One browser tab to command an army of pi agents. Spawn parallel sessions, watch reasoning live, attach OpenSpec changes, ship work — from your laptop or phone.

🌐 Website & demo: blackbelttechnology.github.io/pi-agent-dashboard — animated tour, screenshots, and install guide. 📝 Changelog: CHANGELOG.md

Note: This dashboard only works with pi. Oh My Pi is not supported.

Overview — sessions grouped by folder, branch & OpenSpec context, live cost

Session — chat, OpenSpec apply, interactive ask_user, context gauge

Settings — ports, spawn strategy, zrok tunnel, tool resolution

• Quickstart
• Features
• Prerequisites
• Configuration
• Usage
• Recommended extensions
• Authoring a dashboard plugin
• Troubleshooting
• Architecture
• Monitoring
• Development
• Building the Electron app
• CI/CD & releasing
• License

Three install paths, pick one:

Download a pre-built installer from GitHub Releases:

On first launch a setup wizard walks you through mode selection (standalone vs. power-user), API key / OAuth sign-in, and recommended extensions. The standalone mode bundles Node.js and auto-installs pi + dashboard + openspec into ~/.pi-dashboard/ — no terminal, npm, or Node.js required.

Picking the right macOS DMG: run uname -m in Terminal — arm64 means Apple Silicon (M1/M2/M3/M4), x86_64 means Intel. Or open  Apple menu → About This Mac and read the chip name. Download the matching DMG; if you grab the wrong one macOS will refuse to launch the app with a "cannot be opened" error.

First-run unblocking (unsigned binaries):

• macOS — the DMGs are not yet notarized. Either right-click PI-Dashboard.app → Open the first time, or clear all extended attributes from Terminal:

  xattr -cr /Applications/PI-Dashboard.app

  Use -cr (clear, recursive) rather than -d com.apple.quarantine — it's idempotent and won't print No such xattr: com.apple.quarantine when the attribute isn't there. That message is harmless; it just means quarantine was never set or already cleared.
• Windows — SmartScreen warns on first launch. Click More info → Run anyway, or right-click the downloaded .exe / .zip → Properties → tick Unblock → OK before running. For ZIPs, unblock the archive before extracting.

Note: A future release will rename the macOS DMGs to PI-Dashboard-darwin-arm64-<ver>.dmg and PI-Dashboard-darwin-x64-<ver>.dmg (previously a single PI Dashboard.dmg was produced and silently overwrote one arch on each release). Direct download links pointing at the unsuffixed filename will 404 from that release onward; please link to the Releases page instead. See OpenSpec change fix-darwin-dmg-arch-collision.

If you don't have pi yet, you can install the dashboard directly via npm — pi/openspec/tsx ship as regular npm dependencies, so a single install brings everything in:

npm install -g @blackbelt-technology/pi-agent-dashboard
pi-dashboard
# open http://localhost:8000

If pi is already installed, the bridge-extension flow is equivalent:

pi install npm:@blackbelt-technology/pi-agent-dashboard
pi

The bridge extension auto-starts the dashboard server on first launch:

🌐 Dashboard started at http://localhost:8000

Open http://localhost:8000 in any browser. All active pi sessions appear automatically. See Prerequisites for Node.js / build-tool requirements.

Windows has a few extra one-time setup steps. Run the following in an Administrator PowerShell session:

# 1. Enable long paths (required — npm node_modules nesting exceeds 260 chars)
reg add "HKLM\SYSTEM\CurrentControlSet\Control\FileSystem" /v LongPathsEnabled /t REG_DWORD /d 1 /f

# 2. Install Node.js LTS 22 via winget (ships >= 22.18 so no node-guard refusal)
winget install -e --id OpenJS.NodeJS.LTS --accept-source-agreements --accept-package-agreements

# 3. CLOSE this PowerShell, open a NEW one as Administrator (PATH refresh)

# 4. Verify
node --version    # expect v22.18+ (any 22.x >= 22.18, NOT v22.0–v22.17)
npm --version     # expect 10.x

# 5. Install
npm install -g @blackbelt-technology/pi-agent-dashboard

# 6. Start (foreground first time so you can see any errors)
pi-dashboard start

# 7. From the browser
start http://localhost:8000

C++ build tools are typically not required — node-pty ships a Windows x64 prebuild. Install Visual Studio Build Tools only if the prebuild fails to load. See docs/installation-windows.md for more detail (offline / tarball / nvm-windows caveats).

