mirror of
https://github.com/l0ng-ai/tty7.git
synced 2026-09-21 16:02:20 +00:00
l0ng-ai
dc039e1c65
The hooks tty7 installs into Claude Code already announce every turn over the pty as an OSC 777, and the daemon reads those for the pane's status dot. The same bytes reach the client, where they are worth something else: the byte offset a `prompt-submit` lands on is a *position in the stream*, so advancing the emulator to exactly there and reading the cursor gives the scrollback row that turn began on. That is an outline of the conversation, and a way back into it — which is the one thing a long agent session in a terminal has never had. The Info panel grows a CONVERSATION section: one row per turn, the prompt's first line as its label, a dot that says whether the turn is still running. Clicking a row scrolls the pane so that turn's prompt is the top line. Not a fourth right-panel tab. `RightPanelTab` says out loud why there is no room for one at 260px, and a fourth variant would drop anyone who rolled back to an older build onto Info. This is a fact about the pane, like its shell and its cwd, so it sits with them. ## Why the anchor is not the whole answer The hook is a subprocess writing to the controlling tty while the agent's own renderer writes to it too. Claude Code repaints in place with ink, so the cursor when the hook's bytes land is wherever the last repaint left it — inside the live region, a few rows from where the prompt's echo comes to rest. And once the scrollback limit starts discarding lines, every anchor slides by the discard count at once. So the anchor is a hint, and the prompt's own text is the correction: at click time (by which point it has long been drawn) the row is looked for around the anchor, exact match first — the row that *is* `> hi`, marker stripped — and only then by containment, which keeps its length floor because `hi` appears inside half the rows of any answer. The row that is found is written back, so a second click does not search again and cannot land somewhere else. ## Why rows and not the transcript file Claude Code keeps a JSONL transcript, and reading it would give the assistant's side too. It would also only work for Claude, only when the agent runs on this machine, and only for a path this process may read. An OSC comes back through the pty from wherever the agent actually runs — over ssh, in a container, in a remote workspace — with no file access and no per-agent format. What is lost is the assistant's text; what is kept is every host tty7 supports. ## Mechanics - `OscTokenizer::feed_at` reports each payload's end offset. The client already tokenized OSC 777 on every batch to keep agent events out of desktop notifications, so the scan is free; only a real event now costs a cut, which is what #404 was right to object to about the old per-command mark scanner. - `Cut` is a two-variant enum again (cursor repair, agent turn). Two ascending runs concatenated are not one, so a batch carrying both kinds is sorted — and only such a batch pays for it. - A replayed ring is cut the same way, so reattaching to a pane rebuilds the outline from its own history rather than losing it with the old client. - Turn anchors are dropped where image placements are: `clear_scrollback`, and the grid reset in `adopt_relink`. - A turn that began on the alt screen is listed but not clickable — there is no scrollback behind it to return to. - A turn announced twice is one turn. Hooks are not guaranteed to fire once, and what makes it the same turn is that the one before it never ended: a real repeat can only come after an answer, and an answer brings a `stop`. - The hook forwards the prompt's first line, clamped to 200 characters. The tokenizer *abandons* a payload past 8 KiB rather than truncating it, so a pasted file would otherwise cost the whole event; and a needle spanning a line break matches no single row. No protocol change: the prompt rides in the OSC the hook already sent, and an older client ignores the field.
tty7
A terminal workbench: persistent sessions, remote work, agents.
Pure Rust · GPU rendering on Zed's gpui · VT core from Alacritty
English · 简体中文
Why
- Performance — ~2× the throughput of Alacritty, Ghostty, or Kitty (benchmarks)
- Persistent sessions — quit or reboot; your shells and supported agent sessions keep running, no tmux
- Editor-grade input — suggestions, completion, highlighting, history search
- Remote development — files, repos, panes, and git data stay on the remote machine
- Native SSH — profiles, SFTP, port forwarding, and jump hosts
- Agent-aware — Claude Code, Codex & co.: status, notifications, git context
- CLI + Skills — agents create panes, run commands, and inspect output
Install
Native builds for each platform on Releases:
| macOS | …-macos-arm64.dmg · …-x86_64.dmg |
drag into Applications |
| Windows | …-setup.exe · portable ….zip |
|
| Linux | …-x86_64.AppImage |
chmod +x and run — X11/Wayland libraries bundled |
What's inside
| Editor-grade input | ghost suggestions from history · explained tab completion · syntax highlighting · multi-line editing · click places the caret · ⌃ R fuzzy history |
| Window | tabs & splits · ⌘ P palette · ⌘ F scrollback search · nine themes · IME |
| Agent-aware | per-pane detection (18 CLIs): status dot · notifications · branch + diff · resume after reboot · tray icon when input is needed |
| Remote workspaces | remote files, repos, changes, diffs, worktrees, tabs, and panes · reconnect from any client and continue where you left off |
| CLI + Skills | bundled tty7 CLI · agent skill · pane/workspace control · real PTY commands · output, process, port, and agent status |
| SSH | native russh stack: profiles with keychain secrets · SFTP panel · port forwarding · jump hosts · one-time, unprivileged tty7-server install |
Terminal and keybinding reference: docs/features.md. The agent-facing CLI interface is documented in skills/tty7/SKILL.md.
Install the skill with:
npx skills add l0ng-ai/tty7
Benchmarks
Same machine, same day, same 155×40 grid — Apple M1 Pro, macOS 26.3.1, five-run averages (2026-07-04):
| tty7 | Alacritty | Ghostty | Kitty | |
|---|---|---|---|---|
Plaintext I/O — 11 MB cat (lower = better) |
95 ms | 239 ms | 179 ms | 185 ms |
| DOOM-fire frame rate (higher = better) | 888 fps | 485 fps | 552 fps | 617 fps |
| Cold-launch memory | 116 MB¹ | 105 MB | 128 MB | 130 MB |
¹ GUI 105 MB + the persistent server 11 MB.
Methodology and one-command reproduction: scripts/bench/.
Built on gpui and alacritty_terminal · Apache-2.0 · Discord · Changelog