ProtoLink 0.8.3 owned-agent metrics

Verified against the installed PyPI release protolink==0.8.3. This is an upstream limitation, not an application token estimate.

A root with owned subagents uses the SDK's managed unary execution path. Its native llm.call.completed event retains provider, model and latency metadata but has metrics: null. Direct specialist streaming supplies usage metrics. The same difference occurs before making any consultation: adding an owned child is sufficient. Enabling the LLM's metrics flag does not override the unary engine's explicit event_metrics=False selection.

Minimal offline reproducer

Run with Python 3.12 and ProtoLink 0.8.3. The Mock provider performs no network requests, requires no credentials and incurs no provider charge.

import asyncio
from uuid import uuid4
from protolink import Agent, AgentGroup, Task, create_llm


async def check(owned):
    prefix = f"runtime://metrics-{uuid4().hex}"
    child = Agent(name="child", transport="runtime", url=f"{prefix}/child", verbosity=0)
    root = Agent(
        name="root", transport="runtime", url=f"{prefix}/root", verbosity=0,
        llm_factory=lambda: create_llm("mock", default_response="Hello"),
        subagents=[child] if owned else [],
    )
    group = AgentGroup([root, child])
    await group.start()
    try:
        handle = group.run(root, Task.create_infer("Say hello"))
        async for _ in handle.events():
            pass
        assert (await handle.result()).status == "completed"
        calls = [e for e in handle.report.events if e.type == "llm.call.completed"]
        print("owned=", owned, "metrics=", calls[0].payload["metadata"]["metrics"])
    finally:
        await group.close()


async def main():
    await check(False)
    await check(True)

asyncio.run(main())

Expected observation on 0.8.3:

owned= False metrics= {'usage': {..., 'estimated': True, 'source': 'estimate'}, ...}
owned= True metrics= None

In protolink/agents/engine.py, run_task_streaming takes the managed run_task branch for owned subagents; _run_infer enables event metrics only for self.telemetry or event_callback. That branch has no direct event callback despite having a native run event sink. An upstream fix should preserve native metrics through this path, including inferred owned-child calls, without requiring application telemetry or synthetic usage.

Sibyl displays missing usage/cost as unavailable and retains the native call timing/model attribution. It does not mutate the SDK, attach dummy telemetry, estimate tokens itself or treat missing prices as zero. tests/test_agent_streaming.py checks this behavior. Update that acceptance expectation when a verified SDK release supplies the metrics.