git clone https://github.com/BlackBeltTechnology/pi-agent-dashboard.git
cd pi-agent-dashboard
npm install
npm run build                              # one-time client build
pi install /path/to/pi-agent-dashboard     # global
# or: pi install -l /path/to/pi-agent-dashboard   # project-local only

By default, pi-dashboard on your PATH refers to whatever copy was installed globally (via npm i -g or the Electron bundle). To make it point at your working tree instead — so every edit is live and bridge auto-start uses your changes — link the workspace:

npm run link:local      # symlinks `pi-dashboard` on PATH to packages/server/bin/pi-dashboard.mjs
pi-dashboard status
npm run unlink:local    # restore (removes the global symlink)

The link survives across shells. Every invocation — including pi's bridge auto-spawn — runs packages/server/src/cli.ts via jiti, so you don't need to rebuild the server on edits. The client still requires npm run build (or npm run dev for HMR).

Windows note: symlink creation needs an admin shell or Windows Developer Mode enabled. Everything else works the same as POSIX.

To try the extension in a single pi session without registering it:

pi -e /path/to/pi-agent-dashboard/packages/extension/src/bridge.ts

Remove with pi remove /path/to/pi-agent-dashboard. Alternatively, add the package path directly to ~/.pi/agent/settings.json (global) or .pi/settings.json (project) under "packages": [...].

Sessions & chat

• Real-time session mirroring — all active pi sessions with live streaming messages
• Bidirectional interaction — send prompts and commands from the browser
• Session statistics — token counts, costs, model info, thinking level, context usage bar
• Elapsed time tracking — live ticking counters on running operations, final duration on completed tool calls and reasoning blocks
• Session spawning — launch new pi sessions from the dashboard (headless by default, or via tmux)
• On-demand session loading — browse historical sessions with lazy-loaded content from pi session files
• Force kill escalation — two-click Stop button; first click sends soft abort, second force-kills (SIGTERM → SIGKILL). Session preserved as "ended" for resume/fork.

Workspace & UI

• Workspace management — organize sessions by project folder with pinned directories and drag-to-reorder
• Command autocomplete — / prefix triggers a filtering dropdown
• Mobile-friendly — responsive layout with swipe drawer, touch targets, and mobile action menus
• Markdown preview — rendered markdown views with search, mermaid diagrams, syntax highlighting, and safe handling for raw HTML ref attributes
• Local-image inlining + LaTeX math in chat — agents can reference local screenshots inline as ![alt](/abs/path.png) or ![alt](./relative.png) and they render in chat (the bridge inlines bytes via a streaming-safe pi-asset:<hash> channel — each unique image's bytes ride exactly once per session, no matter how many streaming chunks repeat the token). Math expressions — inline $x = \beta$ and display $$\sum_i^n i$$ (block-level) — are typeset via KaTeX. PNG / JPEG / GIF / WebP / SVG / AVIF / BMP supported with a 5 MB-per-image, 20 MB-per-message cap; oversized or unreadable references render as a visible placeholder rather than a broken-image glyph. The dashboard server adds zero new HTTP routes.
• Searchable select dialogs — keyboard-navigable picker with real-time filtering (OpenSpec changes, flow commands)

