swe_manager task broken: `Command 'jupyter --runtime-dir' returned non-zero exit status XXX`

[chongkong]

This is a rewrite of the already reported #57.

Steps To Reproduce

uv run python run_swelancer.py --issue_ids=14268-manager-0

Or any other swe_manager variant question_id.

Expected to succeed, but fails differently on every invocation; so far I observed:

• Command 'jupyter --runtime-dir' returned non-zero exit status 137
• Command 'jupyter --runtime-dir' returned non-zero exit status 126
• 409 Client Error for http+docker://localhost/v1.41/networks/<container_id>/exec: Conflict ("Container <container_id> is not running

Root cause investigation

Root cause is that in alcatraz, the container has failed during the check on setup_expensify.yml playbook second task (Verify the flow file does not contain PII or unnecessary information)

Traceback (most recent call last):
  File "/app/tests/utils/check_flows.py", line 168, in <module>
    unintended_flow_found = check_flows(args.path, args.exclude_domains, args.prefixes, logger, args.verbose)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/tests/utils/check_flows.py", line 104, in check_flows
    for file_path in filter_flows(path, prefixes):
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/tests/utils/check_flows.py", line 66, in filter_flows
    for file in os.listdir(path):
                ^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/app/tests/issues/14268-manager-0/'

The task requires /app/tests/issues/$ISSUE_ID/ directory should exist but all the manager task IDS (e.g. /app/tests/issues/14268-manager-0/) actually does not exist in the repo issues/ directory, which is copied over during docker build.

Suggested Fix

Current issues directory only contain 398 directories where they seem like a superset of ic_swe tasks.

1. If swe_manager task does not require issue directory, then the setup_expensify.yml second step should not be applied for the manager task.
2. If swe_manager task requires the issue directory, it should be added to the repo.

[chongkong]

Update: there is already an EVAL_VARIANT env var which skips the IC-task-only setup which always fails for the manager task. One should modify the swelancer.py to set EVAL_VARIANT as swe_manager, but this applies globally. Let me work on using the EVAL_VARIANT from the task["variant"] from CSV instead.