Gemma 4 — multimodal (incl. 31B)
  // "vl" not surrounded by other letters: qwen2.5vl, qwen3-vl, qwen3vl, ...-vl-32b
  if (/(^|[^a-z])vl([^a-z]|$)/.test(lower)) return true;
  if (/vision/.test(lower)) return true;              // llama3.2-vision, etc.
  if (/llava|minicpm-?v|pixtral|moondream|internvl|cogvlm/.test(lower)) return true;
  // Cloud multimodal: GPT-4o family, Gemini, and every modern Claude (3.x AND 4.x —
  // sonnet/opus/haiku are all image-capable; the old `claude-3`-only test missed claude-4).
  if (/gpt-?4o|gemini/.test(lower)) return true;
  if (/claude-(3|4|5|sonnet|opus|haiku)/.test(lower)) return true;
  return false;
}

/**
 * Family/size heuristic for context window when the live API doesn't report one.