Problem Description
Currently, Apra Labs Fleet (apra-fleet) supports Claude, Gemini, Codex, and Copilot. However, Google DeepMind has transitioned developer workflows to the Google Antigravity platform, which is invoked via the "agy" CLI command. There is currently no "agy" provider in apra-fleet.
Without "agy" support, users cannot leverage Antigravity's model-agnostic capabilities, remote SSH workflows, and agentic harness natively inside Fleet.
Proposed Solution
To provide both flexibility and power, the "agy" provider should give users control over model selection for their tiers while providing optimized, sane defaults.
-
User-Configurable Model Selection:
- Allow users to map specific models to the "cheap", "standard", and "premium" tiers in their fleet settings (e.g., settings.json).
- This expands the configuration spectrum and lets users customize model selection based on their cost and reasoning requirements.
-
Sane Default Model Tiers for agy:
- Cheap: GPT 120B OSS (for cost-effective, high-reasoning tasks such as log parsing and cheap execution).
- Standard: Gemini 3.5 Flash (for high-speed terminal actions, execution, and large context windows).
- Premium: Claude Sonnet 4.6 (for complex, high-quality code reviews, planning, and design).
-
New Provider Class AgyProvider:
- Create src/providers/agy.ts implementing ProviderAdapter for the agy CLI.
- Define processName as "agy".
- Define authEnvVar as GEMINI_API_KEY and/or ANTIGRAVITY_API_KEY.
- Set credentials path to "~/.gemini/antigravity-cli/" and instructions to "AGY.md".
-
Command and Flag Mapping:
- Implement buildPromptCommand(opts) to build the invocation command using agy.
- Map standard agy CLI flags: --sandbox and --dangerously-skip-permissions.
-
Codebase Integration:
- Update src/types.ts to add "agy" to the LlmProvider type union.
- Register AgyProvider in src/providers/index.ts.
- Update installation/uninstallation routines (src/cli/install.ts, src/cli/uninstall.ts) to handle agy setup.
- Update tests/providers.test.ts and tests/tool-provider.test.ts to include tests for agy.
- Update documentation (README.md, docs/provider-matrix.md, etc.) and skills config to reflect the new provider.
Alternatives Considered
Running Antigravity via the legacy gemini CLI wrapper. However, Google has deprecated the Gemini CLI and is migrating users to agy. Sticking to the old gemini wrapper will fail as APIs and flags diverge.
Additional Context
- Research shows that agy represents Google's unified platform for agentic developer tools (across Antigravity 2.0 Desktop, CLI, and SDK).
- GPT 120B OSS is highly efficient due to its MoE architecture (only 5.1B active parameters per token), making it extremely cost-effective for parsing logs while maintaining high reasoning quality.
- Gemini 3.5 Flash provides exceptional context window length and speed, which is ideal for terminal actions.
- Claude Sonnet 4.6 delivers industry-leading reasoning performance, making it the perfect choice for premium operations like planning and code reviews.
Problem Description
Currently, Apra Labs Fleet (apra-fleet) supports Claude, Gemini, Codex, and Copilot. However, Google DeepMind has transitioned developer workflows to the Google Antigravity platform, which is invoked via the "agy" CLI command. There is currently no "agy" provider in apra-fleet.
Without "agy" support, users cannot leverage Antigravity's model-agnostic capabilities, remote SSH workflows, and agentic harness natively inside Fleet.
Proposed Solution
To provide both flexibility and power, the "agy" provider should give users control over model selection for their tiers while providing optimized, sane defaults.
User-Configurable Model Selection:
Sane Default Model Tiers for agy:
New Provider Class AgyProvider:
Command and Flag Mapping:
Codebase Integration:
Alternatives Considered
Running Antigravity via the legacy gemini CLI wrapper. However, Google has deprecated the Gemini CLI and is migrating users to agy. Sticking to the old gemini wrapper will fail as APIs and flags diverge.
Additional Context