skills/photon/SKILL.md
Build, test, validate, and improve Photon MCPs — single-file TypeScript MCP servers. Use for creating photons with @format annotations (table, chart:bar, slides), stateful photons using this.memory for persistent storage, photons with @readOnly/@destructive annotations, custom UI using @ui tags/HTML templates, photons wrapping APIs (Stripe, payments), task scheduler photons with cron, user-configurable settings (protected settings), this.render() for live output, photon build for standalone binaries, mermaid diagrams for photon architecture, editing .photon.ts files, @auth for MCP OAuth identity (this.caller), identity-aware locks for multiplayer/turn-based photons, @format slides for Marp-style presentations. Also use for validating photon UIs (promise checking, visual QA, functional testing), improving photon quality via autoloop, and auditing whether a UI consumes all backend methods. DO NOT trigger for general TypeScript or non-photon MCP.
npx skillsauth add portel-dev/skills photonInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Photons are single-file TypeScript MCP servers. No compilation — runs directly with tsx.
npm install -g @portel/photon # Install runtime
photon maker new my-weather # Create photon
photon beam # Launch Beam UI
photon mcp my-weather # Run as MCP server
photon cli my-weather current --city London # CLI
Files go in ~/.photon/. See Directory Structure for the full layout. Connect to Claude Desktop:
{ "mcpServers": { "my-weather": { "command": "photon", "args": ["mcp", "my-weather"] } } }
import { Photon } from '@portel/photon-core';
/**
* Weather API
* @version 1.0.0
* @dependencies axios@^1.0.0
* @icon 🌤️
*/
export default class Weather extends Photon {
constructor(private apiKey: string) { super(); }
/**
* Get current weather for a city
* @param city City name {@example London}
* @readOnly
* @title Current Weather
* @format markdown
*/
async current({ city }: { city: string }): Promise<string> {
const res = await fetch(`https://api.weather.com/v1/current?q=${city}&key=${this.apiKey}`);
if (!res.ok) throw new Error(`API error: ${res.statusText}`);
const data = await res.json();
return `**${data.name}** — ${data.temp}°C, ${data.description}`;
}
}
{ success: true, data } wrappers. If it returns, it succeeded.weather current, not weather getCurrentWeather.PHOTON_WEATHER_APIKEY./**
* Brief description (becomes photon description)
*
* @version 1.0.0
* @runtime ^1.5.0
* @dependencies package1, package2@^2.0.0
* @icon 🔧
* @tags api, utility
*/
export default class MyTool extends Photon {
constructor(private apiKey: string) { super(); }
/**
* Method description (becomes tool description)
* @param query Search query {@example "typescript"} {@min 1}
* @readOnly
* @title Search Items
* @format table
*/
async search({ query }: { query: string }) { ... }
}
JSDoc tags map to MCP protocol annotations (spec 2025-11-25):
/**
* @readOnly — no side effects, safe to auto-approve
* @destructive — requires confirmation
* @idempotent — safe to retry
* @openWorld — calls external systems
* @closedWorld — local data only
* @title My Tool — human-readable display name
* @audience user — who sees results: user, assistant, or both
* @priority 0.9 — content importance (0.0–1.0)
*/
UI-only methods: Combine @internal + @audience user for methods the dashboard can call but the LLM never sees:
/**
* Dashboard-only admin panel data.
* @internal
* @audience user
* @readOnly
*/
async metrics() { return { cpu: 42 }; }
@internal hides from tools/list. @audience user marks results as human-only. The UI still calls via window.photon.callTool('metrics', {}).
Auto-generated from TypeScript return types — no tags needed:
async create(params: { title: string }): Promise<{ id: string; done: boolean }> { ... }
For field descriptions, use an interface with JSDoc:
interface Task {
/** Unique identifier */
id: string;
/** Whether complete */
done: boolean;
}
async create(params: { title: string }): Promise<Task> { ... }
Use @format to control rendering. Common values:
| Format | Use For |
|--------|---------|
| table | Array of objects |
| list | Styled list with {@title name, @subtitle email} |
| checklist | Interactive todo list — {text, done}[] with checkboxes, drag reorder |
| article | Magazine-style text flow around images — {text, images?[]} |
| markdown | Rich text, diagrams |
| slides | Marp-style presentation deck |
| chart:bar / chart:line / chart:pie | Data visualization |
| json | Raw JSON |
| dashboard | Composite panels (auto-detected) |
For complete format reference with layout hints, containers, and auto-detection rules, see references/output-formats.md.
