[BUG][tests/test_groupchat.py][Not a valid pytest suite — all 5 tests error on a missing report fixture, leaving GroupChat untested]

[kyegomez]

Summary

tests/structs/test_groupchat.py is not a runnable pytest suite. Every test function takes a report parameter, and no such fixture exists, so all five error at collection. GroupChat therefore has zero working test coverage.

That is how a critical GroupChat defect went unnoticed (see below).

Reproduction

$ PYTHONPATH=. python3 -m pytest tests/structs/test_groupchat.py -q -p no:randomly
E       fixture 'report' not found
...
ERROR tests/structs/test_groupchat.py::test_basic_groupchat
ERROR tests/structs/test_groupchat.py::test_varying_agent_counts
ERROR tests/structs/test_groupchat.py::test_threshold_behavior
ERROR tests/structs/test_groupchat.py::test_idle_timeout
ERROR tests/structs/test_groupchat.py::test_error_cases
5 errors in 1.86s

The signatures are e.g. tests/structs/test_groupchat.py:120:

def test_basic_groupchat(report):

The only report in the repo is a plain helper function in an unrelated file (tests/test_streaming_timing.py:35), not a fixture. The file reads as a script that was converted to test_*.py naming without being converted to pytest.

What it hid

While testing GroupChat live I found that no agent could ever speak: _extract_args did not handle the case where Agent.run returns the forced tool call as a string rather than a list, so every bid parsed as (0.0, "") and the chat ended on turn one.

_ensure_respond_tool injects RESPOND_TOOL and rebuilds the LLM but never sets output_type, so an agent built the documented way (no explicit output_type) returns a string. Confirmed:

agent bid: score=0.92, real message
_extract_args(...) -> (0.0, '')          # default output_type
_extract_args(...) -> (0.84, '...')      # output_type="final"

A single working test that ran a two-agent chat and asserted at least one message was posted would have caught it immediately.

Suggested fix

1. Convert the five tests to real pytest: drop the report parameter, or provide it as a fixture in conftest.py if the reporting output is wanted.
2. Add offline coverage for the bid path specifically — stub swarms.utils.litellm_wrapper.completion to return a forced respond tool call and assert:
   • a bid above threshold results in a posted message,
   • _extract_args parses both the list form and the string form,
   • idle_timeout / lull behaviour ends the chat.

Also note idle_timeout is still a constructor parameter (:198) though the changelog states it is deprecated and unused — one of the tests is named test_idle_timeout, so it may be asserting behaviour that no longer exists.

Found at 3e89f27b (v14.0.2).