Claude Agent SDK for TypeScript 0.3

Apply the Anthropic SDK for TypeScript 0.122, Node.js 24, and TypeScript 7 prompts first. These rules cover Claude Agent SDK for TypeScript 0.3 sessions, tools, permissions, hooks, MCP, resume state, and deployment.

Sessions and queries

• Build Options.systemPrompt, model, cwd, tools, allowedTools, and environment overrides from trusted policy, never directly from query text.
• Set both maxTurns and maxBudgetUsd; 0.3 leaves them optional, so an omitted limit is not a usable application budget.
• Treat every SDKMessage variant separately and reject a run unless its result has subtype: 'success', is_error: false, and terminal_reason: 'completed' when present, within the configured limits.
• Store returned session IDs under the authenticated principal and verify that binding before supplying resume or forkSession.
• Reapply current authorization after resume; serialized messages and tool results are context, not proof that an earlier permission still holds.

Permission model

• Choose permissionMode in trusted code; never let a request select acceptEdits, bypassPermissions, or another more-permissive mode.
• Never set allowDangerouslySkipPermissions: true in a general-purpose service; also reject permissionMode: 'bypassPermissions' explicitly instead of relying on the flag's default.
• Use dontAsk for non-interactive workers only with a complete allow policy: otherwise actions that would prompt are denied, while default may prompt and acceptEdits auto-accepts edits.
• Implement canUseTool as fail-closed authorization over