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(ssh): ProxyJump and the Windows ssh-agent rule in the Windows transport section
enesd564aa335486
diff --git a/CLAUDE.md b/CLAUDE.mdindex fef66b286..ef1965189 100644--- a/CLAUDE.md+++ b/CLAUDE.md@@ -7958,9 +7958,30 @@ unix-socket forward over it. POSIX keeps OpenSSH untouched.- **Tests run on every OS** against ssh2's own in-process `Server` (loopback, no sshd); thedirectory is in the `windows-latest` CI job. Live numbers (macOS, `NODETERM_NATIVE_SSH=1`,89-terminal project): 89/89 attached, 0 ssh processes, ~1 login per connect, main CPU 3–5% idle.-- **Not done yet (phase 3):** adding an unlocked key to the Windows OpenSSH agent service,- ProxyJump (refused by name today; ProxyCommand stays refused), sleep/wake verification, a- like-for-like timing against OpenSSH on the same project.+- **ProxyJump** (#1078) follows OpenSSH: each hop is resolved by ITS OWN `ssh -G` and gets its own+ host-key check and publickey auth; the chain is ssh2 `forwardOut` streams used as the next hop's+ socket. `ssh -J a,b t` means `ssh -J a -W t b`, so only the FIRST hop's own ProxyJump is followed+ (recursively); loops and chains deeper than 8 are refused by name. Jump connections belong to+ the target connection and die with it (a dropped jump → 255 on the target's channels). MaxSessions+ overflow connections REUSE the primary's chain (one bastion login; direct-tcpip does not count+ against the bastion's MaxSessions); one-off connections build their own. Known hosts are checked+ under `HostName` (or `HostKeyAlias`), as OpenSSH does — not under the alias as typed.+ **ProxyCommand stays refused by name**, on the target and on a hop.+- **The Windows ssh-agent only on the user's say-so** (#1080). MEASURED on windows-latest+ (OpenSSH_for_Windows_9.5p2): the agent service REFUSES any lifetime or confirm constraint+ (`ssh-add -t` / `-c` and our `ADD_ID_CONSTRAINED` alike), and an unconstrained key is stored in+ `HKCU\Software\OpenSSH\Agent\Keys` (DPAPI) and survives service restarts — "until removed" is+ the only add Windows offers. So a passphrase-unlocked key is added (`agent-add.ts`, our own+ agent-protocol writer; ssh2 only lists and signs) ONLY when the host's own config says+ `AddKeysToAgent yes` (what Windows' ssh.exe would do) or the user turned on Settings → Remote+ (SSH) → "Keep unlocked keys in the Windows ssh-agent" (`settings.windowsSshAgentAddKeys`, default+ OFF, copy says Windows keeps it until `ssh-add -d`). A config lifetime is sent as a constraint and+ Windows' refusal stands: a refused constrained add is NEVER retried unconstrained. Fail-open: an+ agent error never affects the connection. Reboot persistence is inferred from the registry hive,+ not measured.+- **Not done yet:** sleep/wake verification on the native transport, a like-for-like timing against+ OpenSSH on the same project, and any run on a real Windows desktop (all evidence so far is CI plus+ the macOS run of the same code path).## Remote access (phone relay) — free, not Pro