`modelCapabilities.limits.maxContextWindowTokens` is ignored by the runtime; only the undocumented `provider.maxContextWindowTokens` works

The issue is in the Python SDK's serialization and runtime handling of model capabilities. Start by examining the CopilotClient class, specifically the create_session method and the _convert_provider_to_wire_format function. Look at the ModelCapabilitiesOverride and ModelLimitsOverride classes in the SDK, and compare their field names to the runtime's api.schema.json. The test file tests/copilot_sdk/test_max_context_window.py shows the workaround and can be used to verify a fix. The goal is to ensure the max_context_window_tokens parameter is correctly serialized and honored by the runtime, or to expose the working provider.maxContextWindowTokens field.

create_session(model_capabilities=ModelCapabilitiesOverride(limits=ModelLimitsOverride(max_context_window_tokens=...))) is documented as "Override individual model capabilities resolved by the runtime", but it has no effect on the session's context-window management: session.history.compact keeps reporting contextWindow.tokenLimit = 128000 (the BYOK fallback default), so auto-compaction triggers far too early for long-context BYOK models.

Two additional problems compound this:

1. The Python SDK serializes ModelLimitsOverride to camelCase (maxContextWindowTokens), while the runtime's own api.schema.json declares ModelCapabilitiesOverrideLimits properties in snake_case (max_context_window_tokens) — so even if the runtime honored the override, the wire names would not match. Sending snake_case manually (by monkey-patching _capabilities_to_dict) changes nothing, so the parameter appears to be entirely non-functional.
2. The only working path — the BYOK ProviderConfig top-level maxContextWindowTokens — is not exposed by the Python SDK: ProviderConfig (TypedDict) lacks the field and _convert_provider_to_wire_format drops it from the dict.

Environment

• github-copilot-sdk (Python): 1.0.14 (latest on PyPI at the time)
• Runtime bundle: CLI 1.0.85
• OS: Linux x64

Evidence (all variants, same BYOK model deepseek-flash)

Expected behavior

• session.open's modelCapabilities.limits.maxContextWindowTokens should override the resolved context window used for compaction/truncation/token display (and the Python SDK's wire casing should match what the runtime deserializes); or
• the Python SDK should expose provider.max_context_window_tokens (ProviderConfig field + wire conversion), since that is the only knob the runtime honors today.

Workaround

Monkey-patch CopilotClient._convert_provider_to_wire_format to pass maxContextWindowTokens through on the provider object. Verified working (tokenLimit = 1048576). Repro script: tests/copilot_sdk/test_max_context_window.py.