Integrations

• PromptBus architecture — unified prompt routing with adapters (TUI, dashboard, custom). Interactive dialogs (confirm/select/input/editor/multiselect) survive page refresh and server restart. Multiselect uses the bus-routed browser path exclusively (the dashboard MultiselectRenderer dialog) since pi 0.70's RPC mode has no working terminal-overlay primitive. First-response-wins semantics with cross-adapter dismissal.
• Extension UI System (Phase 1) — extensions can declare slash-command-triggered modal UIs as data, without authoring React or importing an SDK. Listen on pi.events.on("ui:list-modules", probe) and push descriptors into probe.modules; the dashboard renders table / grid / form views with row actions, optional confirm-dialog gates, and MDI icons. Modules survive reconnect via the server-side cache; in pure-pi mode descriptors stay dormant. See openspec/specs/extension-ui-system/spec.md for the protocol contract.
• pi-flows integration — live flow execution dashboard with agent cards, detail views, flow graph, summary, abort/auto controls. Launch flows and design new ones with Flow Architect, all from the browser. Fork decisions and subagent dialogs forwarded via PromptBus.
• OpenSpec integration — browse specs, view archive history, manage changes, create new changes from the sidebar
• Browser-based provider auth — sign in to Anthropic, OpenAI Codex, GitHub Copilot, Gemini CLI, and Antigravity from Settings. Enter API keys for other providers. Credentials saved to ~/.pi/agent/auth.json and live-synced to running sessions.
• Custom LLM providers — add OpenAI-compatible, Anthropic-compatible, or Google Generative AI endpoints (Settings → Providers → LLM Providers). Test button verifies the base URL + API key before saving. Adding / editing / removing takes effect live in every running session — no restart.
• Package management — browse, install, update, remove, and move pi packages between global and project scopes from a single rich-row UI used in both Settings and Pi Resources. Install dialog exposes a Local/Global radio when launched from a per-folder context. Search the npm registry for pi-package extensions/skills/themes; install from npm or git URL. Active sessions auto-reload after changes.

Dev tools

• Integrated terminal — full browser-based terminal emulator (xterm.js + node-pty) with ANSI colors, scrollback, and keep-alive
• Diff viewer — side-by-side and unified diff views with file tree navigation. In Jujutsu workspaces the diff is regime-aware: shows the cumulative changes since the workspace's branch point, not just the working-copy delta.
• Editor integration — open files in VS Code, Cursor, etc. directly from tool call cards
• Jujutsu workspaces (optional) — when jj is on PATH and the session is inside a .jj/ repo, the dashboard surfaces a workspace badge, a + Workspace action that creates jj workspace add + spawns a fresh agent in it, and a Fold back action that drives the jj-workspace-fold-back skill (jj-native rebase + push, never git commit/git merge). Activates silently — zero UI when jj is not installed. See docs/architecture.md for the data flow.

Networking & distribution

• Network discovery — mDNS-based auto-discovery of other dashboard servers on the local network
• Zrok tunnel — optional persistent public URL via reserved shares (see Configuration → Tunnel)

Only needed for Quickstart paths B and C. The Electron app (path A) bundles everything in standalone mode.

Optional:

• Config file: ~/.pi/dashboard/config.json (auto-created with defaults on first run)
• Tool overrides (machine-local): ~/.pi/dashboard/tool-overrides.json — see Tool overrides
• Settings UI: click the ⚙ gear icon in the sidebar header to edit all fields from the browser

CLI flags → environment variables → config file → built-in defaults.

The bridge also honours PI_DASHBOARD_URL=ws://host:port to point at a remote server instead of localhost.

{
  "port": 8000,
  "piPort": 9999,
  "autoStart": true,
  "autoShutdown": false,
  "shutdownIdleSeconds": 300,
  "spawnStrategy": "headless",
  "tunnel": { "enabled": true, "reservedToken": "auto-created-on-first-run" },
  "devBuildOnReload": false,
  "askUserPromptTimeoutSeconds": 300,
  "openspec": {
    "pollIntervalSeconds": 30,
    "maxConcurrentSpawns": 3,
    "changeDetection": "mtime",
    "jitterSeconds": 5
  }
}

OAuth2 authentication guards external (tunnel) access. Localhost is always unguarded.

{
  "auth": {
    "secret": "auto-generated-if-omitted",
    "providers": {
      "github":   { "clientId": "...", "clientSecret": "..." },
      "google":   { "clientId": "...", "clientSecret": "..." },
      "keycloak": { "clientId": "...", "clientSecret": "...", "issuerUrl": "https://keycloak.example.com/realms/myrealm" }
    },
    "allowedUsers": ["octocat", "user@example.com", "*@company.com"]
  }
}

Supported providers: github, google, keycloak, oidc (generic OIDC with issuerUrl).

Callback URL: register https://<tunnel-url>/auth/callback/<provider> in your OAuth provider settings. The tunnel URL is stable across restarts (reserved shares are auto-created).

Security note: /api/spawn-failures is reachable to any caller on deployments without auth; entries contain cwd paths. Enable auth before exposing via tunnel.

