Introduces the capability to count tokens for Antigravity-backed requests. This implementation leverages the `countTokens` endpoint of the Antigravity API, replacing the prior unsupported stub.
Key aspects of this update include:
- **API Integration**: Direct integration with the Antigravity `countTokens` API, including necessary request payload translation and authentication.
- **Resilient Infrastructure**: A fallback mechanism has been established, allowing the system to attempt connections across multiple Antigravity base URLs to ensure request success even in the event of temporary service interruptions.
- **Model Aliasing**: Added mappings for `gemini-3-flash` and `gemini-3-flash-preview` to ensure compatibility with the latest model variants.
- **Robust Error Handling**: Comprehensive error handling and logging are in place to manage failures during API interactions.
Enhances compatibility with the Gemini API by implementing a schema cleaning process.
This includes:
- Centralizing schema cleaning logic for Gemini in a dedicated utility function.
- Converting unsupported schema keywords to hints within the description field.
- Flattening complex schema structures like `anyOf`, `oneOf`, and type arrays to simplify the schema.
- Handling streaming responses with empty tool names, which can occur in subsequent chunks after the initial tool use.
Unify thinking budget-to-effort conversion in a shared helper, handle disabled/default thinking cases in translators, adjust zero-budget mapping, and drop the old OpenAI-specific helper with updated tests.
Expose thinking/effort normalization helpers from the executor package
so conversion tests use production code and stay aligned with runtime
validation behavior.
When budget 0 maps to "none" for models that use thinking levels
but don't support that effort level, strip thinking fields instead
of setting an invalid reasoning_effort value.
Tests now expect removal for this edge case.
- **Thinking Support**:
- Enabled thinking support for all Kiro Claude models, including Haiku 4.5 and agentic variants.
- Updated `model_definitions.go` with thinking configuration (Min: 1024, Max: 32000, ZeroAllowed: true).
- Fixed `extended_thinking` field names in `model_registry.go` (from `min_budget`/`max_budget` to `min`/`max`) to comply with Claude API specs, enabling thinking control in clients like Claude Code.
- **Kiro Executor Fixes**:
- Fixed `budget_tokens` handling: explicitly disable thinking when budget is 0 or negative.
- Removed aggressive duplicate content filtering logic that caused truncation/data loss.
- Enhanced thinking tag parsing with `extractThinkingFromContent` to correctly handle interleaved thinking/text blocks.
- Added EOF handling to flush pending thinking tag characters, preventing data loss at stream end.
- **Performance**:
- Optimized Claude stream handler (v6.2) with reduced buffer size (4KB) and faster flush interval (50ms) to minimize latency and prevent timeouts.
Ensure thinking settings translate correctly across providers:
- Only apply reasoning_effort to level-based models and derive it from numeric
budget suffixes when present
- Strip effort string fields for budget-based models and skip Claude/Gemini
budget resolution for level-based or unsupported models
- Default Gemini include_thoughts when a nonzero budget override is set
- Add cross-protocol conversion and budget range tests
When using OpenAI-compatible providers with model aliases (e.g., glm-4.6-zai -> glm-4.6),
the alias resolution was correctly applied but then immediately overwritten by
ResolveOriginalModel, causing 'Unknown Model' errors from upstream APIs.
This fix skips the ResolveOriginalModel override when a model alias has already
been resolved, ensuring the correct model name is sent to the upstream provider.
Co-authored-by: Amp <amp@ampcode.com>