Python ModelLimits drops max_output_tokens from model metadata

[mowree]

Description

The generated RPC schema includes max_output_tokens in ModelCapabilitiesLimits, but the public Python ModelLimits type does not. CopilotClient.list_models() deserializes model limits through ModelLimits.from_dict(), so the value is discarded.

This follows the same schema/public-type parity principle addressed for TypeScript in #2569, although this issue affects the Python model-list response path rather than the session override path.

Reproduction

from copilot import ModelLimits

limits = ModelLimits.from_dict(
    {
        "max_prompt_tokens": 128000,
        "max_output_tokens": 16384,
        "max_context_window_tokens": 144384,
    }
)

assert limits.max_output_tokens == 16384
assert limits.to_dict()["max_output_tokens"] == 16384

Currently, ModelLimits has no max_output_tokens attribute and to_dict() omits the field.

Expected behavior

ModelLimits should preserve max_output_tokens when deserializing and serializing model metadata.

[github-actions]

Investigated python/copilot/client.py (ModelLimits, ~line 1009) and the generated python/copilot/generated/rpc.py (ModelCapabilitiesLimits, ~line 23547).

Finding: this is a genuine bug. The generated ModelCapabilitiesLimits includes max_output_tokens, and the Python ModelLimits dataclass already exposes max_prompt_tokens, max_context_window_tokens and vision. However, it has no max_output_tokens field, and neither from_dict() nor to_dict() handles it. As a result, list_models() silently drops the value. Related types such as ModelLimitsOverride in session.py do carry it, so this is a parity gap and not intentional.

Fix: add max_output_tokens: int | None = None to ModelLimits, and read and write it in from_dict() and to_dict(). Also add a unit test covering the reproduction from the issue.

Labeled as bug.

Generated by Bug Handler for #2798 · copilot · auto · 16.9 AIC · ⌖ 6.35 AIC · ⊞ 7.3K · ◷