Update dependency MacPaw/OpenAI to from: "0.5.3"

[renovate]

This PR contains the following updates:

Package	Update	Change
MacPaw/OpenAI	minor	from: "0.4.7" → from: "0.5.3"

---

Release Notes

MacPaw/OpenAI (MacPaw/OpenAI)

v0.5.3

Compare Source

Added

• Model.gpt_audio_1_5, Model.gpt_audio and Model.gpt_audio_mini, the audio-in, audio-out models for Chat Completions, also returned by Model.allModels(satisfying:). The README gained an Audio input and output with Chat Completions section covering modalities, audioOptions, audio input parts and streaming (#​397).

Changed

• Model.gpt_4o_audio_preview is deprecated, since OpenAI shut the model down on May 7th, 2026; use Model.gpt_audio_1_5. It is no longer returned by Model.allModels(satisfying:) (#​397).
• swift-openapi-runtime 1.12.0 → 1.12.2, which also moves swift-http-types from 1.3.1 to 1.8.0 (#​439, #​459).
• CI: the API Breakage allowlist is emptied, because the 0.5.2 release is now the comparison baseline for the breaks it listed.

Fixed

• Demo: ResponsesStore.replyFunctionCall included the completed .functionToolCall in the next request along with its .functionCallOutputItemParam while previous_response_id was already set, so the Responses API rejected the follow-up with Duplicate item found with id fc_.... The follow-up now sends only the function-call output, and the pending call stays available if that request fails (#​457).

v0.5.2

Compare Source

Added

• ModelSpec, describing a model's capabilities as its OpenAI docs page lists them: endpoints, features, Responses API tools, supported reasoning efforts, and limitations the docs note, such as Chat Completions function calling being unsupported or requiring reasoning_effort: none (ModelSpec.Endpoint, .Feature, .Tool, .ReasoningEffort and .Limitation). ModelSpec.all currently covers the GPT-6 models; Model.allModels(satisfying:) filters them by their spec. Model.Filter.Endpoint, .Feature and .Tool are now typealiases for the ModelSpec types. ModelSpec.ReasoningEffort converts to Components.Schemas.ReasoningEffort and ChatQuery.ReasoningEffort (xhigh and max become .customValue, as ChatQuery.ReasoningEffort has no cases for them).
• Model.gpt6_astra, Model.gpt6_sol and Model.gpt6_luna, also returned by Model.allModels(satisfying:) for the Chat Completions and Responses endpoints and the MCP tool.
• Model.Filter.Tool.mcp and Model.Filter.requiredTools, so Model.allModels(satisfying:) can narrow results to models that support remote MCP servers in the Responses API.
• ResponseStreamEvent.shellCall, covering the new shell-call streaming events (response.shell_call_command.added/delta/done, response.shell_call_output_content.delta/done) added by the latest OpenAPI spec.
• Tool.programmaticToolCallingTool, the tool case paired with the OutputItem.program/.programOutput response items; without it, decoding a response whose tools array contains a programmatic_tool_calling tool would throw.
• ResponseStreamEvent.customToolCallInput, covering the previously unhandled response.custom_tool_call_input.delta/done streaming events; without it, a custom-tool streaming response would fail the whole stream with an unknownEventType error.
• InputItem.compactionTriggerItemParam, .programItemParam, and .programOutputItemParam, the item cases added by the latest OpenAPI spec; without them, decoding an input item with type: "compaction_trigger", "program", or "program_output" would throw unknownOneOfDiscriminator.
• ResponsesEndpoint.retrieveResponse(query:) and cancelResponse(id:), in callback, async and Combine variants, for GET /responses/{response_id} and POST /responses/{response_id}/cancel. RetrieveModelResponseQuery.include is sent as include[] query parameters. Cancelling takes just the response ID, matching the official Python SDK's client.responses.cancel(response_id), and only applies to responses created with background: true.
• CreateModelResponseQuery.contextManagement, encoded as context_management, for configuring server-side context compaction.
• CI: an API Breakage workflow fails a pull request that changes the public API compared with the latest release tag. Consciously accepted breaks are listed in .github/api-breakage-allowlist.txt with a matching changelog entry.
• CI: the Swift Build workflow now builds and tests on Linux with Swift 5.10, 6.0 and 6.3 containers, tests on macOS and the iOS Simulator, and builds for tvOS, watchOS and visionOS with Xcode. Tests written with Swift Testing only exist on toolchains that ship it (Swift 6); the XCTest suite runs everywhere.
• CONTRIBUTING.md: API stability policy, including how new endpoint groups are added as namespaces and how generated Components.Schemas types are treated.
• This changelog.

