Implements a new provider that uses Google's Antigravity unified gateway API to access Claude, Gemini, and other models through a single OAuth2-authenticated endpoint.
Features:
- OAuth2 token management with automatic refresh
- Multi-account rotation for rate limit distribution
- Support for Claude Opus/Sonnet 4.5 (with/without thinking)
- Support for Gemini 2.5/3 models (Pro/Flash variants)
- Thinking mode support with configurable tokens
- Image processing support
- Dual quota pool tracking (antigravity vs gemini-cli)
- Gemini-style API request format
Authentication:
- Reads from ANTIGRAVITY_REFRESH_TOKEN env var (priority)
- Falls back to ~/.config/opencode/antigravity-accounts.json
- Automatic token refresh with retry logic
- Rate limit tracking per account and quota pool
Files added:
- providers/antigravity.py - Main provider implementation
- providers/antigravity_auth.py - OAuth token manager
- providers/registries/antigravity.py - Model registry
- conf/antigravity_models.json - Model definitions (11 models)
- docs/antigravity_provider.md - Setup and usage docs
- tests/test_antigravity_provider.py - Unit tests (14 pass)
Integration:
- Added to provider priority order after ZEN
- Registered in server.py with auto-detection
- ToS warning logged on first use
OpenRouter's /responses endpoint rejects store:true via Zod validation.
This is an endpoint-level limitation, not model-specific. The fix
conditionally omits the store parameter for OpenRouter while maintaining
it for direct OpenAI and Azure OpenAI providers.
- Add provider type check in _generate_with_responses_endpoint
- Include debug logging when store parameter is omitted
- Add regression tests for both OpenRouter and OpenAI behavior
Fixes#348
model definitions now support a new `allow_code_generation` flag, only to be used with higher reasoning models such as GPT-5-Pro and-Gemini 2.5-Pro
When `true`, the `chat` tool can now request the external model to generate a full implementation / update / instructions etc and then share the implementation with the calling agent.
This effectively allows us to utilize more powerful models such as GPT-5-Pro to generate code for us or entire implementations (which are either API-only or part of the $200 Pro plan from within the ChatGPT app)
feat: Azure OpenAI / Azure AI Foundry support. Models should be defined in conf/azure_models.json (or a custom path). See .env.example for environment variables or see readme. https://github.com/BeehiveInnovations/zen-mcp-server/issues/265
feat: OpenRouter / Custom Models / Azure can separately also use custom config paths now (see .env.example )
refactor: Model registry class made abstract, OpenRouter / Custom Provider / Azure OpenAI now subclass these
refactor: breaking change: `is_custom` property has been removed from model_capabilities.py (and thus custom_models.json) given each models are now read from separate configuration files
fix: restrictions should resolve canonical names for openrouter
fix: tools now correctly return restricted list by presenting model names in schema
fix: tests updated to ensure these manage their expected env vars properly
perf: cache model alias resolution to avoid repeated checks
Adds support for OpenAI's GPT-5-Codex model which uses the new Responses API
endpoint (/v1/responses) instead of the standard Chat Completions API.
Changes:
- Add GPT-5-Codex to MODEL_CAPABILITIES with 400K context, 128K output
- Prioritize GPT-5-Codex for EXTENDED_REASONING tasks
- Add aliases: codex, gpt5-codex, gpt-5-code
- Update tests to expect GPT-5-Codex for extended reasoning
Benefits:
- 40-80% cost savings through Responses API caching
- 3% better performance on coding tasks (SWE-bench)
- Leverages existing dual-API infrastructure
fix: model definition re-introduced into the schema but intelligently and only a summary is generated per tool. Required to ensure CLI calls and uses the correct model
fix: removed `model` param from some tools where this wasn't needed
fix: fixed adherence to `*_ALLOWED_MODELS` by advertising only the allowed models to the CLI
fix: removed duplicates across providers when passing canonical names back to the CLI; the first enabled provider wins
docs: document provider base class
refactor: cleanup custom provider, it should only deal with `is_custom` model configurations
fix: make sure openrouter provider does not load `is_custom` models
fix: listmodels tool cleanup
Fixed issue where OpenAI models appeared twice in listmodels output by:
- Removing self-referencing aliases from OpenAI model definitions (e.g., "gpt-5" no longer includes "gpt-5" in its aliases)
- Adding filter in listmodels.py to skip aliases that match the model name
- Cleaned up inconsistent alias naming (o3-pro -> o3pro)
This ensures each model appears only once in the listing while preserving all useful aliases.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
