An approved tool, a crash, then Resume: a 5-run test, and a wrapper that keeps it to one payment

[salahmakboul]

ADK's Resume page says tools run at least once on resume. I wanted to see what that means for a tool behind require_confirmation=True, so I measured it.

Setup: ADK 2.14.0, Python 3.12, a resumable App, SqliteSessionService, and a fake BaseLlm (no keys) that asks for pay(10, "alice"). pay appends a line to a file. Each phase runs in its own process:

1. The run pauses for confirmation.
2. The yes arrives, the tool pays, and the process exits with os._exit before ADK stores the result.
3. A new process resumes with runner.run_async(..., invocation_id=...).

Five runs per row, and every run in a row gave the same result:

Setup	Payments	On resume
Plain FunctionTool, no crash	1	–
Plain FunctionTool, crash after the payment	2	not asked again
pay wrapped with c4a.effect	1	stops with EffectInDoubt
Same, with reconcile= reading the provider's record	1	finishes normally

The wrapper records the call in its own SQLite file before running it. On resume, a call that already finished returns its recorded result, and a call with no recorded result stops as "in doubt" instead of running again:

import call4agents as c4a

c4a.init(".c4a/ledger.sqlite")
guarded_pay = c4a.effect(pay, name="pay", reconcile=find_payment)
root_agent = Agent(..., tools=[FunctionTool(guarded_pay, require_confirmation=True)])

with c4a.run(session_id):
    async for event in runner.run_async(user_id=user_id, session_id=session_id, invocation_id=invocation_id):
        ...

Scripts and raw results: https://github.com/salahmakboul/call4agents/tree/main/docs/finding/more_frameworks

I maintain Call4Agents (Apache-2.0, no dependencies). Feedback on the test setup is welcome. In particular, if there's a more idiomatic ADK way to record that a tool started, such as a plugin's before_tool_callback, I'd like to compare against it.