Cate CLI

cate is available inside Cate terminals and agent shells. It talks to the current workspace and requires the relevant Settings → CLI permission.

Start by listing panels:

cate panel list

When working repeatedly with one panel, select it for the current agent or terminal session:

cate panel set 1a2b3c4d
cate panel current

The selection is isolated by a per-terminal CLI session, so other agents and terminals keep their own targets. Short ids from panel list are accepted. Use --panel <id> only as a one-command override. Clear the selection to return to Cate's automatic focused/grouped resolution:

cate panel clear

Selections can point to any native panel. Browser and terminal commands reject a selected panel of the wrong type instead of silently controlling another panel. If a selected panel was closed, select another panel before continuing.

Browser workflow

Browser control uses persistent JavaScript with the cua tab API. The old argv actions, selectors, page evaluation, and revisioned string refs have been removed. Start by binding a tab to get its accessibility state, then request a screenshot when visual context is useful:

cate browser run 'var tab = await cua.getTab({panelId:"<full-panel-id>"});'
cate browser run 'await tab.getAXStateAndScreenshot();'

Use full panel IDs inside JavaScript. --panel <id> supports short IDs as an override for CLI panel resolution. Discover tabs with await cua.listTabs(). Create a tab with await cua.createBrowserTab("https://example.com"), or pass {panelId:tab.panelId} as the second argument to choose its panel. Pass {newPanel:true} to create a separate panel; Cate also creates one when needed. Bindings pin both panel and tab; they never silently follow a user's tab switch.

Use numeric IDs from the latest AX observation. For example, after observing a form containing textbox 17 and button 42:

cate browser run 'await tab.setValue(17,"user@example.com"); await tab.click(42); await tab.waitFor({url:"**/dashboard"});'
cate browser run 'await tab.getAXStateAndScreenshot();'

Do not guess IDs or coordinates. Observations contain kind, observationId, documentId, URL, title and viewport. AX observations (kind:"ax") also contain accessibility state and structured elements with role, name, value and states. getAXState() normally emits a concise diff; getAXState({disableDiffing:true}) emits the full tree. getScreenshot() returns only viewport pixels/identity (kind:"image", empty state/elements), avoiding an AX scan. It does not refresh numeric IDs. getAXStateAndScreenshot() refreshes both together. The SDK keeps the last AX observation for numeric targets and the latest visual observation for coordinates. {emit:false} suppresses automatic output. {profile:true} adds phase timings, image bytes and estimated retained-cache usage. Each code cell retains at most 16 million serialized observation characters, including emit:false; split long screenshot loops across cells.

The SDK carries the latest observation through each action and emits fresh state. Numeric IDs persist within one document; navigation requires fresh IDs. Coordinate actions use [x,y] in observed viewport CSS pixels and reject stale viewport coordinates. A dispatched click is not proof that a business operation completed: use waitFor or inspect the resulting state.

await tab.click(42);                         // or [x,y]
await tab.setValue(17, "replacement");
await tab.typeText("insert at selection");
await tab.pressKey("Return");
await tab.selectText(17, "text", {selectionType:"cursor_after"});
await tab.scroll([400,300], "down", 1);
await tab.drag([100,100], [300,200]);
await tab.setChecked(42, true);
await tab.selectOption(42, ["DE"]);
await tab.upload(42, "/authorized/file");
await tab.waitFor({text:"Saved"});
await tab.waitFor({element:42, state:"enabled"});
await tab.goto("https://example.com");
await tab.back(); await tab.forward(); await tab.reload();
await tab.setViewport({width:1280,height:800});
await tab.resize({width:800,height:600});
await tab.downloads(); await tab.close();

Keep deterministic batches short and inspect unexpected changes before continuing. Use var for reusable bindings; top-level await is supported. The session has no Node.js, filesystem, network, or DOM evaluation access. Await every action. nodeRepl.write(value) adds text output. cate browser reset clears JavaScript bindings without closing tabs. Sessions are isolated per terminal/agent through CATE_CLI_SESSION_ID; timed-out sessions reset.

cate browser run 'await tab.getAXStateAndScreenshot();' returns AX state and saves a screenshot to a temporary PNG file. Open the printed path with your image-viewing tool before visual reasoning. Shell output cannot itself attach pixels to the model. --json returns structured content with base64 image data; base64 text is not visual input. The same screenshot output works from tab.getScreenshot() in code. AX reads remain available in code for deterministic branches and extraction.

