autonomous-resume

autonomous-resume — checkpointed resume of an interrupted run

What this shows

Crash-safe autonomous runs, end-to-end with no LLM credentials, in two phases against one SQLite event log:

1. Phase 1 — autonomous.Run capped at MaxTurns(2) to simulate an interruption; the scripted model never gets to call report_done. The driver emits per-turn checkpoint events into the event log.
2. Phase 2 — autonomous.Resume with a autonomous.SessionRef pointing at the same log. It finds the latest non-terminal checkpoint, carries turn count + token totals forward, and a fresh scripted model finishes the task. A session lock (acquired automatically) keeps two resumers from clobbering each other.

Run it

No credentials needed, runs offline:

go run ./examples/autonomous-resume

Key APIs

• autonomous.Run / autonomous.Resume / autonomous.SessionRef — github.com/go-steer/core-agent/v2/pkg/agent/autonomous
• eventlog.Open — github.com/go-steer/core-agent/v2/pkg/eventlog
• agent.WithAppName / agent.WithSession / agent.WithEventLog — github.com/go-steer/core-agent/v2/pkg/agent
• mock.NewScripted — github.com/go-steer/core-agent/v2/pkg/models/mock

What you should see

== Phase 1: autonomous.Run with MaxTurns(2) ==
[Phase 1] reason=max_turns_exceeded turns=2 done_detail="" final_text=""

== Phase 2: autonomous.Resume picks up at the next turn ==
[Phase 2] reason=completed turns=3 done_detail="resumed and finished" final_text=""

(A few "Event from an unknown agent ... checkpoint-..." log lines are expected — that's ADK observing the driver's checkpoint events.)

Phase 2's turn count includes Phase 1's — that's the carried-forward checkpoint state.

Next

• autonomous — the basic loop and report_done.
• autonomous-handle — in-process control of a running loop.
• API reference: docs/site/src/content/docs/embed/api.md