cfg (ast.SandboxV2) is a pointer shared across all concurrent calls to
the same Assistant. GetComputer writes cfg.ID = <random hex> for each
oneshot request, so concurrent tasks overwrite each other's ID.
When cleanup runs in defer, cfg.ID holds whichever task last wrote it —
not the ID of the box this request created — so manager.Remove/Heartbeat
silently fails with "sandbox: not found" and the container leaks.
Fix: derive the box ID from computer.ComputerInfo().BoxID, which is the
immutable b.id set at Create time and never mutated by concurrent callers.
Made-with: Cursor
Three related bugs caused oneshot containers to run indefinitely:
1. watcher.go: switch b.policy had no case OneShot, so even when the
idle timeout fired, no remove action was emitted.
2. watcher.go + manager.go/recoverBoxes: idleTimeoutD was only set for
Session and LongRunning on recovery; OneShot defaulted to 0, which
caused the watcher to hit the `timeout <= 0` early-return and skip
all checks entirely.
3. agent/sandbox/v2/options.go: same gap — opts.IdleTimeout == 0 guard
only filled defaults for Session and LongRunning.
Fix: add DefaultOneShotIdleTimeout (30 min), wire it in recoverBoxes
and options.go, and add case OneShot → Remove in watcher.go.
Made-with: Cursor
- Introduced a new `workspace` field across various robot-related structures, including `CreateRobotRequest`, `UpdateRobotRequest`, and `RobotResponse`, allowing for better organization and management of robots within specific workspaces.
- Updated database queries and response mappings to accommodate the new workspace field, ensuring seamless integration with existing functionalities.
- Enhanced agent execution context to include workspace information, improving the contextual awareness of agents during operations.
- Added tests to validate the creation and updating of robots with workspace data, ensuring robust functionality and backward compatibility.
- Modified the buildEnv function to set environment variables for ANTHROPIC_MODEL and related keys dynamically based on the provided model value, enhancing flexibility and configurability.
- Added detailed logging of environment variables passed to the session in the Stream method, improving visibility into the execution context.
- Implemented conditional logging to display sensitive variables in a secure manner, ensuring that only relevant information is shown.
- Renamed ClaudeRunner to Runner for consistency across the codebase, aligning with the sandbox Runner interface.
- Updated buildCommand and Stream methods to utilize AssistantID directly from StreamRequest, improving clarity and reducing dependency on Config.
- Modified tests to reflect the changes in runner instantiation and argument handling, ensuring compatibility with the new Runner structure.
- Enhanced environment variable setup in buildEnv to include WORKDIR, streamlining the execution context for the runner.
- Renamed `buildInput` to `buildLastUserMessageJSONL` to better reflect its purpose of constructing JSONL from the last user message.
- Updated test cases to use the new function, ensuring they accurately verify the behavior of skipping system messages and only including the last user message.
- Removed the now redundant `buildFirstRequestJSONL` function, streamlining the input building process in the Claude command tests.
- Updated logging in various methods to include detailed traces for better monitoring of execution flow, including assistantID and chatID.
- Improved context management by ensuring proper release of resources and tracking of execution durations in agent calls.
- Added heartbeat logging in the Claude parser to monitor stream processing and prevent potential issues with long-running tasks.
- Refined the .gitignore file to exclude additional markdown files in the sandbox directory.
- Modified the buildAgentContext function to accept assistantID as a parameter, enhancing the context building process for agent calls.
- Updated multiple Call methods to pass the assistantID, ensuring proper context management during agent interactions.
- Introduced session ID and name handling in the buildArgs function, allowing for better tracking of chat sessions.
- Added chatIDToSessionUUID and sanitizeSessionName functions to generate and format session identifiers.
- Updated the Stream method to store session information and manage session lifecycle more effectively.
- Implemented KillSessionCmd for precise process termination based on session names in both Windows and POSIX platforms.
- Enhanced tests to cover new session management features and ensure correct behavior in various scenarios.
- Updated agent context to ensure locale is propagated correctly, defaulting to "en" when no locale is provided. This change enhances internationalization support and ensures consistent behavior across different contexts.
- Adjusted test assertions to reflect the updated execution trigger logic for times mode robots, allowing for a maximum of 2 triggers instead of 1.
- Modified integration tests to utilize the new dry run executor in the manager configuration, ensuring consistent execution behavior during testing.
- Updated multiple test cases to replace direct manager instantiation with the new configuration method, enhancing test reliability and maintainability.
- Increased the timeout duration from 60 seconds to 120 seconds for waiting on execution statuses in multiple E2E test cases, ensuring more reliable test outcomes during longer execution scenarios.
