OpenResponses 3.0 release notes
Version 3.0 was created in App Store Connect on October 19, 2026. Its first code was pushed on November 17, 2026 (7a12b34, Xcode Cloud run 55). On November 19, 2026, at Gunnar’s word, the work first written up as 3.1 (the Settings redesign, Text Size, Connectors and the large-text fixes) was folded into 3.0, and that day’s push (2b70581, Xcode Cloud run 56) became build 56, which replaced build 55 and was submitted for review the same evening. GitHub release: https://github.com/Gunnarguy/OpenResponses/releases/tag/v3.0. The App Store What’s New text is release_notes.txt. The details and their tests are in CHANGELOG.md under 3.0.
Settings, Text Size, Connectors, large text (added November 19, 2026)
| Before | 3.0 |
|---|---|
| Five segmented tabs; Local Python, the hosted shell, multi-agent, tool search, moderation, voice and the API Workbench sat at the bottom of the Model tab | One list of named pages, each row showing its current value, and a search over about 50 options (SettingsHomeView, SettingsSearchIndex) |
| Text size followed the iPhone only | Settings → Text Size: Same as iPhone, or one of the twelve system sizes for this app, applied as a window trait override so sheets follow it (TextSizeSettings.swift) |
| Connected Apps: a featured list behind a category picker, a “Search the MCP Registry” button, and “Provider setup required” services with no button | Connectors: popular services by kind plus the registry’s remote servers paged as the list scrolls, searched while typing; Try Signing In and Connect with a key or token for every listed service; GitHub’s official server; Notion as one connector (MCPConnectionsView) |
| At Accessibility XXXL: “Op / en / AI” in Settings rows, “fet / c…” and “Fai / led” in tool steps, “Welcome…” in onboarding, “Mes…” in the message field | Stacked layouts at the accessibility sizes; both chat-screen audits and the Settings audit pass |
Verified on simulator 54DDDEDE (iOS 27.0), November 19, 2026: 384 unit tests, 3 skipped, 0 failures; all 8 UI tests pass. Not tried: account sign-in or a key connection against a live service.
Tool steps
| What 2.9 showed | Cause | 3.0 |
|---|---|---|
| A function call’s step stayed empty until the call had come in | response.function_call_arguments.delta and .done fell to default in handleStreamChunk, so a row got its arguments only from response.output_item.done (Responses API streaming events reference, read October 19, 2026) | Each delta appends to the row’s rawArguments, and .done fills a row that got no delta. Neither changes the row’s status, and calls still run from output_item.done |
| Nothing, for a custom tool call’s input events | response.custom_tool_call_input.delta and .done also fell to default | Recognized. A custom tool call has no row yet, so nothing shows for it |
Function calls in the chat are the app’s own tools: the browser (browserNavigate and the other browser actions), Calendar, Reminders, Contacts, Notion, Local Python (run_python) and the Custom tool from Settings. Tapping a step shows its arguments under “Arguments:” (ToolTimelineView).
Calendar and Reminders dates
| What 2.9 did | Cause | 3.0 |
|---|---|---|
| A time with no offset (“2026-11-04T18:30:00”) failed: “The provided date is not a valid ISO 8601 timestamp”. On October 21, 2026, all twelve Calendar and Reminders calls for a photographed schedule failed this way | AppleDateUtilities.parseISO8601 (Date.ISO8601FormatStyle) needs an offset or Z; ISO 8601 does not | parseToolDate reads it in the device’s time zone; a value with an offset or Z means that moment, with or without seconds |
| A bare date (“2026-11-04”) failed the same way | The same | As a reminder’s due date, a due day with no time; as a query’s end, the whole day; as an event’s start or end, refused with “Calendar events need a time of day”, since the app makes timed events only |
The tool schemas already ask for the device’s offset “preferably”, so the model’s requests don’t change. parseISO8601 stays strict for the file converter’s date-column check.
