/live-inspect — Live Extension Inspection

Launch a real VS Code instance with GitLens loaded, then inspect UI elements, read logs, interact with views, and evaluate runtime values — all programmatically via Playwright.

MCP Server (Preferred for Iterative Inspection)

The vscode-inspector MCP server provides a persistent, interactive session. It launches VS Code once and exposes tools for screenshot/click/inspect/rebuild cycles — much faster than the batch CLI for agentic feedback loops.

The server is auto-discovered via .mcp.json when Claude Code starts in this repo. When connected, these MCP tools are available:

Reading host logs (gotchas)

• read_logs uses pattern, NOT filter. A wrong/unknown arg is silently ignored and it falls back to pattern: "GitLens" — so you only ever see the activation banner and wrongly conclude "no logs". Always: read_logs({ pattern: "<tag>", last_n }).
• read_logs DOES capture extension-host console.log/warn/error and GitLens Logger.warn/info/error (info+). It reads the GitLens LogOutputChannel, also on disk at .vscode-test/user-data/logs/<TS>/window1/exthost/eamodio.gitlens/GitLens.log. read_console is webview-only (does not see host logs).
• @debug/@trace decorator logs are filtered out. The channel defaults to info; GitLens' rich tracing is debug/trace. Console-mirroring (Logger.isDebugging) is gated on ExtensionMode.Development, but the inspector runs in Test mode → off. gitlens.enableDebugLogging (which runs workbench.action.output.activeOutputLogLevel.debug) does NOT take headless. So for ad-hoc host tracing, instrument with Logger.warn('[tag] …') (info+, always written) and read via read_logs({ pattern: "[tag]" }). gitlens.outputLevel is deprecated — don't rely on it.

Typical Workflow

1. Call launch (once per session — takes ~10s)
2. Call execute_command to open the view you want to inspect
3. Call list_webviews to discover open webviews and their exact titles
4. Call wait_for_webview { webview_title: "<title>" } to wait for Lit hydration
5. Call screenshot { target: "webview", webview_title: "<title>" } or aria_snapshot { webview_title: "<title>" } to see the current state
6. Make code changes, then rebuild and reload (see below)
7. Call screenshot again to verify changes
8. Repeat steps 6-7 as needed
9. Call teardown when done

Rebuilding After Code Changes

Extension host code (commands, providers, services, models, parsers — anything under src/ outside src/webviews/apps/):

rebuild_and_reload { build_command: "pnpm run build:extension" }

This restarts the extension host with the new code on the same VS Code instance. The evaluator bridge does not survive it — it's the --extensionTestsPath entry point, which VS Code invokes only at workbench startup, so the restarted host never re-runs it or re-announces its (ephemeral) port. Screenshot/click/DOM tools keep working, but evaluate, set_account and execute_command's fast path go with it. When you need those after an extension-host change, teardown + launch instead.

Webview code (Lit components, CSS, templates under src/webviews/apps/): No extension host restart needed. Build the webviews, then use the view's refresh command:

rebuild_and_reload { build_command: "pnpm run build:webviews" }
execute_command { command: "gitlens.views.home.refresh" }

Every GitLens webview has a gitlens.views.<name>.refresh command (e.g. gitlens.views.welcome.refresh, gitlens.views.graph.refresh, gitlens.views.commitDetails.refresh). These fully reload the webview with fresh JS/CSS.

Both changed: Use pnpm run build:quick (builds extension + webviews, no linting), then refresh the relevant view.

Quick Examples

Extension host code change:

launch {}
execute_command { command: "gitlens.showHomeView" }
screenshot {}
# ... edit extension host code ...
rebuild_and_reload { build_command: "pnpm run build:extension" }
screenshot {}
teardown

Webview code change:

launch {}
execute_command { command: "gitlens.showWelcomeView" }
inspect_dom { selector: "h1", in_webview: true }
# ... edit webview code ...
rebuild_and_reload { build_command: "pnpm run build:webviews" }
execute_command { command: "gitlens.views.welcome.refresh" }
inspect_dom { selector: "h1", in_webview: true }
teardown

Batch CLI (Fallback for One-Shot Inspection)

scripts/e2e-dev-inspect.mjs — a general-purpose CLI that supports ordered, repeatable actions. Use this when the MCP server is not available or for quick one-off inspections.