- Implemented pre-acquisition of execution slots in the Tick method to prevent race conditions, ensuring that robots do not submit duplicate executions.
- Updated TryAcquireSlot method to support idempotent behavior, allowing for early slot reservation without consuming additional resources.
- Modified worker execution logic to skip pre-checks for robots that have already acquired a slot, streamlining the execution process.
- Improved error handling during execution submission to ensure proper tracking and removal of failed executions.
- Enhanced the extractSandboxVersion function to support multiple input types, including *sandboxTypes.SandboxConfig and map[string]any, for better flexibility in version retrieval.
- Simplified the version extraction process, ensuring consistent handling of sandbox configurations.
- Added support for sandbox configuration in the LoadPath function, allowing for better management of sandbox settings.
- Updated tests to validate the retrieval of assistant tags with type filters, ensuring accurate responses for both assistant and robot types.
- Enhanced the assistant model to differentiate between sandbox versions, improving the handling of sandbox data in the database.
- Improved API responses to include computer filter details for V2 sandboxes, enhancing the information returned to clients.
- Added logic to buffer whitespace when no active text message is present, preventing unnecessary message group creation.
- Ensured proper closure of open text messages before executing new messages to maintain message integrity.
- Enhanced logging in the Stream and runStream methods to provide better visibility into execution flow and errors.
- Implemented shutdown logic to handle process termination gracefully after stream completion, addressing known issues with the Claude CLI.
- Introduced a new global phase agent resolver to streamline agent ID retrieval for various robot pipeline phases, enhancing flexibility in agent configuration.
- Updated existing phase agent retrieval logic to prioritize per-robot configurations, falling back to global settings when necessary.
- Enhanced error handling to provide clearer messages when no agent is configured for specific phases.
- Added tests to validate the new resolution logic and ensure proper functionality across different configurations.
- Introduced Weixin integration support with new configuration options for WeChat iLink Bot.
- Updated existing adapters (DingTalk, Discord, Feishu, Telegram) to include sender_id and app_id in message metadata for improved context handling.
- Enhanced dispatcher logic to accommodate the new Weixin adapter and ensure proper initialization and shutdown processes.
- Improved message handling across integrations to support typing indicators, providing a more interactive user experience.
- Replaced direct database insertion with a structured approach using ExecutionStore for better maintainability.
- Updated time parsing logic to handle both timezone-aware and local time formats, improving robustness.
- Enhanced the insertWatcherExec function to utilize the new execution record structure, ensuring consistency in execution data handling.
- Added GetManager function to retrieve the global manager instance, returning nil if not started.
- Introduced ExecRecovered event to notify about recovered non-terminal executions during manager startup.
- Updated execution store to support querying by multiple statuses with ListByStatuses method.
- Enhanced integration tests to accommodate longer tick intervals and added sleep delays for stability.
- Improved cleanup logic in integration tests to prevent interference from previous execution records.
- Renamed ListRobots function to ListAllRobots for clarity and consistency across the codebase.
- Updated related API endpoints and tests to reflect the new function name.
- Improved caching logic to ensure both autonomous and non-autonomous robots are loaded correctly.
- Enhanced filtering and pagination capabilities within the ListAllRobots function.
- Added new test cases to validate the updated functionality and ensure robust performance.
- Removed the RegisterProxyConfigs function and its related logic from the sandbox initialization, simplifying the setup process.
- Updated the GetComputer function to handle connector configuration injection directly within the ClaudeRunner.Stream method, ensuring a more cohesive approach to configuration management.
- Introduced a new injectA2OConfig function to push connector configurations to the a2o proxy, enhancing flexibility and error handling.
- Improved logging for connector configuration injection to provide better traceability during execution.
- Introduced a new sandboxV2StreamParams struct to group parameters for the executeSandboxV2Stream function, improving code clarity and maintainability.
- Updated the GetComputer function to handle connector configuration injection via the a2o proxy, removing the need for direct connector parameters.
- Enhanced the executeSandboxV2Stream function to utilize the new parameters struct, simplifying the function signature and improving readability.
- Implemented RegisterProxyConfigs to inject OpenAI-compatible connector configurations into the a2o proxy, ensuring proper environment setup for sandbox execution.
- Refactored BuildCreateOptions to remove direct connector handling, aligning with the new configuration injection approach.
- Added detailed trace logging in the Assistant's Stream function to capture computer and workspace information.
- Replaced fmt.Printf statements with kunlog.Trace in the Telegram polling adapter for consistent logging.
- Updated sandbox lifecycle management functions to utilize kunlog for logging workspace and node resolution details.
