- Introduced a new "Heavy" role in the system configuration, allowing for complex reasoning tasks.
- Updated the initialization and environment resolution functions to handle the new Heavy role appropriately.
- Enhanced tests to validate the integration of the Heavy role, ensuring proper connector resolution and model handling.
- Adjusted OpenAPI settings to reflect the addition of the Heavy role, improving overall system capabilities.
- Add `language_model` field to robot data structures for LLM connector overrides.
- Update `AgentCaller` to utilize the robot's language model and include logging capabilities for agent calls.
- Refactor task execution to log task outputs and inputs, improving observability during execution.
- Modify tests to accommodate changes in the runner initialization and ensure proper logging functionality.
- Add handling for Anthropic connector types in the sandbox, allowing direct connections without a proxy.
- Enhance capability retrieval to support both OpenAI and Anthropic formats, ensuring a unified interface.
- Update the executor and command logic to differentiate between OpenAI and Anthropic configurations, streamlining environment setup.
- Modify the provider selection logic to accommodate Anthropic capabilities, improving flexibility in LLM provider management.
- Refactor API detection to include Anthropic, ensuring accurate identification of connector types.
- Modify the expected API URL in the `TestBuildProxyConfig` to include the `/v1` prefix for consistency with the updated URL building logic.
- Remove unnecessary `modified` flags in the `prepareAttachments` method to streamline attachment processing and ensure clarity in the codebase.
This change enhances the accuracy of tests and improves the maintainability of the attachment handling logic.
- Introduce a whitelist mapping for `package.yao` sandbox arguments to corresponding Claude CLI flags, allowing selective argument passing.
- Update `BuildCommandWithContinuation` to iterate over the whitelist and append valid arguments to the Claude CLI command.
This change enhances the flexibility of argument handling in the Claude CLI, ensuring only specified arguments are processed.
- Update `buildAPIURL` function in OpenAI provider to delegate URL construction to `connector.BuildAPIURL`, ensuring consistent URL formatting across the agent LLM and sandbox proxy paths.
- Modify backend URL construction in Claude's `BuildProxyConfig` to utilize the shared `connector.BuildAPIURL` helper, applying the necessary `/v1` prefix for compatibility.
This change enhances code maintainability and consistency in API URL handling across different components.
- Introduce functionality to resolve and manage user-uploaded files in the sandbox environment.
- Implement `prepareAttachments` method to convert attachment URLs to local file paths and handle duplicates.
- Update message processing to replace attachment content with text references, allowing Claude CLI to access files using Read and Bash tools.
- Enhance documentation to inform users about the new attachment handling capabilities.
This change improves the interaction with user-uploaded files, enabling better integration within the Claude CLI environment.
- Ensure the existence of $HOME/.Xauthority for PyAutoGUI/Xlib compatibility.
- Adjust PYTHONPATH to maintain access to user-installed Python packages.
- Explicitly set XAUTHORITY to the correct path for X11 authentication.
- Pass secrets as environment variables for Claude CLI to ensure proper access.
- Refactor VNC image support check to utilize a centralized keyword list for better maintainability.
- Refactor BuildCommand to support session continuation using --continue flag.
- Introduce BuildCommandWithContinuation to handle both initial and continuation requests.
- Add BuildLastUserMessageJSONL for generating JSONL from the last user message in continuation scenarios.
- Enhance Executor to check for existing sessions and adjust command building accordingly.
- Update environment setup to ensure session data is stored correctly for persistence.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Implement secure config file location (/tmp/.yao/proxy.json) instead of user-visible /workspace/
- Add generic options map support for backend-specific parameters (e.g., thinking for Volcengine GLM-4.7)
- Add secrets support for passing sensitive env vars (e.g., GITHUB_TOKEN) to sandbox container
- Remove excessive debug logs, keep critical ones with log.Printf("[Sandbox]...")
- Fix test assertions for system prompt passing via CLI args instead of env var
- Update i18n messages for sandbox loading states
Co-authored-by: Cursor <cursoragent@cursor.com>
- Add --input-format stream-json, --output-format stream-json, --verbose flags
- Use heredoc to pass messages via stdin (no CLI length limit)
- Add BuildInputJSONL function for message conversion
- Add shouldSkipClaudeCLI logic to skip when no prompts/skills/mcp
- Update executor to conditionally start claude-proxy only when needed
- Add SystemPrompt field to Options for skip logic
- Add E2E tests for skip mode and command building
- Fix tests to reflect new command structure
Co-authored-by: Cursor <cursoragent@cursor.com>
- Add claude-proxy: lightweight Go proxy to translate between Anthropic
and OpenAI-compatible APIs with full streaming and tool-calling support
- Update Dockerfile to include claude-proxy binary (multi-arch)
- Add auto-start proxy via entrypoint when env vars are set
- Update executor.go to write proxy config and start proxy
- Simplify command.go to use direct Claude CLI with proxy
- Support both docker run -e and config file for proxy settings
Co-authored-by: Cursor <cursoragent@cursor.com>
- Added functionality to build and manage MCP configuration for sandbox environments, allowing for dynamic tool execution.
- Enhanced the Assistant's Stream method to skip MCP tool calls in sandbox mode, with internal handling by Claude CLI.
- Introduced unit tests for MCP configuration building and skills directory resolution, ensuring robust integration.
- Updated sandbox manager to create IPC sessions and manage tool exposure dynamically, improving interaction with external agents.
- Enhanced documentation to reflect new features and integration points for MCP and skills within the sandbox.
- Added steps to pull necessary Docker images for sandbox testing in both CI workflows.
- Updated the AI test execution to utilize sandbox configurations, ensuring proper environment setup.
- Introduced sandbox initialization in the Assistant's Stream method, allowing for execution of coding agents like Claude and Cursor.
- Enhanced context management to support sandbox execution, improving flexibility in handling agent operations.
