mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-09-25 14:06:18 +00:00
dba3967177
The AIO sandbox's implicit persistent shell session hangs forever when a
command containing a bare 'exit' is executed: exit kills the session's
shell process, and the server's response path for that exec_command
request never completes. Verified against a standalone all-in-one-sandbox
1.11.0 container via the raw SDK:
shell.exec_command('seq 1 1000 | head -n 5') # OK (not a SIGPIPE issue)
shell.exec_command('echo x; exit 0') # HANGS every time
shell.exec_command('echo after') # OK (server recreates shell)
shell.exec_command('( echo x; exit 0 )') # OK, exit code propagates
remote_list_dir_command and remote_search_command both end their probe
scripts with a bare 'exit' (to propagate the find/grep status code), so
every list_dir/grep/glob call deterministically wedges the session —
this is the root cause behind parallel [ls, bash] tool calls deadlocking
an entire run (same defect family as #1433 and #5128).
Fix: keep 'set +e' as the outermost prefix (pinned by existing tests) and
wrap the rest of each probe script in a subshell, so exit only terminates
the subshell. Output and exit codes propagate identically.
Tests: 116 passed (backend/tests/test_aio_sandbox.py, test_remote_list_dir.py,
test_remote_search.py); one endswith assertion updated to the wrapped form.
Co-authored-by: mad_max <mad_max@coscoshipping.local>