eval generate: 'Malformed agent event: missing content' for any agent with before_agent_callback

[sloopec]

Summary

agents-cli eval generate fails on every case with Malformed agent event: missing content. as soon as the evaluated agent (root or nested) has a before_agent_callback set — even on an otherwise trivial, working agent. agents-cli run (interactive) handles the same SSE event fine; the problem is specific to the eval generate code path.

Root cause

google/agents/cli/eval/cmd_generate.py::_parse_sse_event requires a content field on every event received over /run_sse:

def _parse_sse_event(event: dict) -> evals_types.AgentEvent:
    ...
    missing = [field for field in ("author", "content") if not event.get(field)]
    if missing:
        raise ValueError(f"Malformed agent event: missing {' and '.join(missing)}.")

    return evals_types.AgentEvent(
        author=event.get("author"), content=event.get("content")
    )

ADK legitimately emits an event with no content when a before_agent_callback mutates session state: a pure bookkeeping event carrying actions.stateDelta (whatever the callback wrote), with author present but no top-level content. This is the first event of the SSE stream, before any model response. Example, captured directly from /run_sse:

data: {"invocationId":"e-...","author":"root_agent","actions":{"stateDelta":{"foo":"bar"},"artifactDelta":{},"requestedAuthConfigs":{},"requestedToolConfirmations":{}},"nodeInfo":{"path":"root_agent@1"},"id":"...","timestamp":...}

_parse_sse_event raises on this event, which aborts inference for the entire case.

Minimal reproduction (independent of any specific project)

Created via agents-cli scaffold create <name> --adk -y (default scaffold, no project-specific changes). Only change to the generated app/agent.py: add a before_agent_callback that sets a single state variable.

async def _init_state(callback_context) -> None:
    if "foo" not in callback_context.state:
        callback_context.state["foo"] = "bar"


root_agent = Agent(
    name="root_agent",
    model=Gemini(model=MODEL, retry_options=types.HttpRetryOptions(attempts=3)),
    instruction="You are a helpful AI assistant.",
    tools=[get_weather],
    before_agent_callback=_init_state,   # <-- only change
)

tests/eval/datasets/repro.json:

{
  "eval_cases": [
    {
      "eval_case_id": "hello",
      "prompt": {"role": "user", "parts": [{"text": "Hello, what can you help me with?"}]}
    }
  ]
}

agents-cli eval generate --dataset tests/eval/datasets/repro.json -o /tmp/repro_traces

With before_agent_callback:

[generate] case[0] FAILED: Malformed agent event: missing content.
Inference summary: 0/1 succeeded, 1 failed.
Error: Inference failed: 0 of 1 cases succeeded.

Control run (identical code, only before_agent_callback=_init_state, removed):

[generate] case[0] done
Traces saved to /tmp/repro_traces

Isolated via bisection

The following were each tested individually against the same scaffold baseline (real Gemini API, gemini-3.6-flash) and all passed eval generate successfully:

• sub_agents=[<plain Agent>], unused
• sub_agents=[<plain Agent>] + actual delegation via transfer_to_agent
• tools=[AgentTool(<plain Agent>)] + actual tool call
• sub_agents=[<SequentialAgent>], unused, combined with AgentTool + tools
• sub_agents=[<Agent with output_schema>], unused

Only adding before_agent_callback (alone, without any of the above) reliably triggers the failure — this is the minimal sufficient condition.

Expected vs. actual

• Expected: agents-cli eval generate works for any agent with before_agent_callback (a documented, common ADK pattern for state initialization) just as it does without one.
• Actual: The very first event of the stream aborts the entire case before any model response is even generated. Affects every case in the dataset equally (0 of N succeed).

Environment

Component	Version
agents-cli	1.4.0 (also reproducible on 1.3.1 — same missing check present unchanged)
google-adk	2.7.0
Python	3.13.14
Model	gemini-3.6-flash (scaffold default), reproduced independent of model
OS	macOS 26.5.2 (Darwin)
Provider	Gemini API / AI Studio (GEMINI_API_KEY), no Vertex AI

Suggested fix

_parse_sse_event should recognize events with no content but a valid author and (optionally) actions.stateDelta/actions.endOfAgent as pure bookkeeping/callback events, and either (a) skip them (they carry no AgentEvent-relevant model/tool content) or (b) construct the AgentEvent with an empty/synthetic content instead of raising. Alternatively, make content optional in AgentEvent construction if the downstream data structure allows it.

Impact

This blocks agents-cli eval generate — and by extension eval run, which calls generate internally — for any agent using the documented before_agent_callback state-initialization pattern, which is not an uncommon setup.