Web Browser#

Environment-aware browser operations skill. Detects the runtime environment and routes to the appropriate browser backend, then runs one of two workflows: UI testing/verification ( ui-test) or browser-login-assisted credential issuance ( credential-issue).

Topics#

TopicDescriptionGuideui-testSnapshot analysis, click/fill/verify, page-state diagnosisui-test.mdcdp-traceCDP-based closed shadow DOM cascade diagnosis (DOM.getDocument pierce:true + CSS.getMatchedStylesForNode)cdp-trace.mdcredential-issueservice+command param → open login screen → wait for user login → issue access key/token/secret → hand off to automationcredential-issue.md

Topic Dependencies#

text

web-browser (Step 0: environment detection — shared by all topics)
  ├─→ ui-test (UI verification)
  │     └─→ cdp-trace (extends ui-test for closed shadow DOM)
  └─→ credential-issue (browser-login-assisted token/key issuance)
        └─→ chrome-devtools backend preferred (reuses the user's real logged-in session)

• Step 0 (below) is shared — every topic detects the backend first, then runs its workflow.
• ui-test, cdp-trace are the UI-testing family.
• credential-issue reuses the same backend routing + the user-visibility rule, generalized into a service+command parameterized auth flow.
• Authentik SSO verification ( sso-verify) is not included in this skill — it remains in a separate local-only sso-verify skill (user-environment specific, untracked).

CRITICAL — capturing a credential-input screen requires an explicit ask (HARD STOP)#

Before capturing a sign-in / credential-input screen — accessibility snapshot, screenshot, or any full page-content read — call AskUserQuestion and get explicit approval. Applies to every topic in this skill and to every backend.

The reason is not privacy etiquette, it is a measured leak path: a browser profile's saved-password autofill populates the password field, and the accessibility tree renders that field's value in plaintext. The capture therefore carries a live credential into the transcript even though nothing was typed and no screenshot of characters was taken. document.body.innerText does not expose input values, but the accessibility snapshot does.

#Don'tDo1Snapshot a sign-in page to "see what step we're on"Read only location.href and document.title, or specific marker booleans. Ask before any fuller capture2Assume an empty-looking form is safe because you typed nothingProfile autofill fills fields without any typing. Emptiness is not verifiable before the capture that would leak it3Redact only the field you expected to carry the secretApply redaction to every returned field — element text, aria-label, title, placeholder, value. A secret leaked through an unredacted sibling field is the common failure4Return page values when a count or boolean answers the questionPrefer structure over content: hasPasswordField: true, blockMarkers: {...}, buttonCount: 35Treat "the user asked me to test the login" as approval to capture the screenTesting the flow and capturing the credential screen are separate permissions. Ask for the second one explicitly

Self-check (before every snapshot / screenshot / full-text read): does the current page accept a password, token, secret, or OTP? → If yes, or if unsure, restrict to URL + title and ask before capturing more.

CRITICAL — user visibility is the top priority (HARD STOP)#

The primary purpose of browser diagnosis/verification is "the user sees it on their own screen". Screenshot capture is supporting evidence, not a substitute for visibility.

#Don'tDo1Launch with chromium.launch({ headless: true }) and only attach a screenshot in chatchromium.launch({ headless: false, slowMo: 500 }) — let the user follow in real time2"I showed the user a screenshot, so it's fine"screenshot ≠ visible to the user. If the user says "show me", open a visible browser + slowMo3wmux/cmux/Playwright MCP disconnected → fall back to headless CLIEven on CLI fallback, force headless: false. On a Windows desktop OS, a chromium GUI is available4"headless is faster and more stable by default" mindsetSpeed costs user visibility. If the user says "show me", visibility wins5Playwright MCP disconnected → CLI fallback auto-selects headlessCLI fallback is also headless: false. headless is only for explicit non-interactive cases (e.g., CI assertion)6SaaS/API task lacks credentials → fallback to manual user UI operationDo NOT recommend manual user UI clicking when API access is available; fallback to credential-issue topic to issue token/key first

API-capable environment without credentials — fallback to credential-issue (HARD STOP)#

When a task can be performed via API (e.g., Google Forms API, GitHub API, AWS API), but required API tokens or access keys are missing in the environment, do NOT recommend manual user UI clicking or surrender to direct manual UI operation. You MUST recommend credential-issue topic to issue the access key/token via browser login first, then proceed with backend API automation.

