aiterm-mcp 0.7.1 → 0.8.0

package/README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <img src=".github/og.svg" alt="aiterm-mcp — AI holds one persistent terminal as a stdio MCP server (tmux-backed)" width="100%">
+  <img src=".github/og.svg" alt="aiterm-mcp — one persistent MCP terminal your AI drives, and launches other coding agents (Codex/Grok/Composer) into (tmux-backed stdio MCP server)" width="100%">
 </p>
 
 # aiterm-mcp
@@ -12,65 +12,108 @@
 
 > *(日本語: [README.ja.md](README.ja.md))*
 
-> **Let Claude — or any MCP client — drive a real, persistent shell.** One terminal stays open; `ssh`, `docker exec`, a REPL are just text you send into it, so the AI stops reconnecting for every single command. Reads come back token-reduced.
+> **Let your AI orchestrate other AIs.** From Claude Code — or any MCP client — one call spawns a coding agent (Codex, Grok, or Composer) inside a persistent terminal and hands you a session to drive: read what it's doing token-reduced, send it the next instruction.
+>
+> **What it is:** one persistent MCP terminal your AI drives — and can launch other coding agents into. `ssh`, `docker exec`, a REPL, or another agent's TUI all nest inside that one terminal as just text you send in. The mechanism is deliberately plain — your MCP client drives the other agent's terminal turn by turn: no hidden protocol, no shared memory, no autonomous negotiation.
+>
+> **No human at a tmux required.** aiterm is driven programmatically over MCP, so an AI can launch and drive another agent with no one sitting in the terminal — from an orchestration loop, a CI step, or a cron job.
 >
 > *MCP = Model Context Protocol — the open standard that lets tools like Claude Code plug capabilities into an AI.*
 
-Six PTY tools — `pty_open` / `pty_send` / `pty_read` / `pty_key` / `pty_close` / `pty_list` — plus three interactive **agent launchers** (`codex_agent` / `grok_agent` / `composer_agent`) that start a coding-agent TUI inside that same persistent terminal. The backend is **tmux**, so sessions survive even if the MCP server or the AI client restarts.
+Nine tools: six **PTY tools** — `pty_open` / `pty_send` / `pty_read` / `pty_key` / `pty_close` / `pty_list` — to open, drive, and read one persistent terminal, plus three **agent launchers** — `codex_agent` / `grok_agent` / `composer_agent` — that each start another coding agent's TUI inside a fresh one. The backend is **tmux**, so sessions survive even if the MCP server or the AI client restarts.
+
+**Status:** actively maintained · the newcomer here, betting on a different shape (see [vs. the alternatives](#vs-the-alternatives)) · runs on Linux · WSL2 · macOS · native Windows · MIT · see the [CHANGELOG](CHANGELOG.md).
+
+## Why now
 
-**Status:** actively maintained · runs on Linux · WSL2 · macOS · native Windows · MIT · see the [CHANGELOG](CHANGELOG.md).
+A lot of 2026's agent tooling is converging on orchestration: a lead model delegating a mechanical refactor to Codex, running Composer on a bulk edit while it reviews the diff, fanning one task across several agents to spare its own context window. All of those agents already live in a terminal. aiterm makes that terminal a first-class, MCP-native tool — so the model doing the orchestrating can **spawn and steer the others without a human wiring up panes.**
 
-## Why
+## Two ways to use it
 
-Send an AI **one command at a time** and, over SSH, every command becomes its own `connect → authenticate → disconnect`. That stings three ways: you **re-authenticate every time** (passphrase, one-time code, all of it), **short-lived sessions pile up**, and once connections come too fast your own defenses lock you out — `fail2ban` bans you, `MaxStartups`/`MaxSessions` reject you, the account gets locked. The security meant to stop attackers ends up stopping you. (Yes — this bit me on my own box. I built aiterm to drive my homelab from Claude Code without that re-auth hell.)
+### 1. Drive SSH, containers, and REPLs in one persistent terminal — the primitive
 
-aiterm **holds one PTY persistently** and you `ssh host` (or `docker exec -it x bash`) *inside it*, **once**. Every command after that rides the same already-authenticated session: **authenticate once, one session, nothing for the defenses to trip on.** Session kind is never a tool-level distinction.
+This is the base, and it works with just tmux — no other CLI. `pty_open` grabs one local terminal; `ssh host`, `docker exec -it x bash`, or a REPL are just text you `pty_send` into it — **once**. Every command after that rides the same already-authenticated session. Session kind is never a tool-level distinction.
 
 ```
 pty_open()                         → grab one local terminal
 pty_send(id, "ssh 192.168.1.2")    → authenticate once, inside that terminal
 pty_send(id, "uname -a")           → every later command rides the SAME session
-pty_read(id, { wait: true })       → read the reduced output
+pty_read(id, { wait: true })       → read the token-reduced output, completion detected
 ```
 
