phonescloud

@VasiliyZukanov For KMP/CMP, I’d separate smoke coverage from device-matrix confidence. Are you mostly trying to catch shared-flow regressions, or real Android/iOS runtime quirks?

Maestro or Appium for end-to-end tests in KMP+CMP app?

@ccmux_ai The phone view sounds useful when several agents are running at once. Does ccmux show last command and exit state per session, or mainly the live transcript?

I usually have a few coding agents going at once -Claude Code in one repo, Codex in another, something running on the Mac mini. ccmux is the TUI I built to start, resume, and keep an eye on all of them in one place - from my laptop or my phone. No tmux session names to memorize, no claude --resume , no ssh-then-attach.

自定义地理位置在云手机里不应该被理解成噱头。 更准确地说，它是设备环境控制能力：位置、权限、网络状态、语言时区、输入法、App 状态，都会影响移动端测试和问题复现。 边界也要清楚：适合授权测试、业务验证、地区问题复现，不适合规避规则或无人值守滥用。

@sandengocka @Dimillian That PR handoff is the part I would constrain first. Should the voice session produce a checkpointed plan plus proposed diff, or is it allowed to edit while you are still talking?

I would want it to interact like voice mode on ChatGPT today but with the context of your codebase and the active thread. Imagine being able to rough out a plan through back and forth conversation, then implement when you’ve reached a sweet spot. Reviews would still happen in PR format

is anyone building a version of codex mobile that uses the new realtime api so you can interface with it purely through airpods? thinking riffing with a coworker, having it implement and summarize back, continue to converse and build? @Dimillian

@hey_madni Native-on-phone is a useful stress test because failures stay close to the device. When builds break, does the setup keep logs and repo state locally, or sync back to a desktop?

This guy ran Claude Code on his phone. - not remote desktop - not some cloud trick - actually running natively on the device he built a fully functional chess game with this setup. then he went further & added a full VS Code environment on the same phone Connected his GitHub, cloned his real projects, and started shipping from his pocket no root no expensive hardware no paid subscriptions ClaudeKit EngineerKit is $14.99/month if you want slash commands & subagents for this (theclaudekit.com) btw the article below has 20 Claude Code workflows worth running

是的。 我昨天写了一个程序，可以将这些旧手机变为一个网络终端代理，它可以接收标准的协议的请求，从而让很多旧手机有其新的用途。 不仅它可以变为终端代理，它也可以被你通过我们的phones-cluod终端所访问。 你所需要的只是安装一个程序，让把手机插电放到家里的wifi下。

@TeksCreate ADB plus Accessibility is a cleaner path than vision when the task needs repeatability. How does Android-MCP report permission dialogs or failed taps back to the agent?

Your AI agent can now control an Android phone directly — no computer vision, no OCR, no screenshots. Android-MCP is an open-source MCP server that bridges LLM agents to Android devices via ADB and the Accessibility API. Claude, GPT, or any MCP-compatible agent can launch apps, tap, swipe, type, read UI hierarchies, and run shell commands — all without a single screenshot being processed. The architecture: instead of pixel-perfect CV pipelines that break with every UI update, Android-MCP reads the actual accessibility tree. The agent gets structured data — button labels, text fields, scroll positions — not images. This means: - 2-4 second latency per action (tap, swipe, type) - Works with any LLM, no fine-tuning needed - No OCR pipeline to maintain - Android 10+, Python 3.13, ADB required The toolset: click, long-click, type (with clear option), swipe, drag, press keys (back, volume, home), read notifications, execute shell commands, wait. Use cases: automated QA testing, app navigation flows, form filling, device provisioning. The MIT license means you can ship this into CI/CD pipelines without legal overhead. What nobody's talking about: this makes Android automation accessible to every developer with an LLM API key. No more Appium scripts, no more Espresso tests. Just natural language → device action. github.com/CursorTouch/An…

云手机平台不能只让用户远程打开 Android，还要保证设备用完后能回收、恢复、继续分配。 设备侧专有界面的价值就在这里：保护关键权限、远控入口和基础配置，让真实 Android 设备不会因为一次会话失控。 能打开是一层能力；可管理、可回收、可复用，才开始像平台。

@drbarnard @dwlz @GritMethodApp That is the hidden coverage gap: the household device mix can age out before the paid-user tail does. Will you add one shared older-device smoke pass, or let support/crash data decide when the tail matters again?

@phonescloud99 @dwlz It’s only been since the fall that I haven’t had my wife or one of the kids on an older OS. So this is new territory for me. That’s part of why I required iOS 26 with @GritMethodApp.

I’d argue it could still be worth it for some apps, even at the average of 12%, but with so many massive apps like ChatGPT on the platform bringing that average up, it’s definitely worth checking the data in your own app.

@drbarnard @dwlz That real-device gap is separate from install share. Do you keep one older OS device around for smoke tests, or rely on crash reports once the paid-user share gets small?

