Daytona Flow Validator

Daytona Flow Validator

Use this skill to decide whether a Daytona Electron or browser flow actually works. Launching the sandbox is separate. This skill owns the feedback loop.

Core Rule

Never report success from a click, script return value, or recording alone. Validate the same path a human would take, using CDP to drive Chrome or Electron instead of replacing the journey with hidden state changes. Every meaningful action must follow this loop:

1. Observe with browser_snapshot first.
2. Act with browser_click or browser_fill against snapshot UIDs whenever possible.
3. Observe again with browser_snapshot.
4. Assert the expected URL, text, state, process, network, or file result.
5. Capture a screenshot checkpoint when the visible state matters.

If any assertion is missing, the flow is not validated yet.

When a coded eval exists, prefer pnpm evals --flow <flow-id> because the runner binds assertions, screenshots, and validation metadata into one HTML proof. Use manual CDP only to debug or to create a new coded flow.

Use browser_eval, direct API calls, localStorage writes, filesystem edits, or database changes only when a human-visible path is impossible, unavailable in the current product, or needed as setup for data that the UI cannot create yet. When you use one of these shortcuts, say so in the report and do not let it replace the visible click-by-click demo claim.

Human-Visible Demo Standard

For founder, designer, PR, or eval evidence, record the journey as a reviewer would experience it:

• Start recording before the first user-visible action.
• Use Chrome CDP for Den Web and Electron CDP for desktop.
• Prefer accessible snapshot items and visible clicks/fills over synthetic DOM mutation.
• Keep every major transition visible: sign-up, org creation, onboarding, handoff, desktop sign-in, Marketplace, install, and chat response.
• Do not skip screens by setting tokens, editing localStorage, calling APIs, or navigating directly unless the report labels that segment as setup or a known product gap.
• If test data must be created through an API, show the result in the UI immediately after and state that the data creation was invisible setup.
• A recording that cannot be understood without terminal logs, API responses, or explanation is not a full demo pass.

Minimum Pass Evidence

For a UI flow, collect all of these when feasible:

• CDP target proof: browser_list shows the intended target.
• App proof: navigator.userAgent contains Electron/ for desktop flows, or does not for standalone Chrome flows.
• State proof: URL, visible text, selected model/provider, status, or route matches the expected outcome.
• Backend proof: relevant daytona exec process/log/health check for sidecars, Den, worker proxy, or mock servers.
• Frame-by-frame HTML proof: the default deliverable. Named PNGs for each step served as a browseable HTML index on port 8090. See daytona-recording-artifacts for how to produce the index.
• Video clips: only when a step involves motion (streaming text, loading spinners, animations). Embed clips in the same HTML index alongside the static frames.

Frame proof is the default. Video is the exception for interactions that need motion. When the user says "test this on Daytona" and UI is involved, always produce frame-by-frame HTML proof unless the user explicitly asks for video.

Validation Loop Template

Use this structure for each step in the final report:

Step: <what was attempted>
Before: <snapshot/eval showed X>
Action: <tool/selector used>
After: <snapshot/eval showed Y>
Assertion: pass/fail because <observable signal>
Evidence: <screenshot path or artifact URL if captured>

Prefer Accessible Snapshots

Start with browser_snapshot for normal UI controls because it gives stable UIDs for browser_click and browser_fill. Treat browser_eval as an escape hatch, not the default interaction mechanism. Use browser_eval when:

• React dynamic UI makes snapshot UIDs stale.
• Native file pickers must be bypassed.
• Lexical editor input needs a synthetic paste event.
• You need to inspect app state, localStorage, URL, or text quickly.

Even when browser_eval is necessary, keep the user-visible state coherent: observe before, perform the minimal hidden action, then observe the visible result with browser_snapshot or a screenshot.

Lexical Composer

Prefer synthetic paste for the OpenWork composer:

(function pasteComposer(text) {
  var editor = document.querySelector('[contenteditable="true"][data-lexical-editor="true"]');
  if (!editor) return { ok: false, reason: 'composer not found' };
  editor.focus();
  var data = new DataTransfer();
  data.setData('text/plain', text);
  editor.dispatchEvent(new ClipboardEvent('paste', { bubbles: true, cancelable: true, clipboardData: data }));
  return { ok: true, text: editor.innerText };
})('Reply with exactly: Daytona validation OK')

Then assert the Run button is enabled before clicking it.

Linux Desktop Automation

Use CDP for renderer UI first. When the flow opens native Linux UI that CDP cannot control, such as GTK file pickers, OS permission dialogs, XFCE windows, or Electron-native dialogs, switch to desktop automation inside the sandbox.

