Drive Positron through CDP

Use this skill to inspect and interact with a running development build of Positron. Treat the resulting profile and application state as disposable.

This workflow complements automated tests; it does not replace them. Add a test when the verified behavior needs regression coverage (.claude/skills/author-e2e-tests).

Do not use it to hand a persistent Positron to a person: the profile is deleted at cleanup, native dialogs are replaced with in-app ones, and nothing recompiles later source edits. Use the launch-positron command for that.

This file is a guide to choosing and combining the helpers in scripts/. Each helper's header comment is the reference for its flags, output and exit codes: bash .claude/skills/drive-positron/scripts/X.sh --help prints it. To change a helper or add one, read CONTRIBUTING.md.

Know what this changes in your checkout

The disposable profile is isolated. The build state is not.

Before it starts the application, launch.sh runs build/lib/preLaunch.ts against your real checkout, which writes to directories your normal development build also uses:

• .build/builtInExtensions/<name>: deleted and re-downloaded for every built-in extension whose version on disk does not match product.json. A rebase that bumps one is enough.
• .build/electron: deleted and re-downloaded when the Electron version does not match.
• out/: when it is absent, pre-launch runs npm run compile, which competes with the build daemons.

An interrupted pre-launch can leave a built-in extension deleted or half written, and your normal build then fails to start. Do not interrupt the script while it reports pre-launch. To repair:

npm run download-builtin-extensions
npm run electron

Platforms

macOS, Linux, and Windows from Git Bash (not PowerShell or cmd). The helpers need node and curl, rsync or tar, and jq (absent from a bare Git Bash); a header lists any other tool its script needs. On Windows, launch.sh starts the app through WMI on purpose; the comment above that code says why.

Launch, attach, clean up

Launch, from the repository root:

.claude/skills/drive-positron/scripts/launch.sh -- \
	--folder-uri file:///private/tmp/myworkspace --log debug

It prints one JSON line; keep pid, cdpPort, runDir and logFile. Launcher flags go before the -- and app arguments after it; a launcher flag put after it is silently ignored. The header lists both, and the app arguments the launcher supplies. Do not opt out of those unless the scenario is what they suppress.

The profile is a copy of $POSITRON_DEV_USER_DATA_DIR or ~/.positron-dev, or of --source-user-data-dir. To avoid reading your own profile, make a minimal seed (mkdir -p /tmp/seed/User and a settings.json in it) and pass that. A fresh profile tests only the cold start; for a second launch with the state the first one wrote (discovery cache, migrations, recent files), reseed.sh stops the instance and turns its profile into a seed.

To relaunch the same instance on its own profile (workspace storage, open editors, installed extensions and all):

.claude/skills/drive-positron/scripts/palette-run.sh --session positron 'Close Window'  # saves the window's state
./node_modules/.bin/playwright-cli -s=positron close
.claude/skills/drive-positron/scripts/stop.sh --cdp-port "$CDP_PORT"                     # no --run-dir: keep it
.claude/skills/drive-positron/scripts/launch.sh --reuse-profile "$RUN_DIR" -- --folder-uri file:///private/tmp/ws

macOS has no Quit in the palette, and Cmd+Q cannot be sent through CDP: Close Window, then stop.sh. Stop both runs with --run-dir at the end.

Attach Playwright under a literal session name, and use that name in every command and helper (--session positron):

./node_modules/.bin/playwright-cli -s=positron attach --cdp=http://127.0.0.1:"$CDP_PORT"

• Do not derive the name from $$: each shell call has its own PID, so each would get its own session.
• Use a name no other driver on this machine is using: drivers in one checkout share playwright-cli's sessions, so two under one name drive the same window. When $PW_SESSION is set, use it; the helpers read it when --session is left out.
• Run from the repository root and call the binary directly. npx costs about a second per call, and from another folder installs its own copy, which reports The browser 'NAME' is not open.
• Reload the window, open another folder in it, or open a new window with window.sh (reload, open-folder PATH, new-window): it waits for the workbench to come back and attaches the session again if it lost the page. Take a fresh snapshot after: refs restart with a frame prefix (f1e12).