Use @auth for authenticated photons. Enables this.caller in every method.
/**
* Multiplayer chess
* @stateful
* @auth required
*/
export default class Chess extends Photon {
players: Record<string, string> = {};
async join() {
const slot = !this.players.white ? 'white' : 'black';
this.players[slot] = this.caller.id;
if (slot === 'black') {
await this.acquireLock('turn', this.players.white);
}
return { color: slot, name: this.caller.name };
}
/** @locked turn */
async move({ from, to }: { from: string; to: string }) {
// Only reaches here if this.caller.id holds the 'turn' lock
const next = this.turn === 'white' ? this.players.black : this.players.white;
await this.transferLock('turn', next);
return this.board;
}
}
this.caller: { id, name, anonymous, scope, claims } — populated from MCP OAuth JWT.
Identity-aware locks:
this.acquireLock(name, callerId) — assign lock to a callerthis.transferLock(name, toCallerId) — move lock to another callerthis.releaseLock(name) — release, open to anyonethis.getLock(name) — query who holds the lock@locked methods auto-check this.caller.id against lock holder// Lifecycle hooks — receive optional context for hot-reload support
async onInitialize(ctx?: { reason?: string; oldInstance?: any }) {
if (ctx?.reason === 'hot-reload' && ctx.oldInstance) {
// Transfer non-serializable resources (sockets, timers, connections)
this.socket = ctx.oldInstance.socket;
ctx.oldInstance.socket = null; // prevent old instance from using it
return;
}
// Normal first-time initialization
this.socket = await createConnection();
}
async onShutdown(ctx?: { reason?: string }) {
if (ctx?.reason === 'hot-reload') {
return; // DON'T close resources — new instance will take them
}
// Real shutdown: clean up everything
this.socket?.close();
}
Hot-reload rules:
onShutdown({ reason: 'hot-reload' }) → skip resource cleanuponInitialize({ reason: 'hot-reload', oldInstance }) → transfer resources from old instance// Simple emit (local only — goes to current caller's UI)
this.emit({ status: 'processing', progress: 50 });
// Channel emit (cross-photon pub/sub via daemon broker)
// Framework auto-prefixes with photon name: 'messages' → 'whatsapp:messages'
this.emit({ channel: 'messages', type: 'message', data: msg });
// Subscribe to another photon's events (use the full prefixed name)
const broker = getBroker();
const sub = await broker.subscribe('whatsapp:messages', (msg) => {
// handle message
});
sub.unsubscribe(); // when done
Channel naming: Emit with simple names (channel: 'messages'). The framework auto-prefixes with the photon name. Channels with a colon are left as-is.
// Multi-step with user interaction
async *deploy({ env }: { env: string }) {
yield { emit: 'status', message: 'Deploying...' };
const ok = yield { ask: 'confirm', message: `Deploy to ${env}?` };
if (!ok) return 'Cancelled';
return 'Done';
}
// Waiting for external async events (WebSocket, library callbacks, etc.)
async *connect() {
yield { emit: 'status', message: 'Connecting...' };
let resolve: (v: any) => void;
const promise = new Promise(r => { resolve = r; });
this.socket.on('ready', (data) => resolve(data));
await this.initSocket();
const event = await promise; // blocks until external event fires
yield { emit: 'toast', message: 'Connected!', type: 'success' };
return event;
}
Zero-config persistent storage via this.memory:
await this.memory.set('key', value); // photon scope (default)
const val = await this.memory.get<T>('key');
await this.memory.set('shared', data, 'global'); // cross-photon
Three scopes: photon (private), session (per-user), global (shared). Full API: get, set, delete, has, keys, clear, getAll, update.
