Automatically use Ollama model context length instead of default 4096

[Mostbesep]

Summary

When using local Ollama models in Pi Coding, the UI correctly displays the model's advertised context length (for example, Qwen 3.6 shows 262,144 tokens), but the actual runtime context appears to remain at Ollama's default value of 4096 tokens.

This can be misleading because users may believe they are working with the full context window while generation is actually truncated much earlier.

Reproduction

Model:

• qwen3.6:latest
• gemma4:26b

Pi Coding displays:

• Qwen 3.6 → 262,144 context

However during generation:

ollama ps

shows:

CONTEXT 4096

Ollama logs also show:

stop processing: n_tokens = 4095, truncated = 1

indicating that the request hit the runtime context limit.

Expected Behavior

One of the following:

1. Automatically set num_ctx to the model's supported context length (or a configurable value).
2. Expose a context length setting in the UI.
3. Clearly display the actual runtime context being used instead of only the model's maximum supported context.

Why This Matters

Large-context models such as Qwen 3.6 advertise 262K context windows. Users reasonably expect Pi Coding to take advantage of this capability when running locally through Ollama.

Currently, conversations may be truncated around 4096 tokens without any obvious indication that the runtime context differs from the model's advertised capacity.

[tmm6907]

This is an issue with your ollama configuration. You can either:

Use OLLAMA_CONTEXT_LENGTH (in the systemd service config file) to increase the default context length for ollama

OR

Make a Modelfile for your intended model and set PARAMETER num_ctx to whatever context length you need

The endpoint still isn't going to send the correct context length to pi (in my experience at least) so you''ll also need to also either create a ~/.pi/agent/models.json file or using pi's new dynamic provider registration. https://pi.dev/docs/latest/models

I use the Modelfile method with dynamic registration. store the modelfiles with this format: ~/models/qwen3.6:35b-128k/Modelfile . Then I just had my agent use the dynamic registration to make an extension that uses the endpoint and the models folder to set the appropriate context lengths at startup

[Mostbesep]

Yes, the problem was solved exactly like that.