laststance/qa-electron

/qa-electron — Systematic Electron Desktop-App QA

Codex Compatibility

When running this skill in Codex, translate Claude Code-only primitives before acting: AskUserQuestion -> chat/request_user_input, TodoWrite -> update_plan, Task/TaskCreate/TeamCreate/SendMessage -> spawn_agent/send_input/wait_agent when available and allowed, and EnterPlanMode/ExitPlanMode -> a concise chat plan plus explicit approval. Resolve Read/Write/Edit/Bash/WebSearch/WebFetch to Codex file/shell/web tools, and map ~/.claude/... paths to ~/.agents/... or ~/.codex/... unless the task explicitly targets Claude Code.

Cursor Compatibility

When running this skill in Cursor Agent, translate Claude Code-only primitives before acting: AskUserQuestion -> AskQuestion; TodoWrite -> Cursor TodoWrite or an equivalent checklist; Task/TaskCreate/TeamCreate/SendMessage/multi-agent flows -> Cursor Task (subagents), parallel Tasks, or run_in_background when allowed (TeamCreate/SendMessage may have no exact match); EnterPlanMode/ExitPlanMode -> Plan mode (SwitchMode / CreatePlan) plus explicit user approval. Resolve Read/Write/Edit/StrReplace/Bash/web/search/MCP via Cursor Composer or Agent equivalents. MCP names written as mcp__server__tool typically map to call_mcp_tool with configured server identifiers. Map ~/.claude/... to ~/.cursor/skills/, .cursor/skills/, and .cursor/rules/ unless the task explicitly targets Claude Code.

Drive an Electron app running on the host machine, collect evidence (snapshots, screenshots, logs, menu-bar coverage), grade issues, and produce a report. Fixing is out of scope — this skill reports. If the user wants to fix bugs, hand the report to a session that has access to the app's source tree.

Why this exists

Web QA tools reach the renderer, but Electron is two processes and several windows. A bug can live in any of:

• The renderer (Chromium): a button doesn't react, layout breaks at certain widths, a form submits twice. Standard web-QA territory, reached via CDP.
• The main process (Node.js): an IPC handler crashes, auto-updater loops, the app hangs on quit. Visible only in the app's log files or stderr.
• The native OS surface: a menu item is disabled when it shouldn't be, the tray icon menu doesn't reflect state, the file-drop target ignores the drop. Neither CDP nor the DOM knows about these — you need a desktop-automation tool or the user's eyes.
• The platform boundary: the app behaves correctly on macOS and crashes on Windows. Or the window controls are on the wrong side. Or it doesn't open at all on Linux because of a glibc mismatch.

This skill codifies the minimum systematic path that catches the most common classes of Electron bugs without reading the app's Swift/TypeScript source.

On tooling: electron-playwright-cli was the previous primary driver, but its daemon require()s a playwright/lib/mcp/browser/* path that no published playwright artifact ships, so it does not run on a clean host. This skill standardizes on playwright-cli (Microsoft's standalone CLI) attached to Electron's CDP port. The trade-off: CDP attach reaches the renderer (≈95% of practical QA), but native BrowserWindow inventory and main-process IPC introspection are unreachable until upstream ships a working electron-playwright-cli build. For the parts that need them, this skill falls back to osascript / computer-use MCP / log tailing.

Scope and non-goals

In scope: Black-box testing of an Electron app on the current host — driving the renderer via playwright-cli attach --cdp=…, driving the native surface via osascript (macOS) / PowerShell (Windows) / computer-use MCP, reading log files, comparing against cross-platform conventions.

