feat: agent state checkpointing (rollback, restore, session resume)

[EvilFreelancer]

Summary

Adds agent state checkpointing so an agent can snapshot its execution state as it runs, then roll back to an earlier step, restore after a process restart, and resume a session (including over ACP). Checkpointing is opt-in — disabled by default, so existing behavior is unchanged.

Built bottom-up (base → services → config → agent → factory → server → acp), TDD, one layer per commit.

What's included

• AgentContext.to_snapshot()/from_snapshot() + AgentCheckpoint (models.py) — JSON-serializable snapshot; excludes the live asyncio clarification event and re-resolvable skills; serialize_as_any preserves a Pydantic custom_context.
• Pluggable CheckpointStore (services/checkpoint_store.py) — InMemoryCheckpointStore (default) and FileCheckpointStore (JSON on disk, survives restarts), each with optional max_history ring buffer; build_checkpoint_store() helper.
• Config (execution.checkpoint) — enabled / backend (memory|file) / dir / max_history; off by default.
• BaseAgent — checkpoint(), list_checkpoints(), rollback(step); automatic snapshot at the start of every iteration (crash-safe) and at the clarification pause. Auto-checkpoint failures are logged, never abort a run.
• AgentFactory — create(checkpoint_store=…, session_id=…) and restore(checkpoint) (rebinds id + streaming generator, re-applies conversation/context).
• REST API — GET /agents/{id}/checkpoints, POST /agents/{id}/rollback (cancels a running task first), POST /agents/{id}/restore; store wired from config at startup.
• ACP — advertises load_session=true, tags per-prompt checkpoints with the session id, and load_session rebuilds the session from its latest checkpoint; the next prompt resumes the restored agent instead of discarding it.
• Docs & examples — framework/checkpoints.md (en+ru) + nav, three endpoints in the API reference (en+ru), examples/checkpoint_and_rollback.py, and a commented execution.checkpoint block in config.yaml.example.

Testing

• 654 passed, 6 deselected (.venv/bin/pytest); ~66 new tests across models, store, config, agent, factory, server, ACP.
• pre-commit run -a clean (ruff, ruff-format, docformatter, mdformat).

Review

An adversarial review pass ran over the full diff; its findings were folded in as a dedicated fix commit:

• custom_context Pydantic model was silently reduced to {} on snapshot → fixed with serialize_as_any.
• A checkpoint-store failure could fail an otherwise-healthy run → auto-checkpoint isolated.
• ACP resume discarded restored state on the next prompt → prompt() now continues the restored agent; a checkpoint is captured at the clarification pause.
• find_by_session ordered by step across multi-agent sessions → ordered by created_at.
• Server rollback raced the running execute task → task is cancelled first.
• FileCheckpointStore.delete() broke on stray non-JSON files → uses rmtree.

Notes / scope

• Fully backward compatible: with checkpointing disabled the agent takes no snapshots and behaves as before.
• Restoring an agent requires its definition (def_name) to be present in the current configuration.

🤖 Generated with Claude Code