The dashboard auto-connects a zrok tunnel on start when tunnel.enabled is true. Install with brew install zrok (macOS) and run zrok enable <token> to enrol — the dashboard reads zrok's own config (~/.zrok2/environment.json), no keys are stored in the dashboard. Reserved shares provide persistent URLs across restarts.

Tune how often the server polls known directories for OpenSpec updates (openspec block):

Live-reconfigurable via Settings → Advanced → "Background polling (OpenSpec)" or PUT /api/config — no server restart needed. See docs/architecture.md for the cost model.

The dashboard resolves every external tool it calls (pi, pi-coding-agent, openspec, npm, node, tsx, git, zrok, pi-dashboard) through a single ToolRegistry. Each tool has an ordered strategy chain (override → managed install → bare-import / npm-global → PATH search), and every resolution records a diagnostic trail.

Inspecting and overriding — Settings → General → Tools shows every resolved tool, its source, and the trail. You can set a per-tool override path, rescan individually or all at once, and export the full diagnostic report.

Overrides file — ~/.pi/dashboard/tool-overrides.json:

{
  "version": 1,
  "overrides": {
    "pi":              { "path": "C:\\custom\\pi.cmd" },
    "pi-coding-agent": { "path": "D:\\dev\\pi-coding-agent\\dist\\index.js" }
  }
}

The file is deliberately separate from config.json so machine-specific paths don't follow a dotfiles sync. Invalid overrides (path doesn't exist) are recorded in the trail and the registry falls through to the next strategy automatically.

The dashboard exposes an OpenAI-compatible HTTP proxy on the same port as the dashboard UI (/v1/...). Any LLM client that accepts a custom base_url can use it.

# Example: point an OpenAI-compatible client at the dashboard
export OPENAI_BASE_URL=http://localhost:8000/v1
export OPENAI_API_KEY=pi-proxy-<your-proxy-key>

Setup: open Settings → API Proxy in the dashboard UI, enable the proxy, and create an API key.

Endpoints:

• GET /v1/models — list available models (requires models:list scope or all)
• POST /v1/chat/completions — OpenAI chat completions, streaming + non-streaming
• POST /v1/messages — Anthropic messages, streaming + non-streaming

Auth: proxy API keys only (pi-proxy-* prefix). Dashboard JWT is never accepted on /v1/*.

For migration from @blackbelt-technology/pi-model-proxy, see docs/migration/from-pi-model-proxy.md.

The bridge extension automatically starts the dashboard server when pi launches if it's not already running. Disable with "autoStart": false in ~/.pi/dashboard/config.json.

In the Electron app, if the initial launch attempts fail (or the server is stopped externally), the loading page exposes a Start server button, an Open Doctor link, and a collapsible Server log panel showing the last 20 lines of ~/.pi/dashboard/server.log. The system tray menu also includes a Start server / Restart server item that reflects current server state. All entry points share a single idempotent launch routine in the Electron main process.

Debugging the Electron app via CDP. The desktop app accepts an opt-in --debug-cdp[=<port>] CLI flag (and equivalent PI_DEBUG_CDP env var) that exposes Chromium's Chrome DevTools Protocol on a loopback-only port (default 9222) so agent-browser, Playwright, Chrome DevTools, or any CDP client can attach to the installed shell — not just a separate browser pointed at http://localhost:8000. The flag wins if both flag and env are set. CDP is off by default, never binds beyond 127.0.0.1, and logs a one-line [debug-cdp] warning to stderr when active. Single-instance contract: CDP must be enabled at first-instance launch; passing the flag to an already-running instance cannot enable CDP retroactively and only logs a warning — fully quit and relaunch. From this repo, cd packages/electron && npm run dev:cdp launches the dev Electron with CDP on. The bundled universal browser skill (shipped by the bridge extension) contains a worked-example recipe for attaching agent-browser to the Pi Dashboard app.

