翻訳待ち：Rysh – an AI harness where Claude and Codex agents work as a team (Go)

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Uh oh! There was an error while loading. Please reload this page. Notifications You must be signed in to change notification settings Fork 0 Star 2 BranchesTags Open more actions menu Latest commit History 29 Commits 29 Commits Folders and files NameName Last commit message Last commit date .github/workflows .github/workflows action action cmd cmd docs docs evals evals examples examples internal internal packages packages packaging packaging scripts scripts .gitignore .gitignore .golangci.yml .golangci.yml .goreleaser.yml .goreleaser.yml CONTRIBUTING.md CONTRIBUTING.md Dockerfile Dockerfile LICENSE LICENSE Makefile Makefile NOTICE NOTICE README.md README.md SECURITY.md SECURITY.md flake.nix flake.nix go.mod go.mod go.sum go.sum rysh.config.yaml.example rysh.config.yaml.example system-prompt.md system-prompt.md Repository files navigation The Rysh CLI: an agentic terminal multiplexer written in Go. Tabs, panes, splits, and vim/htop working exactly as you expect — except every pane is also an agent that can answer prompts and call tools. This repository holds the CLI itself. It depends on rysh-cli-shared. Building or contributing? Start at rysh-cli-parent — it carries the Makefile, the Go workspace, and CI, and wires this module to a local checkout of the shared one. Video tutorials Five narrated, subtitled walkthroughs of what rysh actually does, in order — each one assumes the previous. Every player below is live on this page; press play, nothing leaves GitHub. The first four are recorded against the build the install line gives you, so every command in them is one you can type. 1 — Install, and your first tabs, lanes and stacked panes · 1:29 01-first-panes.mp4 go install, a first session, a tab, a second lane, a stack of three panes — then ##claude in each pane of that stack, so one window holds three independent Claude sessions. 2 — Stop the session, start it again — the agents are still there · 2:44 02-stop-start.mp4 Why a session is a daemon and not a terminal. Claude and Codex are each told a codeword, the session is stopped outright, and after the restart both resume unprompted and answer with the codeword they were told before it. 3 — Claude and Codex in three stacked panes at once · 2:07 03-claude-codex-parallel.mp4 More than one agent, and more than one vendor, side by side: Ctrl+S and a digit to move around a stack, a task each, three files written in parallel. 4 — A fleet that talks to itself, on a shared board · 2:29 04-fleet-board.mp4 ##board open, rysh ansa prompt, and agents driving each other: roadmap (Claude) sets the goal, fleet-manager (Codex) splits it into work orders, two workers build in parallel, and the manager sends a correction back to a worker before signing off. They ship a browser todo manager — no backend, no build step. 5 — Graph engineering — designing the shape of a fleet · 4:00 05-graph-engineering.mp4 The deep end: why an agent org chart is a graph, what the edges mean, and what changes when you remove one. Orders travel down as messages, results come up on the board, and two workers with no edge between them give independent verdicts. Start at 1 even if you have used tmux or Zellij for years — the panes look familiar and the session model underneath them does not, which is video 2. Secrets stay on your machine SecretNAT is on by default. Secrets are substituted with tokens in the request body before it leaves the machine, and a response carrying a live credential in plaintext is reported into the pane. Responses are not rewritten — by design. The mapping is reversible locally, so ##snat get hands you back the real value in your own pane, and the model only ever saw the token. Install Two binaries exist and they are not interchangeable: this repository the prebuilt distribution binary name rysh ry how you get it go install, or the installer below its own channels, not documented here The names differ so that one machine can carry both. Every install command in this README installs rysh, and a script, alias or doc written for one does not run against the other — which is the whole reason the names differ. Check what you have with rysh version or ry version. This repository is the rysh half: Apache-2.0, with LICENSE and NOTICE beside the source you are reading. macOS and Linux The open-source build, from source: go install github.com/rysh-ai/rysh-cli-code/cmd/rysh@latest The binary lands in $(go env GOPATH)/bin as rysh. Requires Go 1.25.3 or newer — the floor declared in this module's go.mod. The open-source build, prebuilt: curl -fsSL https://packages.rysh.ai/install-rysh.sh | sh Installs rysh from this repository's own releases — the Apache-2.0 binary, not ry. It resolves the newest release at run time; set RYSH_VERSION to pin one instead, or RYSH_INSTALL_DIR to choose where it lands. The prebuilt distribution — installs ry, not rysh. It is the packaged product and a separate distribution; nothing in it is produced by this repository, so ry --help describing a command is not evidence about the source here. Use go install above if you want the build whose source you are reading. Its install commands are deliberately not reproduced here: this README documents the open-source build. The prebuilt distribution has its own channels and its own documentation. Windows WSL2 is the supported path. Native Windows compiles and runs, but it cannot open a pane — those are two different claims and both are true: GOOS=windows GOARCH=amd64 go build ./... succeeds — the windows/amd64 target builds clean from this source. This repository's .goreleaser.yml names that archive cli-only on purpose, so the artifact says what it is. It cannot start a session. Rysh's PTY layer has no ConPTY implementation, so platform.PTYSupported