LangGraph graphs with checkpointed/multi-turn state fail the spawn-safety probe

[ambiorix2099]

Files: src/conductor/ai/agents/runtime/runtime.py (_register_passthrough_worker), src/conductor/ai/agents/runtime/_worker_entries.py (probe_spawn_safety)

Symptom: python examples/agents/langgraph/03_memory.py fails in <1s, client-side, before any server call:

SpawnSafetyError: worker 'memory_agent' is not spawn-safe
  AttributeError("Can't get local object 'CompiledStateGraph.attach_node.<locals>._get_updates'")
Define the callable at module level (importable by qualified name).

Cause: The passthrough worker pickles the compiled graph. LangGraph's CompiledStateGraph contains a local closure, so it can never be picklable. The error's advice is unfollowable — the object belongs to LangGraph, not to user code.

Fix: Either avoid pickling the compiled graph in the framework passthrough path, or detect this graph shape and fail with a message naming a supported alternative.

Verify: python examples/agents/langgraph/03_memory.py runs. Same for 13_multi_turn.py and 27_persistent_memory.py.

Scope: 3 of 46 langgraph examples. All three use checkpointed or multi-turn state; the other 43 pass.

[RemanenetSpy]

Title: Proposal: Resolve Spawn-Safety Probe Failures via Smriti Remote State Proxy

Problem Statement:
LangGraph's multi-turn checkpointed state relies on complex object structures that are incompatible with Python's pickle serialization. This causes conductor-oss/python-sdk to fail the spawn-safety probe when distributing agent processes.

Proposed Solution:
Integrate Smriti as the default external state backend for LangGraph processes within the SDK. Smriti offloads state serialization to a managed PostgreSQL database using deterministic Subject-Verb-Object (SVO) tuples.

Implementation Steps:

1. State Proxy: Implement a custom LangGraph BaseCheckpointSaver that maps state updates to Smriti API calls instead of local memory/disk.
2. Lazy Initialization: During process spawn, the SDK will only serialize the Smriti connection reference (e.g., API key and session ID). The actual state graph remains out-of-band and is not pickled.
3. SVO Tuple Mapping: Translate LangGraph state diffs into immutable SVO tuples stored in PostgreSQL. This provides persistent agent memory and resolves multi-turn data corruption.
4. Vector Extension (Optional): Expose Smriti's pgvector capabilities for users requiring semantic history retrieval across multi-turn sessions, without affecting the critical path of the spawn probe.

Technical Impact:

• Spawn-Safety: 100% bypass of local pickling limitations for complex LangGraph state.
• Scalability: State management is offloaded to a managed database, reducing local memory overhead on worker nodes.
• State Integrity: Deterministic SVO tuples ensure accurate multi-turn state reconstruction.