- Enhanced error handling in the Claude parser with trace logging for JSON unmarshal and scanner errors.
- Improved file handling logging in the workspace API to provide clearer insights into file read operations.
- Updated the BuildIdentifier function to use the ownerID as a fallback for workspaceID, improving identifier generation consistency.
- Enhanced ResolveNodeID to auto-select nodes based on filters and added detailed logging for better traceability.
- Modified GetComputer to include improved logging and streamlined node resolution logic.
- Introduced StringOrArray type for flexible handling of computer filter parameters in the SandboxConfig.
- Enriched workspace response structure to include detailed node information, enhancing API response clarity.
- Implemented default workspace ID generation based on owner and node, ensuring consistent workspace identification.
- Added locale support in CreateOptions for better internationalization.
- Added assertions to ensure the presence and order of `set -e` and `set +e` in the generated bash script when a system prompt is included.
- Updated the test for the scenario without a prompt to confirm that `set -e` is not present, ensuring correct script behavior.
- Improved documentation in the `buildBashScript` function to clarify the use of `set -e` for error handling during script execution.
- Added support for "execute" message type to track tool execution observations within the sandbox environment.
- Implemented handling for execute messages in the stream processing, allowing for real-time updates on tool execution status.
- Enhanced the stream parser to manage the lifecycle of execute messages, including merging input and output data.
- Updated the stream handler to accommodate new message types, improving overall message processing capabilities.
- Refactored related tests to ensure coverage for the new execute message functionality, enhancing reliability.
- Updated the sandbox initialization to resolve both the assistant directory and skills subdirectory, improving directory management.
- Modified the RunPrepareSteps function signature to include an assistantDir parameter, allowing for better handling of file copy operations.
- Adjusted all relevant calls to RunPrepareSteps across the codebase to accommodate the new parameter, ensuring consistency in execution.
- Enhanced test cases to reflect the updated function signature, improving coverage and reliability of the sandbox preparation process.
- Replaced the previous file extension checks with a unified parsing function for sandbox configuration, improving code clarity and maintainability.
- Introduced a new HostExecConfig structure to manage local execution settings, allowing for more granular control over command execution permissions.
- Removed deprecated Moapi API files and related functionality, simplifying the codebase and reducing maintenance overhead.
- Updated the Tai node registration process to ensure local capabilities are accurately reflected based on the environment, enhancing overall system robustness.
- Improved the PowerShell script for killing processes on Windows to ensure child processes are also terminated.
- Introduced a sentinel error for stream completion, allowing for better handling of normal stream terminations.
- Updated the parsing function to close stdout on context cancellation, preventing indefinite blocking on read errors.
- Enhanced logging to provide clearer feedback on stream completion and process management during execution.
- Added UTF-8 encoding settings for input and output streams to prevent JSON corruption on CJK Windows.
- Implemented a search for common installation paths of claude.exe across user profiles to ensure accessibility when running as a different user.
- Updated logging in the Stream function to include the working directory for better context during execution.
- Updated the Makefile and GitHub workflows to include entitlements during the codesigning process for macOS applications, improving security and compliance.
- Added timestamp and runtime options to the codesigning commands, ensuring better handling of application signatures.
- Refactored the signing logic to maintain consistency across different build targets, enhancing the overall release process.
- Changed token generation from "sandbox:mcp" to "grpc:mcp" for both access and refresh tokens, aligning with updated service requirements.
- Enhanced ClaudeRunner to copy skills from the specified directory to the ".claude/skills" path, improving skill management.
- Updated MCP configuration file path to ".claude/mcp.json" for better organization and consistency in file handling.
- Added error logging for skill copying and exit code handling in stream execution, enhancing debugging capabilities.
Made-with: Cursor
- Modified Docker run commands in CI workflows to include the `-direct` flag for enhanced server operation.
- Removed outdated loading status message during sandbox preparation and added new messages for improved user feedback.
- Introduced a loading message for waiting on AI responses, enhancing the user experience during sandbox execution.
- Refactored identity handling in stream processing to streamline message management.
Made-with: Cursor
- Updated Docker run commands in CI workflows to include the `-direct` flag for improved server operation.
- Implemented image existence checks and automatic pulling for sandbox environments, enhancing reliability during initialization.
- Added loading status updates for sandbox operations, providing better feedback during the setup process.
- Refactored lifecycle management to ensure accurate tracking of sandbox states and improved error handling.
Made-with: Cursor
- Updated BuildIdentifier to include assistant ID in session identifiers for better uniqueness.
- Implemented automatic starting of stopped boxes in resolveBox, improving recovery processes.