Dynamic task scheduling via this.schedule — complements static @scheduled/@cron tags:
// Create a scheduled task at runtime
await this.schedule.create({
name: 'nightly-cleanup',
schedule: '0 0 * * *', // 5-field cron or @daily, @hourly, @weekly, @monthly
method: 'purge',
params: { olderThan: 30 },
});
// Manage tasks
await this.schedule.pause(id);
await this.schedule.resume(id);
await this.schedule.cancel(id);
const tasks = await this.schedule.list('active');
Full API: create, get, getByName, list, update, pause, resume, cancel, cancelByName, cancelAll, has. Tasks persist to disk; the daemon executes them.
Use @scheduled/@cron for fixed schedules known at build time. Use this.schedule for dynamic schedules created at runtime (user-configured intervals, conditional jobs, etc.).
Expose configurable options via protected settings. Runtime auto-generates a settings tool and persists to disk:
export default class MyAgent extends Photon {
protected settings = {
/** Polling interval in ms */
pollIntervalMs: 5000,
/** Max concurrent operations */
maxConcurrent: 3,
/** Auto-resume after restart */
autoResume: true,
};
async doWork() {
const interval = this.settings.pollIntervalMs; // read-only Proxy
}
}
Users change settings via CLI (photon cli my-agent settings). Values persist to ~/.photon/state/<name>/<instance>-settings.json.
this.render(format, value) pushes formatted output that replaces the previous render zone (instead of appending):
export default class Monitor {
async status() {
while (true) {
const metrics = await this.collectMetrics();
this.render('table', metrics); // Replaces previous output
await new Promise(r => setTimeout(r, 5000));
}
}
}
Accepts the same format values as @format tags. Call this.render() with no arguments to clear the render zone.
Build standalone executables from any photon — no Node.js required on the target machine:
photon build my-tool # Binary for current platform
photon build my-tool -t bun-linux-x64 # Cross-compile for Linux
photon build my-tool --with-app # Embed Beam UI as a desktop app
Uses Bun's compiler. The binary bundles the photon, its @dependencies, and transitive @photon deps.
photon add — Install a photonphoton add claw # Search all marketplaces by name
photon add my-marketplace/claw # From a specific marketplace
photon add Arul-/photons/claw # From a GitHub repo (owner/repo/name)
If a photon has @photon dependencies, they are auto-installed from the same source recursively.
photon beam / photon cli — Run directly from GitHubphoton beam Arul-/photons/claw # Install + open in Beam
photon cli Arul-/photons/todo add # Install + run method
photon marketplace add Arul-/photons # Add a GitHub repo as marketplace source
photon marketplace update # Refresh all marketplace caches
photon marketplace list # Show configured marketplaces
Link HTML files as interactive result renderers:
/** @ui dashboard */
export default class MyApp extends Photon {
/** @ui dashboard */
async getData({ range }: { range: string }) { return { metrics: 42 }; }
}
Convention: @ui <id> resolves to <photon>/ui/<id>.html. The id MUST match the HTML filename.
@ui dashboard → my-app/ui/dashboard.html ✓
@ui slides → slides/ui/slides.html ✓
Text after the id is treated as a description, NOT a file path.
The UI gets a photon-named global proxy: window.myApp.getData(...), window.myApp.onResult(...).
For full MCP Apps guide, see references/mcp-apps.md.
Photon separates source/assets from runtime data under ~/.photon/:
~/.photon/
├── <name>.photon.ts # Source (or symlink)
├── <name>/ # Assets (@ui templates, images)
├── state/<name>/ # Runtime: this.memory, settings (automatic)
├── data/<name>/ # Runtime: photon-written files (manual)
├── cache/ # Runtime: compiled .mjs cache
└── logs/<name>/ # Runtime: execution logs
Key rule: Photons that write runtime files (auth tokens, downloaded media, databases) MUST use ~/.photon/data/<name>/, never ~/.photon/<name>/. The asset folder is watched for hot-reload — writing there causes reload loops.
// Correct: runtime data goes in data/<name>/
const dataDir = path.join(os.homedir(), '.photon', 'data', 'my-app');
// Wrong: asset folder triggers hot-reload!
const dataDir = path.join(os.homedir(), '.photon', 'my-app', 'downloads');
For the full convention, see Directory Structure.
