wmux

Claude Code panes show the exact running / needs-you / done state Claude reports itself. Claude Code asks the terminal whether it supports the Program Status Protocol (OSC 7501), and wmux now answers. From then on Claude reports its own state: working, waiting on your approval, done, or stopped. The pane follows those reports, so it stays Running through a long silent think, shows Needs you as soon as a permission prompt opens, and settles when Claude says the turn is over. Stop, rewind and stall detection read the same reports. Before, all of this was guessed from the spinner in Claude's window title. That guess remains the fallback when the question cannot be answered, for example before wmux has recognised the agent or when Claude runs over ssh. (#2838)

Agents that drive other apps each get their own colour. On macOS, each agent now has its own coloured ring, name pill (#w1-2 · Claude Code), pointer and typing caret, so several agents working at once can be told apart. Before, every agent shared one orange halo and one cursor. An agent waiting for another one's turn at the real mouse shows as waiting, and a bar at the top of the screen counts the agents and shows the stop keys. The same colour marks the agent's pane as a dot on its tab and sidebar row.

Each agent's own marker on Windows too. On Windows, an agent that drives another app now shows the same coloured ring, name pill, pointer and typing caret as on macOS, so several agents working at once can be told apart. A waiting agent shows as waiting, and a bar at the top of the primary monitor counts the agents and shows the stop keys.

Pair two of your PCs over Tailscale without copying an invite. Remote › Connect a PC has a Tailscale tab, first when this PC is on a tailnet. On one PC, Make this PC discoverable lets your own Tailscale account's devices see its name and wmux version for 5 minutes. On the other, Find PCs lists it and Connect asks it. Both screens show the same 6-digit code, and the discoverable PC's card names the asking device as Tailscale identifies it (device, account, OS and tailnet address) next to Deny and Allow. After Allow, the asking PC confirms the codes match too, and only then do the two PCs pair as an invite would; the other PC sees no pane until you tick one. Before, pairing over a tailnet meant creating an invite on one PC and pasting it on the other. Opening the dialog only reads Tailscale's local status; nothing goes over the network until a button is pressed. (#3046, #3047)

A setting to let agent panes inherit your credentials. Settings > Git safety has a new switch, "Withhold credentials from agent panes", on by default. Turn it off and a pane that an agent opens gets environment variables named like a secret (*_TOKEN, *_KEY, ...) the way a terminal you open yourself does. Commands, supervised panes, bots and fan-out tasks never get them, and their values are never saved to disk. (#3050)

Terminal file links open in a wmux editor tab. Clicking a file path in a terminal (plain paths, wrapped paths and OSC 8 file: links) now opens it in a read-only editor or preview tab of that pane, at the line when the path carries one (src/a.ts:42:7), instead of handing it to the system app. The target line is scrolled into view and marked. A new Settings › Terminal option, "Open file links in", switches back to the system app. (#3057)

WSL paths in terminal links. Paths printed inside a WSL shell, including /home/... paths, file:// links and relative paths, resolve to the matching \\wsl.localhost\<distro> location and open in the viewer. (#3057)

Labs: draw terminals with wongtty, a Rust/WebAssembly terminal, and switch back at any time. Settings › Labs › "Rust terminal front end (wongtty)" rebuilds every open terminal with the other engine as soon as you flip it, and keeps what each one showed: s

A schedule that keeps failing pauses itself. Before, a scheduled run that failed the same way every time kept firing until you noticed. Now, after three scheduled runs in a row fail for the same reason, wmux turns the schedule off and tells you why on the schedule's page. Fix the cause, then turn the schedule back on. Cancelled runs, skipped occurrences, runs that timed out waiting for your answer and Run now never count, and editing what the schedule runs starts the count over. The run that pauses the schedule shows one notification, not two. (Idea adopted, with attribution, from the MIT-licensed cobusgreyling/loop-engineering; no code was copied.)