+<sub>**Origin.** I built aiterm for exactly this. Driving my homelab from Claude Code one command at a time meant every SSH command became its own `connect → authenticate → disconnect`: re-typing the passphrase and one-time code each time, short-lived sessions piling up, and eventually my own defenses (`fail2ban`, `MaxStartups`/`MaxSessions`, account lockout) locking me out — the security meant to stop attackers ended up stopping me. Holding one authenticated session fixes all three at once. That pain is why the persistent terminal exists; launching whole other agents inside it is what it grew into.</sub>
+
+### 2. Launch other coding agents into that terminal — the orchestration flagship
+
+The same primitive hosts another agent's TUI. Three launchers each start one vendor's interactive coding-agent TUI inside a fresh persistent terminal and return a `session_id`. From there you drive it with the same `pty_read` / `pty_send` you'd use on any shell: read its output token-reduced, send it the next step. (The TUIs are full-screen apps, so `pty_read({ screen: true })` gives you the rendered view.) This one needs the vendor's own CLI installed and authenticated — see [Requirements](#requirements).
+
+```text
+codex_agent({ session_name: "codex1", cwd: "/repo",
+              prompt: "port test/legacy.py to vitest" })
+                                    → { session_id: "codex1", … }   # Codex now live in a persistent terminal
+pty_read("codex1", { screen: true })   → read what it's doing (token-reduced)
+pty_send("codex1", "also fix the imports it broke")   → steer it, mid-task — you are driving its TUI
+```
+
+One call per model, so the tool name itself tells you which model you get:
+
+| Tool | Launches | Key args |
+| --- | --- | --- |
+| `codex_agent` | Codex CLI (OpenAI; the CLI's default model) | `prompt?`, `reasoning_effort?`, `cwd?`, `session_name?` |
+| `grok_agent` | Grok Build, model `grok-build` (xAI) | `prompt?`, `reasoning_effort?` (`low`/`medium`/`high`/`xhigh`/`max`), `cwd?`, `session_name?` |
+| `composer_agent` | Grok Build, model `grok-composer-2.5-fast` (xAI) | same as `grok_agent` |
+
+The vendor CLI must be installed and authenticated (`codex` for `codex_agent`; `grok` for both Grok tools). aiterm resolves the binary via `CODEX_BIN` / `GROK_BIN`, then `~/.local/bin/codex` / `~/.grok/bin/grok`, then `PATH`. Prerequisites are checked **before** a session exists: for `grok_agent` / `composer_agent` an out-of-range `reasoning_effort` is rejected up front (the value set is fixed — `low`/`medium`/`high`/`xhigh`/`max`); a missing CLI binary or a nonexistent `cwd` fails for all three. A rejected launch — bad `reasoning_effort`, an unresolvable binary, or a nonexistent `cwd` — leaves **zero leftover session** behind, and if the launch keystroke itself can't be delivered aiterm tears the session down. (aiterm does *not* then verify the vendor CLI actually started or authenticated — `openAgent` returns once the launch command is sent; read the session to confirm the agent came up.) (`codex_agent` forwards `reasoning_effort` to the Codex CLI as a config override — `-c model_reasoning_effort=…` — rather than validating it, since Codex's accepted values vary by version.) Pass an absolute path for `cwd` — `~` is not expanded.
+
+There is deliberately **no Claude launcher**, and no protocol between the agents: aiterm's job is to let your Claude (or any MCP client) reach for *other* agents and drive their terminals. Because a launched agent is just another persistent session, everything else in this README applies to it: token-reduced reads, completion detection, and a human `attach` to watch or take over.
+
 ## Demo
 
-Real captured output from a live session — the token reduction and completion detection are genuine, not mocked. The bracketed meta line is exactly what `pty_read` appends.
+<p align="center">
+  <img src=".github/demo.gif" alt="aiterm-mcp demo: pty_open, a token-reduced grep read, then a nested Python REPL — all in one persistent session" width="100%">
+</p>
 
