- Introduced llm_timeout_secs in ProviderConfig and LLMProviderConfig to specify timeout for LLM requests.
- Updated OpenAIProvider and AnthropicProvider to utilize the timeout setting when creating HTTP clients.
- Enhanced error handling for API responses to include timeout information.
- Modified SessionManager to support agent-specific provider configurations, allowing for more flexible agent management.
- Added tests to verify the correct behavior of timeout settings and agent task validation.
- Introduced ToolMessageState enum to represent tool execution states (Completed, PendingUserAction).
- Updated ChatMessage struct to include tool_state for tracking tool execution status.
- Modified AgentLoop to handle tool results and pending actions, providing appropriate responses to users.
- Enhanced BashTool to detect when commands require user interaction, returning a pending state with hints.
- Updated WebSocket protocol to support tool pending messages, allowing clients to handle pending actions effectively.
- Refactored related tests to ensure proper functionality of new pending state handling.
- Added SessionSummary struct for session metadata.
- Updated ws_handler to create and manage CLI sessions more robustly.
- Implemented session creation, loading, renaming, archiving, and deletion via WebSocket messages.
- Introduced SessionStore for persistent session storage using SQLite.
- Enhanced error handling and logging for session operations.
- Updated protocol definitions for new session-related WebSocket messages.
- Refactored tests to cover new session functionalities and ensure proper serialization.
Implement parallel tool execution in AgentLoop, following the approach
used in Nanobot (_partition_tool_batches) and Zeroclaw (parallel_tools).
Key changes:
- partition_tool_batches(): group tool calls into batches based on
concurrency_safe flag. Safe tools run in parallel via join_all;
exclusive tools (e.g. bash) run in their own sequential batch.
- execute_tools(): now uses batching instead of flat sequential loop.
- CalculatorTool: add read_only() -> true so it participates in
parallel batches (it has no side effects, so concurrency_safe = true).
4 unit tests added covering: mixed safe/exclusive, all-safe single
batch, all-exclusive separate batches, unknown tool defaults.
Byte index slicing like `&text[..100.min(text.len())]` panics when the
byte index falls inside a multi-byte UTF-8 character (e.g., Chinese).
Changed to `text.chars().take(100).collect::<String>()` for safe
character-based truncation.
Feishu API rejects messages with content exceeding ~64KB with error
230001 "invalid message content". Added truncation at 60,000 characters
to prevent this, with a notice appended to truncated content.
- Fetch URL and extract readable text
- HTML to plain text conversion
- Removes scripts, styles, and HTML tags
- Decodes HTML entities
- JSON pretty printing
- SSRF protection
- Includes 6 unit tests
- HTTP client with GET/POST/PUT/DELETE/PATCH support
- Domain allowlist for security
- SSRF protection (blocks private IPs, localhost)
- Response size limit and truncation
- Timeout control
- Includes 8 unit tests
- Execute shell commands with timeout
- Safety guards block dangerous commands (rm -rf, fork bombs)
- Output truncation for large outputs
- Working directory support
- Includes 7 unit tests
