verify-ui
Skill4 starsChanged 7 days agoMITView on GitHub
Use before handing any claude-manager change to the user for verification, and whenever you need to see or click the real app UI. Starts a hidden copy of the app (scripts/agent-app.sh) and drives it with scripts/ui — screenshots, DOM, clicks, typing, console errors — without touching the user's own window or focus.
9 other projects use a version of this.
Adopted by
Most recently taken up by
- crsmithdev/alephtheir own version, first seen Nov 16, 2026
- fimoklei/maestrotheir own version, first seen Nov 16, 2026
- whyashthakker/estacktheir own version, first seen Nov 8, 2026
- rpraharaj/ai-sdlctheir own version, first seen Nov 1, 2026
- rpraharaj/ai-engineertheir own version, first seen Nov 1, 2026
- jonmumm/skillstheir own version, first seen Oct 21, 2026
Used with
Projects with it also use cut-release.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
MicrosoftSkillcut-release
Use this skill to cut an APM release from the current worktree: assess whether the cycle since the last tag warrants a patch or minor bump (semver discipline against the merged-since-last-tag diff), sanitize the [Unreleased] CHANGELOG block into a dated version block with one concise "so what" entry per merged PR (drop internal-only churn, consolidate duplicates), bump pyproject.toml + uv.lock, run the CI-mirror lint chain, and open the release PR. Activate on "ship a release", "cut v0.x", "release prep", "bump and PR", "open release PR", "what kind of release do we need", or any phrasing that ends in opening a release PR -- even when the user does not say "skill". Stops BEFORE tagging; tagging stays a human gate that triggers the release workflow. Refuses to bump to a major (>= 1.0.0) version without explicit operator confirmation.
Also in 3 projects that use this
Other versions9 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 9
- ThisIsSadeghi/KMPilot★ 56same name
- oliver-kriska/prmarmot★ 10same name
- jonmumm/skills★ 3same name
- fimoklei/maestro★ 1same name
- rpraharaj/ai-engineer★ 1same name
- ThisIsSadeghi/Loupe★ 1same name
- crsmithdev/aleph★ 0same name
- rpraharaj/ai-sdlc★ 0same name
- whyashthakker/estack★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository3 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- session-harvestSkill
- session-surgerySkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 1 is distinctive, and 3 are very specific to this project.
0
Common
Most of the 7 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
never activates, no Dock icon, its window is fully transparent, ignores the mouse and sits behind every window, so it never steals focus or …
- Very specific: line 28
Never run make dev yourself — that launches the user's own window.
- Very specific: line 60
After each step, scripts/ui snap and look at the PNG: layout, overflow, clipping, alignment, light text on light ground.
- Very specific: line 75
When done, or when abandoning the hand-off early, scripts/guided.sh release.
“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Verify a change in the real app
The user does only the final check. Before you stop and ask for it, exercise the change yourself in the hidden agent copy of the app and fix what you find.
The agent app
scripts/agent-app.sh start builds the debug binary and launches it with CM_AGENT=1 (see crates/app/src/agent_bridge.rs):
- never activates, no Dock icon, its window is fully transparent, ignores the mouse and sits behind every window, so it never steals focus or moves the user's Space;
- runs its own Vite server from this checkout, so frontend edits hot-reload into it;
- uses a throwaway config dir (
/tmp/ccm-agent-<hash>/config), so the user's config, bookmarks, session meta and status file are untouched.
It reads the user's real ~/.claude sessions. Treat them as read-only: never press delete, fork, rename, surgery or anything that writes a session file, and don't start claude in the terminal tab unless the change is about it. As a backstop the agent copy refuses delete, fork and the status-plugin toggle (agent_bridge::refuse_in_agent), showing an error instead; guard any new command that writes ~/.claude the same way.
| command | when |
scripts/agent-app.sh start | once; a no-op when already running |
scripts/agent-app.sh restart | after any Rust change (frontend changes hot-reload) |
scripts/agent-app.sh restart --sandbox DIR (also start) | run on a sandbox Claude config dir (CLAUDE_CONFIG_DIR=DIR, e.g. from scripts/guided.sh) instead of the real ~/.claude |
scripts/agent-app.sh stop | when done |
scripts/agent-app.sh logs [n] | backend stderr, panics |
Never run make dev yourself — that launches the user's own window.
