Repository navigation
Python ModelLimits drops max_output_tokens from model metadata #2798
Description
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"] == 16384Currently, 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.
Activity
github-actions commentedon Sep 30, 2026
Investigated
python/copilot/client.py(ModelLimits, ~line 1009) and the generatedpython/copilot/generated/rpc.py(ModelCapabilitiesLimits, ~line 23547).Finding: this is a genuine bug. The generated
ModelCapabilitiesLimitsincludesmax_output_tokens, and the PythonModelLimitsdataclass already exposesmax_prompt_tokens,max_context_window_tokensandvision. However, it has nomax_output_tokensfield, and neitherfrom_dict()norto_dict()handles it. As a result,list_models()silently drops the value. Related types such asModelLimitsOverrideinsession.pydo carry it, so this is a parity gap and not intentional.Fix: add
max_output_tokens: int | None = NonetoModelLimits, and read and write it infrom_dict()andto_dict(). Also add a unit test covering the reproduction from the issue.Labeled as
bug.