osFoundry: screen rules beat byte-tail for Claude Code prompts

Claude Code 2.1.303 builds its permission prompt by jumping the cursor to fixed columns with CSI sequences like \x1b[5G rather than writing spaces, and it sends no newline bytes at all — all 11 saved captures (2,099 to 7,672 bytes) contain zero. That design defeats the obvious shortcut of grepping the end of the output. A Python port of osFoundry's rules scored the last 8 KB of each capture after stripping escape sequences and keeping text after the final carriage return, and produced idle for every case, including the four that were truly blocked; the apparent prompt text also ran together into 'Doyouwanttoproceed?'. Feeding the same bytes into a minimal screen emulator that handles only characters, \r \n \b, cursor movement and erase recovered all 11 states correctly.

The osFoundry desktop app follows that conclusion. In terminal_pty.rs each session holds a vt100::Parser, and detection uses only screen_text() from the emulated screen. The 200-line scrollback above it is deliberately excluded, because a prompt that has already been answered and scrolled up would otherwise keep the pane marked blocked forever.

On a live setup with a stand-in agent that copies Claude Code's wording and redraw habits, a 700 ms sweep of all panes cost a median 0.26 s and up to 0.70 s of extra delay, with up to one cycle of lingering state — fine for a badge a person reads. The project's guidance for anyone building their own: classify the emulated screen, match prompt-specific text and footers, put blocked rules above working rules, keep the program that launches agents from editing those rules, and treat the process table as a way to see who is running, not who is waiting.