Codex users compare iOS dictation, multi-thread UX, and long-context prompts

• NickADobos's dev loop turned Codex into a phone-first build loop: talk into the iPhone, keep testing the app, then send a batch of fixes and rebuild to device.
• OpenAIDevs on Build iOS Apps pushed more of that loop inside Codex itself, with an in-app browser, SwiftUI previews, and hot reload, while OpenAI's iOS simulator guide shows the same agent flow wired through XcodeBuildMCP.
• The UX split is getting sharper, because scaling01's CLI-tabs take says five Codex CLI tabs still beat the app for parallel work, while theo's reply says project-thread sorting is exactly why the app is easier.
• Long runs look less fragile than they did a few weeks ago: Vtrivedy10 on compaction said ultra-long Codex threads now survive many compactions, and aibuilderclub_'s six-part /goal structure frames prompt design, not model capability, as the main reason goals drift.
• Rough edges are still visible, because embirico's Windows-app reply said OpenAI was still unhappy with the Windows app, while a GitHub issue on iOS scrolling and a mobile reconnect issue show the phone loop is real but not smooth yet.

You can skim the official settings doc, the changelog, and the Build iOS Apps use case. There is also a live paper trail of the annoying parts, including slow long-thread scrolling on iOS, visible reconnects after app resume, and Windows 11 computer-use gaps.

The clearest workflow shift in this batch of posts is that some Codex users barely touch a laptop during app work.

According to NickADobos's dictation tip, Codex on iOS keeps recording when the app is backgrounded, so the loop becomes: open voice mode, switch into the app under test, narrate what feels wrong, then jump back just long enough to send the transcript. In follow-ups, NickADobos on background voice said this worked through ChatGPT's voice mode, and NickADobos on stack support said the stack was Swift in his case but not limited to Swift.

That workflow sounds extreme until you line it up with koltregaskes on Codex on phone, who said the phone mattered precisely because it captured ideas away from the laptop.

OpenAI is not just watching that phone-native loop happen. It is productizing it.

The shipped plugin exposes three concrete surfaces inside Codex:

• View and test the iOS app in the in-app browser, per OpenAIDevs on Build iOS Apps
• Open SwiftUI previews, again per OpenAIDevs on Build iOS Apps
• Hot reload edits without leaving Codex, per OpenAIDevs on Build iOS Apps

OpenAI's own iOS simulator debugging guide extends that same pattern into a fuller agent loop: select the project and scheme, drive the simulator, capture screenshots and logs, and keep the setup stable for the rest of the session. The companion post in OpenAIDevs on serve-sim and SnapshotPreviews also names the open source pieces behind the feature, serve-sim for browser streaming and SnapshotPreviews for SwiftUI preview extraction.

Codex's app UX is landing unevenly once work stops being a single thread.

scaling01's post argued that the app adds too much friction when you are juggling multiple things at once, and that five CLI tabs are still better. theo's reply pushed the opposite view, saying thread sorting by project beats managing a bunch of terminals.

That argument lines up with the official product changes. WesRoth's QoL roundup and OpenAIDevs on restored state both point to more stateful desktop behavior, including restored drafts, unread indicators, worktree context, and archived-section state. The official settings page adds more evidence that Codex is leaning into app-style organization, with configurable file-opening behavior, terminal-tab defaults, archived threads, profiles, memories, and keyboard shortcuts.

The more interesting long-context story is not bigger windows. It is that users are starting to trust longer autonomous runs again.

The aibuilderclub_ thread breaks a strong /goal into six parts:

1. Outcome
2. Verification
3. Constraints
4. Boundaries
5. Iteration policy
6. Stop condition

That structure comes from aibuilderclub_'s six-part /goal structure, while aibuilderclub_'s checklist adds an eight-item pre-launch check for objective acceptance criteria, a clear verification command, failure-retry instructions, and a final report. The linked GoalBuddy repository packages the same idea as a local Codex companion that turns a vague request into a goal.md charter, a state.yaml board, and role-tagged tasks.

The prompt-minimization angle shows up elsewhere too. dkundel on more context said extra context reduced prompts to shorthand like “fix the thing Katia messaged me about,” while gdb on missing context said tasks usually fail because context or skills are missing, not because the model cannot do them.

The happy-path demos are strong enough that the failure cases stand out more.

OpenAI's public reply on Windows was blunt: it is still not where they want it. The GitHub issue tracker filled in the specifics during the same window, with one Windows 11 report saying Chrome backend and Computer Use appeared on one machine but not another, and another Windows freeze report describing a renderer lockup in code review.

Mobile has its own tax. One iOS issue described long conversations becoming almost unusable to scroll, while another resume issue said returning to a backgrounded mobile thread causes a visible reconnect instead of a quiet resume. Reliability also bled into quota and session management: thsottiaux on undercounted tokens disclosed a token undercount bug that affected fewer than 15 percent of Pro and Plus accounts, and bridgemindai on the reset reported that paid-plan usage meters were reset after three reliability incidents in 24 hours.

The quiet product change underneath all of this is that Codex is behaving less like a disposable chat pane and more like a persistent work surface.

OpenAI added profile pages with activity graphs, streaks, lifetime tokens, peak daily tokens, and top features like plugins and /fast mode, according to OpenAIDevs on profiles and the matching changelog entry. Sharing is private-by-default and limited to consumer ChatGPT plans, per the official settings documentation.

That shipped alongside the less flashy persistence work: restored unread markers, drafts, zoom level, disabled hotkey preferences, archived-section state, and worktree context, plus searchable settings and cleaner fullscreen behavior in OpenAIDevs on settings search and OpenAIDevs on app polish. Codex users spent the week arguing about voice loops and CLI tabs, but the deeper pattern is that OpenAI keeps moving state, history, and telemetry into the app itself.