#Don'tDo1API token missing → "Please edit/click manually on the website"Recommend credential-issue topic to issue API token/key via browser login2Direct UI automation fails → fallback to manual user operationCheck if API automation is available → issue credential via credential-issue → execute API

Self-check (every time before launching Playwright/chromium)#

1. Did the user use a visibility request keyword such as "show me", "open it", "web-browser", or "browser test"? → If yes, force headless: false
2. Is this work interactive verification or diagnosis for the user? → If yes, headless: false
3. headless is justified only when (a) CI assertion (b) the user explicitly said "in headless" (c) Playwright MCP is used (the UI shows itself)
4. screenshot is supporting evidence — it can be attached to a chat report, but it does not replace user visibility

Violation case (2026-05-28, 1st)#

During a closed shadow DOM ak-library cascade investigation, used a npx playwright Bash invocation + chromium.launch({ headless: true }) and only attached a screenshot in chat. The user requested "show it via web-ui-test" and no visible browser was provided. The user reacted angrily that the Chromium UI never appeared.

Login wall mid-capture — ask before stopping, don't silently defer (HARD STOP)#

When a capture/documentation task (report evidence, purchase/registration flow guide, etc.) hits a screen that requires login, and completing that login would reveal materially different information than what's already captured (e.g., the real final price vs. a promotional pre-login price, actual post-login UI state vs. an assumption), do NOT silently stop and paper over the gap with a deferral disclaimer. Ask the user via AskUserQuestion whether to continue (via interactive login in a visible backend) or whether the pre-login capture is sufficient for the purpose at hand.

#Don'tDo1Hit a login wall → write "please have finance/ops enter payment details themselves for security" and stop, without askingDecompose the remaining flow: payment/credential entry should be deferred to the user/business owner, but login + viewing the resulting screen is often just informational — ask which is actually needed before deciding to stop2Treat "login" and "entering payment info" as one bundled decision to skip togetherThey are different risk levels. Login-then-observe (e.g., see the real cart/checkout price) does not require entering card/account credentials — only the latter needs deferral3Report a pre-login/promotional price or state as if it were final, without flagging the gapIf the login-gated final screen wasn't verified, explicitly flag it ("actual payment screen not verified — may differ from the listed price") instead of presenting the pre-login figure as authoritative4Assume the backend can't support interactive login without checkingCheck chrome-devtools connection + visibility (per credential-issue.md "Fresh-login flow") first; if visible, open the page there and have the user sign in in that same window, then continue capturing5Decide unilaterally that "this is good enough" when the report's factual accuracy depends on the gated screenIf the gap could make a delivered report/guide factually wrong (e.g., a payment-request report citing a price that turns out incorrect), the stop-vs-continue decision belongs to the user, not the assistant

Violation case (2026-07-22, 1st)#

While building a domain-registration payment-request report, captured the domain-search-result page (showing a promotional price) and the login screen, then stopped at the login wall with a disclaimer ("have finance/ops enter payment details"), never asking whether to continue via login to verify the real checkout price. The report's stated price differed from the actual payment-screen price. User feedback (paraphrased): "don't arbitrarily skip capturing screens that require login — ask first."

Known Automation Limitations — SaaS Portal Action-Level CAPTCHA Gates#

Some SaaS portals allow full browser login automation but selectively trigger CAPTCHA challenges on creation/mutation actions (not just on login). Document confirmed cases here so agents do not repeat failed automation attempts.

ServiceAutomatableCAPTCHA-blockedFallbackDiscord Developer PortalLogin (via persistent profile with saved credentials)New application creation, bot token resetKeep browser visible ( headless: false); user handles hCaptcha manually; script polls page.url() for /bot URL and auto-captures token once user navigates thereDiscord Developer PortalReading existing app info, navigating between tabs_(same)__(same)_

Discord Developer Portal — specific notes (2026-08-18, 1st confirmed)#

• Login: Playwright persistent context ( launchPersistentContext) with a saved user data directory retains Discord session cookies. Navigation to discord.com/developers/applications succeeds without re-authentication.