Out of scope:

• Source-code reading or editing (delegate fixes elsewhere)
• Cross-OS testing on a single host — this skill tests the OS it runs on. If the user needs all three OSes covered, run the skill once per host.
• Auto-updater full E2E (requires a staged release channel — out of a single QA run's scope)
• Performance profiling (Chromium DevTools Performance panel, heap snapshots)
• Notarization / code-signing certificate chain audit
• Production-telemetry review (Sentry, Crashpad aggregated reports)

Required context before starting

If any of these are unknown, use AskUserQuestion to collect them:

1. Launch command — how to start the app with CDP exposed on a known port. For project apps, this is usually pnpm dev configured to pass --remote-debugging-port=9222 to Electron. For installed third-party apps, run the binary manually (e.g. open -a "Slack" --args --remote-debugging-port=9222).
2. CDP port — defaults to 9222. Confirm via lsof -i :9222 after launch. If the app is already running without the flag, it must be restarted — CDP cannot be turned on at runtime.
3. Scope — "the whole app" vs "only the onboarding flow" vs "only the settings window". A scoped run is cheaper and usually more useful.
4. Host OS — the skill adapts its checks to macOS / Windows / Linux. If unclear, run uname -s.
5. Log location — where does the app write its log file? Common defaults:
   • macOS: ~/Library/Logs/<AppName>/main.log
   • Windows: %APPDATA%\<AppName>\logs\main.log
   • Linux: ~/.config/<AppName>/logs/main.log If the user doesn't know, ask or inspect the app's userData folder.

Do not guess. Missing context causes the skill to automate the wrong app or miss real issues.

Before any browser interaction: invoke /dnd to load the drag-and-drop verification protocol. Required even when DnD is not yet known to be involved — ref-based drag returns false success on dnd-kit and similar libraries running inside the Electron renderer.

---

Phase 0: Launch and baseline

Goal: clean start, app launched with CDP exposed, playwright-cli attached, baseline screenshot + DOM snapshot recorded, log tail running.

1. Confirm no stale playwright-cli session is bound:

   playwright-cli list
   

   If default (or a previously named session) is still attached from an earlier run, detach it before re-attaching:

   playwright-cli --s=default detach
   # for a wedged daemon:
   playwright-cli kill-all
   

2. Launch the app with CDP enabled (port 9222 by default). For a project in dev mode this is usually:

   pnpm dev          # the project must pass --remote-debugging-port=9222 to Electron
   

   For an installed third-party app:

   open -a "<AppName>" --args --remote-debugging-port=9222   # macOS
   

   Verify the port is listening before attaching:

   lsof -i :9222
   

3. Start a log tail in the background:

   mkdir -p /tmp/qa-electron-session
   tail -F "<log-path>" > /tmp/qa-electron-session/app.log 2>&1 &
   echo $! > /tmp/qa-electron-session/log.pid
   

   Replace <log-path> with the log file from the required-context step. On macOS, log stream --process "<AppName>" is an alternative that also catches stderr (requires sudo on some versions — skip if it prompts).

4. Attach playwright-cli to the running CDP port. The attach is one-shot per session; subsequent commands target it via --s=default:

   playwright-cli attach --cdp=http://localhost:9222
   playwright-cli --s=default snapshot --filename=/tmp/qa-electron-session/00-snapshot.yaml
   

   If attach errors with ECONNREFUSED, the app didn't start with CDP enabled (re-launch) or the port is wrong (lsof -i :9222). If snapshot returns nothing useful, the renderer may be Canvas/WebGL with no accessible DOM — note it and fall back to screenshot-only.

5. List every webview/tab target reachable via CDP:

   playwright-cli --s=default tab-list           # webview/tab inventory across attached pages
   

   Record the output — this is the target inventory for Phase 2 (matrix generation) and Phase 4 (multi-window coverage). Typical Electron app: 1 main page + 0–N webviews. If you see >5, confirm with the user that's expected.

   Limitation: CDP attach does NOT enumerate native BrowserWindow instances directly — it sees their attached pages. If the app uses hidden / unattached BrowserWindows, those won't appear here. For secondary windows, open them via the UI / menu and they will show up in tab-list once their renderer is alive.

6. Baseline — main window only for now:

   playwright-cli --s=default tab-select 0
   playwright-cli --s=default screenshot --filename=/tmp/qa-electron-session/00-baseline.png
   

Phase 1: Surface mapping

Goal: inventory every surface that needs testing before diving deep.