Two Modes

Use Development mode when you need gitkraken.env (e.g. testing feature flags against dev API). Use Test mode when you need evaluate() to inspect runtime values (e.g. vscode.env.machineId).

Common Recipes

Inspect any view's DOM content

node scripts/e2e-dev-inspect.mjs --command gitlens.showWelcomeView --query-frame h1

The --query-frame action searches all frames (including nested webview iframes) for matching elements and prints their text content.

Get the full accessibility tree of a view

node scripts/e2e-dev-inspect.mjs --command gitlens.showHomeView --aria

Inspect a specific DOM element

node scripts/e2e-dev-inspect.mjs --command gitlens.showWelcomeView --aria-selector "[class*='header']"

Click something, then inspect the result

node scripts/e2e-dev-inspect.mjs \
  --command gitlens.showHomeView \
  --click-frame "button.start-work" \
  --pause 2000 \
  --query-frame ".dialog-content h2"

Read runtime values (requires --with-evaluator)

node scripts/e2e-dev-inspect.mjs --with-evaluator \
  --eval "vscode.env.machineId" \
  --eval "vscode.version" \
  --eval "vscode.env.appName"

Check feature flag behavior with dev environment

node scripts/e2e-dev-inspect.mjs --env dev \
  --command gitlens.showWelcomeView \
  --query-frame h1 \
  --logs FeatureFlagService

Search extension logs for any pattern

node scripts/e2e-dev-inspect.mjs --logs "error"
node scripts/e2e-dev-inspect.mjs --env dev --logs ConfigCat

Take a screenshot

node scripts/e2e-dev-inspect.mjs --command gitlens.showGraphView --screenshot /tmp/graph.png

Keep VS Code open for manual interaction

node scripts/e2e-dev-inspect.mjs --env dev --keep-open

Add custom settings

node scripts/e2e-dev-inspect.mjs \
  --setting "gitlens.currentLine.enabled=true" \
  --setting "gitlens.hovers.currentLine.over=line" \
  --command gitlens.showWelcomeView --aria

WSL / SSH / Headless Linux

If VS Code is not installed natively in your Linux environment, use --download-vscode to download a portable binary. Xvfb is started automatically if no $DISPLAY is set.

node scripts/e2e-dev-inspect.mjs --download-vscode --command gitlens.showHomeView --aria

Requires xvfb package for headless environments: sudo apt-get install xvfb

How AI Agents Should Use This

Prefer the MCP server for iterative work. Call launch once (use download_vscode: true on WSL/SSH/headless Linux), then use tools in a loop. No output parsing needed — tools return structured results directly.

Token & round-trip discipline

How you drive inspection decides whether it's cheap or ruinous. Internalize these defaults:

• Prefer text/measured evidence over screenshots. A full-window screenshot costs ~1.7K image tokens (image cost scales with resolution) and is re-shipped on every later turn, so it compounds. Answer "what is the state / is it correct" with evaluate_in_webview (geometry via getBoundingClientRect(), computed styles, text, counts) or aria_snapshot({ selector }). Reserve screenshot for when the pixels themselves are the question (visual polish, overlap, alignment), and scope it to a webview.
• Batch probes into one call. Don't fire N evaluate_in_webview calls reading one field each — return a structured object in a single call: evaluate_in_webview({ expression: "(() => { const el = document.querySelector('gl-graph-app').shadowRoot.querySelector('…'); return { top: el.getBoundingClientRect().top, color: getComputedStyle(el).color, count: … }; })()" }). Project only the fields you need (returns are soft-capped at 20K chars); never return whole innerHTML.
• Filter every read. read_console({ level: "error", last_n }) and read_logs({ pattern: "<tag>", last_n }) — the arg is pattern, NOT filter (a wrong key silently dumps everything). Both default-cap at 200 lines; use read_console({ clear: true }) as a cursor so the next read only sees new messages.
• Fold setup into launch. launch({ commands: ["gitlens.showGraphView", …], account: "pro", log_level: "info" }) opens views and unlocks Pro features inside the launch call (saves a round-trip each) and keeps on-disk logs small.
• Reuse the session. launch once, then drive in a loop — it persists. For webview-only edits, build:webviews + the view's .refresh is ~3–5× cheaper than anything that restarts the host.
• Extension-host code changes cost a relaunch. rebuild_and_reload restarts the host, but the evaluator bridge does not come back: it's the --extensionTestsPath entry point, which VS Code runs only at workbench startup, so the restarted host never re-announces its port. Everything bridge-backed (evaluate, set_account, execute_command's fast path) degrades or dies afterwards, and the tool now says so explicitly instead of failing silently. Use teardown + launch when you need a live bridge after an extension-host change. (workbench.action.reloadWindow is not a workaround — under --extensionTestsPath the host exit reads as "tests finished" and the whole instance quits.)