• Bot creation blocked: Clicking "New Application" and submitting the modal triggers an hCaptcha dialog (e.g. "Hold on! You are human, right?"). The force: true checkbox click and JS dispatchEvent workarounds successfully activate the Create button, but Discord's backend detects the automated browser and intercepts submission with CAPTCHA.

• Recommended hybrid flow:

  1. Launch Playwright with headless: false + launchPersistentContext (reuses login session).
  2. Navigate to the applications page.
  3. Set up a polling loop watching page.url() for the /bot path (every 2s, max ~4min timeout).
  4. Inform user to manually create the application (handle hCaptcha) and navigate to the Bot tab.
  5. When /bot URL is detected, script resumes: click "Reset Token" → capture input[readonly] value → enable [role="switch"] intents → save changes.
  6. Write token to a temp file → hand off to next automation (K8s Secret injection, etc.).

---

Step 0: Environment & Tool Priority Resolution (MANDATORY — before any browser action)#

Resolve which browser capability backend to invoke based on the current environment, prioritizing native/editor plugins over MCP servers.

Step 0a: Host OS layer — WSL detection (MANDATORY, runs before the Tool Priority Matrix)#

The multiplexer matrix below answers which tool drives the browser. It does not answer which OS the browser process runs on — and for the CDP-hostile services in credential-issue.md that second axis decides whether automation works at all. A Playwright MCP server started inside WSL drives a Linux Chrome, a different fingerprint surface from chrome-devtools-mcp attaching to the user's Windows Chrome. Resolve this layer first, record the result, and never carry a CDP-hostile verdict across it.

bash

# WSL detection — either signal is sufficient
[[ -n "$WSL_DISTRO_NAME" ]] && echo "host=wsl"
grep -qi microsoft /proc/version 2>/dev/null && echo "host=wsl"

Host layerRecord asConsequence for CDP-hostile servicesWSL (either signal true)host=wslPlaywright MCP (Linux Chrome) is a first-class candidate, not a skipped backend — measure before escalatingWindows, no WSL signalhost=windowsThe documented Windows chrome-devtools-mcp findings apply as writtenmacOS / Linux nativehost=nativeUnmeasured for these services — treat as unknown and measure before asserting#Don'tDo1Jump straight to the Tool Priority Matrix and pick a backendResolve host= first — it gates the CDP-hostile escalation ladder2Apply a CDP-hostile verdict recorded on one host layer to anotherVerdicts are per host layer. A Windows block is not a WSL block3Escalate a CDP-hostile service to wmux/cmux/OS-launch while host=wsl is untestedOn host=wsl, try Playwright MCP first and record the outcome before escalating

Tool Priority Matrix#

Environment1st Priority (Native Plugin / CLI)2nd Priority (MCP Server Fallback)wmuxwmux browser commands via Bash (User-visible)Playwright MCP ( mcp__playwright__*)cmuxcmux browser commands via Bash (User-visible)Playwright MCP ( mcp__playwright__*)Plain / tmuxPlaywright MCP ( mcp__playwright__*) (Headless default)Headless Playwright CLI via Bash

Backend Availability Check#

bash

# wmux check
[[ -n "$WMUX" ]] || command -v wmux >/dev/null 2>&1

# cmux check (detect via ANY of these; CMUX_SESSION is NOT set by cmux app)
[[ -n "$CMUX_BUNDLE_ID" || -n "$CMUX_PANEL_ID" || -n "$CMUX_BUNDLED_CLI_PATH" ]] || command -v cmux >/dev/null 2>&1

