Daytona Recording Artifacts Agent Skill

Daytona Recording Artifacts

Use this skill to collect proof that a Daytona UI flow works.

Use daytona-flow-validator before declaring the flow passed. If the user asks

to do e2e tests for a feature, frame proof is required unless they explicitly

ask for non-UI or mock-only validation.

Default: Frame-by-Frame HTML Proof

The default proof format is a browseable HTML page with named PNG screenshots

for each step of the flow. This is faster to produce, easier to review, and

works on any device.

Use video (MP4) only when the proof requires motion: streaming text, loading

spinners, animations, drag-and-drop, or real-time interactions that a static

frame cannot capture. When video is used, embed it inside the frame-by-frame

HTML page alongside the static frames.

When a coded eval exists, prefer the eval runner frame output first:

pnpm evals --flow <flow-id> --cdp-url <printed-electron-cdp-url>

The runner writes evals/results/<run-id>/index.html plus validated PNG evidence

from ctx.prove(...) and ctx.screenshot(...). Serve or copy that result

directory for PR evidence.

How to produce frame proof

• Serve a directory from the sandbox on port 8090:

daytona exec "$SANDBOX" -- 'bash -lc "mkdir -p /workspace/proof-frames; nohup python3 -m http.server 8090 --directory /workspace/proof-frames >/dev/null 2>&1 &"'

• At each important state, capture a browser_screenshot locally and upload

it to the sandbox with a numbered name like 01-auth-landing.png,

02-org-filled.png, etc.

• Generate a browseable index.html with all frames as labeled images in a

responsive grid. Include the branch, commit, and sandbox name.

• Get the public URL:

FRAMES_URL=$(daytona preview-url "$SANDBOX" -p 8090 2>/dev/null | grep -v "^time=")
echo "${FRAMES_URL}/index.html"

• Post the index URL in the PR body. Individual frame PNGs are also directly

linkable: ${FRAMES_URL}/01-auth-landing.png.

Fallback when /daytona-artifacts is unavailable

The preferred path is test-on-daytona.sh --artifacts-volume, which mounts and

serves /daytona-artifacts. If that path is unavailable or not writable, do not

stop at local screenshots. Use /workspace/proof-frames as a fallback:

daytona exec "$SANDBOX" -- 'bash -lc "mkdir -p /workspace/proof-frames; nohup python3 -m http.server 8090 --bind 0.0.0.0 --directory /workspace/proof-frames >/tmp/proof-frames-http.log 2>&1 &"'

Upload the generated index.html and PNG frames there, then verify every file is

non-zero before sharing:

daytona exec "$SANDBOX" -- 'bash -lc "ls -lh /workspace/proof-frames && curl -s -I http://127.0.0.1:8090/index.html | sed -n \"1,8p\""'
daytona preview-url "$SANDBOX" -p 8090

If you use this fallback, say so in the PR/eval report because the files are not

on the persistent artifacts volume.

When to include video clips

Embed short MP4 clips in the same proof directory when the step involves:

• Streaming text appearing in real-time (chat responses).
• Loading/progress indicators that resolve.
• Animations or transitions between states.
• Drag-and-drop or multi-step interactions.

Reference them from the HTML index alongside the static frames.

Recording Standard (video, when needed)

A useful Daytona recording should look like a person using the product, even

though CDP is driving the browser or Electron window.

• Record the entire relevant journey, not only the final state.
• Start before the first visible click and stop after the final visible success state.
• Drive Chrome/Electron through visible controls with browser_snapshot, browser_click, and browser_fill wherever possible.
• Keep API calls, localStorage writes, direct navigation, and filesystem checks out of the recorded path unless they are unavoidable setup.
• If invisible setup is unavoidable, label it in the PR/eval and resume the recording at the next visible user step.
• Prefer slower, understandable click-by-click recordings over faster scripts that jump between states.
• The recording should be understandable without terminal output; logs and API checks are supporting evidence only.

The Volume

The reusable Daytona volume is:

ipollowork-eval-artifacts:/daytona-artifacts

The helper serves it on port 8090 when --artifacts-volume or

--record-video is used.