An Electron app's surfaces:

• Main window — the primary UI
• Secondary windows — preferences/settings, about, onboarding, command palette, modal editors
• Menu bar (macOS) or hamburger menu (Windows/Linux) — File, Edit, View, Window, Help, plus app-specific menus
• Tray icon (system tray / menu bar extra) — right-click menu
• Notifications — only visible when fired
• Deep-link protocol handlers — slack://, vscode://, etc.

Walk them breadth-first:

1. Main window routes — use playwright-cli --s=default snapshot to see what's interactive, click through top-level nav, screenshot each route.
2. Every menu bar top-level item (macOS):

   osascript -e 'tell application "System Events" to tell process "<AppName>" to get title of every menu of menu bar 1'
   

   For each top-level menu, enumerate its items:

   osascript -e 'tell application "System Events" to tell process "<AppName>" to get title of every menu item of menu "File" of menu bar 1'
   

   Record the full menu tree. This is the checklist for Phase 5 (native OS integration).
3. Secondary windows — open each via the main window's UI (Preferences, About, etc.) or via the menu (Cmd+, for preferences is a macOS convention). After each open: playwright-cli --s=default tab-list → note any new webview/tab target → switch to it with tab-select → screenshot + snapshot. Hidden / unattached BrowserWindows won't appear in tab-list; for those, capture via mcp__computer-use__screenshot while they're foregrounded.
4. Tray icon — if the app has one, right-click it (use computer-use MCP right_click on the tray icon's screen coordinates) and screenshot the menu. Record the items.

Budget: 3–6 minutes. If the app has >20 menu items across >5 top-level menus, ask the user which are load-bearing for this run.

Phase 2: Generate the test matrix

Goal: before touching anything else, produce a checksheet of every happy-path case this run will execute. The matrix forces completeness (no feature slipped), makes progress legible to the user (N/M cases passed), and produces a coverage number worth quoting in the report.

Scope of the matrix: happy path only — every advertised feature has at least one case exercising its primary success path. Edge cases, race conditions, and creative "what-if" scenarios belong to Phase 9 (Exploratory testing), not the matrix.

1. Read references/test-matrix-generation.md — gives the nine derivation rules for turning Phase 1's inventory into matrix rows.
2. Read templates/test-matrix-template.md — the matrix skeleton.
3. Walk the Phase 1 inventory surface by surface, applying the nine rules in order. Assign TC-001, TC-002, ... in execution order (group by surface so the agent isn't thrashing between windows).
4. Save the matrix to ./qa-reports/electron-<date>-<app>-<os>/test-matrix.md.
5. Present a coverage summary to the user before executing:

   "Generated N cases across S surfaces. Estimated execution: M minutes. Plan to execute: (a) Full, (b) Core only, (c) Smoke, (d) adjust?"

6. On user confirmation, proceed to Phase 3.

Happy-path coverage target: ≥ 95% of matrix cases executed + passing. Below 95% means either the app has real bugs (Phase 10 will triage them) or cases were blocked (document the blockers in the matrix).

Do not execute the matrix during Phase 2. Generation and execution are separate phases — mixing them produces under-derived matrices because the agent gets distracted running the first interesting case.

Phase 3: Execute the test matrix

For every row in the matrix from Phase 2, work this checklist. The matrix tells you what to run; the subsections below are the how per-row — what to notice beyond pass/fail. For each row:

1. Verify preconditions (cold launch or the named prior state)
2. Execute the Steps column
3. Compare the UI outcome against Expected
4. Screenshot before + after; save snapshot to snapshots/TC-NNN.txt
5. Mark Status: PASS / FAIL / BLOCKED in the matrix
6. On FAIL: fill the Failures detail block in the matrix AND flag a candidate issue for Phase 10 (Triage). On PASS: move on — don't gold-plate a passing case.

Re-snapshot (playwright-cli --s=default snapshot) after any action that mutates the DOM — eN refs are valid only for the most recent snapshot.

Most bugs fall out of a matrix run plus the lenses below — the point is to be systematic, not clever.