-A noisy `git log`, read back token-reduced (458 → 273 tokens):
+Real captured output — each block below was just run through aiterm in this repo; the numbers, the elision marker, and every `is_complete` verdict are the tool's own, not mocked. The bracketed meta line is what `pty_read` appends; its labels are Japanese in the actual output, translated here for readability (the [Japanese README](README.ja.md) shows them verbatim).
+
+A long output folded head+tail — the middle is elided by the reducer, not by me (166 → 56 tokens):
 
 ```text
-→ pty_send("demo", "git log --oneline -12")
+→ pty_send("demo", "seq 1 150")
 → pty_read("demo", { wait: true })
-← 3ce487e (HEAD -> main, origin/main) docs(readme): lead the Why with the SSH pain …
-  39a9668 (tag: v0.4.0) release: v0.4.0 — nested completion early-return …
-  c1ed87b feat(completion): early-return nested status when nested + no until …
-  … 9 more commits …
-  [aiterm demo: 13 lines / ~273 tok (raw 13 lines / ~458 tok)] [is_complete=True via quiescent]
+← 1
+  2
+  3
+  ⋮  (head runs to line 29 — abbreviated in this README)
+  … ⟨102 lines elided · full=true, or line_range="A:B"⟩ …    ← the tool's own marker
+  ⋮  (tail resumes at line 132 — abbreviated in this README)
+  149
+  150
+  [aiterm demo: 51 lines / ~56 tok (raw 152 lines / ~166 tok); 102 lines hidden] [is_complete=True via quiescent]
 ```
 
-A `grep`, folded by the per-command reducer to just the hits (127 → 46 tokens):
+A `grep`, folded by the per-command reducer to a count header plus just the hits:
 
 ```text
 → pty_send("demo", "grep -rn capture-pane src/ test/")
 → pty_read("demo", { wait: true, rtk: true })
 ← 2 matches in 1 files:
-  src/core.ts:159: // maxBuffer … capture-pane (large scrollback) …
-  src/core.ts:329: const args = ["capture-pane", "-p", "-J", "-t", name];
-  [aiterm demo: rtk:grep applied / ~46 tok (raw ~127 tok)] [is_complete=True via quiescent]
+
+  src/core.ts:159:// maxBuffer defaults to 1 MiB; capture-pane (large scrollback) … (line truncated here)
+  src/core.ts:335:const args = ["capture-pane", "-p", "-J", "-t", name];
+  [aiterm demo: rtk:grep applied / ~46 tok (raw ~53 tok)] [is_complete=True via quiescent]
 ```
 
-Nesting is just text you send in — here a Python REPL *inside* the same PTY:
+Nesting is just text you send in — here a Python REPL *inside* the same PTY (an `ssh host`, a `docker exec -it … bash`, or a launched coding-agent TUI nests exactly the same way):
 
 ```text
 → pty_send("demo", "python3")
-→ pty_read("demo", { until: ">>> " })             # nested prompt = "the inner shell is ready"
+→ pty_read("demo", { until: ">>>" })                # nested prompt = "the inner shell is ready"
 → pty_send("demo", "print(sum(range(1_000_000)))")
-→ pty_read("demo", { until: ">>> " })
-← 499999500000                                     [is_complete=True via until]
+→ pty_read("demo", { wait: true, until: ">>>" })
+← 499999500000                                      [is_complete=True via until]
 ```
 
-`ssh host` and `docker exec -it … bash` nest exactly the same way (see [Why](#why)) — an animated GIF of the full SSH flow is on the way; everything above is real output, not a script. While nested, pass `until` (the inner prompt) or `mark: true`, because quiescence cannot fire there by design — see [Completion detection](#completion-detection-4-layers) and [Known constraints](#known-constraints-by-design-not-bugs). A human can `attach` to the same tmux socket and watch any of this live (see [A human can watch](#a-human-can-watch)).
+The only edits to the captures above are the two `⋮` lines (a long head/tail run abbreviated for the README) and one over-long grep line truncated to fit — the `⟨…⟩` marker, the token counts, and every `is_complete` verdict are exactly what the tool printed. (Use `until: ">>>"` without a trailing space — the captured prompt is trimmed, so `">>> "` would miss and fall through to `timeout`.) While nested, pass `until` (the inner prompt) or `mark: true`, because quiescence cannot fire there by design — see [Completion detection](#completion-detection-5-layers) and [Known constraints](#known-constraints-by-design-not-bugs). A human can `attach` to the same tmux socket and watch any of this live (see [A human can watch](#a-human-can-watch)).
 