Qwen Code gets the wmux MCP server. If you use Qwen Code, wmux now adds its MCP server to your existing ~/.qwen/settings.json, the same way it already does for Claude Code, Codex and Gemini CLI, so Qwen can drive panes, the browser and agent messaging without manual setup. wmux never creates that file; if Qwen Code is not set up, nothing is written. Qwen's hook events are also mapped for later status tracking, but the hooks are not installed yet. (#2824)

Codex panes show the exact running / needs-you state Codex reports itself. Codex CLI writes its own status (working, waiting on you, idle) with the Program Status Protocol (OSC 7501), and wmux now reads it: a pane stays Running through a long, quiet reasoning stretch, shows Needs you the moment an approval box opens even when the question wraps on a narrow pane, and settles when Codex says the turn is over. Before, the status was guessed from the window title, output volume and lines on the screen, which read a quiet turn as idle and a repainted footer as finished. Those guesses remain the fallback for agents that do not report. (#2836)

Choose My Hermes or wmux's Hermes per bot. Add bot and the bot profile now offer My Hermes, which is your own Hermes (~/.hermes or one of its profiles), next to Claude and wmux's managed Hermes. Before you pick a profile, wmux shows whether it could serve a bot: the version, a model, approvals, and no wmux MCP entry or messenger platforms. These checks are read-only, and wmux never writes to your Hermes. One bot can use a profile at a time, and wmux asks you in its own dialog before it binds one. My Hermes works on macOS and Linux only.

Read and answer a Hermes Agent pane from your phone. Before, the phone's Chat view had nothing to show for a pane running Hermes Agent. Now, once you have sent Hermes its first message in the terminal, the Chat view shows the conversation, tool calls included, and you can send the next message from the phone while Hermes is waiting at its prompt. wmux reads Hermes' own session history without changing it, and only while Hermes is running. Sending is closed during a turn and while Hermes asks you something in the terminal. This works with Hermes' standard terminal interface, not with its --tui interface, a profile started with -p, or two Hermes panes in the same folder.

Bots run on My Hermes. A bot set to My Hermes now answers in its 1:1 chat with your own Hermes profile, using that profile's tools and approval settings as you configured them. wmux sends it only your own messages from that chat: nothing from other bots, guests or the rest of the room. Its only wmux tool is posting in that chat. Hermes's permission requests come to you as approval cards. Stopping the bot cancels its turn and never stops your Hermes gateway. Before every start wmux checks the profile again, and it never writes to it. Schedule actions and the Run button are not sent to My Hermes yet (routines you write are; see below); the chat says so instead. macOS and Linux only.

Routines for My Hermes. A bot on My Hermes can now run on a schedule, from routines you write for it yourse

An approved release is published. Before, a bot's approved request to publish a release was only recorded. Now wmux carries it out with your gh login after the Undo window. In a repository whose workflows run on a tag push, it pushes the tag at the commit the card showed, and that workflow builds and publishes the release. Elsewhere it makes the release as a draft and then publishes it. Before anything is written, wmux checks that the tag is still free and the commit is on the default branch with its checks green, and, in a project with a package.json, that the version matches the tag. An approval while the checks have not passed is refused, not held: the bot asks again once they pass. Only the card's own Approve button approves a release, on the desktop or the phone. Typed words, "Approve N pending" and "Always allow" never approve one.

Codex in the chat-only view. A new chat's composer has an Agent chip: pick Claude Code or Codex before the first message. A Codex chat starts on Codex's own default model and always asks before it runs a command or edits a file; its approvals come to the same cards, and to the phone, as Claude's. Moving the conversation to and from the terminal stays Claude-only. Needs Codex 0.162.1 or later. (#2731)

Amp, Droid, Pi, Cursor CLI and DeepSeek Harness are known agents. wmux now recognises them in a pane by their process, so a pane running one is labelled with its name. Once one of Amp, Droid, Pi or Cursor CLI is installed, the first-run "Agent ready" group shows its card too; an agent that is not installed gets no card there. Pi and Cursor CLI resume through their own session pickers. Fan-out lists all five as not yet verified.

GitHub Copilot CLI gets a setup card and resume. Once Copilot is installed, the first-run group shows its card with the same checks as the other agents, and a pane that ran Copilot can reopen its session picker (copilot --resume).

Custom agents. Settings › Roles & fan-out has a new Custom agents section for agent CLIs wmux does not know yet. Give one a name, a command (a program on PATH or an absolute path) and, if it needs them, arguments, a first prompt flag, a model flag and unattended flags. It can then be picked for a role or a fan-out preset row, and it launches as typed. wmux does not read its screen, so it has no status, approval cards or resume, and it is never given unattended or permission flags you did not type. An agent a role or preset still uses cannot be deleted; its row says what uses it.

