Code agents

Push supports exactly four code agents today: Claude Code, OpenAI Codex, OpenCode, and Hermes. Any other CLI runs fine as a plain terminal — it just doesn't get the agent affordances below.

It's your agent: your install, your login, your model choice, your bill. Push launches the runtime's own CLI in a managed tab and never proxies, meters, or marks up its traffic — when your agent talks to its provider, that's between you and them.

What "supported" means

• Naming — tabs title themselves from what the agent is doing.
• Resume — quit Push and come back; the session continues via the runtime's own resume mechanism:

• History — Push reads the runtime's own local session store (Claude Code's files in ~/.claude, Hermes' database in ~/.hermes, and so on) to render your session history. Locally. It never uploads them.

Claude Code is the deepest integration — auto-resume, /clear handling, and theme-respect when your Mac switches appearance. The other three are first-class for naming, resume, and history.

What Push adds around your agent — the honest list

Push is transparent to your agent's work, but not invisible. Here is everything it does, all of it observable:

1. Pins the tab to a stable session identity in the agent's environment, so resume works.
2. Adds a short note to the agent's context telling it it's running inside Push and can use the push CLI to track sessions, file issues, and report back. It's a capability note, not steering — your prompts and the agent's behavior are otherwise untouched. (Hermes is the exception: it gets no context note at all; Push only observes its lifecycle.)
3. Installs and maintains the push skill in each agent, so asking your agent to "track this session" just works. You'll see Push type a visible /p first turn in a fresh tab ($p on Codex; /push is the same thing) — that's the skill announcing itself, in the open.
4. Reads, never writes, the runtime's session store for history.

Nothing is hidden in scrollback you can't see, and none of it phones home — the data page covers what does.

Bypass Permissions

Each runtime row in Settings → Terminal has a Bypass Permission switch for unattended work. The confirmation states its exact scope: "Push will start future X terminal sessions in unattended mode. Existing sessions and agent configuration files are unchanged." The same control appears per-tab as Bypass Permissions in the tab menu. Use it deliberately — it's your agent's own permission model being told yes in advance.

Keeping runtimes current

Automatically update code agent runtimes (Settings → Terminal) runs each installed runtime's own updater in the background while that runtime is idle — or leave it off and update your CLIs yourself.