bridge-eyes
Skill7 starsChanged 8 days agoMITView on GitHub
Use when frontend work in this repo needs to be looked at rather than asserted — after changing anything under bridge/dashboard/web, bridge/miniapp/web or site/, or when the user says "screenshot it", "does it look right", "show me the dashboard", or asks you to verify a UI change. Covers headless capture in WSL and driving the live UIs over CDP.
- Reads credentials
In this repository27 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- accessibility-auditSkill
- algorithmic-artSkill
- api-docsSkill
- beautiful-shadowsSkill
- brainstormingSkill
- bridge-feature-sliceSkill
- bridge-shipSkill
- bridge-worktreeSkill
- changelogSkill
- color-paletteSkill
- content-strategySkill
- design-firstSkill
- design-implementSkill
- design-reviewSkill
- design-tokensSkill
- emil-design-engSkill
- empty-statesSkill
- frontend-designSkill
- landing-pageSkill
- micro-interactionsSkill
- mystical-assistant-designSkill
- ogilvy-copywritingSkill
- onboarding-guideSkill
- typography-scaleSkill
- visual-style-presetsSkill
- writing-guidelinesSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 44
Use the bundled CDP script instead, which navigates, waits a fixed settle, and grabs Page.captureScreenshot:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Looking at the UIs headlessly
There is no display here. Every visual claim has to come from a captured pixel, and the obvious way to capture one hangs forever on this project's own UIs.
Which port is which
| Port | Surface | Notes |
| 8790 | Dashboard | Desktop. ?skipboot=1 bypasses the 4.25s boot intro. |
| 8787 | Mini App | 390px wide. /api/* needs Telegram init-data. |
| 8791 | Landing page | Static. |
DASH_TOKEN is empty in this machine's .env, so local GETs need no auth. If it
is set, append ?token=….
Static pages: use the bridge's own helper
from bridge import screenshot
open("/tmp/shot.png", "wb").write(screenshot.capture("http://127.0.0.1:8791/", 1280, 800))bridge/screenshot.py already owns the WSL specifics — it finds the newest
chrome-headless-shell under ~/.cache/ms-playwright/ (the build number changes
on every Playwright update, so never hardcode the path) and puts the cached
libasound.so.2 stub on LD_LIBRARY_PATH. Don't hand-roll chrome flags.
Then Read the PNG to actually look at it.
Live app pages: capture() will hang — use CDP
Chrome's --screenshot mode waits for the page to finish loading. The dashboard
and Mini App hold SSE streams open forever, so it never fires: no image, no
error, no timeout worth waiting for. Measured on the live dashboard — still
nothing at 75s, while the landing page returns in 0.8s.
Use the bundled CDP script instead, which navigates, waits a fixed settle, and
grabs Page.captureScreenshot:
node .claude/skills/bridge-eyes/shot.mjs 'http://127.0.0.1:8790/?skipboot=1' /tmp/dash.png 1280 800
# → /tmp/dash.png 169KB, ~3.4sThen Read /tmp/dash.png.
Driving, not just looking
The script is deliberately small; when you need to click through to a state,
extend the same CDP session with Runtime.evaluate rather than reaching for
Playwright (the npm module is not installed, and the chrome-devtools MCP fails
with "Target closed" in this WSL box).
- Launch with your own
--user-data-dir— without it chrome hands off to an already-running instance and your debug port never listens. (shot.mjsdoes.) - Dashboard selectors: session rows are
.sessrow; several.mscrollelements exist — the transcript scroller is the one containing[data-key]rows. Turn rows carrydata-index/data-key;[data-prompt-idx]was removed in the virtualization rewrite and now silently matches nothing. - The editor is inside
AnalyzeModal, not a top-level tab: clickMAPon a project row, then theEDITORtab. - Mini App with real data needs a forged
X-Telegram-Init-Data:secret = HMAC_SHA256("WebAppData", TELEGRAM_BOT_TOKEN), thenhash = HMAC_SHA256(secret, "\n".join(sorted("k=v")))overauth_date,query_id,user— the user id must be inALLOWED_CHAT_IDS. Block*telegram-web-app.js*viaNetwork.setBlockedURLsor the real CDN script overwrites your stubwindow.Telegram. - Opening a session from another repo calls
api.selectand changes the bridge's active project. Stay in the active project, or restore it withPOST /local/select {"dir": …}on 8790.
Recording, not just stills
A still cannot show a transition, a hover, or a race. bridge/rec.mjs
is shot2.mjs with Page.startScreencast in place of the single capture, and
takes the same arguments plus an output fps:
node bridge/rec.mjs 'http://127.0.0.1:8790/?skipboot=1' /tmp/clip.webm \
1280 800 3000 '{}' "$ASYNC_JS" 10
# → /tmp/clip.webm 800KB 490 frames -> 102 @10fps 10.2sThe 7th argument is the same Runtime.evaluate string shot2.mjs takes, run
with awaitPromise, so an async IIFE with await sleep() between clicks is the
whole step language — no separate DSL.
- Output is VP8/webm, never mp4. Playwright's bundled ffmpeg (the only one
on this machine) is built
--disable-everything: mjpeg in, libvpx out, webm muxed. No h264, no concat demuxer, no gif. Telegram accepts it viasendVideo(measured 2026-08-31);send.pyfalls back tosendDocumentif that ever stops being true. - Chrome emits ~50 frames/s and only when pixels change, so the recorder
drops the surplus and pads gaps with the last frame, against an absolute
video clock. Check the printed real-time figure against the container's
Duration— they should match within a frame or two. If they don't, the video is lying about how long the UI took. - Watch it yourself before sending:
send.pynow takes.webmas well as PNGs.
Isolation needs nothing new. Build in a worktree, serve that dist with
.mystical/probe/probe.py (static + read-only GET proxy), and record against
that port — the recorder already gets a fresh mkdtemp chrome profile per
run, so nothing touches the live bridge's state.
In-session alternative
A run started by the bridge gets a Screenshot MCP tool
(bridge/verify_mcp.py) that returns the image as a content block — no file, no
node. Use it when it's there; it caps at 1600×1600 to bound token cost.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

