Summary

The Ollama-compatible bridge can list the newer OpenAI models such as gpt-5.5, but /api/show reports an incorrect context length for them.

Expected effective context length for the current GPT-5.5 / GPT-5.4 family should be exposed as 400000 tokens to Ollama clients. The current bridge reports smaller values:

• gpt-5.5: falls back to the default 131072
• gpt-5.4, gpt-5.4-pro, gpt-5.4-mini, gpt-5.4-nano: currently use 272000

This causes clients that rely on Ollama model metadata to under-estimate the available context window.

Where it happens

In src/ollama/bridge.ts, inferContextWindow() returns values from CONTEXT_WINDOW_OVERRIDES, otherwise it falls back to 131072.

Current override table includes GPT-5.4-family entries at 272000, but does not include gpt-5.5, so GPT-5.5 receives the fallback value.

The value is then surfaced by /api/show in:

• parameters: num_ctx <contextLength>
• model_info["<architecture>.context_length"]
• model_info.context_length

Reproduction

1. Enable the Ollama bridge.
2. Query model metadata, for example:

curl http://localhost:<ollama-port>/api/show \
  -d "{\"model\":\"gpt-5.5\"}"

3. Check parameters / model_info.context_length.

Actual: 131072 for gpt-5.5.

Expected: 400000.

For gpt-5.4, actual is currently 272000; expected is also 400000.

Suggested fix

Update the Ollama bridge context override table so current GPT-5.5 / GPT-5.4 models expose 400000, e.g.:

• gpt-5.5 -> 400000
• gpt-5.4 -> 400000
• gpt-5.4-pro -> 400000
• gpt-5.4-mini -> 400000
• gpt-5.4-nano -> 400000

It may also be worth adding a small unit test around /api/show or inferContextWindow() so newly added model IDs do not silently fall back to 131072.