Changed

• Breaking: ResponsesEndpointProtocol, ResponsesEndpointAsync and ResponsesEndpointCombine gained retrieveResponse and cancelResponse requirements. Code that only calls openAI.responses is unaffected, but custom conformances (e.g. test doubles) need to implement them.
• Breaking: GetModelResponseQuery was renamed to RetrieveModelResponseQuery to match retrieveResponse(query:). It had no public initializer before, so it couldn't be constructed outside the package.
• CreateModelResponseQuery.init gained a contextManagement parameter (defaulting to nil). Existing call sites compile unchanged; only code referring to the initializer by its full name, such as CreateModelResponseQuery.init(input:model:...) used as a function value, needs updating.
• Breaking: Regenerated Components.Schemas from the latest OpenAPI spec. Most of the reported breaks are mechanical (new cases added to generated enums, and generated memberwise initializers gaining parameters for new optional fields) and are allowlisted in .github/api-breakage-allowlist.txt without individual call-outs, matching how the 0.5.1 regeneration was documented. The changes worth knowing about if you read or construct these types directly:
  • ResponseObject.instructions changed from String? to ResponseObject.Instructions?, letting instructions be either a plain string or a list of input items. Instructions is a generated two-case enum (.case1(String) for the string form, .case2([InputItem]) for the list form); code that read instructions as a String needs to switch over it instead. This mirrors how the official Python SDK models the same field (Union[str, List[ResponseInputItem], None]) rather than adding a second property under a new name.
  • Components.Schemas.ServiceTier was renamed to ServiceTierResponses.
  • Components.Schemas.Conversation2 was removed; Response.conversation and ResponseProperties-derived types now use ResponseConversation instead.
  • MCPToolCall.error changed from String? to MCPToolCallError?.
  • ResponseOutputTextAnnotationAddedEvent.annotation changed from OpenAPIObjectContainer to Annotation?, matching the stricter Annotation schema.
  • FunctionToolCallOutput.callId and FunctionCallOutputItemParam.callId changed from String to String?.
  • ResponseStreamEvent.reasoning was renamed to .reasoningText, with its payload changing from ReasoningEvent to ReasoningTextEvent.
  • ResponseObject.toolChoice changed from Schemas.ToolChoiceParam? to Schemas.ToolChoiceParam. The generated ResponseProperties.toolChoice is optional only because that schema is shared with request bodies, where tool_choice may be omitted; an actual Response always includes it, which we verified against the API docs and the official Python SDK (Response.tool_choice: ToolChoice, non-optional). This corrects a facade type that didn't match the real response shape, but is still a source-breaking type change for callers who read toolChoice as optional.
  • OutputItem.webSearchToolCall's payload changed from Components.Schemas.WebSearchToolCall to a new hand-written WebSearchToolCall type whose action is optional. The spec marks action required, but the live API can emit a web_search_call output item without one (e.g. a response.output_item.added event while status is still in_progress), and decoding threw a keyNotFound error for action whenever that happened. This is an acknowledged upstream spec bug (openai/openai-openapi#572), not something specific to any one status. Callers that read .webSearchToolCall(let call) need to unwrap call.action instead of accessing it directly.