Visual scan

• Screenshot → eyeball for overlaps, truncation, missing assets, broken layout
• playwright-cli --s=default snapshot → look for elements with empty name (probably unlabeled buttons), elements with role button but no keyboard focus marker, obvious placeholder text ("Lorem ipsum", "TODO", "Untitled")
• Compare against references/cross-platform-conventions.md — window controls on the right side? Menu bar lives at top? Standard shortcut keys?

Interactive surface

Enumerate every interactive element from the snapshot. For each:

• playwright-cli --s=default click eN → screenshot → check: expected action, no unhandled exception in the log tail, no new window opened unexpectedly
• If the click opens a modal: can it be dismissed? Esc, click outside, an explicit Close/Cancel button — at least one of these should work. If none do → critical (focus trap)
• If the click opens an external URL: does it go to the OS default browser (correct for most apps) or does it open inside the Electron window (a security issue unless the app intentionally embeds that URL)?

Forms and keyboard

• Every input field: empty submit, valid input, invalid input, overflow input (2000+ chars), paste-heavy input. Use playwright-cli --s=default fill eN "..." and playwright-cli --s=default press Enter (or Tab).
• Shortcut coverage — standard shortcuts MUST work:
  • Cmd/Ctrl+C/V/X/A/Z in any text field
  • Cmd/Ctrl+W closes the focused window (or tab, depending on app)
  • Cmd/Ctrl+Q quits (macOS) / Alt+F4 closes (Windows)
  • Cmd/Ctrl+, opens preferences (macOS convention, less strict on Win/Linux)
  • Tab cycles focus through interactive elements in reading order

Per-window sanity

For each secondary window found in Phase 1:

• Can it be resized? Minimum size reasonable?
• Can it be closed without orphaning a parent?
• Does it remember position/size across relaunches? (close app → relaunch → reopen window → check)

Phase 4: Multi-window + webview coverage

Run every primary flow in both the main window AND any secondary windows or webviews that share surface. Common bugs:

• A feature works in the main window but crashes when invoked from the preferences window (different webview = different JS context)
• State saved in the main window is not visible in the secondary window until relaunch (stale cached props)
• A webview receives events that the main window swallows (or vice versa)

Tool loop:

playwright-cli --s=default tab-list               # list webview/tab targets
playwright-cli --s=default tab-select 2           # switch to webview/tab index 2
playwright-cli --s=default snapshot               # new context
playwright-cli --s=default click e3
# ...
playwright-cli --s=default tab-select 0           # back to main

If the app spawns a new window during a flow (OAuth popup, file picker preview, deep link), playwright-cli --s=default tab-list will show it on the next call only if the new window has an attached renderer. Native BrowserWindow instances without a CDP-attached page are invisible to this path — switch to mcp__computer-use__screenshot plus osascript for those.

Phase 5: Native OS integration

These are the parts neither the DOM nor CDP can see.

| Check | macOS | Windows | Linux | |-------|-------|---------|-------| | Menu bar items all work | osascript click each | keyboard nav via Alt+letter | keyboard nav | | Tray icon menu | right-click menu bar extra | right-click system tray | right-click status-icon | | Notifications appear & click routes correctly | Notification Center | Action Center | varies | | File open dialog returns valid path | trigger + pick file | same | same | | Drag-and-drop receiver accepts files | drag from Finder | from Explorer | from Files | | Clipboard paste from other app | Cmd+V after copy from TextEdit | Ctrl+V from Notepad | Ctrl+V | | Deep link opens app from other context | open "<proto>://..." | start <proto>://... | xdg-open | | App appears in Dock/Taskbar with correct icon | eyeball | eyeball | eyeball | | Quit from Dock right-click fully exits | right-click → Quit | right-click → Close | right-click |

macOS menu-bar driving example:

# Click File → New Window
osascript -e 'tell application "System Events" to tell process "<AppName>" to click menu item "New Window" of menu "File" of menu bar 1'

Screenshot after each trigger using computer-use MCP (since the menu is native, not in the renderer):

mcp__computer-use__screenshot  # captures the entire screen including the native menu

For tray icons and other screen-coordinate-based interactions, use computer-use MCP. Request access first:

mcp__computer-use__request_access({ apps: ["<AppName>"] })

See references/cross-platform-conventions.md for the full list of platform-specific expectations (window controls, shortcut variants, menu structure, notification UX).

