Skip to content
Python

fix: detect Ollama vision models from naming conventions (fixes #337)

Manav.Shrivastavabfbbeee48cbc
diff --git a/coworker/providers/capabilities.py b/coworker/providers/capabilities.py
index f096ff3a..2548c1b4 100644
--- a/coworker/providers/capabilities.py
+++ b/coworker/providers/capabilities.py
@@ -24,9 +24,12 @@ def capabilities_for(model: str) -> ModelCapabilities:
# Ollama (local) models vary widely and many fake/mishandle parallel tool calls — assume
# tools work (we only point at tool-capable models) but stay conservative otherwise.
+ # Vision is detected from common model naming conventions (-vl, vision, llava, etc.).
if provider == "ollama":
+ _vision_patterns = ("-vl", "vision", "llava", "bakllava", "cogvlm", "minicpm-v")
+ has_vision = any(p in name for p in _vision_patterns)
return ModelCapabilities(
- tools=True, vision=False, parallel_tool_calls=False, streaming=True
+ tools=True, vision=has_vision, parallel_tool_calls=False, streaming=True
)
# Cloud-account providers (custom-added ids; curated ones answered from the matrix).