Qwen Code panes show their own status. A pane running Qwen Code was labelled Gemini CLI, because Qwen draws Gemini's input row, and its working, idle and permission states were never read. wmux now recognises Qwen Code by its window title and banner, shows it working while a reply streams and ready once the turn ends, and marks the pane as needing you while a shell or file-write permission prompt waits. English UI; the idle state on Windows is not read yet. (#2740)

Pi is recognised on screen. A pane running Pi (@earendil-works/pi-coding-agent) is now identified from Pi's own window title and labelled as Pi. When Pi asks whether to trust a project folder, the pane shows that it is waiting for you. Before, a Pi pane looked like a plain shell and that question went unnoticed. (#2745)

Custom agents are named on their pane. A pane running one of your custom agents now shows that agent's name. wmux recognises it by its process: the Command's own name, or the process names you list in the new Process names field. It is a label only; the agent still gets no status, approval cards or resume. (#2751)

Self-host wmux for a team. deploy/docker/ now holds a container image for one person's headless wmux. It comes with a compose example that gives each person their own isolated instance behind Traefik, plus company sign-i

Pasting an invite says what comes next. After a PC joins another with its invite (Remote › Connect a PC › Paste an invite), the dialog now says that the other PC chooses which panes it shows this one, or shares one from a pane's menu. Before, it stopped at "Paired with …" and left the person to find out why there was nothing to link. (#2678)

Chat view asks once whether to show an agent as a chat. The first time Claude or Codex runs in a pane, a small card asks "Show this as a chat?" (Show / Keep terminal). The answer is remembered for each agent. Before, the chat view had to be found in Settings. Anyone who already set the chat view option is never asked, and Settings can still change it. (#2680)

Each chat view says which chat it is. A small label reads "Terminal-linked chat" or "Chat only", and hovering it explains the difference. Before, the two looked alike. (#2680)

A new agent pane from the phone can start in bypass or YOLO mode. POST /api/sessions {agentLaunch} now accepts the mode and confirm that chat/launch takes, under the same rules: the agent's own mode only, the transcript grant, the wmux web --allow-dangerous-launch ceiling and the exact <agent>:<mode> confirm, with the same 403 and 428 refusals. /api/config lists the allowed modes as agentLaunchModes. The mode applies to that start only: a recovered pane comes back in the default mode, and the desktop shows the same notice a dangerous chat launch raises. Before this, the mode could be chosen only when launching into an existing pane.

wmux now says when an agent is signed out. When Claude Code or Codex loses its sign-in, wmux shows one notice for that agent and account, with a Sign in again button that opens the agent's own sign-in. The workspace's row reads "Signed out" until the account works again. This happens when a turn fails on sign-in or when an agent starts on a signed-out account. Before, a run in that state failed with a generic error or waited on the agent's sign-in screen. (#2686)

Continue a terminal Claude in chat. On macOS and Linux, a Claude Code conversation in a terminal pane can now move into the chat view and continue there under the same session id. Before, a conversation could only move from chat to the terminal. wmux first checks that Claude is idle: no turn running, nothing waiting for approval, and no dialog or draft on screen. A confirmation then lists the processes Claude started that stop with it. wmux stops Claude and makes sure it has exited before the chat resumes the conversation, so the two are never running it at the same time. If the move fails after Claude was stopped, the terminal's Resume reopens the conversation. A chat that was moved to the terminal can come back the same way. (#2688)

Drop a file or paste an image into another PC's pane. On a pane attached from another PC, a dropped file or a pasted image is uploaded to that PC and its path there is typed into the pane, ready to send. That PC must have Photo & file upload turned on; files other than JPEG and PNG need it to run this version too. When something stops the send (uploads off, a prompt waiting on the pane, a file over 10 MB, a full upload folder), a notice says why and nothing is typed. (#2689)

Continue a usage-limited conversation on another PC. The usage-limit card now lists the paired PCs that let this one start an agent there. Choosing one opens the "Start an agent…" dialog with a note about the work as the first prompt, replies going to the held pane, and a wait that ends before the limit resets (four hours at most). Quoted parts of the note are fenced and marked as reference only. The note never carries the transcript or local user and machine names: in quoted text the folder, home folders, file: URLs and network paths are replaced, and its file lists leave out paths outside the