fix: switch alibaba-coding-plan to OpenAI-compat endpoint with proper compat flags (#1003)
The alibaba-coding-plan provider was using the Anthropic-compatible
endpoint (/apps/anthropic) with anthropic-messages API, which caused
issues with thinking mode on several models (MiniMax-M2.5 thinking
loop, missing thinkingFormat for Qwen/GLM models).
Changes for all 8 models:
- API: anthropic-messages → openai-completions
- Endpoint: /apps/anthropic → /v1 (OpenAI-compatible)
- Added per-model compat flags:
- Qwen models: thinkingFormat: 'qwen', supportsDeveloperRole: false
- GLM models: thinkingFormat: 'qwen', supportsDeveloperRole: false
- MiniMax-M2.5: supportsReasoningEffort: true, maxTokensField: 'max_tokens'
- Kimi K2.5: thinkingFormat: 'zai', supportsDeveloperRole: false
- Enabled reasoning for qwen3-max (was incorrectly false)
- Fixed context windows to match tested values
- Fixed MiniMax-M2.5 maxTokens: 24576 → 65536
d252168de5