Summary I successfully identified and fixed the

RecallObservation

working_dir bug in local runtime headless mode. Problem: When using local runtime with

SANDBOX_VOLUMES={real_dir}:/workspace:rw

, the `RecallObservation`'s

working_dir

was incorrectly set to

/workspace

instead of the actual host directory. Root Cause: Both headless mode (

openhands/core/main.py

) and server mode (

openhands/server/session/agent_session.py

) were using

config.workspace_mount_path_in_sandbox

(always

/workspace

) instead of the actual working directory. Solution: Changed both files to use

config.workspace_mount_path or config.workspace_mount_path_in_sandbox

, which: • Uses the actual host directory when

SANDBOX_VOLUMES

is set • Falls back to

/workspace

when no custom mount is configured Changes Made: 1. Fixed headless mode in

openhands/core/main.py

2. Fixed server mode in

openhands/server/session/agent_session.py

3. Added comprehensive unit tests in

tests/unit/test_recall_observation_working_dir.py

Verification: All tests pass, including new unit tests that reproduce the bug and verify the fix works correctly with various SANDBOX_VOLUMES configurations. Alternative Solutions: I analyzed 6 alternative approaches, with the current solution providing the best balance of simplicity, correctness, and maintainability. The fix ensures that workspace context in LLM instructions now correctly shows the real working directory path instead of the hardcoded

/workspace

when using local runtime with custom volume mounts. Would you like me to push these changes to a new branch and create a pull request? _{View full conversation}