Parallel sessions
Each worktree runs its own copy: the state dir, socket, Vite port and build all follow the checkout path, so a worktree's scripts/ui only ever drives that worktree's copy, even while other sessions build or drive theirs. A fresh worktree's first start clones the newest sibling checkout's target/ and node_modules (APFS copy-on-write, no extra disk) and is ready in about 15 s instead of a full cold build. Each running copy holds about 1 GB of memory, so stop it when you finish.
Driving it: scripts/ui
Targets are a CSS selector, text=Label (innermost visible element with that text) or X Y in CSS px.
scripts/ui snap --width 1000 # PNG path (2x pixels); open it with Read and look
scripts/ui text [selector] # visible text, cheaper than a screenshot
scripts/ui find 'text=Compact' # matches with boxes
scripts/ui click 'text=Settings' # real, trusted clicks (scrolls into view first)
scripts/ui click 'input[placeholder^="Filter"]'; scripts/ui type 'query'
scripts/ui key Escape; scripts/ui key cmd+k; scripts/ui key shift+Tab
scripts/ui key cmd+a; scripts/ui key Backspace # ⌘ keys the page doesn't take run the app menu item, as in the real app
scripts/ui wait 'text=Loading…' --gone --timeout 30
scripts/ui eval 'document.querySelectorAll(".row").length'
scripts/ui eval --body 'const r = await something(); return r'
scripts/ui logs [--clear] # console errors/warnings, uncaught errors, rejections
scripts/ui resize 900 700 # layout at other sizes
scripts/ui windows; scripts/ui -w win-1 snapKnown limits: hover is synthesized in JS (React enter/over handlers run, CSS :hover does not; hover-revealed buttons still take clicks); document.hasFocus() is false; right-click and native menus/dialogs are not driven.
The loop
scripts/agent-app.sh start(orrestartafter Rust edits), thenscripts/ui logs --clear.- Walk the feature the way the user would: every new control, the empty/loading/error states you can reach, and one neighbouring feature it could break.
- After each step,
scripts/ui snapand look at the PNG: layout, overflow, clipping, alignment, light text on light ground. Checkscripts/ui logsfor errors. - Fix, re-run, repeat until clean. A blank screen or a render loop shows up here, not in vitest.
- Close the agent-app pass with a short report (not the user hand-off; that is the next section, "Hand off to the user"): what you exercised, what you fixed, the final screenshot(s), and only what you could not check (native menus, dialogs, real hover styling, multi-monitor, the PTY with a live claude).
Hand off to the user
After the agent-app pass, the user still checks the feature by hand in a visible dev window on a sandbox (${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-manager/guided, holding claude/ and app/), so their real sessions and config are never touched. You prepare it and walk them through; they run the app. One guided test at a time.
Skip it for backend-only changes (nothing visible to click). When a change only shows on real data, say so and give plain make dev steps for the user to run on their own data instead.
scripts/guided.sh reset. If it refuses (exit 2), it names the worktree that holds the sandbox: tell the user which one and stop.- Prepare the sandbox:
scripts/guided.sh copy <session-id> [--cwd DIR]for each real session the test needs (a standing permission; name those sessions in the hand-off message), hand-write small sessions for exact cases, andscripts/guided.sh settings '<json>'for app settings. - Tell the user to run exactly
make -C <this worktree's absolute path> dev-guided(absolute: their terminal is rarely in the repo root) and wait for "it's up". Never run it yourself, as withmake dev. The first run asks for a one-timeclaude auth logininto the sandbox; launching never wipes anything. - Walk one step per message ("do X; you should see Y"). After each,
scripts/ui --guided snapand look at their window before the next step. - On a bug: fix it, then
scripts/guided.sh resetand re-prepare only if the setup changed (say so); for Rust changes ask for ⌘⇧R, which rebuilds and keeps the sandbox. After anyresetwhile their window is open, also ask for ⌘⇧R (or a relaunch): the running app still holds the old config and session state and would write it back. - When done, or when abandoning the hand-off early,
scripts/guided.sh release.
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.

