Fix agent output sanitization to prevent internal ReAct fields from leaking

Browse Source

Fixes #3873

This commit addresses a bug where internal ReAct-style fields like 'Thought:',
'Action:', 'Action Input:', and 'Observation:' were leaking into the final
agent output, particularly in hierarchical crews when delegated tasks failed
or when agents hit max iterations.

Changes:
- Added sanitize_react_output() utility function to strip internal ReAct fields
- Applied sanitization in handle_max_iterations_exceeded() when converting
  AgentAction to AgentFinish
- Applied sanitization in format_answer() exception handling to prevent leaks
  when LLM responses cannot be parsed
- Added comprehensive unit tests to verify the fix and prevent regressions
- Updated pyproject.toml to use pytest-vcr instead of pytest-recording to
  resolve plugin conflicts

The sanitization function removes lines starting with ReAct field markers
while preserving legitimate content, ensuring clean user-facing outputs.

Co-Authored-By: João <joao@crewai.com>

...

This commit is contained in:

Devin AI

2025-11-10 11:32:48 +00:00

parent 0f1c173d02

commit efb005afaf

4 changed files with 4288 additions and 4015 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

8086

uv.lock generated

File diff suppressed because it is too large Load Diff