Every @ui photon makes implicit promises through its backend methods. The UI must deliver on them.
A photon's backend methods are its capabilities. The UI is the interface to those capabilities. If a method exists but the UI has no way to trigger it, that's a broken promise.
Three signal levels:
| Signal | Meaning |
|--------|---------|
| @ui <id> on a method | This method's result is rendered by the UI — the UI MUST handle it |
| @audience user | Result is for the human, not the LLM — the UI SHOULD surface it |
| Public method (no @internal) | Capability exists — the UI or CLI should make it accessible |
If a method is public but the UI doesn't use it, either:
@internal to be honest about what's exposedEvery @ui photon should have a ## UI Promises section in its class docblock:
/**
* Slides — AI-Native Presentation Tool
*
* ## UI Promises
*
* - Filmstrip with thumbnails for slide navigation
* - Drag-and-drop to reorder slides
* - Theme selector (default, gaia, uncover)
* - Fullscreen presentation with scaled slides
* - Markdown editor with live preview
* - Speaker notes editor
* - Deck picker to switch or create presentations
*/
These are the acceptance criteria. Each line is a testable claim. Validation means checking every claim is true.
To validate, compare two lists:
.photon.ts, especially those tagged @ui or @audience userwindow["slides"].move(...))Any method in list 1 but not in list 2 is a gap. Any promise in the docblock that has no corresponding UI control is a broken promise.
| Layer | What it catches | What it misses | |-------|----------------|----------------| | Cosmetic (screenshot review) | Broken layout, contrast, overflow, clipping | Missing features, interaction bugs | | Functional (promise check) | Missing features, unconsumed methods, state gaps | Subtle visual bugs |
Both layers are needed. A UI can look perfect while missing half its features.
| Topic | When to Read |
|-------|-------------|
| Directory Structure | Need the ~/.photon/ layout rules for assets vs runtime data |
| Docblock Tags | Need the complete tag reference (class, method, inline, daemon, MCP) |
| Output Formats | Need layout hints, chart mapping, containers, or auto-detection rules |
| Dependency Injection | Using @mcp, @photon, or this.call() for cross-photon communication |
| Daemon Features | Setting up webhooks, cron jobs, or distributed locks |
| User Settings | protected settings, persistence, auto-resume patterns |
| MCP Apps | Building custom HTML UIs with the photon bridge |
| Visualization | Generating Mermaid diagrams from photons |
| Mermaid Syntax | Flowchart shapes, arrows, subgraphs |
| Photon Patterns | Common emit/ask/yield patterns, path resolution (this.callerCwd), Mermaid equivalents |
| Examples | Complete Photon-to-Mermaid conversion examples |
| Validation | Promise checking, gap detection, UI testing checklist |
development
Automated visual UI issue detection from screenshots. Use EVERY TIME you take a screenshot with agent-browser during UI development. Runs local Qwen3-VL model on Apple Silicon — zero API cost. Catches layout bugs, contrast issues, missing content, alignment problems, and accessibility violations that you'd otherwise miss while focused on one specific fix.
development
Platform-aware design rules for vertical video (Reels, Shorts, TikTok, LinkedIn mobile). Covers safe zones, caption sizing/placement, aspect ratios, thumb-stopper hooks, effects that survive mobile compression, and per-platform cheatsheets. Use when building 9:16 / 4:5 / 1:1 video for social media - any time you need to decide where to place text, how big captions should be, what aspect ratio to export, or which effects survive mobile playback.
tools
Use when work should span one or more detached tasks but still behave like one job with a single owner context. TaskFlow is the durable flow substrate under authoring layers like Lobster, ACPX, plugins, or plain code. Keep conditional logic in the caller; use TaskFlow for flow identity, child-task linkage, waiting state, revision-checked mutations, and user-facing emergence.
tools
# Lobster Lobster executes multi-step workflows with approval checkpoints. Use it when: - User wants a repeatable automation (triage, monitor, sync) - Actions need human approval before executing (send, post, delete) - Multiple tool calls should run as one deterministic operation ## When to use Lobster | User intent | Use Lobster? | | ------------------------------------------------------ | --------------------------