LangChain for TypeScript 1.5

Apply the Node.js 24 and TypeScript 7 prompts first. These rules cover LangChain for TypeScript 1.5 models, prompts, tools, agents, retrieval, structured output, middleware, memory, and operations.

Models and prompts

• Build createAgent models, system prompts, tools, middleware, and checkpointers from trusted registries, not request-controlled names.
• Allowlist context, configurable, tags, and metadata: callbacks and integrations may treat them as routing or storage controls, and context can override model-call limits and HITL policy.
• Set provider token/request limits and LangChain model-call limits; creating an agent imposes no budget.
• Keep origin labels on user, retrieved, memory, and tool messages instead of concatenating them into authoritative instructions.
• Validate returned messages and structured output at the destination; schema-valid values are still model-controlled.
• Deserialize only trusted payloads with @langchain/core/load or langchain/load; keep secretsFromEnv disabled and secret/import maps narrow.

Tools and agents

• Give each tool() a bounded schema and dispatch only the registered implementation; descriptions guide the model but grant no authority.
• Authorize parsed tool arguments against the current principal, tenant, resource, and state immediately before side effects.
• Configure toolCallLimitMiddleware with explicit runLimit and, for persisted threads, threadLimit; each is unlimited when omitted, though omitting both throws at construction.
• Choose exitBehavior deliberately: the tool-call limiter defaults to