Phase 6: States

Force the conditions the OS + network control:

| State | How to force | What to check | |-------|--------------|---------------| | OS Light Mode | macOS System Settings → Appearance → Light | renderer respects it (if app claims to) | | OS Dark Mode | macOS System Settings → Appearance → Dark | renderer respects it, no hardcoded white | | Offline | Network Link Conditioner → 100% loss, or unplug ethernet+wifi | offline banners, retry UI, cached-data visibility, error copy | | Background | click another app to defocus | app pauses timers / animations, doesn't keep CPU hot, doesn't keep network busy when idle | | Window resize (min) | drag the corner to the smallest allowed size | no layout break, no content cut off without scroll | | Window resize (max) | fullscreen | no wasted whitespace, no stretched assets | | Multi-display | drag window to a second monitor | window remembers position, layout adapts to DPI | | OS suspend / sleep | pmset sleepnow (macOS) | app reconnects network on wake, doesn't spin on reconnect loop | | OS account switch / lock | lock screen + reopen | state preserved, sockets resume cleanly |

At minimum: OS theme change + Offline + Window resize to small. These three catch the most state bugs for desktop apps.

Phase 7: Accessibility

• Every interactive element has a non-empty accessible name. Use playwright-cli --s=default snapshot — empty-name buttons / inputs are flagged.
• Keyboard-only navigation — starting from the app's first focusable element, Tab through the entire primary flow. Every action the user might take with the mouse must be reachable via keyboard.
• Screen reader smoke test — macOS: turn on VoiceOver (Cmd+F5) and arrow through a core screen. Does each element read sensibly? Images have alt text? Role is correct (button vs link vs heading)?
• High contrast / Reduce transparency — macOS System Settings → Accessibility → Display → Increase contrast. Does the app stay legible? Any element disappear because it relied on a semi-transparent background?
• Zoom — Chromium Ctrl+= / Cmd+= to zoom the renderer. Does layout survive at 200%? Are any elements clipped at 400%?

See references/issue-taxonomy-electron.md for the full category list.

Phase 8: Security spot-check

Electron has no gatekeeper — the app's author decides what's safe. Spot-check:

• DevTools access — Cmd+Option+I (macOS) / Ctrl+Shift+I. If DevTools opens in a production build, that's usually a misconfiguration (dev builds should have it; production builds usually disable it). Note in report.
• External-link handling — find a user-generated-content area (chat message, note body, etc.) if one exists. Enter a link https://example.com/test. Click it. Does it open in the user's default browser (correct) or inside the Electron window (suspicious — could indicate nodeIntegration exposure)?
• nodeIntegration smoke test — in a renderer console (via CDP):

  playwright-cli --s=default eval 'typeof require'
  

  If the result is "function", nodeIntegration is enabled on this renderer. This is a critical security finding unless the app author has a deliberate reason (and for the main window, there's almost never a good reason — it should be false with contextIsolation: true).
• CSP presence — playwright-cli --s=default eval 'document.querySelector("meta[http-equiv=Content-Security-Policy]")?.content' on the main window. No CSP is not automatically critical, but note it.
• Renderer console — playwright-cli --s=default eval "console.error('qa probe')" plus check the app log for security warnings Electron itself logs (Electron Security Warning messages in the devtools console when running in development). These tell the user their own app is warning them.

Full security audit is out of scope — the spot-check surfaces the most-common misconfigurations.