 ## Quickstart (≈60 seconds)
 
@@ -95,7 +138,7 @@ pty_read("t1", { wait: true })      → "hello"   (token-reduced, completion det
 pty_close("t1")                     → terminal released
 ```
 
-That's it. The terminal in `t1` is real and persistent — `ssh`, `docker exec`, a REPL are just text you `pty_send` into it.
+That's it. The terminal in `t1` is real and persistent — `ssh`, `docker exec`, a REPL, or a launched agent's TUI are just things that live inside it. To launch a worker agent instead, one call does it: `codex_agent()` returns a `session_id` you drive with the same `pty_read` / `pty_send`.
 
 **Prefer a global install, or a different client?**
 
@@ -105,37 +148,58 @@ npm i -g aiterm-mcp
 claude mcp add --scope user --transport stdio aiterm -- aiterm-mcp
 ```
 
-This registers it in `~/.claude.json`; you'll get an approval prompt the first time. **Any other MCP client** (Cursor, Cline, Claude Desktop, …) works too — just launch `npx -y aiterm-mcp` (or `aiterm-mcp`) over stdio. Needs **Node ≥ 18** and **tmux** — see [Requirements](#requirements).
+This registers it in `~/.claude.json`; you'll get an approval prompt the first time. **Any other MCP client** (Cursor, Cline, Claude Desktop, …) should work too — just launch `npx -y aiterm-mcp` (or `aiterm-mcp`) over stdio. Needs **Node ≥ 18** and **tmux** — see [Requirements](#requirements).
+
+## Headless: no human at the terminal
+
+Because an MCP client drives aiterm programmatically over stdio, everything above can run with **nobody sitting at a tmux**. Your Claude Code session can `codex_agent()` a task, `pty_read` the result, and act on it — unattended. That makes aiterm a fit for exactly the places a human-driven terminal isn't:
+
+- **Multi-agent orchestration** — an orchestrator hands sub-tasks to Codex / Grok / Composer, each in its own persistent session, and reads them all back.
+- **CI** — a job step can spin up an agent, drive it, and tear it down.
+- **cron** — a scheduled run can launch an agent and collect its output.
+
+The terminal is real and shared, so a human *can* jump in ([A human can watch](#a-human-can-watch)) — but nothing requires one to.
 
 ## How it works
 
 ```mermaid
 flowchart LR
-    AI["AI / MCP client"] -->|"pty_send"| S["aiterm-mcp<br/>stdio MCP · 9 tools"]
+    AI["AI / MCP client<br/>(the orchestrator)"] -->|"pty_send · codex_agent<br/>grok_agent · composer_agent"| S["aiterm-mcp<br/>stdio MCP · 9 tools"]
     S -->|"pty_read<br/>token-reduced"| AI
-    S -->|"tmux send-keys<br/>capture-pane"| P["one local PTY<br/>tmux · persistent"]
+    S -->|"tmux send-keys<br/>capture-pane"| P["persistent PTYs<br/>tmux · survive restarts"]
     P -->|"ssh · docker · repl"| R["nested<br/>remote · container · REPL"]
+    P -->|"launches a fresh PTY per agent"| A["another coding-agent TUI<br/>Codex · Grok · Composer"]
 ```
 
-One PTY is the only primitive. Everything else — SSH, containers, REPLs — is just text you `pty_send` into it. Because the PTY lives in tmux, sessions outlive the MCP server and the AI client.
+One PTY is the only primitive. Everything else — SSH, containers, REPLs, and the launched agent TUIs — is just something interactive running inside a persistent terminal, driven with the same `pty_send` / `pty_read`. Each launcher opens its own fresh PTY. Because the PTYs live in tmux, sessions outlive the MCP server and the AI client.
 
 ## vs. the alternatives
 
-|  | **aiterm-mcp** | one-shot shell MCP<br/>(e.g. `mcp-server-commands`) | terminal / SSH / tmux MCPs<br/>(e.g. `iterm-mcp`, `ssh-mcp`, `tmux-mcp`) |
-| --- | --- | --- | --- |
-| Persistent session | ✅ tmux, survives restarts | ❌ new shell every call | ⚠️ varies |
-| SSH / containers | nest with one `pty_send` | reconnect every command | ⚠️ often separate tools / per-call connect |
-| Token-reduced reads | ✅ per-command reducers | ❌ raw output | ⚠️ rarely |
-| Completion detection | 4-layer: exit / `until` / quiescence / timeout | n/a (blocks per call) | ⚠️ prompt-match, fragile |
-| Human can co-drive | ✅ shared tmux socket (`attach`) | ❌ | ⚠️ varies |
+aiterm sits at the intersection of two families: terminal-driving MCP servers, and the newer "agents talk to each other through a shared terminal" idea (see [Where aiterm fits](#where-aiterm-fits)). Here's how the axes line up — honestly, including where the others are strong.
 
