[Bug] ConcurrentWorkflow silently drops one agent's answer, and which one changes every run

[ayaangazali]

Problem

Conversation.return_all_except_first slices [2:]:

def return_all_except_first(self):
    """Return all messages except the first one."""
    return self.conversation_history[2:]

The docstring says one; the code skips two. It hardcodes the assumption that every history begins [System, User, ...], which is not true in two ways:

• Agent.short_memory adds the System row conditionally — _initialize_new_conversation only calls add("System", ...) if self.system_prompt is not None.
• Swarms build their own Conversation with no system prompt at all. ConcurrentWorkflow's history is [User, agent, agent, agent].

So on a swarm the slice eats the first agent's answer, not a preamble.

Impact

dict-all-except-first is ConcurrentWorkflow's default output_type, so this is default behaviour, not an opt-in path. Three agents in, two answers out:

class Fake:
    def __init__(s, n): s.agent_name = n
    def run(s, task=None, img=None, imgs=None, **k): return f"answer-from-{s.agent_name}"

cw = ConcurrentWorkflow(agents=[Fake("A"), Fake("B"), Fake("C")], autosave=False)
cw.run("q")

history  ['User', 'B', 'C', 'A']
returned [{'role': 'C', ...}, {'role': 'A', ...}]        # B is gone

Results are collected as they complete, so the answer that disappears is whichever agent finished first — a different agent on each run. No exception, no warning; the output is still well-formed, just missing one contributor.

AgentRearrange is affected conditionally: team_awareness=True inserts a System message and makes [2:] correct, team_awareness=False makes it drop an agent. An unrelated flag changes how much of the output survives.

Fix

Derive the offset instead of hardcoding it: skip a leading System row when one exists, then skip the first real message, which is the input task. Both slicing methods share one helper so they cannot drift apart.

Verification

swarm, 3 agents      returned ['A', 'B', 'C']   count 3      (was 2)
Agent-shaped history ['a1', 'a2']                            unchanged
no system prompt     ['a1', 'a2']                            task only dropped

The Agent-shaped case is the one that must not move: with a System row present it still drops both the System row and the task, exactly as before.

tests/structs/test_conversation.py — 51 passed. tests/structs/test_concurrent_workflow.py — 10 passed.

One file, 24 lines.

(Unrelated to this PR: black . --check currently fails on examples/mcp/agents/07_huggingface_model_search.py and 12_semgrep_security_scan.py, which arrived with 47d94eb. The file this PR touches is clean.)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.