Summary

• Problem: discoverOllamaModels() hard-codes contextWindow: 128000 for every model. While this doesn't trigger the context-window guard, it shows misleading info in the TUI (e.g. "tokens ?/128k" for a 32k model). Additionally, users configuring Ollama as a Custom Provider get DEFAULT_CONTEXT_WINDOW=4096, which triggers "Model context window too small (4096 tokens)" and blocks the agent entirely.
• Fix: After discovering models via /api/tags, query each model's actual context_length via POST /api/show and use it as contextWindow. Queries are parallelised with a 3s per-model timeout; on failure, the 128k fallback is preserved.

Change Type

• Bug fix

Scope

• Providers / Ollama

Linked Issue

• Closes [Bug]: Context window not detected correctly from Ollama API (detects 4096 instead of actual value) #24068

Changes

• src/agents/models-config.providers.ts:
  • Added queryOllamaContextWindow() — queries /api/show and extracts context_length from model_info (key varies by architecture, e.g. qwen2.context_length, llama.context_length)
  • Updated discoverOllamaModels() to query each model in parallel and use real context window values

Security Impact

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? Yes — additional POST /api/show per model during Ollama discovery (local network only, 3s timeout)
• Command/tool execution surface changed? No
• Data access scope changed? No

Compatibility

• Backward compatible? Yes — falls back to 128k default if /api/show fails
• Config/env changes? No
• Migration needed? No

Greptile Summary

This PR fixes Ollama model discovery by querying each model's actual context window via /api/show instead of hardcoding 128k for all models. The fix resolves two issues: misleading TUI display for models with smaller context windows, and blocking errors when Ollama is configured as a Custom Provider (which defaults to 4096 tokens, below the 16k minimum).

• Adds queryOllamaContextWindow() that queries /api/show and extracts context_length from model_info (architecture-specific key like qwen2.context_length)
• Updates discoverOllamaModels() to query models in parallel with a 3s timeout per model
• Falls back to the existing 128k default if the query fails
• Uses Promise.allSettled to ensure one failing model doesn't block discovery of others

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The implementation is well-designed with proper error handling, graceful fallbacks, and parallel execution with timeouts. It addresses a real bug without introducing breaking changes or security risks. The 3s timeout and Promise.allSettled pattern ensure resilient discovery even if some models fail.
• No files require special attention

_{Last reviewed commit: c6abbbc}