-## Requirements
+|  | **aiterm-mcp** | one-shot shell MCP<br/>(e.g. `mcp-server-commands`) | terminal / SSH / tmux MCPs<br/>(e.g. `iterm-mcp`, `ssh-mcp`, `tmux-mcp`) | shared-tmux agent-to-agent<br/>(e.g. `smux`) |
+| --- | --- | --- | --- | --- |
+| Persistent session | ✅ tmux, survives restarts | ❌ new shell every call | ⚠️ varies | ✅ tmux |
+| SSH / containers / REPLs | nest with one `pty_send` | reconnect every command | ⚠️ often separate tools | ✅ tmux (human drives) |
+| Launch another agent in one call | ✅ `codex_agent` / `grok_agent` / `composer_agent` | ❌ | ❌ | ⚠️ agents join a human-run tmux via a CLI + skills |
+| Headless (no human at a tmux) | ✅ MCP-driven, programmatic | ✅ | ⚠️ varies | ❌ built around a human in the tmux |
+| MCP-native (any MCP client) | ✅ one `claude mcp add` | ✅ | ✅ (they are MCPs) | ❌ tmux config + CLI + Agent Skills |
+| Token-reduced reads | ✅ per-command reducers | ❌ raw output | ⚠️ rarely | ❌ raw tmux |
+| Completion detection | 5-layer: exit / `mark` / `until` / quiescence / timeout | n/a (blocks per call) | ⚠️ prompt-match, fragile | ❌ agent reads the pane |
+| Destructive-command gate | ✅ tripwire (override with `force`) | ❌ | ⚠️ varies | ❌ |
+| Human can co-drive | ✅ shared tmux socket (`attach`) | ❌ | ⚠️ varies | ✅ (its core model) |
 
-- **Node.js >= 18**
-- **tmux** (runtime prerequisite; check with `tmux -V`. Install with `apt install tmux` / `brew install tmux`)
-  - **macOS / Linux / WSL2** run tmux directly. On macOS install it with `brew install tmux` (stock macOS ships none). If your MCP client is launched from the **GUI** rather than a terminal, Homebrew's bin (`/opt/homebrew/bin` on Apple Silicon, `/usr/local/bin` on Intel) may be off its `PATH`; aiterm auto-searches those locations, or set **`AITERM_TMUX=/path/to/tmux`** to point at it explicitly.
-  - **Native Windows** has no tmux, so aiterm transparently runs tmux **inside WSL**. It needs [WSL](https://learn.microsoft.com/windows/wsl/) installed and initialized, with **tmux installed inside your WSL distro** (`sudo apt install tmux`); verify with `wsl tmux -V`. Sessions, the socket, and human `attach` all live on the WSL side — the AI just drives them from the Windows-side command. (You reach Windows tools the same way you reach SSH: `pty_send "powershell.exe …"` nests into PowerShell.)
-- Optional: the [`rtk`](https://github.com/rtk-ai/rtk) binary (used by `pty_send`'s `rtk: true` delegation; works fine without it)
+## Where aiterm fits
+
+"AIs talking to each other through a shared terminal" is becoming its own category — and it's a genuinely good idea. The terminal is a universal interface every coding agent already speaks, so no bespoke agent-to-agent protocol is needed; the shell *is* the shared surface. `smux` (by @shawn_pana) popularized this framing as a one-command shared tmux environment a human sets up, that agents then join via a `tmux-bridge` CLI and Agent Skills. It's good at the in-the-loop, shared-pane workflow it's built for, and it has real traction.
+
+aiterm takes the same core insight — the terminal as the meeting point — and makes three deliberate, different choices:
+
+1. **Headless by construction.** Because aiterm is driven programmatically over MCP, an AI can launch and drive another agent with *no human sitting in the tmux* — from an orchestration loop, a CI step, or a cron job. The shared-tmux tools lead with a human at the keyboard (their docs center on interactive pane navigation), so unattended operation isn't their native mode; aiterm's is.
+2. **MCP-native, not a workflow you adopt.** aiterm is a stdio MCP server: one `claude mcp add` line and it works as structured tools in any MCP client that speaks stdio (tested in Claude Code; Cursor, Cline, and Claude Desktop speak the same protocol and should work the same way). It doesn't ask you to adopt a tmux config, learn pane navigation, or install skills into your setup — the client already knows how to call tools.
+3. **Launching an agent is one tool call — an orchestration primitive.** `codex_agent()` spawns Codex in a persistent terminal and returns a session you drive immediately. You don't arrange panes or paste between them by hand; the launch, the steering, and the reads are all tool calls the orchestrating model can make on its own.
+
+On top of that sits a productized layer a raw tmux bridge doesn't have: **token-reduced reads**, **5-layer completion detection**, and a **destructive-command tripwire**. None of this makes the human-in-the-tmux model wrong — it's a different, complementary bet on where the human is standing.
 
