OpenAI Agents SDK for TypeScript 0.21

Last reviewed Oct 21, 2026

Export as PDF

Prompt Details

Prompt Composition

5prompts
  1. JavaScript ES2026102 lines · 3.1K tokens
  2. Node.js 24101 lines · 5.3K tokens
  3. TypeScript 746 lines · 1.5K tokens
  4. OpenAI SDK for TypeScript 7.980 lines · 3.4K tokens
  5. OpenAI Agents SDK for TypeScript 0.2166 lines · 3.7K tokens
Total395 lines · 17K tokens

CWE Coverage

  • 15
  • 20
  • 22
  • 59
  • 73
  • 78
  • 79
  • 88
  • 89
  • 95
  • 113
  • 117
  • 200
  • 208
  • 209
  • 214
  • 250
  • 252
  • 295
  • 297
  • 306
  • 319
  • 323
  • 330
  • 338
  • 346
  • 347
  • 352
  • 354
  • 362
  • 367
  • 377
  • 426
  • 444
  • 459
  • 471
  • 488
  • 494
  • 502
  • 522
  • 524
  • 532
  • 602
  • 639
  • 653
  • 732
  • 770
  • 772
  • 807
  • 829
  • 838
  • 841
  • 862
  • 863
  • 908
  • 915
  • 916
  • 918
  • 922
  • 942
  • 943
  • 1004
  • 1321
  • 1333
  • 1426
  • 1427

Prompt Preview

Apply the OpenAI SDK for TypeScript 7.9 prompt first, including its Node.js 24 and TypeScript 7 baselines. These rules cover the 0.21 line's text and Realtime runs, tools, handoffs, MCP, saved state, tracing, and sandbox capabilities.

Instructions and capability boundaries

  • Keep directly installed @openai/agents and @openai/agents-* packages on one linked release set in the lockfile. A caret range on this pre-stable SDK stays within one minor; review the current release policy and breaking changes when changing minor lines, without assuming older minors receive backports.
  • Prevent prompt injection by keeping user input, retrieved documents, tool results, workspace files, and memory out of trusted Agent.instructions, SandboxAgent.baseInstructions, and policy-bearing configuration. Summaries and compaction output remain untrusted data; never let them grant capabilities or approvals.
  • Build RunContext identity and tenant fields from authenticated application state. Keep callModelInputFilter and dynamic instruction callbacks application-owned; neither model arguments nor stored conversation content may replace current authorization context.
  • Use bounded Zod schemas or supported synchronous Standard Schema values for tool().parameters, agent.asTool().parameters, Agent.outputType, and handoff().inputType. Raw JSON Schema only describes the contract and parses JSON; validate it locally before use. Keep strict mode enabled and enforce domain constraints after transforms.
  • For tool authorization, check the current principal, operation, normalized