Phase 9: Exploratory testing (creativity-driven)

The matrix (Phase 3) covered the happy path: every advertised feature passed through its primary success path. What it could NOT cover:

• The gaps between matrix steps (double-click, Esc mid-animation, rapid repeats)
• Sequences the designer didn't anticipate (undo→redo→delete→undo)
• Moments of transition (app sleep, focus loss, network flap)
• Truth vs. Appearance: the UI claims X changed, but did the filesystem / DB / main process actually change?

The last category is the Electron-specific killer. Users lose trust when they hit "Delete" and the item comes back on restart, or when they toggle a preference that doesn't stick. A matrix cannot catch these by design — it checks the UI state, not the underlying store.

Budget

15-25% of total QA time. For a 60-minute run, that's 10-15 minutes. If you blow past 25 minutes without finding anything interesting, stop.

Process

1. Read references/exploratory-heuristics.md — nine tours (lenses) for generating creative test cases on the fly.
2. Pick 2-4 tours based on the app's shape (the reference has a selection cheat sheet matched to app type).
3. The Truth-vs-Appearance tour is mandatory for any Electron app that writes to disk, DB, or network. Keep a shell open alongside the app and verify UI claims against the underlying store.
4. For each tour: state the question out loud, run it, note what happened even if nothing interesting (a "boring" answer is still evidence of a well-protected invariant).
5. Log each finding with: tour used, the specific prompt, expected vs. actual, severity hint, minimal repro, evidence paths.
6. Move findings into the main triage list in Phase 10.

Electron-specific shell probes

Keep these one-liners ready for Truth-vs-Appearance verification:

# Filesystem-backed state
ls -la <path>; readlink <symlink>; stat <file>

# SQLite-backed state (common in Electron apps)
sqlite3 "~/Library/Application Support/<AppName>/db.sqlite" \
  "SELECT COUNT(*) FROM <table>"

# JSON/YAML config
jq '.thatField' "~/Library/Application Support/<AppName>/config.json"

Electron userData locations:

• macOS: ~/Library/Application Support/<AppName>
• Windows: %APPDATA%\<AppName>
• Linux: ~/.config/<AppName>

Stop criteria

Stop exploratory testing when any of:

• 2-4 tours executed with reflective notes on each
• Time budget hit (25% of total)
• Phase 10 (Triage) will run long if you pile on more findings

Do NOT stop after just one tour produced nothing — try a second tour first (different tour = different bug class).

Phase 10: Triage

Take every issue collected across phases and classify:

• severity: critical / high / medium / low (definitions in references/issue-taxonomy-electron.md)
• category: one of security / functional / visual / accessibility / state / content / platform / crash
• scope: which window / process (main / renderer / webview-N)
• repro: minimum steps, starting from a cold launch
• evidence: screenshot path + snapshot snippet + log excerpt if any

De-duplicate aggressively. If the same missing aria-label appears on the same reused component across six screens, that's one issue with a list of affected screens — not six.

Phase 11: Report

Use templates/qa-report-template-electron.md as the skeleton. Fill in:

• App metadata (name, version, Electron version if determinable from playwright-cli --s=default eval 'process.versions.electron', host OS)
• Health score per category (see the template)
• Top 3 things to fix, with issue IDs linking below
• Full issue list, severity-grouped

Save the report to ./qa-reports/electron-<date>-<app>-<os>.md relative to the user's current directory. Save all screenshots and snapshots alongside in ./qa-reports/electron-<date>-<app>-<os>/.

Don't embed full DOM snapshots in the report — they're noisy. Save them alongside and link. Only embed the specific snippet relevant to an issue.

Phase 12: Session cleanup

Always, even if the report isn't finished:

# Kill the log tail
[ -f /tmp/qa-electron-session/log.pid ] && kill "$(cat /tmp/qa-electron-session/log.pid)" 2>/dev/null

# Detach playwright-cli — this does NOT kill the Electron app, only the
# CDP-attached session.
playwright-cli --s=default detach 2>/dev/null || true

# If a daemon is wedged:
# playwright-cli kill-all

Note: playwright-cli attaches via CDP — it does NOT own the Electron process. The dev server (pnpm dev) or whoever launched the app still holds the process; quitting the app is up to the user. This is the opposite of the old electron-playwright-cli model where the daemon owned the process.

