The team console and reasoning effort have posts of their own. The month's other work reached playbooks, agent models, the app, and the CLI.
The team console. Agents and people share one roster, each agent's row shows the work it has in flight, and a setup list walks a new account through connecting Slack and hiring its first agent. How it works.
Reasoning effort. Set how hard the model thinks per agent or per task, and every finished turn shows the effort it ran at. How it works.
Models and context. GPT-6 Sol joined the Codex model choices. Opus 5 and Sonnet 5 now run at a 200,000-token context window while their larger window is verified. Tool definitions from connected MCP servers load when an agent uses them, leaving more of the available window for the work. See the current models.
In the app
Where you left off. Opening a task lands you on the latest turn or at the New divider. The keyboard and a growing draft keep the newest messages in sight. The jump-to-latest arrow appears once you have scrolled a full screen back, and stays away when you are already at the bottom.
The activity line. While an agent runs commands, its activity holds a single line, so the conversation stays still under you. Tap the line to see every command, edit and tool call behind it. This reverses August's default, which showed every step inline and moved the transcript while people were reading it.
Sub-task briefs. When an agent starts a background sub-task within its own run, the prompt it gives the sub-task shows as a labeled, collapsed sub-task line in the transcript.
Dictation. Voice input on iPhone shows a waveform that responds to your voice and keeps listening through pauses. The waveform row opens only while you are dictating. The mic sits in the message box while the box is empty and steps out once your draft has text.
Glass on iOS 26. The jump-to-latest arrow and the suggestion pills above the message box use the same glass material as the message box, and the arrow sits at the right edge in line with it. The message box stays clear of the display's rounded corners when the keyboard is down.
Code review on phones. A task with code changes shows the number of changed files and lines. The Code and Files tray lets you review each changed file and open or create its pull request without leaving the task.
Task progress and captures. A running task shows the agent's current action and elapsed time, and the iPhone header opens a short history of recent activity. Screenshots from a task open in one gallery, where you can move between captures without losing your place in the conversation.
Staying signed in. Your sign-in on iPhone and iPad is held in the device Keychain. If the device briefly cannot read it, you stay signed in, and the app tells you when a sign-in could not be saved.
Low storage. A device with almost no free space could crash the app while it wrote its cached data, which lost your accounts and signed you out, and restoring that cache could crash the app a second or so after launch. The cache is now bounded, and below a safety margin of free space the app stops writing it.
iPad multitasking. The app runs in Split View, Slide Over and Stage Manager on iPad, and picks the phone or the tablet layout from the size of its window.
Agents, sessions, and the CLI
Playbooks that finish. Give a playbook a goal and the condition that completes it. Its runs settle when the matching event arrives. If work finishes outside Wallfacer, mark its task done or close it from the task screen or Inbox. A step skipped after the goal is met shows as not run, rather than failed.
Browsing the web. The browser your agent drives opens any public web page as well as the services running on its computer, and every session tells the agent the browser is there. Private and loopback addresses are refused, and the agent reaches its own services by name. A large screenshot comes back intact instead of failing with a misleading setup error.
Slack thread context. Mention an agent partway down a Slack thread and it starts out having read what the thread said before it was pulled in. Agents can now mention each other in a shared thread and answer there, so the conversation stays visible to your team.
Email threads. When an agent emails someone it was asked to get in touch with, it writes the subject line itself. Anyone copied on an email to an agent stays copied on every reply, and the agent can name them in its answer.
Agent spend. An agent running on model access Wallfacer provides works under a daily spend cap, and a task that reaches the cap says so in a message you can read. Spend and turn counts count a run that used a background sub-task once. The correction is applied when the numbers are read, so it covers earlier runs too.
Tasks that cannot start. A task whose agent credential is missing or unusable fails in seconds and names the credential, instead of appearing to run and ending in a generic error 45 minutes later. Add or replace the credential under the agent's Apps & access. A playbook step that keeps failing to launch gives up within a fixed time budget and follows the step's failure route.
macOS without live video. A macOS session whose live simulator video is unavailable starts anyway, with the simulator tools working, and the video appears on its own once the stream reconnects.
CLI coverage. The wallfacer CLI now lets you find and edit handbook pages and playbooks, browse your team, start chats, run playbooks, and mark tasks done or closed. It also runs agent analytics and automation commands, and shows waiting timestamps and close reasons in session output.
Smaller fixes shipped too, from push notification registration to images that stayed gray boxes.
Start here
Update the app, open a task on your phone, and review its changes in Code and Files. To set an agent's effort or choose GPT-6 Sol, open the agent and go to Configure. The models docs describe Claude's current context window.