Delegate the driving to a Sonnet driver (default)

When the session model is Opus or Fable, default to dispatching the mechanical driving to the inspector-driver subagent (pinned to Sonnet 5) — it costs ~1/5 per token and a smoke test showed quality parity on mechanical inspection (DOM reads, extension-host API reads, measurements). You stay the orchestrator: you decide the probe list, you interpret the returned evidence, you decide fixes.

• Dispatch: Agent({ subagent_type: "inspector-driver", model: "sonnet", prompt: <setup + the exact probes/steps> }). The agent's system prompt already encodes the driving discipline above, so give it a concrete step list, not a vague goal.
• You own the instance lifecycle. Launch once yourself (or in the first dispatch), keep it alive, and tell drivers not to launch/teardown — they reuse your running instance. Teardown yourself when done.
• Batch the ask. One dispatch should collect evidence across many states/probes — that's what amortizes the subagent's fixed overhead (~35K tokens). Don't dispatch a driver for a single trivial probe; run that inline.
• Screenshots stay with you. A driver can't hand an image back. For pixel judgment, either take the screenshot yourself, or have the driver return the measurable facts (bounding rects, computed styles, overflow booleans) and judge those.
• Haiku for pure evidence collection. For a fixed probe list with no screenshots (DOM/API reads, measurements, filtered logs), drop the driver to model: "haiku" — validated at parity with Sonnet and ~1/3 cheaper. Keep Sonnet (the default) when the driver may hit fine on-screen text, an ambiguous probe, or any visual call; Haiku reliably makes coarse visual calls ("rendered vs empty/broken") but misreads small text.

Choosing the right tool

GitLens Webview Reference

Root element tag convention: gl-<name>-app. Use these for inspect_dom selectors and evaluate_in_webview queries.

Inspecting Webview Content

GitLens webviews use Lit web components with Shadow DOM. Here's the recommended approach:

1. Discover: list_webviews to find open webviews. Output includes index, id (e.g. gitlens.views.commitDetails), title, url, dimensions, and content status. (Or use the reference table above.)
2. Wait: wait_for_webview { webview_title: "Home" } to ensure Lit hydration is complete
3. Structure: aria_snapshot { webview_title: "Home" } for the accessibility tree
4. Shadow DOM: inspect_dom { selector: "gl-home-app", property: "shadowDOM", in_webview: true, webview_title: "Home" } to see rendered Lit templates
5. JS state: evaluate_in_webview { expression: "document.querySelector('gl-home-app').shadowRoot.querySelector('.my-element').textContent" } to read shadow DOM content. Use .shadowRoot.querySelector() to reach elements inside Lit shadow roots — plain document.querySelector() cannot cross shadow boundaries.
6. Styles: evaluate_in_webview { expression: "getComputedStyle(document.querySelector('gl-home-app').shadowRoot.querySelector('.my-element')).color" } for computed styles
7. Errors: read_console { level: "error" } to check for JS errors in the main process. For webview-specific errors, use evaluate_in_webview to inspect state directly.

Targeting a specific webview when multiple are open

When more than one webview is visible (e.g. Graph + Commit Details), webview_title matching can silently fail because outer-frame titles are often empty for unfocused webviews, and the tool falls back to "first webview with content" — which is usually the wrong one. All webview-targeting tools (evaluate_in_webview, wait_for_webview, inspect_dom, aria_snapshot, screenshot, click) accept three matchers:

• webview_title — outer-frame title (best when titles are reliable; matches the table above)
• webview_url — case-insensitive substring of the webview URL. Matches the id/purpose/extensionId query params in vscode-webview:// URLs. Use a fragment like "commitDetails" or "graph".
• webview_index — 0-based index from list_webviews. Deterministic fallback when title and URL matching are insufficient.

