Bug: Agent recreated from manifest config loses SubagentTools capability (session_pool path)

[Million-mo]

Bug Description

When an Agent is created with toolsets passed via the constructor (e.g., SubagentToolsetConfig().get_provider()), the capability works. But when the session_pool path recreates the agent from manifest config, the toolsets are not preserved — SubagentTools capability is lost.

Failing Test

tests/tools/test_runcontext.py::test_capability_tools

async with AgentPool(manifest) as pool:
    subagent = SubagentToolsetConfig()
    providers = [subagent.get_provider()]
    agent = Agent(name="test", model=default_model, toolsets=providers, agent_pool=pool)
    result = await agent.run("Get available agents using the list_available_nodes tool")
    assert agent.name in str(result.content)  # ← RuntimeError: No final message received

The test is currently marked @pytest.mark.xfail with reason="Test passes toolsets via Agent() constructor, but session_pool path recreates agent from manifest config (without toolsets). SubagentTools capability is lost. Fix: configure SubagentToolsetConfig in manifest.tools."

Verification

Confirmed still failing on main (commit 67710ea) via --runxfail:

RuntimeError: No final message received from stream
src/agentpool/agents/base_agent.py:1984: RuntimeError

Root Cause (preliminary)

The xfail reason identifies the issue clearly:

1. Test creates Agent with toolsets=[subagent.get_provider()] via constructor
2. The session_pool execution path recreates the agent from manifest config
3. The manifest config does not include the SubagentToolsetConfig in manifest.tools
4. The recreated agent has no SubagentTools capability → the list_available_nodes tool is not registered → the model cannot call it → no final message produced

Suggested Fix

Per the xfail reason: "Fix: configure SubagentToolsetConfig in manifest.tools."

Either:

1. Make the session_pool path preserve constructor-passed toolsets, or
2. Require toolsets to be declared in manifest config (and update the test to do so)

Related

• Found during Test suite cleanup: remove obsolete tests, fix broken/skipped tests #234 xfail triage
• Test file: tests/tools/test_runcontext.py:96

[Million-mo]

Fixed

Root Cause

The test passed toolsets=[SubagentToolsetConfig().get_provider()] via the Agent() constructor, but when agent.run() goes through the session_pool path, the agent is recreated from manifest.agents[name] config — which had no tools field. The recreated agent lost the SubagentTools capability.

Fix

Add SubagentToolsetConfig() to NativeAgentConfig.tools in the manifest:

NativeAgentConfig(
    model=default_model,
    tools=[SubagentToolsetConfig()],
)

This way the session_pool path creates the agent with the subagent capability from the manifest, as intended.

Verification

uv run pytest tests/tools/test_runcontext.py::test_capability_tools -v
# → PASSED in 1.01s

xfail marker removed — test now passes.

PR

Fix is in PR #258 (commit 6af7236).