Snapshot and Restore Skill

Process steps in order. Do not skip ahead.

Step 1 — Choose Manual or Automatic Checkpoints

Koog 1.3 uses Persistence from ai.koog.agents.snapshot.feature for both modes. There is no separate Snapshot feature, snapshot() or runFromSnapshot() API.

For automatic crash recovery without explicit save calls, invoke Skill(skill: "add-persistence"), deliver its durable backend and stable-session run path, and explain that automatic checkpointing is the requested mode. Finish here. For chosen save points and forks, proceed immediately to Step 2.

Step 2 — Implement the Fork

Add ai.koog:agents-features-snapshot:1.3.0 to build.gradle.kts. Read skills/snapshot-and-restore/references/checkpoint-fork.md and write its factory and comparison helper to src/main/kotlin/com/example/CheckpointFork.kt, adapting the named prefix and continuation nodes to the developer's existing graph.

Run the initial prefix once and retain the AgentCheckpointData supplied to onCheckpoint. Pass it to compareCheckpointBranches. Replay resumes after the saved node. The two copies replace its serialized output for the continuation; changing the ordinary input argument alone does not change the restored output. Use distinct session IDs for independent branches and stable matching graph/node names. Forks preserve the saved prefix history and typed storage. The capture run finishes without executing the continuation; only the two restored branches execute it.

In-memory storage is suitable for same-process comparison. For later process restarts, replace it with a durable provider and retrieve the saved checkpoint by session and checkpoint ID. Increment checkpoint versions for repeated saves in a session. Node outputs and typed storage must serialize successfully; custom data classes require a supported serializer. Invoke Skill(skill: "manage-state") for typed storage. Replay does not undo external side effects; isolate branch effects or make them idempotent.

Finish here.