LangGraph for Python 1.2

Apply the LangChain for Python 1.3 and Python 3 prompts first. These rules cover LangGraph 1.2 graphs, state, reducers, persistence, threads, interrupts, durable execution, replay, subgraphs, and deployment.

Graph topology and routing

• Define nodes, edges, conditional routes, entry points, and termination conditions in trusted code.
• Map model-produced route labels through an exact allowlist and reject unknown destinations.
• Keep privileged nodes unreachable unless deterministic policy confirms the current identity and workflow state.
• Enforce routing in code: destinations only affects diagrams, and Command(goto=...) does not cancel static edges, so keep static edges from bypassing approval routes.
• Set a finite recursion_limit in the invocation config and independently bound branches, subgraphs, parallel Sends, wall time, and aggregate model or tool budget.
• Treat every node output as untrusted before it changes routing, capabilities, or external state.

State and reducers

• Define a narrow typed state schema and keep credentials and service clients outside serializable graph state.
• Store authoritative identity, permissions, approvals, and tenant scope in protected runtime context or application state.
• Use reducers that enforce size limits, provenance, and deterministic merge semantics for concurrent updates.
• Prevent untrusted state updates from replacing protected fields or injecting executable objects.
• Authorize update_state() and bulk_update_state() separately from ordinary input; as_node can change pending execution, while