 ## Tools
 
@@ -143,14 +207,14 @@ One PTY is the only primitive. Everything else — SSH, containers, REPLs — is
 | --- | --- | --- |
 | `pty_open` | Grab one terminal, return a `session_id` | `name?`, `shell="bash"` |
 | `pty_send` | Send text (a command) | `session_id`, `text`, `enter=true`, `mark`, `force`, `rtk`, `raw` |
-| `pty_read` | Read output, token-reduced (incremental by default) | `session_id`, `wait`, `until`, `timeout`, `screen`, `full`, `lines`, `line_range`, `raw`, `rtk` |
+| `pty_read` | Read output, token-reduced (incremental by default) | `session_id`, `wait`, `until`, `until_regex`, `timeout`, `screen`, `full`, `lines`, `line_range`, `raw`, `rtk` |
 | `pty_key` | Send a control key | `session_id`, `key` (`C-c`/`Enter`/`Up`…) |
 | `pty_close` | Close a session | `session_id` |
 | `pty_list` | List sessions | (none) |
 
 ### Interactive agent launchers
 
-Each launcher starts a specific vendor's interactive coding-agent TUI inside a fresh persistent PTY and returns its `session_id` — from there you drive it with plain `pty_read` / `pty_send`, exactly like any other session. One tool per model, so the tool name itself tells you which model you get.
+Each launcher starts a specific vendor's interactive coding-agent TUI inside a fresh persistent PTY and returns its `session_id` — from there you drive it with plain `pty_read` / `pty_send`, exactly like any other session. One tool per model, so the tool name itself tells you which model you get. The TUI is a full-screen app, so read it with `pty_read({ screen: true })` for the rendered view.
 
 | Tool | Launches | Key args |
 | --- | --- | --- |
@@ -158,11 +222,11 @@ Each launcher starts a specific vendor's interactive coding-agent TUI inside a f
 | `grok_agent` | Grok Build, model `grok-build` (xAI) | `prompt?`, `reasoning_effort?` (`low`/`medium`/`high`/`xhigh`/`max`), `cwd?`, `session_name?` |
 | `composer_agent` | Grok Build, model `grok-composer-2.5-fast` (xAI) | same as `grok_agent` |
 
-The vendor CLI must be installed and authenticated (`codex` for `codex_agent`; `grok` for both Grok tools). aiterm resolves the binary via `CODEX_BIN` / `GROK_BIN`, then `~/.local/bin/codex` / `~/.grok/bin/grok`, then `PATH`. All prerequisites are validated **before** a session is created — an invalid `reasoning_effort`, a missing CLI, or a nonexistent `cwd` fails with an explicit error and leaves no session behind.
+The vendor CLI must be installed and authenticated (`codex` for `codex_agent`; `grok` for both Grok tools). aiterm resolves the binary via `CODEX_BIN` / `GROK_BIN`, then `~/.local/bin/codex` / `~/.grok/bin/grok`, then `PATH`. Prerequisites are validated before a session is created (grok/composer reject an out-of-range `reasoning_effort`; a missing CLI or a nonexistent `cwd` fails for all three), and a failed launch leaves no session behind. Full details under [Launch other coding agents into that terminal](#2-launch-other-coding-agents-into-that-terminal--the-orchestration-flagship). Pass an absolute path for `cwd` — `~` is not expanded.
 
-### Completion detection (4 layers)
+### Completion detection (5 layers)
 
-`pty_read({ wait: true })` decides "is the command done?" via four layers: process exit / `until` regex match / output is quiescent ∧ the shell is back (quiescence) / timeout. While nested (inside SSH), the "shell is back" check cannot fire, so pass `until` with the remote prompt for a clean decision.
+`pty_read({ wait: true })` decides "is the command done?" via five layers: process exit / a `mark:true` sentinel (auto-detected — see below) / an `until` match (a literal substring by default; pass `until_regex: true` for a regex) / output is quiescent ∧ the shell is back (quiescence) / timeout. While nested (inside SSH, a container, a REPL, or a launched agent's TUI), the "shell is back" check cannot fire, so pass `until` with the inner prompt — or send with `mark: true` and `pty_read({ wait: true })` auto-detects the completion sentinel (no `until` needed, works nested too) — or, for a full-screen agent TUI, read `{ screen: true }` once its output settles.
 