Precedence: index → url → title → first-with-content. Pick webview_url first (intent-revealing); fall back to webview_index if needed.

list_webviews
# returns [{ index: 0, id: "gitlens.views.graph", ... }, { index: 1, id: "gitlens.views.commitDetails", ... }]
evaluate_in_webview { webview_url: "commitDetails", expression: "performance.now()" }

Screenshot Best Practices

Always target a specific webview instead of taking full-window screenshots:

screenshot { target: "webview", webview_title: "Home" }

This captures just the webview content instead of the entire VS Code window. Screenshots are downscaled to 1280px longest side and WebP-compressed by default (image cost scales with resolution); raise max_dimension (up to 1920) or pass format: "png" only when you need fine pixel detail. Better still, skip the screenshot when a text/geometry answer will do — see "Token & round-trip discipline" above.

Use resize_window only when explicitly testing a responsive breakpoint — it resizes the actual Electron window and is clamped by the host display. For larger headless render surfaces, use launch({ screen_resolution }) instead.

Troubleshooting

Webview frame access fails / "not found" errors: Try launch { disable_site_isolation: true }. This disables OOPIF site isolation so Playwright can access webview iframes directly. Note: CORS/CSP are also disabled, so webview behavior may differ slightly from production.

Headless screenshots too small: Use launch { screen_resolution: "2560x1440x24" } for a larger Xvfb display (default: 1920x1080x24).

All Options

Exercising Pro-gated features

Pro-gated features (the Commit Graph beyond local repos, Launchpad, Worktrees beyond 1, Cloud Patches, Composer, all AI features, Drafts, Workspaces, etc.) check the user's subscription before unlocking. You can't exercise these without a Paid/Trial subscription on the session.

The extension ships a subscription simulator in DEBUG builds that overrides the session's subscription state without touching the real account. Drive it with the set_account tool.

Setup

set_account { plan: "pro" }

That's the whole setup for most Pro-feature testing. Onboarding is pre-dismissed by default (dismiss_onboarding) — every tour/banner (composer welcome, home walkthrough, MCP banner, rebase-editor warning, integration banner, SCM-grouped welcome) is a full-screen overlay that intercepts clicks during automation. The tool reports the resulting account, so you can confirm the gate actually opened instead of assuming it.

You can also bootstrap at launch, saving a round-trip:

launch { session: "dev", account: "pro" }

Plan reference

Modifiers: reactivated: true (trial plans only — reactivated rather than fresh); feature_preview_day / feature_preview_seconds (community only — the Pro Preview window; day 0 = day 1, 1 = day 2, 2 = day 3, 3 = expired). Both are rejected with a clear error if paired with a plan they don't apply to.

Common recipes

Community gate (paywall UX): set_account { plan: "community" }

Feature-preview countdown: set_account { plan: "community", feature_preview_day: 0 }

Plan-tier differences: set_account { plan: "business" } (note: "Business" is teams on the wire — the tool handles that mapping)

Stop simulation (mandatory teardown)

set_account { plan: "none" }

Restores the prior subscription, feature previews, and any onboarding flags that were pre-dismissed. Re-calling with another plan also clears any prior simulation state.

Don't hand-roll the payload. The underlying command (gitlens.plus.simulate.subscription) takes a numeric SubscriptionState const enum, and nothing coerces names to numbers. execute_command with { "state": "Paid" } returns true and silently lands you on trial-expired — verified, not theoretical. set_account owns the name→number mapping; use it.

Using a real account instead

For real entitlements and organizations rather than a simulated plan, launch a persisted session and sign in once:

launch { session: "dev" }
sign_in {}

The session's VS Code user data lives under .vscode-test/agent-sessions/<name>/ and survives teardown, so later launch { session: "dev" } calls start already signed in. sign_in is human-in-the-loop and needs a visible window: it opens a browser, and because the vscode:// callback resolves against the default user-data-dir it usually lands in the user's main VS Code, leaving GitLens' "paste the authorization code" input box as the way to finish. Ask the user to complete it, then wait. It refuses on a throwaway session, where the account would be discarded on teardown.