Doctor diagnostics. Help → Doctor (or the loading-page link) opens a styled BrowserWindow (doctor.html) that runs the same checks the Electron app already performed — grouped into sections (Runtime, Pi, Server, Bundles, Diagnostics) with status pills, paths, and per-row suggestion callouts; toolbar offers Re-run, Copy as Markdown / Plain, Open server log, Open doctor log, Run setup wizard. The web client exposes the portable subset at Settings → Diagnostics, which fetches /api/doctor and renders the same sections (Electron-only rows omitted). Both surfaces share packages/shared/src/doctor-core.ts, so a check defined once shows up everywhere.

pi-dashboard start           # background daemon (production)
pi-dashboard start --dev     # dev mode (proxy to Vite, fallback to production build)
pi-dashboard stop            # stop daemon (also kills stale port holders)
pi-dashboard restart         # restart (production)
pi-dashboard restart --dev   # restart in dev mode
pi-dashboard status          # daemon status

Daemon stdout/stderr is logged to ~/.pi/dashboard/server.log (append mode with timestamped headers per start).

npx tsx packages/server/src/cli.ts
npx tsx packages/server/src/cli.ts --port 8000 --pi-port 9999
npx tsx packages/server/src/cli.ts --dev   # proxy to Vite dev server

# Restart in the same mode
curl -X POST http://localhost:8000/api/restart

# Switch to dev / production
curl -X POST http://localhost:8000/api/restart -H 'Content-Type: application/json' -d '{"dev":true}'
curl -X POST http://localhost:8000/api/restart -H 'Content-Type: application/json' -d '{"dev":false}'

# Check current mode
curl -s http://localhost:8000/api/health | jq .mode

The restart endpoint waits for the old server to exit, starts the new one, and verifies health. It works identically on Windows, macOS, and Linux (no sh / lsof / curl dependency).

When started with --dev, the server proxies client requests to the Vite dev server for HMR. If Vite is not running, it automatically falls back to serving the production build from dist/client/:

• pi-dashboard start --dev always works — no 502 errors
• Start/stop Vite independently without restarting the dashboard
• Start Vite later and refresh the browser to get HMR

Set "devBuildOnReload": true in config.json for a one-command full-stack refresh:

/reload → build client → stop server → reload extension → auto-start fresh server

Blocks pi for ~2–5s during the build. The server shutdown affects all connected sessions — they auto-reconnect when one restarts the server.

Headless (default) — runs pi as a background process with no terminal attached. Interaction through the web UI.

tmux — runs pi inside a tmux session named pi-dashboard, each spawned session as a new window:

tmux attach -t pi-dashboard                # attach
tmux list-windows -t pi-dashboard          # list windows
# inside tmux: Ctrl-b n / p / w           # next / prev / picker

Switch with "spawnStrategy": "tmux" in ~/.pi/dashboard/config.json.

Bash-style history recall and per-session draft persistence:

Drafts (typed-but-unsent text) are persisted per session in localStorage under chat-draft:<sessionId> and survive navigation (Settings, OpenSpec preview, diffs, …) and full page reloads. Drafts never leak between sessions.

The dashboard integrates tightly with a small, curated set of pi extensions — for custom tool rendering, the Flow dashboard, and anthropic-messages protocol compatibility. The Electron wizard installs them in one go; the Packages tab and a top-of-page banner keep them discoverable afterwards.

Authoritative source: packages/shared/src/recommended-extensions.ts. Descriptions, versions, and installed-state are enriched live via GET /api/packages/recommended (offline fallback).

The dashboard's UI is composed of named slots that plugins claim with React components. To create a new plugin, install the scaffolding skill:

npm i -g @blackbelt-technology/pi-dashboard-plugin-skill

Then, from any pi session:

/skill dashboard-plugin-scaffold

The skill has two modes:

• new — scaffold a fresh packages/<id>-plugin/ inside this monorepo. Pick which of the 10 React slots to claim (session-card-badge, content-view, settings-section, tool-renderer, …); the skill renders package.json (with pi-dashboard-plugin manifest), src/client.tsx with stubs, optional src/server/index.ts, optional src/bridge/index.ts, configSchema.json, and tests.
• augment — retrofit an existing pi-extension project on disk. The skill greps for TUI surface (ctx.ui.*, pi.registerTool, …), drives the agent through a canonical mapping table, asks per-callsite what to port, then injects a manifest field into package.json and adds src/dashboard/. Purely additive — your existing TUI keeps working in pure-pi sessions.