Check/install the tools:

daytona exec "$SANDBOX" -- "bash -lc 'if ! command -v xdotool >/dev/null 2>&1; then sudo apt-get update && sudo apt-get install -y xdotool wmctrl; fi'"

Inspect the real desktop window state before acting:

daytona exec "$SANDBOX" -- "bash -lc 'DISPLAY=:99 wmctrl -l; DISPLAY=:99 xdotool getactivewindow getwindowname 2>/dev/null || true'"

Native file picker pattern:

daytona exec "$SANDBOX" -- "bash -lc 'DISPLAY=:99 xdotool search --name \"Authorize folder\" windowactivate; DISPLAY=:99 xdotool mousemove 760 151 click 1 key ctrl+a type --delay 1 -- \"/workspace/hello\" key Return; sleep 1; DISPLAY=:99 xdotool mousemove 1465 927 click 1'"

Rules for native desktop automation:

• Use wmctrl -l before and after to prove the expected native window opened or closed.
• Prefer window titles such as Authorize folder over coordinates when possible.
• Use coordinates only after inspecting a screenshot/noVNC state; coordinate clicks are display-size dependent.
• Close stale dialogs with Escape before capturing evidence.
• After native automation, reassert app state with CDP and inspect a fresh screenshot. Native dialogs commonly remain on top and invalidate evidence.

Close stale native dialogs before recording or screenshots:

daytona exec "$SANDBOX" -- 'bash -lc '''DISPLAY=:99 xdotool search --name "Authorize folder" windowclose %@ 2>/dev/null || true; sleep 1; DISPLAY=:99 wmctrl -l''''

Screenshots

Use browser screenshots for renderer state:

browser_screenshot({ browser_url: CDP_URL, target_id: TARGET_ID })

Use Daytona display screenshots for noVNC/window state:

daytona exec "$SANDBOX" -- 'bash .devcontainer/capture-daytona-screenshot.sh'

Do not treat screenshots as the only assertion. Inspect text/state with CDP too.

Before publishing, commenting, or reporting a screenshot URL, open the saved image and visually verify it matches the claim. Use webfetch on the artifact URL, Read on the local PNG path, or another image-capable viewer. A screenshot is not valid evidence until the image itself has been inspected.

For every screenshot, assert these visual checks:

• The target OpenWork/Chrome window is visible and not covered by a native file picker, modal, toast, desktop window, or unrelated overlay.
• The exact claimed state is visible in the image, not just present in DOM or command output.
• Important labels, selected rows, dropdown options, status indicators, artifact cards, or output panes are legible enough for a reviewer.
• The screenshot URL belongs to the sandbox and flow being reported.

If any check fails, mark the evidence as failed, fix the visible state, capture a new screenshot, inspect the new image, and only then share it. If bad evidence was already posted, post a superseding correction that clearly says the earlier screenshot was invalid.

Repair Loop

If a frame does not support the claim, repair before reporting a verdict:

1. Diagnose why the frame is invalid: wrong route, hidden target, duplicate screenshot, error state, native dialog, stale modal, or unreadable content.
2. Repair the visible state: close blockers, activate the intended window, wait for text/route stability, scroll the target into view, or rerun the visible action.
3. Reassert the expected state with CDP.
4. Capture a new frame with a new name.
5. Include repair attempts in the final report. If repair fails, report Incomplete or Failed, not Passed.

Before every Daytona display screenshot, run a native-window check and fail fast if a picker is present:

daytona exec "$SANDBOX" -- 'bash -lc '''DISPLAY=:99 wmctrl -l | tee /tmp/windows-before-shot.txt; ! grep -q "Authorize folder" /tmp/windows-before-shot.txt; DISPLAY=:99 .devcontainer/capture-daytona-screenshot.sh --output /daytona-artifacts/screenshots/<flow>/<step>.png''''

If a Chromium or Electron window is intentionally part of the shot, activate the right window first with wmctrl -a "OpenWork - Dev" or wmctrl -a "OpenWork Cloud - Chromium" so the screenshot shows the intended journey step.

Failure Handling

When a step fails:

• Capture current browser_snapshot or document.body.innerText.
• Capture a screenshot.
• Inspect the relevant logs before retrying.
• Retry only after naming the suspected cause.

Common logs:

daytona exec "$SANDBOX" -- 'tail -120 /tmp/electron.log'
daytona exec "$SANDBOX" -- 'tail -120 /tmp/vite.log'
daytona exec "$SERVER_SANDBOX" -- 'tail -120 /tmp/den-api.log'

For Den Web flows specifically:

• If the page remains on Checking account, Loading your workspace, or Checking workspace access, verify whether the client hydrated by trying a real browser_click/browser_fill, not only browser_eval.
• If Next dev is behind a Daytona proxy, switch Den Web to next build + next start before declaring the app broken.
• Validate direct Den API auth independently. A direct API pass plus Den Web proxy failure is an incomplete handoff, not a full pass.
• If you bridge auth by writing localStorage in Electron, state that explicitly in the final report and limit the pass claim to the downstream desktop flow.

Final Verdict

Use one of these verdicts:

• Passed: every expected outcome has an observable assertion and frame-by-frame proof is published.
• Failed: at least one assertion disproves the expected outcome.
• Incomplete: the sandbox/tooling failed, evidence is missing, or only a recording/screenshot was collected without frame proof.

Source: different-ai/openwork → .opencode/skills/daytona-flow-validator/SKILL.md