PTY intelligence

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

elanous can work a terminal the way a person does: it reads the screen, types, presses keys, clicks with the mouse, waits, and decides what to do next. It uses this to drive a plain shell, other coding agents such as Codex and Claude Code, and another copy of elanous itself — and you can watch all of it live in the web app.

It does not always reach for a coding agent. For each mission it measures the machine first and starts from the lowest rung that can do the job.

Everything runs on your machine, as your user, with your own subscriptions. elanous never types a password and never reads cookies or credential files.

Three directions

Direction	What happens	Example
elanous → agent	elanous opens a terminal (PTY), starts a shell or an agent CLI, and drives it: typing, keys, mouse, reading the screen, deciding when it is done.	Start Codex, answer its trust prompt, paste the mission, wait for the diff.
agent → file → next agent	One agent leaves its result somewhere — a diff, a summary, a file — and that becomes the input of elanous or of another agent.	Codex writes the change; a second agent reviews the diff file; elanous gates it.
agent → elanous	An agent calls elanous through its official integration (MCP), for example to start a harness run, read logs or control a terminal.	Codex or Claude Code calls elanous tools from inside its own session.

The three combine: elanous drives one agent, hands its output to another, and that agent can call back into elanous to run the gated harness and open the PR.

The tool ladder

Rung	Tool	When
0	Environment check	Always first — OS, shell, package managers, which CLIs are installed, whether agents are signed in (yes/no only).
1	Shell and text tools	One or two lines of jq, awk, sed, rg, curl, ffmpeg.
2	Domain CLIs	git, gh, docker, kubectl, elanous …
3	elanous capabilities (MCP/ACP)	Memory, logs, notes, skills, terminal control.
4	elanous harness	A code change that needs gates, review and a PR.
5	Coding agents	Large builds or a second pair of eyes — Codex, Claude Code.

See where a mission would start, without running anything:

elanous env profile            # what this machine has (read only)
elanous agent-mission plan "sum the amount column of this CSV as JSON"
# Starting rung: 1 · tool: jq

If a tool the rung needs is missing, elanous installs it by typing the package-manager line in the shell and then checks it with the smallest real job (not --version). A line that needs sudo is handed to you instead.

Research before it acts

When a mission needs outside knowledge — an unfamiliar CLI, a library's docs, an error message — elanous searches several engines at once and shows the fan-out as it happens:

elanous research "Miller mlr convert csv to json" --limit 3
elanous research "…" --json      # results with per-engine timing and hit counts

Research results are treated as data. elanous never runs a command it found on a web page as-is.

Mouse

Some terminal apps turn on mouse mode. elanous can click and scroll in them:

elanous pty mouse <ref> <column> <row>           # left click (1-based)
elanous pty mouse <ref> 10 5 --button right
elanous pty mouse <ref> 1 1 --scroll down

If the app has not turned mouse mode on, the click is refused rather than sent as stray characters.

Drive a coding agent on a mission

elanous agent-mission mission --backend codex "add retries to the payment client"
elanous agent-mission mission --backend claude --evidence test --test-path src/pay.test.ts "…"
elanous agent-mission mission --backend codex --plugin elanous-basics@elanous "make this PDF folder searchable"
elanous agent-mission mission --chain codex,claude,elanous "…"   # build → review → evidence and pull request

elanous opens the agent in a PTY inside a separate git worktree, drives it to the end, and closes the mission only when the evidence gate passes (doc, tsc or test). Claude Code is driven through its terminal by default; only the opt-in agent-mission mission --headless flag runs Claude's non-interactive mode instead. When an agent first opens a folder and asks whether to trust it, elanous answers that screen before it types the mission.

• --plugin <name>@<market> — elanous opens the agent's own plugin screen (/plugins in Codex, /plugin install in Claude Code), installs the plugin from that marketplace, and then runs the mission with its skill. The official marketplace is elanous (see Plugins).
• --chain codex,claude,elanous (or codex,claude,codex,elanous) — Codex builds, Claude Code reviews, (Codex fixes,) and elanous checks the evidence and opens the pull request. Other orders are refused, and --chain cannot be combined with --headless or --no-commit. The reviewer gets the builder's diff and the next agent gets a short summary — never the raw screen.
• --backend also accepts gemini, grok and aside — see Recommended programs.

Watch it

Every terminal elanous drives appears in the web app's Terminal menu, with the agent's name and a strip showing what elanous just did and why. Several terminals can sit side by side. See Terminal and Live.

Sign-in

When an agent shows a sign-in screen, elanous stops and sends you the link or device code — it does not sign in for you.

If you have signed in once in a browser that elanous can drive, you can approve a waiting sign-in by hand:

elanous agent-mission browser-login --pty <ref> --provider codex   # or claude
elanous agent-mission browser-login --pty <ref> --provider codex --dry-run   # only look

It approves only a plain "authorize" or "enter code" page. Any password, two-step, captcha or passkey prompt goes back to you. Missions do not do this on their own yet.

Status

Piece	Status
Environment check, tool ladder, research, mouse, installing missing tools	Available
Driving Codex / Claude Code / elanous in a PTY	Available (Claude Code needs you signed in)
Handing results from one agent to the next in one mission (--chain)	Available
Installing a plugin inside the agent's own screen (--plugin)	Available for Codex (recorded end to end) · Claude Code uses /plugin install
Browser-assisted sign-in	Manual (agent-mission browser-login) · automatic use during a mission is in progress