Log every action. With these set, every helper appends one line per action to the log, with the key values it returned after "->" (the plot shown, the cursor, the cell's output), every read helper one line with the values it read ("read ..."), every failed call one line with its error ("FAILED ..."), so a negative check has its evidence too, and shot.sh saves into the folder and logs the shot:

export DRIVE_POSITRON_LOG="$RUN/actions.log" DRIVE_POSITRON_SHOTS="$RUN/shots"

Log yourself what the helpers cannot see: a raw playwright-cli click or key, a shell command outside the app, a wait.

Clean up. Positron can hold several gigabytes, so always stop the instance:

./node_modules/.bin/playwright-cli -s=positron close
.claude/skills/drive-positron/scripts/stop.sh --cdp-port "$CDP_PORT" --run-dir "$RUN_DIR"

• Before stop.sh, and before closing a window: copy any logs you need from runDir (the instance's logs are in <runDir>/logs), and read the browser console (playwright-cli console), which exists only in a live window; confirm every screenshot exists, and run listeners.sh --diff, which needs the instance running to know its process tree.
• Run stop.sh in a command of its own: a failure earlier in the same chain cannot be retried once the instance is gone.
• Never kill "$PID" (on Windows it is the MSYS shell, not the app), and never remove .playwright-cli, which other runs share.

Principles

• Never press Escape while a .qmd is in front. With its kernel busy, Escape is Quarto: Interrupt Kernel wherever focus is: in the editor, on the workbench, over a completion list or a hover. Only a quick input and a menu that has focus take the key first. The helpers close things without it (editor.sh suggest and hover close their widget by moving focus); a raw press Escape in a .qmd can stop the cell under test.

• Locate by role and name. Find controls by the accessible role and name a screen reader announces, not CSS classes or coordinates. ui.sh read VIEW shows them; copy the name from there.

• Fail loud. Every helper exits non-zero with ok: false and an error when it could not do what was asked, and refuses rather than guessing: a command not in the palette, focus not where the keys would go, a file that is not the active editor. Check the exit status. A refusal is often a fact about the app worth recording. An argument a helper does not take is a usage error (exit 2), logged like any failure. When a modal dialog is open, the error of any failing helper names it and its buttons ("dialogs"): it is usually why keys and clicks went nowhere.

• Never type into the Command Palette or Quick Open and press Enter. When no row matches exactly, the highlighted row can be a different command. palette-run.sh, open-file.sh and ui.sh pick run a row only when its label matches exactly.

• One call per command. Each helper is one Playwright call (about a second). Batch independent steps into one Bash call, and use a helper before a string of raw playwright-cli commands.

• Read before you act. Read the view, then act on what it shows; actions report what changed, so read the answer before the next step. Take a screenshot early when the UI does not match expectations: it shows a blocking dialog, an unopened workspace, or focus in the wrong place faster than DOM reads.

To do X, use Y

All in .claude/skills/drive-positron/scripts/.

ui.sh works in any view, but where the accessibility tree is thin, use the area's own helper: the Variables pane and the Data Explorer grid read as one line of text, and Quarto inline output is not in the tree at all.

Raw playwright-cli, when no helper fits

./node_modules/.bin/playwright-cli -s=positron snapshot          # ~250 lines; snapshot a container's ref instead
./node_modules/.bin/playwright-cli -s=positron find "Run Cell"   # matching nodes and their context
./node_modules/.bin/playwright-cli -s=positron click e153        # `click e980 right` for a right-click
./node_modules/.bin/playwright-cli -s=positron press ArrowDown   # Playwright key names, exact case
./node_modules/.bin/playwright-cli -s=positron highlight e153    # overlay before a shot; --hide clears
./node_modules/.bin/playwright-cli -s=positron console warning   # the renderer console, not logFile

• The agent shell is often zsh, which does not split a variable into words: PW="playwright-cli -s=x" then $PW click ... runs as one word and fails. Write the command out, or use a function or an array.
• Use refs from the latest snapshot, never screen coordinates. A selector must match one element, or strict mode fails the step.
• An unknown key name (BackSpace) is rejected and nothing is pressed.
• Do not type or fill into Monaco: editor.sh for files, monaco-paste.sh for chat inputs (it replaces the whole text).
• To capture a ref in a script, query --json snapshot with jq on .role and .name rather than grepping the YAML.
• A screenshot is taken with shot.sh, which logs it; pass --hires to a raw screenshot only for detail finer than a CSS pixel.