Agent actions display a cursor/highlight in the browser panel. User input takes control back and cancels pending automation. Responsive viewport size and canvas panel size are independent; resize applies only to canvas panels with a 400×300 minimum.

Other surfaces

cate editor open src/app.tsx:42
cate panel create terminal
cate panel create canvas
cate panel set <id>
cate panel current
cate panel clear
cate panel close <id>

Read a terminal before sending input. type does not append Enter:

cate panel set 1a2b3c4d
cate terminal read
cate terminal type npm test
cate terminal press enter

Terminal input goes to whatever currently owns that PTY, including foreground TUIs. Never send keys until the panel id and current screen are verified.

Agent orchestration

Use cate agent when a task benefits from visible, persistent delegation: independent parallel work, cross-provider review, or isolated implementation in a Cate worktree. Keep small, tightly coupled edits in the current agent.

Discover registered runs before acting on an older mission or after context compaction:

cate agent list

Create a worker with a bounded, self-contained prompt and concrete success criteria. Cate chooses the first hook-ready registered agent when --agent is omitted:

cate agent create "Inspect the API boundary and report risks" --title "API scout"
cate agent create "Implement the parser and run its focused tests" \
  --agent codex --title "Parser" --new-worktree agent/parser
cate agent create "Review the current worktree changes" --worktree <worktree-id>
cate agent create "Continue in this terminal" --terminal <terminal-panel-id>

--terminal reuses an idle terminal panel by restarting its PTY with the selected agent. It rejects busy terminals, agent-owned terminals, and the calling terminal. Omit it to create a new terminal panel.

Workers may recursively create and supervise their own workers with the same commands. This naturally forms an agent tree: each terminal owns the workers it creates, and each parent normally communicates with its direct children. Use recursion when another level of decomposition is genuinely useful, not merely to relay a simple instruction.

Supervise workers through the agent lifecycle rather than typing into their terminals:

cate agent wait <run-id> [<run-id>...] --wait-timeout 10000
cate agent inspect <run-id>
cate agent send <run-id> "Please add the missing regression test"
cate agent review <run-id>
cate agent apply <run-id>
cate agent keep <run-id>
cate agent discard <run-id>
cate agent stop <run-id>

Run ids may be the unique short ids printed by cate agent list. wait accepts 5000–60000 milliseconds and may be called with no ids to monitor all live direct children. Call it again while workers remain active. inspect includes recent terminal output; use cate terminal read --panel <panel-id> only as a lower-level diagnostic fallback.

Prefer send for follow-up work on the same responsibility. If followUpSupported is false, create a fresh worker instead. When a worker fails, inspect failureReason; a provider-specific authentication, quota, or service failure can justify retrying with a different registered --agent.

For an isolated worker, ask it to run relevant checks and commit completed work, then use review before choosing apply, keep, or discard. Apply rechecks that the worktree is clean and mergeable. Discard permanently removes a worker-owned worktree and its branch, including uncommitted changes, without an interactive confirmation. Keep records that the worktree should remain for later. Review is read-only: a finished process or successful review does not mean its branch has been integrated. The parent remains responsible for verification and for reporting any uncommitted or unintegrated work.

Review panels

Select a Review Panel once, inspect its comparison, and record structured findings. --panel <id> is an optional one-command override for every review command.

cate panel set <review-panel-id>
cate review inspect
cate review note add --file src/app.ts --line 42 --side new \
  --severity error --body "Handle the rejected request"
cate review note resolve <note-id>
cate review complete

Use complete only when running as the review agent assigned by that Review Panel. Review commands record findings; they do not modify files, stage, commit, or push changes.