Every Codex session, at a glance.

Yes, with Perch Pro — the subscription bought in the iPhone app; the perpetual $29 licence does not include it. Codex was the second agent to get it. You can type a full message, multi-line included, straight from the iPhone app with your Mac locked, and stop a turn that's going wrong. In Codex's default mode Perch reaches the session through Codex's own background service, so it works whichever terminal Codex is in, and Stop works even while Codex is writing its answer. A session started with --no-daemon needs to be running in tmux, herdr or Rex, and there Stop does nothing in the few seconds while Codex writes out its final answer — Perch cannot tell that apart from the rest of the turn, so a Stop tapped right at the end may not land.

Because Codex won't run them until you say so. Codex requires every non-managed hook to be reviewed and trusted by hash — run /hooks inside Codex and approve them. Until you do, the hooks are installed but inert. Perch asks Codex directly which hooks it will actually execute, so it can tell you this instead of just looking broken.

The rollout scan, which is most of the panel: sessions, model, context, cost, usage bars and Stats all keep working. What you lose is the live half — instant state changes, the “needs you” alert, and Allow/Deny from a notification.

No. Perch writes its hooks in and chains onto anything already there, so an existing setup keeps firing. Removing the integration from Settings takes them back out. The one step Perch can't do for you is the trust approval — that's Codex's, by design.

Yes, and Perch labels which is which on every row. Codex Desktop gets the better jump of the two: clicking a session opens that exact conversation in the app, rather than merely raising it.

No — for Codex it's an estimate derived from token counts, and Perch presents it as one. It is priced against OpenAI's published prices, each part of a session at the model that did it, and when some of the work ran on a model with no published price the figure ends in “+”. Rate-limit windows, your plan label and your credit balance are read values, not estimates.

Because replies and Stop need proof that Perch's hooks run in that session. At a Codex approval prompt a typed reply approves the command and Escape denies it, so Perch offers neither until Codex confirms that Perch's approval hook is installed, switched on and trusted — usually about a second — and, once the session has run a turn, until that session has sent a hook event. A session started with --disable hooks, -c features.hooks=false or a profile that turns hooks off says so on its row and gets no replies or Stop. So does a project whose own .codex/config.toml turns hooks off.

Yes. Since Codex 0.157, sessions run in Codex's background service by default. Perch keeps them on the list for as long as they're open, and sends your phone's replies and Stop through that service, so they need no tmux or herdr. Starting Codex with --no-daemon runs it as an ordinary terminal session instead.

Yes, both. Perch reads OpenAI's official status page, and when the part Codex uses is down, a red banner in the dropdown says so and links to the incident; a slowdown or maintenance shows as a small icon beside Codex's usage. Settings → General → Integrations shows your Codex version, and when a newer one is out on the channel you installed it from, how to update there. Perch never installs anything itself. Both checks are anonymous reads of public pages, and each can be turned off. Seeing an error and no banner? Is Codex down, or is it you?

No. Session state, rollouts and usage stay local. Data leaves only for features that require it: your paired iPhone (encrypted, over your LAN or your own tailnet), the push relay, and Polar for license checks.