Problem

The bundled GitHub Copilot provider currently publishes an empty provider catalog (models: []) during discovery. As a result, newly available Copilot models such as gpt-5.5 can be selected only through the dynamic fallback path, which uses conservative synthetic defaults instead of the capabilities advertised by the Copilot /models API.

For example, Copilot /models advertises gpt-5.5 with:

• max_context_window_tokens: 400000
• max_output_tokens: 128000
• vision support
• reasoning effort support including xhigh

But because the provider catalog does not ingest those fields, OpenClaw can treat the model as a generic synthetic Copilot model and show/use a much smaller effective context window.

Impact

Users configuring GitHub Copilot models shortly after they become available may get stale or incorrect model metadata until OpenClaw is updated manually. In particular, GPT-5.5 may be usable but not budgeted with its advertised context/output limits.

Expected behavior

The GitHub Copilot provider should fetch ${baseUrl}/models during catalog discovery and map capability metadata into model definitions, including:

• capabilities.limits.max_context_window_tokens → contextWindow
• capabilities.limits.max_output_tokens → maxTokens
• capabilities.supports.vision / vision limits → input
• capabilities.supports.reasoning_effort → reasoning / thinking profile support
• supported_endpoints → provider runtime API selection

Proposed fix

PR #71924 implements this for the GitHub Copilot provider catalog and adds coverage for GPT-5.5 capability mapping.

Linked PR: https://github.com/openclaw/openclaw/pull/71924