OpenAI Agents SDK v0.20.0：默认模型改 GPT-5.6 Luna

Key Changes

Version 0.20.0 is a minor release because it changes the implicit default model and includes a potentially breaking MCP dependency migration for applications with customized local HTTP transports. MCP v1 remains supported, and affected applications can migrate to httpx2 or pin mcp<2.

Highlights:

• The default model is now gpt-5.6-luna; explicit models, run-level overrides, and OPENAI_DEFAULT_MODEL continue to take precedence.
• Local MCP connections support MCP Python SDK v1 and v2 across stdio, SSE, and Streamable HTTP transports.
• Applications using custom MCP HTTP authentication or client factories must use the HTTP types owned by the installed MCP major version, or pin mcp<2.
• RunState.add_input() can stage durable user input before a resumed model call, with guardrail, persistence, and serialization support.
• Sandbox mount validation adds explicit credential-exposure acknowledgements and preserves redacted error contracts without serializing credential authority.
• Realtime input transcription supports the GA transcription settings for gpt-live-transcribe, gpt-transcribe, and gpt-realtime-whisper.

What's Changed

• feat: use GPT-5.6 Luna as the default model by @seratch in #4282
• feat: add durable pending input to RunState by @seratch in #4325
• feat(mcp): support MCP Python SDK v1 and v2 by @seratch in #4106
• feat(sessions): pass run context to custom sessions by @seratch in #4209
• feat(realtime): support GA transcription options by @seratch in #4351
• feat: preserve raw usage payloads by @seratch in #4279
• feat: allow applications to approve unsafe replays by @seratch in #4319
• fix: falsy optional reference handling by @seratch in #4305
• fix: prevent queue consumer deadlocks by @seratch in #4201
• fix: publish completed streamed guardrail results on failure paths by @adityasingh2400 in #4223
• fix: annotate the bare tool guardrail decorator overloads by @adityasingh2400 in #4238
• fix: tool approval is not honored on resume when Runner.run is given a context by @ojassharma7 in #4245
• fix: redact JSON validation errors by @seratch in #4211
• fix: copy raw_responses when building a RunState by @adityasingh2400 in #4237
• fix: keep empty turns in the nested conversation history by @abhay-codes07 in #4230
• fix: keep tool guardrail results when a resumed run interrupts again by @adityasingh2400 in #4239
• fix: work on copies of the lists a resumed run adopts from RunState by @hsusul in #4251
• fix: strip placeholder item IDs without provider_data on the Responses path by @LeSingh1 in #4266
• fix: bind tool approvals to concrete invocations by @seratch in #4257
• fix: normalize optional fields in tool invocation identity by @seratch in #4289
• fix: count requests when providers omit usage by @seratch in #4290
• fix: suppress pydantic serializer warnings by @seratch in #4291
• fix: keep FunctionTool subclasses copyable by @seratch in #4311
• fix: preserve streamed guardrail session state by @seratch in #4312
• fix: keep acknowledged safety checks serializable after restore by @ribhavjain in #4316
• fix: preserve default-valued fields in serialized tool output by @hsusul in #4307
• fix: strip output-only created_by when replaying model output as input by @hsusul in #4308
• fix: preserve local shell outputs across RunState resume by @seratch in #4320
• fix: serialize containers of models and dataclasses as structured data by @ribhavjain in #4339
• fix: reject schemas that exceed safe recursion depth by @seratch in #4358
• fix: reject unsafe strict-schema ref siblings by @seratch in #4356
• fix(sandbox): reject unsafe mount credential configurations by @seratch in #4255
• fix(extensions): settle snapshot directory mount transitions for Modal by @seratch in #4315
• fix(sandbox): preserve sandbox error contracts during mount redaction by @seratch in #4344
• feat(sandbox): add explicit mount credential exposure acknowledgements by @seratch in #4321
• fix(sandbox): cancel sibling env resolvers when one fails by @PranavMishra28 in #4195
• fix(sandbox): encode Blaxel terminal WebSocket URL query values by @abhay-codes07 in #4217
• fix(sandbox): stop splitting Cloudflare SSE events at chunk boundaries by @abhay-codes07 in #4215
• fix(sandbox): keep move_to when coercing apply_patch operation mappings by @hsusul in #4242
• fix(sandbox): retain program items in memory rollouts by @seratch in #4276
• fix(sandbox): restore archived file modes when extracting a workspace tar by @hsusul in #4287
• fix(sandbox): keep source entry metadata when lazily loading a skill by @hsusul in #4294
• fix(mcp): hosted MCP approval identity by @seratch in #4256
• fix(mcp): keep MCP error content when structured output is enabled by @adityasingh2400 in #4224
• fix(mcp): build the active server list without re-consuming the iterable by @adityasingh2400 in #4235
• fix(mcp): preserve free-form MCP object schemas by @seratch in #4310
• fix(mcp): serialize non-text content blocks as JSON instead of Python repr by @ribhavjain in #4338
• fix(mcp): serialize manager lifecycle operations by @seratch in #4340
• fix(mcp): bound lifecycle and CI waits by @seratch in #4342
• fix(chat-completions): keep url citations when converting chat completions output by @adityasingh2400 in #4222
• fix(chat-completions): assemble chat completions content parts in content index order by @adityasingh2400 in #4236
• fix(chat-completions): propagate the OpenAI request ID on the Chat Completions path by @hsusul in #4243
• fix(chat-completions): keep url citations on the streamed chat completions path by @Luccacvb in #4252
• fix(chat-completions): announce the chat completions assistant message once by @hsusul in #4275
• fix(chat-completions): convert input_file items that reference a file_id on the Chat Completions path by @hsusul in #4295
• fix(chat-completions): raise on audio output in the streamed chat completions path by @Luccacvb in #4309
• fix(chat-completions): omit parallel_tool_calls without tools on the Chat Completions path by @Luccacvb in #4359
• fix(sessions): release SQLite engine config entries when engines are collected by @adityasingh2400 in #4210
• fix(sessions): enforce closed state on SQLite session empty add_items by @adityasingh2400 in #4231
• fix(sessions): make session mutations atomic by @seratch in #4212
• fix(sessions): preserve required program item ids for OpenAI conversations by @hsusul in #4253
• fix(sessions): apply the reasoning item id policy to stored session history by @LeSingh1 in #4278
• fix(sessions): preserve DaprSession created_at across writes by @adityasingh2400 in #4213
• fix(sessions): strip placeholder item IDs before Conversations persistence by @hsusul in #4288
• fix(sessions): preserve falsey compaction decision hooks by @hsusul in #4299
• fix(sessions): restore session history when compaction replacement is cancelled by @nothariharan in #4298
• fix(sessions): prune orphaned tool outputs from limited sessions by @seratch in #4324
• fix(realtime): apply tool call item updates to session history by @hsusul in #4284
• fix(realtime): preserve falsey custom models by @hsusul in #4286
• fix(realtime): compute G.711 audio length for typed and mapping format spellings by @abhay-codes07 in #4292
• fix(voice): forward TTSModelSettings.speed to the OpenAI speech API by @adityasingh2400 in #4220
• fix(voice): keep the word separator when flushing streamed sentences by @abhay-codes07 in #4227
• fix(voice): start the dispatcher when a session ends with no turns by @adityasingh2400 in #4259
• fix(voice): report transcription session close failures to the consumer by @Luccacvb in #4343
• fix(tracing): release the trace scope when a generator is closed by @adityasingh2400 in #4221
• fix(tracing): release the span scope when a generator is closed by @adityasingh2400 in #4233