1.14.4a1
📦 crewaiView on GitHub →
🐛 8 fixes🔧 1 symbols
Summary
This release focuses on stability and security by fixing several bugs related to LLM interactions, executor state management, and dependency updates, including a bump to litellm for an SSTI fix. New documentation for various sandbox tools has also been added.
Migration Steps
- If you were relying on the previous behavior of `force_final_answer` in `agent_executor`, review its usage due to renaming.
🐛 Bug Fixes
- Fixed crew chat description helpers to be resilient against LLM failures.
- Ensured messages and iterations are reset between invocations in the executor.
- Forwarded the trained-agents file through replay and test operations in the CLI.
- Ensured custom trained-agents file is honored during inference in the agent.
- Bound task-only agents to the crew to guarantee multimodal input_files reach the LLM.
- Serialized guardrail callables as null during JSON checkpointing.
- Renamed `force_final_answer` in agent_executor to prevent self-referential router issues.
- Bumped `litellm` dependency to address an SSTI vulnerability and ignore an unfixable pip CVE.