is false on Windows and every session-opening command is refused up front with WSL guidance — rather than starting a session and then failing on your first pane (cmd/rysh/pty_preflight.go). What it can do is the pane-less command set, including talking to a session running in WSL from the Windows side: rysh send, rysh exec, rysh prompt, rysh list-sessions, rysh install, rysh eval, rysh doctor. For a real session, install WSL2 and use the Linux instructions inside it — the Linux build runs unmodified: wsl --install # inside WSL go install github.com/rysh-ai/rysh-cli-code/cmd/rysh@latest # rysh Native panes need ConPTY behind the same seam; it is not implemented yet. Build from scratch This module alone builds standalone — it carries no replace directive, so the shared module resolves from its published version: git clone https://github.com/rysh-ai/rysh-cli-code cd rysh-cli-code go build ./cmd/rysh # -> ./rysh From the superrepo, which is how to develop against both modules at once — its go.work wires the shared module to the sibling checkout, so an edit there is picked up without a release round-trip: git clone --recursive https://github.com/rysh-ai/rysh-cli-parent cd rysh-cli-parent make build # -> bin/rysh make install # -> ~/.local/bin/rysh (override with PREFIX=) make test # both modules Go 1.25.3 or newer either way. First run export ANTHROPIC_API_KEY=sk-ant-... rysh onboard --provider anthropic --key-env ANTHROPIC_API_KEY rysh doctor onboard validates the key, writes a project-local rysh.config.yaml, and opens your session. rysh --help lists the full command surface. Each pane has an input mode and Esc Esc cycles it: shell (a real PTY) → prompt (goes to the LLM) → rysh (multiplexer commands) → chat (conversation, no tools). Those four are enabled by default; ##mode list shows what a pane has and ##mode new adds the rest. The parent repo README has the keybindings, sessions, agents and humanoids. Three tours Commands starting with ## are typed into a pane in rysh mode. From outside the session, rysh exec -- '##' runs the same thing and prints its output. 1. Create a session rysh create work # create and attach; add -d to leave it detached rysh list-sessions # what is running rysh attach work # come back to it later rysh detach work # leave it running rysh stop work # shut the daemon down A session is a daemon plus its panes. It outlives your terminal, so closing the window detaches rather than kills. 2. Tabs, lanes, panes and stacks A tab holds lanes (columns); a lane holds stacks; a stack holds panes, Zellij-style — one expanded, the rest collapsed to a title bar showing [N/M]. ##new tab a new tab ##new lane a new lane (column) in the active tab ##new pane a pane at the bottom of the active lane ##new grid 4 4 panes stacked in the active lane ##new grid 3x4 3 lanes x 4 panes, in the active tab ##new grid 2x3x4 2 tabs x 3 lanes x 4 panes ##new stack 4 4 more stacked panes in the active stack Look at what you built, and move things around: ##tab list ##lane list ##pane list ##panegroup layout the lane layout, whole ##move pane up|down|left|right reorder in the stack, or cross to the next lane ##move pane to-lane put a live pane somewhere else ##tab name build ##lane name left ##pane name builder A stack with an agent in every pane is video 3 of the tutorials at the top of this page. 3. The agents board, fleets, and a fleet's own board Panes can run an interactive Claude or Codex, and rysh remembers which, so a session stop/start resumes the same conversation: ##claude run claude in this pane, resumed automatically ##codex run codex in this pane, same deal ##pane new --claude "start on the parser" a new pane, already working The agents board is where those panes talk to each other and to you: ##board open open the board pane ##board post post a milestone ##board reply An agent inside a pane posts without stealing focus, and reads back: rysh board post --as "$RYSH_PANE" -- 'parser wiring done' rysh board tail --limit 20 ##board agent up starts a board agent — a hidden pane running Claude that sits in the path, routes messages between agents, and refuses a request it knows is stale. ##board agent visible draws it without moving your focus; ##board agent invisible puts it away while it keeps running. A fleet is a named group of panes with a board of its own. Registering one opens no panes — it is the cheap half: ##fleet register epic-07 --board epic-07 ##fleet state epic-07 up ##fleet list ##fleet show epic-07 ##fleet forget epic-07 drop it; the panes keep running rysh board tail --fleet epic-07 # that fleet's stream, not the session's A fleet may be claudes, codexes, or both — the commands are the same either way. Layout Path What cmd/rysh the main package — entry point and command surface internal/tui the terminal UI internal/actors workspace / tab / pane / agent actors, proto.actor over NATS internal/vterm terminal emulation, including a vt10x fork with scrollback internal/provider LLM provider adapters internal/platform host capabilities that change what rysh can do (e.g. PTY support) action/ the setup-rysh composite GitHub Action Contributing and security This repository is a one-way export of a tree developed elsewhere, so a commit pushed straight here is overwritten by the next export rather than kept. That does not make patches pointless — CONTRIBUTING.md explains where one actually lands. Found a vulnerability? Do not open a public issue, PR or discussion. Rysh runs shell commands and holds provider credentials on a developer's machine, so a public report is an exploit notice. SECURITY.md has the private channels. License Apache-2.0. See LICENSE and NOTICE. Resources Readme Apache-2.0 license Contributing Contributing Security policy Security policy Activity Custom properties Stars 2 stars Watchers 0 watching Forks 0 forks Report repository