The assistant.

One chat rail, standing across every project you’ve added. It reads the app through a fixed set of read-only tools, and anything that would change project state waits on a card you confirm.

01What it is

One conversation, with sight of every project, run, and quick session in the workspace. The rail is titled cyboflow assistant, and its header carries a Global chip beside a settings gear. Global is literal.

Where it showsEvery landing surface: home, project overview, backlog, review queue, Insights, workflows, verify queue, and the A/B comparison view

Where it doesn’tInside a session workspace, and in the launch wizard

ToggleToggle right rail, ⌘]

CollapsedA narrow strip with an Expand cyboflow assistant chevron

Windows uses Ctrl in place of ⌘ throughout this page. Inside a session that same chord toggles the session’s own right rail instead, and it does nothing when the assistant is switched off.

Drag the rail’s left edge to resize it. The width you set and the collapsed state both survive a restart.

After the guided set-up finishes, the rail opens with a one-off greeting naming the project you added. The greeting is display only. No model turn runs, and the line stays out of the transcript.

02Proposals

The assistant cannot change project state on its own. Launching a run, reprioritizing the backlog, adding backlog items, editing or creating a flow, and opening a session all land as a proposal card first. Confirming one runs the action through the same write paths every other change in the app goes through, stamped as you.

A card shows a head bar reading Proposed action · <kind>, an amber needs confirm, and a Confirm / Dismiss footer. Six kinds exist: launch run, reprioritize backlog, edit workflow, open session, add to backlog, and create workflow.

Create workflow can mint an entire custom flow, along with the custom agents its steps bind to. It shipped in 0.4.2.
One add-to-backlog proposal creates at most 20 entities, and one create-workflow proposal at most eight agents.
A card that resolves while you watch stays on screen for 45 seconds and then disappears. A card that resolved before the rail mounted (after an app restart, for example) never renders.

Exception Saving a custom widget to your own library is the one action the assistant takes with no confirm card. Custom widgets has the detail.

03What it can read

Everything the assistant knows about the app arrives through one fixed read-only tool family.

A cross-project digest of sessions and runs, the backlog, one entity’s full body, the review queue, flows, and a project’s agents.
Read-only queries against the app database, and a product reference it can quote.
Scoped file read, list, and grep, bounded by the Folder Access list in Settings.
Its own transcript. The history tool searches every past turn, so it can answer what you discussed last week after the live context has been cleared.

That read access is what makes it worth asking which flow to use. Describe a goal and it maps it onto Launch, Planner, Sprint, Ship, or Compound. It checks the real backlog and queue before it recommends one, and it flags a project as ready for a Compound run once roughly five or more open findings pile up there.

04Using it

Three chips sit above the composer: Status update, Triage the backlog, and Modify a workflow. Each one sends its own label as the message.

Send⌘⏎

ComposerGrows to four lines

ImagesPaperclip, paste, or drag. Four per message, 5 MB each

Image typesPNG, JPEG, GIF, and WebP. Anything else is refused at attach time, in the rail

Model badgemodel · <model>, under the composer, naming the model that answered the last turn

There’s no stop button for a turn in flight. The composer and the chips are disabled while one runs, and come back when it ends.

The assistant answers when you ask it something, and it doesn’t spend a turn of its own at start-up.

05Settings

The gear in the rail header opens Settings → Assistant, a top-level tab. Changes there take a Save Changes click.

Enable assistantOn by default. Switch it off and the rail hides, and the assistant spends no more tokens from the next turn. No restart

RuntimeFollow default runtime, Claude, or Codex. The Codex option appears once you’ve turned Codex on under Settings → Integrations

ModelSits under Runtime, defaults to App default, and lists models for whichever provider the runtime resolves to

Context retentionContext Retention: Clear daily (the default), Compact daily, or Auto-compact

Folder accessFolder Access: every registered project folder, with a per-project switch and an Extra folders box, one path per line. When the assistant runs on Codex, the tab notes that folder access applies to the Claude assistant

Session summariesSummarize idle sessions: after a quick session sits idle for 5 minutes, a Haiku call refreshes its summary on the session canvas. Codex and OMP CLI sessions are not covered

Switching the runtime starts a fresh conversation rather than replaying one provider’s history into the other, and it clears the model you had picked. Retention decides what the model carries into a new day, not what you can see: whichever option you pick, your visible chat history stays. On Codex, Compact daily behaves as Clear daily, because there’s no compaction call to make.

Session summaries shares the tab but is a separate feature. Summarize idle sessions stays available with Enable assistant off.

06Custom widgets

The custom widgets on the review queue and the project overview are written by the assistant.

Customize → Add widget → Create a custom widget, and a placed widget’s Edit with assistant, expand the rail and pre-fill its composer with a kickoff message. Both live in customize mode. You send that message yourself.

The assistant then writes the widget, and saving it to your library is the one action that needs no confirm card. A published widget appears under Customize → Add widget → Mine. A draft save stays in the authoring slot until it is published.

07What it will not do

It runs no shell commands and edits no files. On Claude the assistant is given no built-in tools at all, only the scoped Cyboflow tool family.

On Codex it runs in a read-only sandbox, with the shell tool off, every patch rejected by the sandbox, and your own Codex MCP servers switched off by name. That leaves the Cyboflow family as the only way it reads anything. Image generation and image viewing are off there too.

Sub-agent spawning is one of a few surfaces Codex does not let the app remove. The assistant’s own instructions forbid it, and a spawned sub-agent inherits the same read-only, no-shell configuration.

File reads stay inside the Folder Access list, and secret files such as .env and private-key files are refused.