Problem#
#53 says chat and text capabilities use OpenRouter's parameter names as string[]. BytePlus chat rows still return the host-native object:
"capabilities": {
"structured_outputs": { "json_object": true, "json_schema": false },
"tools": { "function_calling": true },
"cache": { "prefix_cache": true, "session_cache": true },
"maxReasoningTokens": 131072
}
All 13 BytePlus chat rows have this shape (for example glm-5-3-flash-260828 and deepseek-v4-1-flash-260910). Every other chat provider returns string[].
Impact#
Consumers need a BytePlus-only branch to read capabilities. TanStack AI has one. It maps json_object: true to structured_outputs, which is wrong: json_schema: false means no structured outputs.
Ask#
Project BytePlus chat rows onto the string[] from #53: tools, tool_choice, reasoning, response_format, and structured_outputs only when json_schema is true. Keep the host-native object under a different field if it is useful, for example hostCapabilities.
Acceptance#
GET /v1/models?provider=byteplus&activity=chat returns capabilities as string[] on every row.
glm-5-3-flash-260828 has response_format but not structured_outputs.