Fixed

• Demo: a non-streaming Responses request to a reasoning model (e.g. gpt-5) always includes a .reasoning output item, and ResponsesStore.createResponse(query:) didn't handle it (or the MCP output item cases), so it threw unhandledOutputItem instead of completing.
• Components.Schemas.InputItem and the InputItem facade failed to decode a valid { "type": "item_reference", "id": "..." } payload with unknownOneOfDiscriminator. The generator only matched discriminator values declared as a plain string enum, but ItemReferenceParam.type is declared as a nullable anyOf: [<string enum>, {type: null}], which it didn't unwrap; the InputItem facade also fixed the same bug independently, plus a stale "ItemReference" discriminator string (missing the Param suffix) and a missing "message" match for EasyInputMessage.
• ResponseStreamEvent and OutputItem didn't round-trip through JSONEncoder/JSONDecoder: both have a hand-written init(from:) that decodes the API's type discriminator, but no encode(to:), so the compiler-synthesized one wrapped values by enum case name instead of matching the API's shape; the hand-written ResponseOutputItemAddedEvent/ResponseOutputItemDoneEvent also gained a sequenceNumber field they were missing, since re-encoding through them dropped sequence_number.
• ResponseStreamEvent.reasoningText streaming events never decoded: ModelResponseStreamEventType listened for response.reasoning.delta/.done, but the API sends response.reasoning_text.delta/.done, so reasoning-text deltas always failed with an unknownEventType error.
• make download-spec could leave the tracked openapi.yaml truncated if curl was interrupted mid-transfer; it now downloads to a temp file and moves it into place only on success.
• Build warning in ModelResponseEventsStreamInterpreter when logging a failed stream event decode in debug builds.
• The test target compiles for the package's minimum iOS deployment target again; it used Regex, which requires iOS 16.
• Building on Linux with Swift 5.10 works again. swift-corelibs-foundation gained the async URLSession APIs only in Swift 6, so the async client now bridges the completion-handler API on older Linux toolchains.

v0.5.1

Compare Source

Added

• refusal field on the streaming chat completion delta (#​408).
• reasoningContent on assistant chat messages for providers that return reasoning (#​402).
• annotations on ChatStreamResult (#​399).

Changed

• Regenerated Components.Schemas from the updated OpenAPI spec (#​433). This changed public generated types: transcription segment and word timestamps moved from Float to Double, ImageGenTool.SizePayload lost its fixed cases, responseId was removed from the response audio events, and ChatResult.Choice.Message.Annotation moved to a top-level Annotation type. Two ChatCompletionMessageParam initializers were removed.
• swift-openapi-runtime 1.11.0 → 1.12.0 (#​424).

v0.5.0

Compare Source

Added

• model is optional in CreateModelResponseQuery for the stored-prompt (prompt_id) flow (#​419).

Changed

• Regenerated Components.Schemas from the fresh OpenAPI spec (#​423); the release notes flagged possible breaking changes.
• swift-openapi-runtime 1.8.2 → 1.11.0 (#​416).

Fixed

• Missing name field on ResponseFunctionCallArgumentsDoneEvent (#​420).

v0.4.9

Compare Source

Added

• completion_tokens_details on CompletionUsage (#​411).
• prompt_cache_key on CreateModelResponseQuery (#​412).
• Dependabot configuration for GitHub Actions and Swift dependencies (#​414).

Changed

• ChoiceDeltaToolCall.index is required again (#​388).
• ChatQuery description output is easier to read (#​394).
• Documentation: response_format versus text for structured outputs (#​413); security policy points to the private reporting channel (#​410).

Older releases are documented on the GitHub Releases page.

v0.4.8

Compare Source

What's Changed

• Add none reasoning effort, add gpt-5.1 and gpt-5.1-chat-latest models by @​neelvirdy in #​393
• Add verbosity support for Responses API by @​neelvirdy in #​395
• Added priority to service tier by @​tzdesign in #​405

New Contributors

• @​tzdesign made their first contribution in #​405

Full Changelog: MacPaw/OpenAI@0.4.7...0.4.8

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.