When streaming responses with tool calls, the finish_reason was being
overwritten. The upstream sends functionCall in chunk 1, then
finishReason: STOP in chunk 2. The old code would set finish_reason
from every chunk, causing "tool_calls" to be overwritten by "stop".
This broke clients like Claude Code that rely on finish_reason to
detect when tool calls are complete.
Changes:
- Add SawToolCall bool to track tool calls across entire stream
- Add UpstreamFinishReason to cache the finish reason
- Only emit finish_reason on final chunk (has both finishReason + usage)
- Priority: tool_calls > max_tokens > stop
Includes 5 unit tests covering:
- Tool calls not overwritten by subsequent STOP
- Normal text gets "stop"
- MAX_TOKENS without tool calls gets "max_tokens"
- Tool calls take priority over MAX_TOKENS
- Intermediate chunks have no finish_reason
Fixes streaming tool call detection for Claude Code + Gemini models.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The prompt for the gpt-5.2 codex model has been updated with more comprehensive instructions. This includes detailed guidelines on general usage, editing constraints, the plan tool, sandboxing configurations, handling special user requests, frontend task considerations, and final message presentation. The updates aim to improve the model's understanding and execution of complex coding tasks by providing clearer directives and constraints.
This change introduces specific logic to load and use instructions for the 'gpt-5.2-codex' model variant by recognizing the 'gpt-5.2-codex_prompt.md' filename. This ensures the correct prompts are used when the '5.2-codex' model is identified, complementing the recent addition of its definition.
Introduce a centralized OAuth session store with TTL-based expiration
to replace the previous simple map-based status tracking. Add a new
/api/oauth/callback endpoint that allows remote clients to relay OAuth
callback data back to the CLI proxy, enabling OAuth flows when the
callback cannot reach the local machine directly.
- Add oauth_sessions.go with thread-safe session store and validation
- Add oauth_callback.go with POST handler for remote callback relay
- Refactor auth_files.go to use new session management APIs
- Register new callback route in server.go
The previous commit added thinkingLevel support but didn't apply it
when the reasoning effort came from model name suffix (e.g., model(minimal)).
This was because ResolveThinkingConfigFromMetadata returns nil for
level-based models, bypassing the metadata application.
Changes:
- Add ApplyGemini3ThinkingLevelFromMetadata for standard Gemini API
- Add ApplyGemini3ThinkingLevelFromMetadataCLI for CLI API format
- Update gemini_cli_executor to apply Gemini 3 thinkingLevel from metadata
- Update antigravity_executor to apply Gemini 3 thinkingLevel from metadata
- Update aistudio_executor to apply Gemini 3 thinkingLevel from metadata
- Add comprehensive test coverage for Gemini 3 thinkingLevel functions
Per Google's official documentation, Gemini 3 models should use
thinkingLevel (string) instead of thinkingBudget (number) for
optimal performance.
From Google's Gemini Thinking docs:
> Use the thinkingLevel parameter with Gemini 3 models. While
> thinkingBudget is accepted for backwards compatibility, using
> it with Gemini 3 Pro may result in suboptimal performance.
Changes:
- Add model family detection functions (IsGemini3Model, IsGemini25Model,
IsGemini3ProModel, IsGemini3FlashModel)
- Add ApplyGeminiThinkingLevel and ApplyGeminiCLIThinkingLevel functions
for applying thinkingLevel config
- Add ValidateGemini3ThinkingLevel for model-specific level validation
- Add ThinkingBudgetToGemini3Level for backward compatibility conversion
- Update NormalizeGeminiThinkingBudget to convert budget to level for
Gemini 3 models
- Update ApplyDefaultThinkingIfNeeded to not set a default level for
Gemini 3 (lets API use its dynamic default "high")
- Update ConvertThinkingLevelToBudget to preserve thinkingLevel for
Gemini 3 models
- Add Levels field to all Gemini 3 model definitions:
- Gemini 3 Pro: ["low", "high"]
- Gemini 3 Flash: ["minimal", "low", "medium", "high"]
Backward compatibility:
- Gemini 2.5 models continue to use thinkingBudget as before
- If thinkingBudget is provided for Gemini 3, it's converted to the
appropriate thinkingLevel
- Existing configurations continue to work
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.
Updates schema flattening logic to handle multiple non-null types, providing a more descriptive "Accepts" hint.
Removes redundant tracking of the current tool name in `Params` as it's no longer needed for streaming limits, simplifying the structure.
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.
Enhance the configuration diff logic to include detection and reporting of `prefix` changes for all model types. Update related struct naming for consistency across the watcher module.