 ### Token reduction
 
@@ -172,11 +236,31 @@ The vendor CLI must be installed and authenticated (`codex` for `codex_agent`; `
 
 ### Safety
 
-Before sending, `pty_send` blocks destructive commands (`rm -rf /`, `mkfs`, `dd of=/dev/…`, `DROP TABLE`, …) — pass `force: true` to override — and sanitizes ESC / bracketed-paste terminators. `pty_read` neutralizes control characters in what it returns.
+Before sending, `pty_send` blocks destructive commands (`rm -rf /`, `mkfs`, `dd of=/dev/…`, `DROP TABLE`, …) — pass `force: true` to override — and sanitizes ESC / bracketed-paste terminators. `pty_read` neutralizes control characters in what it returns by default (`raw: true` returns the bytes verbatim). This is a **tripwire, not a sandbox** (see [Known constraints](#known-constraints-by-design-not-bugs)).
 
 ## A human can watch
 
-Sessions live on a shared tmux socket. The `tmux -S … attach -t <id>` line printed by `pty_open` lets a human attach to the same terminal and intervene (`Ctrl-b d` to detach). On native Windows the printed line is the WSL form — `wsl tmux -S … attach -t <id>` — since the session lives inside WSL.
+Sessions live on a shared tmux socket. The `tmux -S … attach -t <id>` line printed by `pty_open` (and by each agent launcher) lets a human attach to the same terminal and intervene (`Ctrl-b d` to detach) — including watching a launched Codex/Grok/Composer session run and taking the keyboard from your AI mid-task. On native Windows the printed line is the WSL form — `wsl tmux -S … attach -t <id>` — since the session lives inside WSL.
+
+## Requirements
+
+- **Node.js >= 18**
+- **tmux** (runtime prerequisite; check with `tmux -V`. Install with `apt install tmux` / `brew install tmux`)
+  - **macOS / Linux / WSL2** run tmux directly. On macOS install it with `brew install tmux` (stock macOS ships none). If your MCP client is launched from the **GUI** rather than a terminal, Homebrew's bin (`/opt/homebrew/bin` on Apple Silicon, `/usr/local/bin` on Intel) may be off its `PATH`; aiterm auto-searches those locations, or set **`AITERM_TMUX=/path/to/tmux`** to point at it explicitly.
+  - **Native Windows** has no tmux, so aiterm transparently runs tmux **inside WSL**. It needs [WSL](https://learn.microsoft.com/windows/wsl/) installed and initialized, with **tmux installed inside your WSL distro** (`sudo apt install tmux`); verify with `wsl tmux -V`. Sessions, the socket, and human `attach` all live on the WSL side — the AI just drives them from the Windows-side command. (You reach Windows tools the same way you reach SSH: `pty_send "powershell.exe …"` nests into PowerShell.)
+- For the **agent launchers**: the corresponding vendor CLI, installed and authenticated — `codex` for `codex_agent`, `grok` for `grok_agent` / `composer_agent`. (Not needed if you only use the PTY tools.)
+- Optional: the [`rtk`](https://github.com/rtk-ai/rtk) binary (used by `pty_send`'s `rtk: true` delegation; works fine without it)
+
+## Known constraints (by design, not bugs)
+
+- **While nested (ssh / docker / REPL / a launched agent TUI), quiescence cannot fire by design**, because the foreground command is no longer in the shell set (bash/sh/zsh/fish/dash). When nested with no `until` and no `mark`, `pty_read({ wait: true })` returns early as `is_complete=False via nested` (rather than burning the full `timeout`, since no signal can confirm completion there) with a note to pass `until` (a literal substring by default; `until_regex: true` for a regex) or `mark: true` (an exit-code sentinel, auto-detected) for a confirmed completion. For a full-screen agent TUI, read `{ screen: true }` once its output settles.
+- **`is_complete=False` is not a failure.** It means "completion was not observed within `timeout`." For long commands, raise `timeout` or use `until`/`mark`.
+- **The destructive gate is a tripwire, not a sandbox.** It blocks common destructive forms only. It does **not** catch relative-path `rm`, things that become dangerous after `$VAR` expansion, or commands run on the far side of an SSH session — and it does not police what a launched coding agent does inside its own session.
+- **The agent launchers spawn a vendor TUI; they don't wrap or proxy it.** aiterm validates prerequisites and starts the CLI in a persistent PTY — the model, auth, and behavior are the vendor CLI's. There is no `claude` launcher and no protocol between agents; "conversation" is your MCP client driving the TUI (send input, read output).
+- **`pty_send({ rtk: true })` is single-line only and needs the external `rtk` binary** (passthrough without it). The `pty_read({ rtk: true })` reducer, by contrast, is self-contained and rtk-independent.
+- **The `pytest` reducer matches rtk 0.42.0** on test counts, the rule line, and `FAILURES`-block formatting (locked by regression tests). It **deliberately preserves the full failure reason** on the `FAILED` summary lines (emitted under `-ra`/`-rf`), whereas rtk 0.42.0 truncates the reason at the first `" - "` — a readability choice, so those lines are intentionally not byte-identical to rtk. The `[full output: …]` tee-pointer line rtk appends on large output is not reproduced on the read side.
+- **tmux is started with `-f /dev/null`**, so it does not read `~/.tmux.conf` (to keep behavior reproducible across machines).
+- **All sessions live on a single socket (`claude.sock` on POSIX).** `tmux … kill-server` removes them all.
 
