ci-triage
Diagnose a failing GitHub Actions run — find the first real error in the logs, tell a flake apart from a genuine failure, and identify the commit that broke…
Operate the user's real desktop apps and windows on macOS, Windows, or Linux. Use when the user asks to open or drive a native app, click or type in a desktop UI, inspect or arrange a window, use their signed-in system browser, or verify an end-to-end GUI flow, including "open
$ npx -y skills add superset-sh/superset --skill computer --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/computerContext preview
The summary Claude sees to decide when to auto-load this skill.
Operate the user's real desktop apps and windows on macOS, Windows, or Linux. Use when the user asks to open or drive a native app, click or type in a desktop UI, inspect or arrange a window, use their signed-in system browser, or verify an end-to-end GUI flow, including "open
name: computer description: Operate the user's real desktop apps and windows on macOS, Windows, or Linux. Use when the user asks to open or drive a native app, click or type in a desktop UI, inspect or arrange a window, use their signed-in system browser, or verify an end-to-end GUI flow, including "open Settings and turn on dark mode", "click Save in that window", "what's on screen". Uses an accessibility-first driver (Cua Driver by default, Peekaboo on macOS). Not for Superset's in-app browser pane, headless scraping, or anything an API or CLI can do directly. argument-hint: the app and what to do in it allowed-tools: Bash(cua-driver:*) Bash(peekaboo:*) Bash(superset pages publish:*)
Operate the user's real desktop through an accessibility-first driver: a local CLI that exposes accessibility snapshots, exact window screenshots, native menu operations, and targeted input without making the agent guess at stale screen coordinates. The contract in this skill (preflight, task-scoped session, snapshot, act, verify, semantic operations first, conservative failure handling, clean teardown) is driver-agnostic. The command reference below is for Cua Driver, the default.
when the requested result does not require GUI interaction.
[Cua Driver](https://cua.ai/docs/cua-driver) (`cua-driver`, cross-platform, the default) or [Peekaboo](https://peekaboo.sh) (`peekaboo`, macOS only). If both are installed, use Cua Driver and mention the alternative once.
anything; recommend Cua Driver unless the user is macOS-only and prefers Peekaboo.
driver's own command surface; discover it with the driver's help output and documentation rather than assuming the Cua command shapes below.
1. Confirm the installed executable with `command -v cua-driver` on macOS or Linux, or `Get-Command cua-driver` in Windows PowerShell, then inspect `cua-driver --version`. Do not silently install or upgrade it. If it is missing, ask the user to authorize an installer from the [Cua Driver installation guide](https://cua.ai/docs/how-to-guides/driver/install). On macOS or Linux, use:
/bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"
On Windows, use PowerShell:
irm https://cua.ai/driver/install.ps1 | iex cua-driver autostart enable cua-driver autostart kick
2. Run `cua-driver doctor` and `cua-driver status`. On macOS, also run `cua-driver permissions status`. If Accessibility or Screen Recording is missing, ask the user to run `cua-driver permissions grant` and complete the OS prompts. Never work around missing OS permissions with AppleScript, synthetic shell input, or another GUI driver. 3. Preserve an already-running daemon's permission mode; do not restart it merely to broaden permissions. If the environment needs explicit service setup, follow the platform-specific guidance reported by `doctor`. 4. Inspect the installed tool surface instead of assuming a version-specific schema:
cua-driver list-tools cua-driver describe get_window_state cua-driver describe click
Invoke tools with `cua-driver call <tool> '<json>'`. Use `describe` whenever an argument is uncertain.
Create a unique session before the first GUI observation:
cua-driver call start_session '{"session":"superset-computer-<unique>"}'Pass the same `session` to every tool that accepts it. It owns this run's agent cursor and lifecycle cleanup. It does not select the capture mode or grant access beyond the daemon's fixed permission mode.
Follow this loop for every GUI action:
1. State the exact postcondition, such as "Settings shows Dark mode selected." 2. Discover the target with `get_accessibility_tree` or another semantic tool, then take a fresh `get_window_state` snapshot for the exact PID and window. 3. Prefer the snapshot's `element_token` over `element_index`, and prefer both over pixel coordinates. Tokens identify the control and fail closed when the snapshot becomes stale. 4. Perform one action. Default to background delivery; use foreground delivery only when fresh evidence shows the background action did not land. 5. Take a new snapshot and verify the postcondition before continuing. A successful tool response proves delivery, not the resulting UI state.
Minimal shape:
cua-driver call get_accessibility_tree '{}'
cua-driver call get_window_state '{"pid":844,"window_id":10725,"session":"superset-computer-<unique>"}'
cua-driver call click '{"pid":844,"element_token":"s0000002a:14","session":"superset-computer-<unique>"}'
cua-driver call verify_state '{"pid":844,"window_id":10725,"session":"superset-computer-<unique>","expect":[{"element":{"selector":{"label_contains":"Saved"},"exists":true}}]}'Re-snapshot before every later action. Never reuse a token after a snapshot of the same window, navigation, modal transition, or substantial repaint.
`hotkey` for non-text keys.
window ID, then call `get_browser_state`.
Superset is an agentic IDE to orchestrate 100+ coding agents in parallel. Run any agent with your own subscription.
Diagnose a failing GitHub Actions run — find the first real error in the logs, tell a flake apart from a genuine failure, and identify the commit that broke…
Triage a GitHub issue into something actionable — reproduce the claim, find duplicates, judge severity, and apply labels, milestone, and assignee. Use when the…
Find and merge duplicate Linear issues — group reports of the same underlying bug, pick the survivor, and move the evidence across. Use when the backlog has…
Turn a rough report into a Linear issue someone can pick up — reproduce the claim, check for duplicates, and fill in team, priority, and labels. Use when the…
Draft a Linear project status update from what actually moved — progress, risks, and the one decision that needs making. Use when someone asks for a project…
Personalized audit that teaches the advanced Superset features the user isn't using yet (automations, parallel agents, tasks, multi-host, terminal remote…