Three related improvements to workflow inside ac-native: 1. Persistent PTY sessions (keep claude alive on piece-leave) terminal.leave() no longer calls system.pty.kill(). The PTY child lives as long as ac-native is running, so bouncing out to the prompt and back via `code` re-attaches to the same claude session instead of cold-starting the CLI every time. Matters because claude's initial project scan + OAuth handshake + model handshake takes several seconds, and losing it on every piece swap makes "check something quickly in the prompt then come back" expensive. terminal.boot() checks system.pty.active && lastCmd match on entry; if so it just resizes + clears the grid cache and returns (no fresh spawn). Fresh spawn only on first entry or after the PTY child has exited on its own. 2. Plain Escape exits the terminal piece (no kill, no send-to-PTY) Was passed through as \x1b to claude, which cancels the current operation there but still left the user stuck in the terminal. Now esc jumps to prompt directly, PTY stays warm for next visit. Users who need to send a real ESC byte to claude (e.g. to cancel a prompt without leaving) can use Ctrl+[ which produces \x1b through the existing Ctrl+key handler. 3. Git identity + credentials derived from /mnt/config.json pty.c used to hardcode "Jeffrey Alan Scudder" + "whistlegraph" for git author + GitHub HTTPS credential helper. Now it reads the baked config.json handle + email and uses those, falling back to a generic ac-native / noreply identity only if the config is unreadable. Paired with a `compush` git alias (add-all + commit -m + push origin HEAD) so the on-device workflow matches the repo's CLAUDE.md notation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>