Exercising AI features

GitLens AI features (Generate Commit Message, Explain *, Generate Changelog, Composer, Review Changes, Generate Search Query) cannot use real provider calls during automated inspection — they cost money, require keys, and produce non-deterministic outputs you can't assert against. The extension ships a deterministic AI simulator in DEBUG builds that you control via VS Code commands.

AI features are also Pro-gated, so two simulators must be enabled in order:

1. Subscription simulator — see Exercising Pro-gated features above. Use set_account { plan: "pro" }.
2. AI simulator (gitlens.plus.simulate.ai) — replaces the AI provider with a stub that returns content the agent injects. Suppresses the first-run ToS modal and the AI All-Access promo notification automatically.

The AI simulator dispatches on a discriminated op arg: enable, disable, inject, clear, lastMessages. Calling without args opens a QuickPick.

Setup (one-time per session)

set_account { plan: "pro" }
execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "enable" }] }

set_account already pre-dismisses onboarding, so the AI simulator doesn't need dismissOnboarding as well. Both share the same snapshot/restore pattern.

Inject-then-trigger pattern

The agent authors the response content, pushes it onto the simulator's stash, then triggers the AI command. The next sendRequest for that action consumes the inject:

execute_command { command: "gitlens.plus.simulate.ai", args: [{
  "op": "inject",
  "action": "generate-commitMessage",
  "content": "<summary>Stable test summary</summary><body>Deterministic body content.</body>"
}] }
execute_command { command: "gitlens.ai.generateCommitMessage:scm" }
# read SCM input — assert it contains "Stable test summary"

inject payload: { op: "inject"; action?: AIActionType; content: string; sticky?: boolean }. Omit action to inject for the next call regardless of action. Set sticky: true to keep the same content for every call of that action.

Authoring response content

The schemas are enforced in packages/plus/ai/src/utils/results.utils.ts — read it once for the canonical tag set. The simulator returns a sensible default per action when no inject is queued, so smoke calls without explicit injects still produce predictable output.

Mode shortcuts (negative-path UX)

Without injecting per-call, you can globally force a failure mode by re-enabling with a different mode:

execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "enable", "mode": "error" }] }    # provider error UX
execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "enable", "mode": "cancel" }] }   # cancellation UX
execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "enable", "mode": "slow" }] }     # progress-indicator UX
execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "enable", "mode": "invalid" }] }  # composer 4-attempt validation-failure UX

Mode invalid only triggers retry behavior for generate-commits (the Composer) — other actions tolerate malformed content and just render it.

Composer authoring (reflection workflow)

The Composer's generate-commits validator demands every input hunk index appear exactly once in the response. The agent can't author a valid response without first seeing the hunks the prompt sent:

1. Trigger Composer with no inject queued (default response will fail validation predictably)
2. Read the messages the simulator received:

   execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "lastMessages" }] }
   

   Returns AIChatMessage[]. Parse the user message's hunks JSON to learn the hunk index space.
3. Author a valid {"commits":[{"message":"...","explanation":"...","hunks":[{"hunk":0},{"hunk":1},...]}, ...]} response covering every hunk index exactly once
4. Inject for action generate-commits. The Composer's automatic 4-attempt retry loop will consume the inject on the next attempt.

Cleanup

Mandatory between scenarios — leftover injects from one test will leak into the next:

execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "clear" }] }

Mandatory teardown (restores the prior gitlens.ai.model, gitlens.ai.enabled, and AI All-Access flags):

execute_command { command: "gitlens.plus.simulate.ai", args: [{ "op": "disable" }] }

{op: "enable"} always clears the stash first, so re-enabling between tests is also safe.

Build requirement (both simulators)

Both simulators are DEBUG-only (gitlens:debugging context). Standard dev launch via launch {} and pnpm run build:extension includes them. Production bundles strip them.

Related skills

• /live-exercise — agent-driven iterative working rhythm for UI-bearing work (audit + fix loop), which uses this skill's tools as its primitive.
• /live-perf — agent-driven performance measurement + improvement skill, also built on these primitives.
• /live-pair — user-driven interactive pair-programming session; the user gives feedback, the agent edits/rebuilds/refreshes live.

Use /live-exercise, /live-perf, or /live-pair when touching UI; use this skill on its own for one-off inspection.