What the harness changes

A finding that depends on one of these is about the harness, not the product:

• Modal message boxes are in-app (window.dialogStyle: "custom"), so CDP can click them. A real user sees the native dialog: judge its wording here, not its look.
• The app runs with umask 077, so files it saves are mode 600.
• Workspace trust is off, the keychain is mocked, Welcome is skipped, and the first-run prompts are suppressed (see the launch.sh header).
• A terminal created while hidden, or a window resized through CDP, can draw text at the wrong size. Compare with a terminal opened by hand at the same size before reporting it.

Positron behavior to account for

• Keyboard shortcuts do nothing while focus is in a webview (the Viewer, an HTML output, an app). The helpers move focus out first; before a raw key press, click the editor or a pane.
• With several sessions, the active console is the one used last; code typed into "the console" lands in the wrong one. console-run.sh names the language. Keys typed into "the terminal" go wherever focus is; terminal-run.sh checks.
• In a .qmd with a cell running, Escape is Quarto: Interrupt Kernel, wherever focus is (see Principles). Do not press it to close something.
• In a notebook in command mode, typed keys edit cells (3 makes a cell Markdown). Restore focus to the notebook before a notebook action: opening an output in a plot tab moves it away.
• positron.notebook.enabled defaults to true, so an .ipynb opens in the Positron notebook editor. Reopen Editor With sticks to the file. Notebook cells are [role=article], not <article>.
• Notebook: Run All Cells is hidden while a cell runs; the toolbar shows Stop Execution in its place.
• A fresh window starts sessions by itself: Python when the workspace has a venv, R when it has .R files, some seconds after launch. panel.sh sessions lists what is open; start-session.sh reports an open session of the language rather than starting a second (--new starts one anyway).
• A fresh profile may find a bare interpreter without matplotlib or pandas. Give the run a venv (run-venv.sh, linked as the workspace .venv), and let interpreter discovery and extension installs finish before deciding a kernel is missing.
• Lists, trees and quick picks draw only the rows in view, moved with a transform, and closed pickers stay in the DOM. Scroll or filter before saying a row is missing, and read references/reading-ui-state.md before trusting any other reading of one.
• The Viewer's frames get new refs after every reload or app restart; a frame that failed to load shows only as iframe, the same as one still loading.
• Two things are called a modal: Positron's React .positron-modal-dialog-box (what the e2e Modals page object matches) and the upstream .monaco-dialog-box, which a showInformationMessage(..., { modal: true }) raises even from inside the first.
• Expect selectors to change. The helpers keep theirs in scripts/selectors.ts; the page objects under test/e2e/pages/ are maintained too; otherwise take a fresh snapshot.

Troubleshoot

• Attach reports connect ECONNREFUSED: the app exited after opening CDP. Read logFile.
• listen EINVAL, or an IPC path over 103 characters (macOS, Linux): the run-directory base is too long. Unset $POSITRON_LAUNCH_TMP or shorten it.
• Snapshot refs disappear: look for a modal dialog in a screenshot, then snapshot again. After a reload, attach again.
• Keys stop working in every editor (editor.sh key fails with "the key changed nothing"; cursor, text and unsaved state stay as they were): window.sh reload clears it.
• A built-in extension does not load: npm run gulp compile-extensions. Do not rely on watch-extensions when another extension stops the watch task.
• Confirm an instance is gone: curl -sf -o /dev/null "http://127.0.0.1:$CDP_PORT/json/version" && echo running || echo stopped works on every platform.

launch.sh is a fork of .agents/skills/launch/scripts/launch.sh and does not inherit its fixes; when the upstream changes, port what applies and keep the Positron profile, path, isolation and liveness behavior. Read .agents/skills/launch/SKILL.md for the debug-port mapping, dap-cli breakpoints and parallel instances.