For the slot taxonomy, the manifest schema, and the plugin context API, see the skill's reference docs (or the runtime: @blackbelt-technology/dashboard-plugin-runtime). The reference fixture is packages/demo-plugin/.

If pi launches but the dashboard never becomes reachable, inspect the launch log:

cat ~/.pi/dashboard/server.log                          # Linux / macOS
type %USERPROFILE%\.pi\dashboard\server.log             # Windows

The log is append-mode with timestamped headers per start attempt, so previous crashes are preserved. Common issues:

• ERR_UNSUPPORTED_ESM_URL_SCHEME on Windows — fully fixed in 0.4.0+. The 0.2.10 release wrapped the --import loader position as a file:// URL, but the entry-script position stayed a raw Windows path — which crashed Node on non-C: drives (A:\, B:\, …) because the drive-letter heuristic has gaps there. 0.4.0 routes all four server-spawn call sites through a single spawnNodeScript / toFileUrl helper that wraps both positions unconditionally, and a repo-level lint test prevents regression. Upgrade the package.
• Cannot find pi's TypeScript loader — pi is not installed globally. Run npm install -g @mariozechner/pi-coding-agent.
• Fastify crash at startup — you're on Node 22.0.0–22.17.x or 24.1.0–24.2.x which are affected by nodejs/node#58515. Upgrade to 22.18+ or 24.3+.

• Windows: netstat -ano | findstr :8000 then taskkill /F /PID <pid>
• Unix: lsof -t -i :8000 | xargs kill
• Or change port in ~/.pi/dashboard/config.json.

Since the safe-server-switch release, switching servers via the header dropdown is transactional: the UI verifies the target through a short-lived staging WebSocket (5 s timeout) before clearing state or writing localStorage. If the target is unreachable, nothing changes — a toast appears and you stay on the previous server.

If the currently-active server drops for more than 3 s, a yellow “Disconnected from <host>. Retrying…” banner appears at the top with a Switch server button — use it to pick a reachable server.

You should no longer need to manually localStorage.removeItem("pi-dashboard-last-server") to recover from a bad switch. If you still get stuck, please file an issue.

Since the consolidate-windows-spawn-and-platform-handlers release, pi sessions on Windows survive dashboard restart, matching macOS/Linux behaviour. Previously, killing the dashboard process (Task Manager, Ctrl+C, /api/restart, crash) terminated every running pi session because the children were in the server's libuv kill-on-close Job Object. The fix uses detached: true so children are excluded from the parent's job.

If you previously relied on "closing the dashboard cleans everything up," use the per-session Force Kill action instead (or POST /api/session/:id/force-kill).

Install Windows Terminal (wt.exe) for tabbed interactive sessions — the dashboard prefers it over WSL tmux / headless when available. Windows 11 ships with it; on Windows 10 install from the Microsoft Store.

If wt.exe is on PATH but launching does nothing, check Settings → Apps → Advanced app settings → App execution aliases. If the "wt" alias is disabled, wt.exe is found but can't be executed. Enable the alias or uninstall/reinstall Windows Terminal.

Since v0.3+, session grouping uses OS-aware path equality (platform/paths.ts). Sessions group correctly under a pinned folder even across trailing-separator, separator-style, or case differences (on Windows and macOS).

If you still see two entries for what should be one folder, the paths are likely on different Windows drives (A:\Foo and B:\Foo are different filesystems and never merge) — that's correct behaviour. If the paths really are the same filesystem, file an issue with both the pinned path (Settings → Tools → Export diagnostics) and the session cwd from /api/sessions.

Open Settings → General → Tools, click the chevron next to the failing tool to see the full tried[] trail, then either (a) install the missing tool on PATH / in the managed location shown in the trail, or (b) set an explicit override via the row's path input. Hit Rescan to pick up the change without a server restart.

pi-flows and pi-anthropic-messages install via SSH (git@github.com:…). If your system has no GitHub SSH key, set one up following GitHub's SSH docs, or substitute the equivalent HTTPS URL in the manifest if your fork is public.