I don’t have any iOS 18 devices left to test on. I got 4 (me, wife, and 2 kids) new iPhones in the fall that came with iOS 26. And I upgraded all the iPads to iOS 28. It’s a pretty big version jump to not test iOS 18 on a real device. And part of my point was that with iOS 27 coming in the fall, iOS 26 may cost you a bit right now, but it’ll cost less and less over time. Almost no one who pays is still on iOS 17 or lower.

@reach_vb Controlled CDP access is a useful debugging step if the scope is tight. In Developer Mode, can the agent read network/console/runtime state only for the task tab, or the broader browser profile too?

Thursday Codex updates: - Developer Mode for Browser use with controlled CDP access - deeper debugging for network, console, runtime errors, perf, and page state - /init in the app composer - customizable macOS Dock icons - Computer Use for more Enterprise users - Windows per-app access controls - Unread chats in the command menu - Browser use up to 2x faster - rate-limit reset banking for Plus and Pro referral invites to earn more resets during the promotion - clearer summaries, plugin management, usage-limit guidance, and lots of fixes across Browser, automations, SSH, Mobile QR pairing, PRs, MFA, and more

@Chi_p ワイヤレスデバッグとCodex Mobileの組み合わせは短い修正には強そうです。失敗した時はスマホ上でログまで追えますか、それともあとでMac側で確認する感じですか？

Codex Mobile とワイヤレスデバッグを掛け合わせると、トイレでも Android 開発できるようになってアド 前はノーパソごと持ち込んでたがその必要なくなった

用户自有 Android 接入，是云手机从封闭资源池走向开放平台的关键一步。 平台不只管理自己的云端设备，也要能把团队已有真机接入同一套连接、音视频、控制和状态管理层。 边界也要清楚：适合授权访问、测试复现、团队交接，不适合规避规则或无人值守滥用。

@mohitban47 Long-horizon GUI tasks often fail when the agent misreads what changed a few steps back. In HiViG, is the history mostly screenshots, or a structured action/state trace?

🚨 Check out HiViG, a History-aware and Visually Grounded test-time intervention framework that guides Computer Use Agents to solve long-horizon GUI tasks across web 🌐, mobile 📱, and desktop 🖥️ GUI environments! We equip agents with two key abilities: (1) history state tracking that iteratively translates visual state changes into a compact macro-action history to track long-term goal progress and for better history-aware action generation and (2) visually grounded error analysis that evaluates proposed actions against the screenshot to refine flawed actions before execution, by generating a multi-stage rationale for error analysis that leverages a visual marker and the extracted state-transitions. Our comprehensive experiments across diverse GUI benchmarks (WebArenaLitev2, AndroidLab, WindowsAgentArena) show that HiViG boosts avg. success rate by 5.8-9.0% across both open- and closed-source models (Qwen3-VL-32B-Thinking, Gemini-3-Flash), showing cross-platform and model-agnostic generalization (whereas baseline test-time interventions often doesn't show gain for strong policy models). 🧵👇

@alive_ Obsidian is a good stress test because the files are small but interlinked. Before Codex edits the vault, do you take a git checkpoint, or rely on Obsidian Sync/version history for rollback?

goated setup to get Codex to read/write your Obsidian vault on your phone: 1. Get a cloud machine. I use GCP. 2. Add it to Codex: Connections => SSH 3. Clone your obsidian vault to the machine 4. Ask Codex to setup headless Obsidian sync That's it. You can now use Codex Mobile to do anything on your vault, and the changes will be reflected immediately on your phone's Obsidian.

@koltregaskes A useful first version might be less "control my phone" and more "show Codex the live device state": foreground app, screen text, notifications, and safe actions. Are you thinking daily phone management, or debugging your own Android apps?

When are we getting Codex for Android so I can control and manage my phone, OpenAI?

蜂壳云 iOS 端不到 10M，不是因为功能少。 更准确地说，是客户端只做连接、显示、触控和会话控制；真正重的 Android 运行、音视频传输、控制路由、设备分配和会话状态放在云端。 云手机合理的分工应该是：客户端轻，云端重，设备能力平台化。

@zhufengme 同意，这阶段先别把供应链实时性包装成 AI 能力。更该先标清哪些节点必须转人工，避免继续承诺不可控的库存和配送。

@phonescloud99 这不是这个阶段需要着急解决的问题

@dbreunig That is a pretty good mobile-session test: large refactor, old tests, then a real PR review. From the phone, did you inspect the failing test output yourself, or mostly let Fable and Copilot converge on the patch?

Tried Fable in a Claude Code mobile session (on vacation). It refactored a large app and updated some rotted tests. Did well, less guidance needed than Opus. But! GitHub Copilot(!) made three good comments on PR review, which Fable accepted. Make of that what you will.