 ## Development
 
@@ -187,17 +271,7 @@ npm test           # build, then the node:test regression suite (requires tmux)
 npm link           # put `aiterm-mcp` on PATH locally
 ```
 
-Logic lives in `src/core.ts` (tmux control, reduction, completion detection, safety) and `src/rtk.ts` (per-command reducers); `src/index.ts` is the MCP surface. The design origin and the reducer's porting source (the pytest reducer is ported to be byte-exact with upstream rtk 0.42.0, locked by regression tests) are in `prototype/python/`.
-
-## Known constraints (by design, not bugs)
-
-- **While nested (ssh / docker / REPL), quiescence cannot fire by design**, because the foreground command is no longer in the shell set (bash/sh/zsh/fish/dash). When nested with no `until`, `pty_read({ wait: true })` returns early as `is_complete=False via nested` (rather than burning the full `timeout`, since no signal can confirm completion there) with a note to pass `until` (a regex for the prompt) or `mark: true` (an exit-code sentinel) for a confirmed completion.
-- **`is_complete=False` is not a failure.** It means "completion was not observed within `timeout`." For long commands, raise `timeout` or use `until`/`mark`.
-- **The destructive gate is a tripwire, not a sandbox.** It blocks common destructive forms only. It does **not** catch relative-path `rm`, things that become dangerous after `$VAR` expansion, or commands run on the far side of an SSH session.
-- **`pty_send({ rtk: true })` is single-line only and needs the external `rtk` binary** (passthrough without it). The `pty_read({ rtk: true })` reducer, by contrast, is self-contained and rtk-independent.
-- **The `pytest` reducer matches rtk 0.42.0** on test counts, the rule line, and `FAILURES`-block formatting (locked by regression tests). It **deliberately preserves the full failure reason** on the `FAILED` summary lines (emitted under `-ra`/`-rf`), whereas rtk 0.42.0 truncates the reason at the first `" - "` — a readability choice, so those lines are intentionally not byte-identical to rtk. The `[full output: …]` tee-pointer line rtk appends on large output is not reproduced on the read side.
-- **tmux is started with `-f /dev/null`**, so it does not read `~/.tmux.conf` (to keep behavior reproducible across machines).
-- **All sessions live on a single socket (`claude.sock`).** `tmux … kill-server` removes them all.
+Logic lives in `src/core.ts` (tmux control, reduction, completion detection, safety, agent launch) and `src/rtk.ts` (per-command reducers); `src/index.ts` is the MCP surface. The design origin and the reducer's porting source (the pytest reducer is ported to match upstream rtk 0.42.0, except the deliberate `FAILED`-line difference noted above, and is locked by regression tests) are in `prototype/python/`.
 
 ## Try it
 
@@ -207,7 +281,7 @@ One command, no clone, no build:
 claude mcp add --scope user --transport stdio aiterm -- npx -y aiterm-mcp
 ```
 
-If aiterm saved you a round-trip of tokens, **[star the repo](https://github.com/kitepon-rgb/aiterm-mcp)** — it's the cheapest way to help others find it.
+If aiterm let your AI hand a task to another agent — or saved you a round-trip of tokens — **[star the repo](https://github.com/kitepon-rgb/aiterm-mcp)**. It's the cheapest way to help others find it.
 
 - **npm:** https://www.npmjs.com/package/aiterm-mcp
 - **Issues / bug reports:** https://github.com/kitepon-rgb/aiterm-mcp/issues