Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.
TypeScript adhdagent-orchestrationai-agentscanvasclaude-codecodexcoding-agentsdeveloper-toolselectrongeminilinuxmacosmultiplexeropencodeparallel-agentsterminalterminal-emulatorterminal-multiplexertmuxworkspace-manager
docs: browser guard, 0700 host dir, pinned host key type
review09bcc241e369
diff --git a/CLAUDE.md b/CLAUDE.mdindex 0de76dc3c..4fd7f5be0 100644--- a/CLAUDE.md+++ b/CLAUDE.md@@ -4022,6 +4022,25 @@ command-bearing opens; this does not add a human-confirm dialog or change mobilereconnecting terminals. These changes share the core listener in Desktop and Server Edition;the mobile wire protocol and node-identity rules are unchanged.+ - **A browser is refused before the bearer is looked at** (`core/agents/hook-browser-guard.ts`).+ The bearer stays the real gate; this is the belt beside it, against the one caller that can+ reach a loopback port without being a process of ours. A request carrying ANY `Origin` header,+ or a `Host` that is not `127.0.0.1` / `localhost` / `[::1]` (any port), gets 403 on both+ listeners, before the 403-vs-421 bearer distinction, so a DNS-rebound page learns nothing.+ MEASURED, every shipped client passes: curl over the socket sends `Host: localhost` (the codex+ launcher's socket URL sends `localhost:0`), curl over TCP `127.0.0.1:<port>` (the SSH TCP+ fallback arrives with the HOST's port — hence any port), Node's `http` over the socket+ `localhost`, Node's fetch no Origin (but `sec-fetch-mode: cors` — never refuse on that), Bun's+ fetch (opencode's runtime) no Origin. The SSH reverse tunnel forwards bytes, so a host-side+ curl's Host arrives unchanged. A request with no Host is let through (Node already 400s an+ HTTP/1.1 one; no browser sends one). **A new client must send no Origin and a loopback Host** —+ `hook-server.browser-guard.test.ts` drives the real transports and goes red otherwise.+ - **`~/.nodeterm` on an SSH host is 0700** (`privateRemoteDirCommand`, run by `setup()` on every+ connect and tunnel repair): created under `umask 077` and an existing one tightened, never+ through a symlink, never a directory the login user does not own (`-O`, which only a root login+ needs). The secrets inside were already 0600 one by one; the umbrella is so a future file written+ there without a mode is not born world-readable in a traversable directory.+- **No keyring is not "no identity" (#1088).** Electron 42 on a Linux session without SecretService/kwallet picks `basic_text` and `safeStorage.encryptString` THROWS (measured under xvfb).The desktop secret load used to reject on every boot ⇒ no node token was ever written ⇒ `send` /@@ -9548,7 +9567,14 @@ unix-socket forward and dev-server `-L` forward over it. POSIX keeps OpenSSH untmaster check. A new ssh call site owes the same routing.- **Semantics are OpenSSH's:** ControlMaster auto/no, `-O check|exit|forward|cancel`,`StrictHostKeyChecking=accept-new` over the user's own known_hosts (hashed entries included —- that is why HMAC-SHA1 appears; CodeQL's alert on it is dismissed with the reason), publickey+ that is why HMAC-SHA1 appears; CodeQL's alert on it is dismissed with the reason) — with the+ recorded host key TYPE pinned: the types on record lead ssh2's `algorithms.serverHostKey`+ (OpenSSH's order_hostkeyalgs, `hostKeyAlgorithmsFor`), and a host that HAS a recorded key but+ presents another type is refused like a changed key, naming both types and `ssh-keygen -R`+ (stricter than OpenSSH's accept-new, which records the new type; type-blind, a MITM holding only+ an ed25519 key against a host recorded under ECDSA was silently recorded and trusted). The+ list handed to ssh2 is always a reordering of `SSH2_DEFAULT_SERVER_HOST_KEY`, pinned to the+ package by test, because an exact list naming anything ssh2 lacks makes `connect()` throw; publickeyonly (agent, then key files, passphrase through the existing dialog, never in BatchMode), theuser's `~/.ssh/config` via `ssh -G` (never a second parser of it), a dropped connection endsevery channel with 255 (what `SshReconnector` reads).diff --git a/CONTRIBUTING.md b/CONTRIBUTING.mdindex 092a298fe..9ea23685c 100644--- a/CONTRIBUTING.md+++ b/CONTRIBUTING.md@@ -975,6 +975,12 @@ installation-qualified endpoint only after bearer verification. A wrong bearer aany handler runs; only that explicit wrong-owner response (or transport failure) permits endpointfailover. A node-identity 403 stays final. Test with disposable sockets, never a running user's tunnel.+**A hook-server client sends no `Origin` and a loopback `Host`.** The server refuses both before+the bearer (a browser page, DNS rebinding included, is the caller this keeps out). curl, Node's+`http` and fetch, and Bun's fetch all comply as they are used today; a new client that sets an+`Origin`, or names the server by anything but `127.0.0.1` / `localhost` / `[::1]`, gets a 403.+`hook-server.browser-guard.test.ts` drives the real transports — add yours there.+**Both raw listeners change together** — `src/main/index.ts` and `src/server/agent-status.ts`. A newfield on a hook event that reaches only the desktop leaves the Server Edition quietly without thefeature, and the boundary tests can only tell you an import is wrong, never that a field is missing.