Expected layout:

/daytona-artifacts/recordings
/daytona-artifacts/screenshots
/daytona-artifacts/validation

Start With Artifacts

For screenshots and validation notes without video:

bash .devcontainer/test-on-daytona.sh [branch-or-commit] --artifacts-volume

For full human-review evidence:

bash .devcontainer/test-on-daytona.sh [branch-or-commit] --record-video --recording-name <name>

--record-video implies --artifacts-volume.

Capture Screenshot Checkpoints

Capture a persistent screenshot from the Daytona display:

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

Use this after important states: welcome screen, workspace created, settings

connected, task response visible, error state reproduced, or final success.

Before sharing any screenshot URL, follow daytona-flow-validator and inspect

the saved PNG itself. Confirm the visible image shows the claimed state and is

not covered by a native picker, modal, toast, desktop window, or unrelated

overlay. If the screenshot does not match, recapture and inspect a replacement.

Stop Recording

Always stop with the helper so ffmpeg finalizes the MP4 cleanly:

daytona exec "$SANDBOX" -- 'bash .devcontainer/stop-daytona-recording.sh'

Do not use kill -9; it can corrupt the file.

After stopping, verify the recording exists and has duration:

daytona exec "$SANDBOX" -- 'ls -lh /daytona-artifacts/recordings'
daytona exec "$SANDBOX" -- 'ffprobe -v error -show_entries format=duration -of default=nw=1:nk=1 /daytona-artifacts/recordings/<name>.mp4'

If the duration is near zero, missing, or the file is absent, the recording is

not usable evidence.

Get Artifact URLs

Get the artifacts base URL:

ARTIFACTS_URL=$(daytona preview-url "$SANDBOX" -p 8090 2>/dev/null | grep -v "^time=")

Then append paths:

echo "${ARTIFACTS_URL}/recordings/<name>.mp4"
echo "${ARTIFACTS_URL}/screenshots/<name>.png"

Artifact proxy URLs are not permanent. If the sandbox stops, the old

daytonaproxy URL will fail even when files still exist in

/daytona-artifacts. Restart the sandbox and artifact server, then generate a

fresh URL:

daytona sandbox start "$SANDBOX"
daytona exec "$SANDBOX" -- 'bash -lc '\''cd /daytona-artifacts && nohup python3 -m http.server 8090 --bind 0.0.0.0 > /tmp/daytona-artifacts-http.log 2>&1 &'\'''
daytona exec "$SANDBOX" -- 'curl -s -I http://127.0.0.1:8090/recordings/<name>.mp4 | sed -n "1,8p"'
daytona preview-url "$SANDBOX" -p 8090

Only share the refreshed URL after the local curl -I returns 200 OK with a

non-zero Content-Length.

Before And After Flow

Use before/after recordings for UI regressions or design changes:

bash .devcontainer/test-on-daytona.sh dev --record-video --recording-name my-feature-before
daytona exec "$SANDBOX" -- 'bash .devcontainer/stop-daytona-recording.sh'
daytona exec "$SANDBOX" -- "bash -lc 'cd /workspace && git fetch origin feat/my-branch:feat/my-branch && git checkout feat/my-branch'"
daytona exec "$SANDBOX" -- "bash -lc 'cd /workspace && DISPLAY=:99 .devcontainer/start-daytona-recording.sh --detach --output /daytona-artifacts/recordings/my-feature-after.mp4'"
daytona exec "$SANDBOX" -- 'bash .devcontainer/stop-daytona-recording.sh'

Validation Standard

Use these layers in order of priority:

• Frame-by-frame HTML proof (default): named PNGs in a browseable index.

This is the primary evidence format for UI flows.

• CDP/browser assertions: prove URL, text, state, accessibility tree, and

process state alongside each frame capture.

• Video clips (when needed): short MP4s embedded in the frame index for

steps that require motion proof (streaming, animations, loading states).

Do not report success from a recording alone. The AI should inspect state with

browser tools and use screenshots to validate visible behavior before declaring

the flow passed.

Do not use a video as the primary demo if most of the flow happened through

hidden automation. In that case, mark the run as technical validation only and

produce a new frame-by-frame proof for human review.

If you discover invalid evidence after the fact, do not reuse the same URL as

if it were valid. Produce new frames with new names and explain in the

PR/comment that the earlier artifact was superseded.