graph LR
    subgraph "Per pi session"
        B[Bridge Extension]
    end

    subgraph "Dashboard Server (Node.js)"
        PG[Pi Gateway :9999]
        BG[Browser Gateway :8000]
        HTTP[HTTP / Static Files]
        MEM[(In-Memory Store)]
        JSON[(JSON Files)]
    end

    subgraph "Browser"
        UI[React Web Client]
    end

    B <-->|WebSocket| PG
    UI <-->|WebSocket| BG
    UI -->|HTTP| HTTP
    PG --- MEM
    PG --- JSON
    BG --- MEM

First-party features can live as separate monorepo packages by declaring a pi-dashboard-plugin field in their package.json. The Vite plugin auto-discovers these manifests at build time and generates a static registry (packages/client/src/generated/plugin-registry.tsx) using named imports for tree-shaking. During server startup, loadServerEntries() dynamic-imports each plugin's server entry and invokes registerPlugin(ctx: ServerPluginContext). See docs/architecture.md for the full plugin data flow.

See docs/architecture.md for detailed data flows, reconnection logic, and persistence model.

flowchart TD
    A[pi session starts] --> B[ensureConfig]
    B --> C[loadConfig]
    C --> D{TCP probe :piPort}
    D -->|Port open| E[Connect to server]
    D -->|Port closed| F{autoStart?}
    F -->|false| G[Skip]
    F -->|true| H[Spawn server detached]
    H --> I["Notify: 🌐 Dashboard started"]
    I --> E

The server is spawned detached (child_process.spawn with detached: true, unref()), so it outlives the pi session. Duplicate spawn attempts from concurrent pi sessions fail harmlessly with EADDRINUSE.

The health endpoint provides server and agent process metrics:

curl -s http://localhost:8000/api/health | jq

Returns:

• mode — "dev" or "production"
• server.rss, server.heapUsed, server.heapTotal — server memory
• server.activeSessions, server.totalSessions — session counts
• agents[] — per-agent metrics (CPU%, RSS, heap, event loop max delay, system load)

Agent metrics are collected every 15s via heartbeats and include eventLoopMaxMs — useful for diagnosing connection drops during long-running operations.

npm install          # Install dependencies
npm test             # Run all tests (vitest)
npm run test:watch   # Watch mode
npm run build        # Build web client (Vite)
npm run dev          # Start Vite dev server (HMR)
npm run lint         # Type-check (tsc --noEmit)
npm run reload       # Reload all connected pi sessions
npm run reload:check # Type-check + reload all pi sessions

# Terminal 1: Dashboard server in dev mode
npx tsx packages/server/src/cli.ts --dev

# Terminal 2: Vite dev server (HMR for the web client)
npm run dev

# Terminal 3: pi with the bridge extension
pi -e packages/extension/src/bridge.ts   # or just `pi` if installed

# Open http://localhost:8000 (server proxies to Vite for SPA routes + assets)
# Or http://localhost:3000 (Vite directly, proxies API/WS to :8000)

# After client changes (production mode)
npm run build
curl -X POST http://localhost:8000/api/restart

# After server changes (runs TypeScript directly, no build needed)
curl -X POST http://localhost:8000/api/restart

# After bridge extension changes
npm run reload

# Full rebuild (e.g., after pulling updates)
npm run build
curl -X POST http://localhost:8000/api/restart
npm run reload

Your own extensions can broadcast UI events to the dashboard:

pi.events.emit("dashboard:ui", {
  method: "notify",
  message: "Deployment complete!",
  level: "success",
});

Supported methods: confirm, select, input, notify.

Monorepo with npm workspaces — top-level layout only. See AGENTS.md for the full file-by-file index.

packages/
├── shared/      # Shared TypeScript types, protocols, config, session-meta helpers
├── extension/   # Bridge extension (runs inside pi) — WS client, PromptBus, event forwarding, auto-start
├── server/      # Dashboard server — HTTP + dual WebSocket gateways, in-memory store, terminals, auth, tunnel
├── client/      # React + Tailwind web client — 80+ components, hooks, event reducer
└── electron/    # Electron desktop wrapper — wizard, system tray, auto-update, bundled Node.js

Prerequisites: Node.js 22.12+; platform-specific tools handled by Electron Forge automatically.

npm run electron:build                    # Build for current platform & arch
npm run electron:build -- --arch x64      # Override architecture
npm run electron:build -- --skip-client   # Skip client rebuild

Or step by step:

npm run build                         # Build web client
cd packages/electron
bash scripts/download-node.sh         # Download Node.js for bundling
npm run make                          # Build installer