Decisions
OpenAI’s Decisions API (POST /v1/decisions) is generally available: it returns typed answers instead of text, and runs only on gpt-6-luna at $0.10 per million input tokens with output not charged (Decisions guide and API reference, read November 13, 2026).
| Where | What it does |
|---|---|
Settings → Decisions (under Developer) (DecisionsView) | Text input and an optional photo (sent as a JPEG data URL, at most 1,536 pixels on its longest side). Up to 400 questions: yes or no (predicate, a probability), choice (2 to 255 unique values, a probability each) and score (2 to 10 ordered levels; the score is a probability-weighted average of the level numbers). Answers show as bars with their confidence; a refusal and an unknown answer type are shown, not dropped. The footer shows input tokens, cached tokens and the cost. Three examples and Copy Request JSON are in the ⋯ menu. Questions are saved on the device (UserDefaults decisionsQuestions.v1); the input and photo are not. The same data-sharing consent as the Workbench, and offline in Demo Mode. |
| API Workbench → Create decision | POST /decisions with the support-ticket example. |
scripts/decisions_probe.sh [Keychain service] | One live call with the same example; the key goes from the Keychain to curl on stdin. |
Request building, validation and answer decoding are in DecisionsModels.swift (DecisionsRequest.problem names what OpenAI would reject before the request is sent). The Agents API (/v1/agents/sessions, beta) was considered and not added: the app already does multi-agent turns, programmatic tool calling, the hosted shell and MCP through the Responses API, and the Agents API has no Zero Data Retention.
Rating request
Until 3.0 the app never asked for a rating: on November 14, 2026 it had 359 first-time downloads and 1 US rating. 3.0 asks with Apple’s own sheet (SwiftUI requestReview; RequestReviewAction and the Human Interface Guidelines’ Ratings and reviews page, read November 14, 2026).
| Rule | Value |
|---|---|
| Asks on | the first chat answer that finished with no failed or cancelled tool step, or the third finished answer of any kind |
| When | two seconds after the answer finishes, only while the app is in front, nothing is streaming and onboarding is closed; otherwise held until the next return to the app |
| Per version | once |
| Between asks | at least 14 days |
| Never | at launch, in onboarding, in Explore Demo, from a button, under tests; an error, a stopped answer or one waiting for an approval doesn’t count |
| Recorded as asked | only after the request was made. Apple decides whether the sheet shows (at most three times in 365 days, never in TestFlight) |
About has a “Rate OpenResponses” link to https://apps.apple.com/app/id6757338355?action=write-review. The rules are ReviewPromptService.shouldAsk and outcome; the request is made in ContentView.askForRatingWhenInFront.
Verification
November 17, 2026, the tree that was pushed (7a12b34, with every 3.0 change): the full unit suite (-only-testing:OpenResponsesTests) on simulator “OpenResponses tests” 7AFAC899 (iPhone 18 Pro, iOS 27.0): 384 executed, 4 skipped, 0 failures, including the 14 in ReviewPromptServiceTests. Xcode Cloud run 55 archives that commit. Not seen yet: the rating sheet itself, and a live Decisions answer.
ToolRowsTextAndFilesTests and StreamingEventDecodingTests on simulator “OpenResponses tests” 96574C1C (created October 19, 2026): 23 executed, 0 failures, including the new testFunctionCallArgumentsStreamIntoTheirRow and testArgumentsDoneFillsAnEmptyRowAndCustomInputHasNoRow. The full suite CI runs has not been run on this change yet. Code: commit 8836c39.
Calendar and Reminders dates (October 21, 2026, uncommitted when run): AppleDateUtilitiesTests and DateFormattingTests on simulator “OpenResponses tests” D4947B13 (iPhone 18 Pro, iOS 27.0, created that day; Reminders access granted with simctl privacy for the EventKit test): 26 executed, 0 failures. The full suite CI runs (-only-testing:OpenResponsesTests -skip-testing:OpenResponsesTests/BrowserLiveSiteTests) on the same simulator: 360 executed, 2 skipped (the live docs-page test without its variable, and a search-by-meaning test that needs Apple’s on-device model), 0 failures. Two read-only review passes; the first found a skipped midnight, bare event dates, untested call sites, non-ASCII digits and a release-notes claim, all fixed; the second found only low-severity items, also fixed.
Decisions (November 13, 2026, uncommitted when run): DecisionsTests (9, at the end of OpenAIServiceTests.swift, one of them draws the answer cards with ImageRenderer) and the full unit suite (-only-testing:OpenResponsesTests) on simulator “OpenResponses tests” 2460D513 (iPhone 18 Pro, iOS 27.0): 370 executed, 3 skipped, 0 failures. In that simulator’s Explore Demo the screen opens from Settings → Advanced → Developer Operations, Ask shows the Demo Mode message, and the question editor lists the example’s choices. Not yet sent to the live API.