Do & Don't — Browser Backend Selection#

EnvironmentDetect (ANY true → environment matches)Do (use this)Don't (forbidden)wmux$WMUX set OR command -v wmux succeedswmux browser open/snapshot/click/type commands via BashPlaywright MCP — user cannot see the invisible Playwright windowcmux$CMUX_BUNDLE_ID set OR $CMUX_PANEL_ID set OR $CMUX_BUNDLED_CLI_PATH set OR command -v cmux succeeds (e.g. /Applications/cmux.app/Contents/Resources/bin/cmux)cmux browser panel commandsPlaywright MCP — same reasonPlain / tmuxNone of wmux/cmux signals presentPlaywright MCP (Step 1 below)—

cmux detection — multi-var OR rationale

cmux app sets several env vars when launching a shell, but CMUX_SESSION is NOT one of them (a legacy guess by analogy with WMUX). Real vars observed in a cmux-launched shell:

• CMUX_BUNDLE_ID (e.g. com.cmuxterm.app)
• CMUX_PANEL_ID (UUID per panel)
• CMUX_BUNDLED_CLI_PATH (CLI absolute path)
• CMUX_SHELL_INTEGRATION_DIR
• CMUX_AGENT_LAUNCH_*
• GHOSTTY_RESOURCES_DIR (cmux uses Ghostty-based terminal)

CMUX_SOCKET is set but often empty — do not use it as the sole signal. Use the OR matrix above.

#Don't (single-var assumption)Do (multi-var OR)1[ -n "$CMUX_SESSION" ] only check → false negative on cmux appOR across CMUX_BUNDLE_ID / CMUX_PANEL_ID / CMUX_BUNDLED_CLI_PATH2Use CMUX_SOCKET as detection (empty in many cases)Treat empty CMUX_SOCKET as no-signal; rely on the 3 vars above + CLI presence3Assume cmux env var name mirrors wmux ( *_SESSION)Verify against actual cmux app shell environment — vars differ per terminal multiplexer

wmux Browser Commands Reference#

When $WMUX is set, use these instead of Playwright MCP.

Invocation form: the rest of this document uses the bare wmux browser … form, which is what runs when wmux is on PATH (the common case). If wmux is not on PATH in the current environment, substitute node "$WMUX_CLI" for wmux in every command below — $WMUX_CLI points to the same entry point. The two forms are interchangeable; pick whichever resolves on the current shell and use it consistently.

bash

wmux browser open <url>          # navigate (= playwright navigate)
wmux browser snapshot            # get accessibility tree with @eN refs
wmux browser click @eN           # click element
wmux browser type @eN <text>     # type into element
wmux browser fill @eN <value>    # set input value
wmux browser get-text            # get page text
wmux browser screenshot          # capture screenshot
wmux browser eval <js>           # run JavaScript
wmux browser back                # go back
wmux browser forward             # go forward
wmux browser reload              # reload page

Workflow: browser open <url> → browser snapshot → read tree → browser click/type @eN → browser snapshot again.

Refs ( @e1, @e2...) expire after page changes — always re-snapshot.

Do & Don't — wmux vs Playwright Mapping#

Actionwmux (Do)Playwright MCP (Don't in wmux)NavigateBash("wmux browser open <url>")``mcp__playwright__browser_navigateSnapshotBash("wmux browser snapshot")``mcp__playwright__browser_snapshotClickBash("wmux browser click @eN")``mcp__playwright__browser_clickTypeBash("wmux browser type @eN text")``mcp__playwright__browser_typeScreenshotBash("wmux browser screenshot")``mcp__playwright__browser_take_screenshotEvaluate JSBash("wmux browser eval <js>")``mcp__playwright__browser_evaluateWait for textRe-snapshot + checkmcp__playwright__browser_wait_for

Key difference: wmux browser is visible to the user in real-time on the right panel. Playwright opens an invisible window the user cannot see.

---

Quick Reference#

After Step 0 backend detection, route to the topic:

GoalTopicEntryVerify a UI change, snapshot, click/fillui-testui-test.mdDiagnose ::part not applying / closed shadow DOM cascadecdp-tracecdp-trace.mdOpen a service login → wait for user login → issue access key/tokencredential-issuecredential-issue.md

Step execution order: Step 0 (this file — detect backend + user-visibility rule) → read the target topic .md → follow its procedure. The topic .md files hold the actual procedures; this file is the shared backend-detection + index.