@termiq/cli 3.47.1 → 3.48.0

package/README.md

@@ -8,7 +8,7 @@ End-to-end encrypted — the TermIQ server is a blind relay that streams your ou
 
 The headless half of a long-running task is exactly where you can't sit and watch — a remote box over SSH, a CI runner, an unattended host. Start the work, leave, and let TermIQ tell you when it needs you:
 
-- **Push the moment it matters** — get notified when a task **completes**, is **waiting** for a confirmation, or hits an **error**, even with your phone in your pocket.
+- **Push the moment it matters** — get notified when a task **completes**, is **waiting** for a confirmation, or an AI agent **fails**, even with your phone in your pocket.
 - **Answer from the notification** — approve a `Y/n` prompt without opening anything; the answer goes straight to the terminal.
 - **Reliable input** — what you type from your phone reaches the terminal in order, even on a flaky mobile connection.
 - **Real-time sync** — output streams to your phone with sub-second latency and reconnects on its own when the network changes.
@@ -69,7 +69,7 @@ Check the connection with `termiq status`; sign out with `termiq logout`.
 | `--log-to-file`   | Write logs to `~/.termiq/logs/cli.log`                              |
 | `--force`         | Start even when already inside a TermIQ session (nested run)        |
 | `--multi-viewer`  | Latest-wins geometry: the pty follows whoever typed or resized last, and other viewers mirror it. Off by default (the size is the smallest viewer's). Persists in `config.json` |
-| `--push-prompt-text` | Put the prompt's text in push notifications. Off by default — see [Notifications and privacy](#notifications-and-privacy). Persists in `config.json` |
+| `--push-prompt-text` | No effect, still accepted so existing scripts keep working: notifications never carry the prompt's text — see [Notifications and privacy](#notifications-and-privacy) |
 | `--no-signal-waiting` | Ignore the agent's own "waiting" signal. On by default: Codex and Gemini announce that they are waiting in the window title, so you still get notified even when TermIQ's own parsing misses the prompt (you get a plain input box rather than buttons). Persists in `config.json` |
 | `--version`, `-V` | Print version and exit                                              |
 
@@ -84,11 +84,32 @@ terminal is waiting — not what it asked. The question and its buttons reach
 your phone over the encrypted channel and are readable only by your own
 devices; the server relays ciphertext and cannot open it.
 
-If you would rather read the question straight from the notification shade,
-`--push-prompt-text` sends that text over the plain event channel instead. It
-is off unless you turn it on, because turning it on means the server sees the
-question and the button labels in order to put them in the notification body.
-Everything inside the app works the same either way.
+Notifications never carry the question, whatever the settings: the server puts
+no terminal text into them. `--push-prompt-text`, which used to put the question
+in the notification body, no longer has any effect; the CLI still accepts it so
+that existing scripts keep working.
+
+### When an agent fails
+
+You get a notification when an AI agent (Claude Code, Codex, Gemini CLI,
+Copilot, Cursor Agent, Aider and others) exits with an error, and when the
+shell itself is killed, by the out-of-memory killer for example. Stopping an
+agent with `Ctrl+C`, a clean exit, a failing command that is not an agent (a
+build, `grep`, `git`) and the shell staying open after the agent are not
+failures. The notification says only that the task failed, never the error. One
+failure gives one notification, and a terminal gives at most one a minute.
+
+Only the shell knows how a command inside it ended, so `termiq` starts bash and
+zsh with a small hook that reports the exit code and the program name of each
+command to the CLI, in a sequence your terminal ignores. Each hook signs its
+reports with a random value only its shell knows, so text that merely passes
+through the terminal, a file you `cat` or the output of a remote host, cannot
+fake a failed agent. The hook only adds itself: your `~/.bashrc` or `.zshrc`
+load as usual and are never changed, and its file lives in a private temporary
+directory while the shell runs. Other shells, and every shell on Windows, start
+as before; there a failed agent is noticed only if your own shell integration
+reports command ends (OSC 133). Set `TERMIQ_SHELL_INTEGRATION=0` to start bash
+and zsh without the hook.
 
 ### When the buttons don't appear