Greptile Summary

Fixes two related bugs affecting Ollama models configured with api: "openai-completions":

Root Cause 1: Missing num_ctx parameter - When using Ollama via OpenAI-compatible API, the num_ctx parameter was never sent, causing Ollama's server to default to 4096 tokens regardless of the configured contextWindow. The fix detects Ollama providers (by name or baseUrl heuristics) and injects num_ctx into the request payload via the onPayload callback, matching the behavior of the native Ollama API path.

Root Cause 2: Incorrect fallback model resolution - When a model wasn't found in the registry, the fallback code blindly used models[0] config instead of matching by model ID, potentially using the wrong contextWindow value. The fix now attempts to find the correct model by ID before falling back to models[0].

The implementation is backward-compatible, follows existing patterns (consistent with ollama-stream.ts:431), and properly chains existing onPayload callbacks. The detection logic uses provider name and baseUrl heuristics (port :11434 or path /ollama) which could theoretically false-positive, but the injected num_ctx parameter is harmless to non-Ollama providers.

Confidence Score: 5/5

• Safe to merge - well-structured bug fix following existing patterns with minimal risk
• The changes are narrowly scoped, backward-compatible, and consistent with existing codebase patterns. The Ollama detection logic is defensive, the payload injection properly chains callbacks, and the model resolution fix is a clear improvement. No new dependencies, security issues, or breaking changes.
• No files require special attention

_{Last reviewed commit: d2cd136}

Superseded by #29205, which carries forward the num_ctx OpenAI-compatible Ollama fix together with fallback model-ID token-limit resolution and additional tests.

Credit: @Sid-Qin for the core num_ctx direction and problem report captured in this PR.