Output by platform:

From macOS or Linux, build installers for all platforms:

npm run electron:build -- --all              # macOS (native) + Linux + Windows (Docker)
npm run electron:build -- --linux            # Linux .deb + .AppImage only
npm run electron:build -- --windows          # Windows .zip only
npm run electron:build -- --linux --windows  # Both, skip native

On an Apple Silicon mac, produce both the arm64 and Intel x64 DMGs in one invocation:

npm run electron:build -- --mac-both

Requires Rosetta 2 (softwareupdate --install-rosetta --agree-to-license) so node-pty's x64 prebuilt binary can be unpacked during the cross-arch run. The script wipes per-arch caches between the two builds (resources/.last-arch sentinel) so back-to-back runs don't accidentally ship arm64 binaries inside an x64 DMG. Intel macs cannot cross-build arm64 locally (Rosetta is one-way) — use CI for arm64 validation.

Docker builds use a Node 22 Debian container for Windows cross-compilation. Output goes to packages/electron/out/make/.

# Start the dashboard server and Vite dev server first
pi-dashboard start --dev
npm run dev

# Then launch Electron pointing at the dev server
cd packages/electron
npm run start:dev

All platform icon variants are generated from the master icon at packages/electron/resources/icon.png:

cd packages/electron
npm run icons    # Generates .icns (macOS), .ico (Windows), and resized PNGs

See docs/release-process.md for the full cut-a-release workflow.

Every push to develop and every pull request against develop triggers ci.yml:

1. npm ci — install dependencies
2. npm run lint — type check
3. npm test — run tests
4. npm run build — build web client

The publish workflow (publish.yml) triggers on v* tags:

npm version patch   # or minor / major
git push --follow-tags

This runs CI, publishes to npm with --provenance for supply-chain transparency, and builds Electron installers for all platforms on native runners:

All artifacts are uploaded to a draft GitHub Release. Release notes are extracted automatically from the matching ## [<version>] section of CHANGELOG.md.

Build a one-off installer for a feature branch. No release, no publish, no tag.

Workflow: .github/workflows/ci-electron.yml. Trigger from GitHub Actions tab → CI Electron (on-demand) → Run workflow button.

Optional legs input narrows the matrix (default all; accepts darwin, linux, win32, or comma-list like darwin-arm64,linux-x64 for cheap iteration).

Version slug: <base>-ci.<UTC-stamp>.<branch-slug>.<sha7> (e.g. 0.5.3-ci.20260525-143000.feature-foo-bar.abc1234). Prerelease segment SemVer-ranks below <base>.

Download installers from the Actions run page → Artifacts section. 14-day retention.

Safe by construction: no npm publish, no GitHub Release, no auto-update impact. electron-updater default allowPrerelease: false skips -ci. slugs — installed users unaffected.

The publish workflow uses npm Trusted Publishers over OIDC — no NPM_TOKEN secret required. Short-lived, workflow-scoped credentials are exchanged between GitHub and npm at publish time, and every release carries automatic npm provenance tying the published artifact to the exact workflow run.

Requirements (already configured in publish.yml):

permissions:
  contents: write   # draft GitHub Release + tag push
  id-token: write   # OIDC token exchange with the npm registry
environment: npm-publish

Trusted Publishing requires npm CLI ≥ 11.5.1. The workflow upgrades npm automatically (npm install -g npm@latest) before publishing.

One-time npm-side setup — repeat once per published package (five scoped workspaces; @blackbelt-technology/pi-dashboard-electron is private and skipped):

For each:

1. Go to npmjs.com → the package → Settings → Trusted Publisher → GitHub Actions
2. Fill in:
   • Organization or user: BlackBeltTechnology
   • Repository: pi-agent-dashboard
   • Workflow filename: publish.yml (filename only, not the full path)
   • Environment name: npm-publish (must match the environment: field in the workflow)
3. Save

GitHub Environment (recommended) — configures an optional human-approval gate on every release:

1. GitHub repo → Settings → Environments → New environment → name npm-publish
2. Optionally add required reviewers and/or deployment branch/tag protection rules (e.g. restrict to v* tags)

No secrets to rotate, no tokens to leak.

MIT