feat: add CDP-based interactive live view for headless Chromium (Approach 2 probably what BB is doing)#176
Open
hiroTamada wants to merge 9 commits intomainfrom
Open
feat: add CDP-based interactive live view for headless Chromium (Approach 2 probably what BB is doing)#176hiroTamada wants to merge 9 commits intomainfrom
hiroTamada wants to merge 9 commits intomainfrom
Conversation
Add a lightweight Go service (cdp-live-view) that streams browser frames via CDP Page.startScreencast and forwards mouse/keyboard input via Input.dispatchMouseEvent/dispatchKeyEvent. This replaces the previous noVNC/Xvfb approach with near-zero overhead (~40 MB memory under load, negligible CPU impact) while keeping Chromium in true headless mode (--headless=new --ozone-platform=headless). The service is opt-in via ENABLE_LIVE_VIEW=true and serves an HTML5 canvas-based viewer on port 8080. Made-with: Cursor
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Risk assessment: Medium-High.
Why:
- Adds a new production runtime component (
cdp-live-view) with substantial new behavior (~900 LOC) inserver/cmd. - Introduces a new externally reachable live-view surface (
/+/ws) and browser input forwarding path (Input.dispatchMouseEvent/Input.dispatchKeyEvent). - Changes container/image startup and deployment scripts (
Dockerfile,wrapper.sh,run-docker.sh,run-unikernel.sh) including additional port exposure and conditional service startup.
Decision:
- Code review is required.
- Assigned 2 reviewers with history in these codepaths.
- Not self-approving due risk level.
hiroTamada
changed the title
- Protect sessionID/targetID/currentURL/pageTitle with sync.RWMutex to prevent data races between concurrent goroutines - Capture sessionID before goroutine in screencast frame ack to avoid sending ack to wrong CDP session after target switch - Add Chrome-like browser toolbar with back/forward/reload and address bar - Fix double-typing by using rawKeyDown + char instead of keyDown + char - Set Emulation.setDeviceMetricsOverride for full 1920x1080 viewport Made-with: Cursor
- Reset s.cdp to nil when CDP readLoop exits so connectCDP can reconnect on next viewer action instead of staying broken forever - Clear sessions map and target state on disconnect - Capture msgId before setTimeout closure in benchmark-local.mjs to avoid checking/deleting the wrong pending entry Made-with: Cursor
- Return proper errors for CDP protocol errors instead of nil - Drain all pending call channels when readLoop exits to prevent goroutine leaks on reconnection cycles - Add nil checks for s.cdp in screencast ack goroutine and switchToTarget to prevent panics during concurrent disconnect Made-with: Cursor
These were ad-hoc local/unikernel benchmark scripts not intended for the repository. Made-with: Cursor
- Hold viewer mutex for both frame_meta text and JPEG binary writes to prevent url_update messages from interleaving between the pair - Also mutex-protect text writes in broadcastURLUpdate - Clean up pending map entry on context cancellation in call/callSession to prevent memory leak from accumulated timed-out entries Made-with: Cursor
Use the sessionId returned by Target.attachToTarget instead of busy-waiting up to 5s polling the sessions map. Also remove the unused sendBinary method on the viewer struct. Made-with: Cursor
Only allow http/https schemes in the live view URL bar to prevent file:// and javascript: URI abuse. Default ENABLE_LIVE_VIEW to false in run-docker.sh so the feature is truly opt-in. Made-with: Cursor
…ents Only null s.cdp in onDisconnect if it still points to the disconnected instance, preventing a late-firing callback from clobbering a newer connection. Use e.detail for clickCount in mousedown/mouseup so the browser sees correct double-click sequences, and remove the redundant dblclick handler that sent phantom extra events. Made-with: Cursor
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
hiroTamada
added a commit
that referenced
this pull request
Mar 11, 2026
Benchmark tool and results comparing CDP operation latency across four image variants: headless baseline, Approach 1 (Xvfb/noVNC, PR #174), Approach 2 (CDP screencast, PR #176), and headful. Covers 40+ CDP operations across 9 categories (screenshot, JS eval, DOM, input, network, page, emulation, target, composite) plus concurrent load testing. Includes results from Docker (4 vCPU / 1 GB headless, 8 vCPU / 8 GB headful), Docker with constrained headful (4 vCPU / 1 GB), and KraftCloud (Unikraft) environments. Key findings: - Approach 2 (CDP screencast) adds near-zero overhead vs baseline - Approach 1 (Xvfb/noVNC) adds ~30% overhead on input/screenshot ops - Headful under headless constraints is not viable (39% idle memory) - On Unikraft, Approach 2 remains the better choice for live view Made-with: Cursor
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
cdp-live-view) that streams browser frames via CDPPage.startScreencastand forwards mouse/keyboard input viaInput.dispatchMouseEvent/Input.dispatchKeyEvent--headless=new --ozone-platform=headless)ENABLE_LIVE_VIEW=trueenv var; serves an HTML5 canvas-based viewer on port 8080Changes
server/cmd/cdp-live-view/main.go: Go server that connects to CDP, discovers the browser WebSocket URL dynamically, tracks active targets viaTarget.setDiscoverTargets, streams screencast frames to connected viewers, and dispatches input events back to the browserserver/cmd/cdp-live-view/viewer.html: HTML5 canvas client that renders streamed JPEG frames and captures mouse/keyboard events for interactive controlimages/chromium-headless/image/supervisor/services/cdp-live-view.conf: Supervisor config for the new serviceimages/chromium-headless/image/Dockerfile: Build and copy thecdp-live-viewbinaryimages/chromium-headless/image/wrapper.sh: Conditional startup ofcdp-live-viewbehindENABLE_LIVE_VIEW=trueimages/chromium-headless/run-docker.sh: Expose port 8080 and passENABLE_LIVE_VIEWenvimages/chromium-headless/run-unikernel.sh: Expose port 443:8080/http+tls when live view is enabledBenchmark (unikernel, 1 vCPU / 3 GB)
The live view adds ~40 MB memory overhead under load with negligible CPU impact. Screenshots, JS evaluation, and other CDP operations are unaffected.
Checklist
Made with Cursor
Note
Medium Risk
Introduces a new network-exposed service that can drive browser input via CDP and adds new port mappings, so misconfiguration could widen access even though it’s opt-in via
ENABLE_LIVE_VIEW.Overview
Adds an opt-in interactive live view for headless Chromium by introducing a new Go service (
cdp-live-view) that streams CDP screencast frames over WebSockets and forwards mouse/keyboard/navigation events back to the browser.The container image now builds and ships the new binary, registers it under supervisord, and conditionally starts/stops it from
wrapper.shwhenENABLE_LIVE_VIEW=true. Local Docker and Unikernel run scripts expose the viewer on port8080(and map443to8080for Unikernel when enabled).Written by Cursor Bugbot for commit 2d9527e. This will update automatically on new commits. Configure here.