Restore any OS-level toggles you changed (Dark Mode, Network Link Conditioner, Reduce Motion). Leaving them flipped creates confusing "why is my screen dark now" moments.

---

Deliverables checklist

Before telling the user "done":

• [ ] qa-reports/electron-<date>-<app>-<os>.md exists and opens cleanly
• [ ] Test matrix saved at qa-reports/electron-<date>-<app>-<os>/test-matrix.md
• [ ] Matrix coverage ≥ 95% (or deferred/blocked cases each documented with a reason)
• [ ] Phase 9 exploratory ran — Truth-vs-Appearance tour executed whenever the app writes to disk / DB / network, findings moved into triage
• [ ] At least one screenshot per issue (not one of the whole app)
• [ ] Each issue has severity, category, scope, repro
• [ ] Health score table is filled (no {SCORE} placeholders left)
• [ ] "Top 3 Things to Fix" has actual issue titles
• [ ] Cross-platform findings are tagged with which OS was tested
• [ ] Native OS integration phase ran, not skipped
• [ ] Log tail process was killed
• [ ] No OS-level toggles left flipped

---

Reference files

• references/issue-taxonomy-electron.md — severity + category definitions
• references/electron-agent-browser-reference.md — playwright-cli + CDP commands used by this skill (file name is historical; contents are the current cheat sheet)
• references/cross-platform-conventions.md — macOS / Windows / Linux expectations the skill checks against
• references/test-matrix-generation.md — nine derivation rules for Phase 2 matrix rows (near-100% happy-path coverage)
• references/exploratory-heuristics.md — nine tours for Phase 9 creativity-driven testing (Truth-vs-Appearance, Goldfish, Wrong-Order, etc.)
• templates/qa-report-template-electron.md — fill-in report skeleton
• templates/test-matrix-template.md — matrix checksheet skeleton (coverage table + per-surface case rows)

Load the references on demand (they're not in context by default). Read test-matrix-generation.md at the start of Phase 2, exploratory-heuristics.md at the start of Phase 9, and the taxonomy / conventions when a triage question needs them.

---

Escape hatches

• Daemon stuck / commands hang: playwright-cli list to see active sessions. playwright-cli --s=default detach for graceful, or playwright-cli kill-all to force-terminate every session.
• attach --cdp fails with ECONNREFUSED: the app didn't start with --remote-debugging-port=9222 (re-launch with the flag) or the port is already taken (lsof -i :9222). For project apps using pnpm dev, confirm the dev script passes the flag through to Electron.
• App launches but tab-list is empty: the renderer hasn't reached first paint yet — wait a few seconds and re-run, or watch for splash screens that delay attaching the main page. If still empty, the app may have crashed silently — check the log tail.
• playwright-cli snapshot returns nothing useful: the renderer might be rendering with Canvas or WebGL with no accessible DOM. Fall back to visual-only (playwright-cli --s=default screenshot) and note AX coverage is limited.
• Native BrowserWindow not visible to CDP: unattached / hidden BrowserWindow instances don't appear in tab-list. Capture them with mcp__computer-use__screenshot while foregrounded; introspect main-process state via the app's logs or by adding temporary IPC probes (out of scope for a report-only run).
• Menu-bar osascript returns "not authorized": macOS System Settings → Privacy & Security → Accessibility → add Terminal (or the relevant agent host). Do not work around this — it's the user's consent boundary.
• Crash on launch, no obvious cause: check the app log, then macOS Crashlytics-style dumps at ~/Library/Logs/DiagnosticReports/<AppName>*. Quote the first 20 lines + crashed thread into the report.
• Third-party installed app (Slack, VS Code, etc.): launch the app's binary directly with --remote-debugging-port=9222. macOS: open -a "Slack" --args --remote-debugging-port=9222. Some apps strip custom flags; in that case launch the inner binary directly (/Applications/Slack.app/Contents/MacOS/Slack --remote-debugging-port=9222).