- Introduced status management for boxes, allowing for accurate tracking of their state (running, exited, stopped).
- Added idle timeout defaults based on lifecycle policy in BuildCreateOptions, enhancing configuration flexibility.
- Refactored tests to validate new box status behavior and lifecycle management improvements.
Made-with: Cursor
- Introduced a human-readable display name for the sandbox, constructed from the agent and workspace names.
- Updated the BuildIdentifier function to include workspace ID in identifiers for long-running and persistent lifecycles.
- Modified sandbox creation options to incorporate the display name, improving clarity in sandbox management.
- Refactored related tests to ensure compatibility with the new identifier structure and display name integration.
Made-with: Cursor
- Added support for issuing and passing a sandbox token during stream execution in the Assistant.
- Updated StreamRequest to include a Token field for managing user authentication.
- Enhanced ClaudeRunner to set environment variables for the sandbox token and refresh token.
- Refactored SandboxToken structure to clarify its purpose and manage token credentials effectively.
- Modified sandbox management routes to enforce OAuth guard for improved security.
Made-with: Cursor
- Implemented a new endpoint to retrieve the absolute path of a workspace's root directory.
- Enhanced the workspace interface with a GetRoot method to facilitate this functionality.
- Updated the workspace manager to utilize the new method for improved path resolution.
Made-with: Cursor
- Updated the AssistantInfo struct to include new fields: Connector, ConnectorOptions, Modes, DefaultMode, Sandbox, and ComputerFilter for improved assistant configuration.
- Enhanced the loading process to extract Sandbox flag and ComputerFilter from V2 sandbox configuration.
- Refactored GetInfo method to return comprehensive assistant details for better UI integration.
- Introduced new endpoint for workspace options to streamline InputArea selector functionality.
Made-with: Cursor
- Refactored the sandbox initialization process to resolve the connector before obtaining the Computer, allowing for the injection of OPENAI_PROXY_* environment variables.
- Updated the GetComputer and BuildCreateOptions functions to accept an optional connector parameter for improved environment variable management.
- Standardized the VNC configuration by replacing SANDBOX_VNC_ENABLED with VNC_ENABLED across Dockerfiles and related scripts for consistency.
- Enhanced the VNC service startup script to check the new VNC_ENABLED variable, ensuring proper service initialization.
Made-with: Cursor
- Updated the registration process to utilize NodeID instead of TaiID, allowing for server-generated TaiIDs.
- Implemented additional fields in the registration request, including DisplayName and improved error handling for missing parameters.
- Enhanced the connection logic for registered nodes, ensuring proper client binding and logging for better traceability.
- Introduced new methods for extracting and managing extra claims in OAuth tokens, improving the flexibility of user identification.
Made-with: Cursor
- Implemented V2 sandbox initialization in the assistant loading process, allowing for standalone sandbox.yao configuration.
- Added support for V2 sandbox execution paths in the Assistant's Stream method, differentiating between V1 and V2 sandboxes.
- Introduced comprehensive tests for V2 sandbox configurations, ensuring correct loading and execution behavior.
- Updated the context and types to accommodate V2 sandbox features, including system information and workspace management.
Made-with: Cursor
- Refactor GPT-5 test cases to improve clarity and maintainability.
- Comment out tests for temperature handling in GPT-5, indicating they are temporarily disabled.
- Remove the obsolete hostexec test file to clean up the codebase.
- Enhance the sandbox manager to support host execution capabilities and improve lifecycle management.
Made-with: Cursor
The bubbletea-based TUI for agent request visualization was not
practical and added significant binary size. This removes the TUI
entirely: the --tui flag, the agent context TUI model/messages, and
all SendTUI/GetTUIProgram call sites. Development-mode logging now
always prints ANSI-colored output to stdout.
Drops direct dependencies on charmbracelet/bubbletea and lipgloss.
Made-with: Cursor
- Add gRPC server configuration to the application, allowing for gRPC communication.
- Introduce new Makefile targets for gRPC unit testing and proto code generation.
- Update CI workflows to include gRPC tests with SQLite as the transport layer.
- Refactor the sandbox design to support multi-node capabilities and improve isolation.
- Enhance the service layer to facilitate internal request forwarding for gRPC APIs.
This commit lays the groundwork for integrating gRPC into the Yao SDK, improving performance and scalability.
- Introduce a fallback mechanism in the GetConnector method to use a default connector if the specified connector is not found.
- Add logging to notify when falling back to the default connector, improving debugging and traceability.
- Update comments for clarity on connector selection priority, ensuring better understanding of the logic flow.
