UAI Copilot Cluster Intelligence (VS Code Extension)

A VS Code extension that connects to the UAI Copilot backend and provides cluster intelligence, diagnostics, and operational guidance directly from within the editor.

Versioning policy: publish flow always increments from Marketplace by exactly +0.0.1.

Shared Backend Contract (UAI-Extension + UAI-Android)

The extension and Android client intentionally share the same backend API surface for health/system/cluster telemetry.

• One backend, two clients.
• Endpoint set is shared by design to keep behavior and diagnostics consistent across desktop and mobile.
• Extension status/tooltip now shows cluster average ping across online nodes and local node ping separately.

For operator access, the wallet panel shows an Open Admin Panel button for the UAI-Leader account only.

Getting Started

1. Install the extension from the VSIX or via the Marketplace (once published).
2. Ensure the backend API is running and accessible (default: http://127.0.0.1:8081).
3. Open the extension from the command palette: UAI Copilot: Open.
4. For cluster/local LLM chat, open UAI Chat from the activity bar or UAI Copilot: Open UAI Chat (uses uaiCopilot.chat.* + OpenAI-compatible completions).

Offline-first (no public internet)

The default API URL is loopback (http://127.0.0.1:8081). After container images are on disk, you can run the stack and extension without outbound internet: start the backend from the automation tool repository root with docker compose up -d (see root docker-compose.yml). Wallet/demo auth use VS Code local storage and do not require cloud sign-in. Optional LAN discovery only scans private IPv4 subnets on your network. See modules/cluster_intelligence/README.md for how this package maps to UAI-Cluster-Intelligence and for mesh op...

Configuration

The extension supports configuration under uaiCopilot in VS Code settings. You can also place a .vscode-extension-config.json file in the workspace root to override settings per-project.

• uaiCopilot.apiUrl (default: http://127.0.0.1:8081): the base URL of the backend API.
• uaiCopilot.enableHealthCheck (default: true): whether to validate connectivity when the extension starts and to run periodic health checks.
• uaiCopilot.healthCheckInterval (default: 30000): interval (ms) between periodic health check pings when health checks are enabled.
• uaiCopilot.timeout (default: 5000): request timeout (ms) for the health check.
• uaiCopilot.retryAttempts (default: 3): number of retry attempts when connecting to the backend.
• uaiCopilot.retryDelay (default: 1000): delay (ms) between retry attempts.
• uaiCopilot.healthCheckPaths (default: [/health, /database/health]): array of endpoint paths to attempt during health checks.
• uaiCopilot.enableDebugLogging (default: false): enable verbose logging for debugging.
• uaiCopilot.notificationFailureThreshold (default: 1): consecutive failed checks before showing a disconnection notification.
• uaiCopilot.healthCheckHistorySize (default: 25): maximum number of health check results retained in the diagnostics history.

UAI Chat (uaiCopilot.chat.*): Activity bar (UAI Chat) webview sends messages to uaiCopilot.chat.chatCompletionsUrl, or to {uaiCopilot.apiUrl}/v1/chat/completions when that URL is empty. Optional chat.systemPrompt is prepended as a system message. uaiCopilot.chat.disableCopilotWorkbenchChat (default true) prevents the extension from calling VS Code workbench.action.chat.changeModel — use UAI Chat for local/cluster inference instead of Copilot’s Chat UI.

Copilot Auto lock (uaiCopilot.copilotChatAutoLock.*): keeps inline completions on github.copilot.selectedCompletionModel = Auto (Vendor Default) when enabled; the poll timer is inline-only. When disableCopilotWorkbenchChat is false, the same subsystem can re-apply workbench.action.chat.changeModel per reapplyChatIntervalMs (≥ 15000; 0 disables). modelVendor must stay copilot for Copilot Auto chat when you are not using cluster LLM (auto is normalized to copilot). GitHub can still cap usage server-side — configure billing if you need higher limits.

Cluster / Copilot-chat steering (uaiCopilot.clusterLlm.*): when disableCopilotWorkbenchChat is false, useForChatAutoLock routes changeModel to clusterLlm.chatVendor / clusterLlm.chatModelId. Ignored while disableCopilotWorkbenchChat is true. Configure VS Code’s BYOK/OpenAI-compatible provider when you use Copilot Chat with BYOK (http://<host>:11089/v1 when running ../../UAI-Cluster-Intelligence/scripts/llm_service.py); see README_LLM.md. Preflight docker-compose step 0c merges github.copilot.chat.byok.ollamaEndpoint, clusterLlm.*, chat.disableCopilotWorkbenchChat, and (when UAI_LLM_GATEWAY_URL is set) chat.chatCompletionsUrl into .vscode/settings.json unless UAI_SKIP_VSCODE_SETTINGS_SYNC=1.

Cluster & Routing

The extension can load-balance and fail over across multiple cluster endpoints automatically.

As of 3.10.0, node discovery is dynamic at runtime:

• the extension performs periodic LAN discovery across local private subnets,

• new reachable nodes are merged into the router immediately,

• metadata is collected directly from node health payloads (hostname, CPU, RAM, GPU, TF compute),

• no local state file is required for node onboarding.

• uaiCopilot.clusterEndpoints (default: []): list of cluster endpoint URLs to probe in parallel. When non-empty, overrides apiUrl for routing. Example: [http://127.0.0.1:8081, http://192.168.1.81:8081].

• uaiCopilot.routingStrategy (default: "round-robin"): how the next endpoint is chosen. Options: round-robin (even distribution), failover (always prefer the first healthy node), fastest (lowest-latency node wins).

The status bar shows live cluster health: CPU%, RAM, compute label, and a connected/discovered node ratio (e.g. 2/9 nodes). A hardware stats item shows total CPU, GPU/CPU mode, RAM, and estimated Teraflops.

Live updates (WebSocket)

As of 3.11.0, the extension also subscribes to the backend's WebSocket telemetry channels (/clusters, /system, /analytics) for near-real-time status updates, instead of waiting for the next polling tick.

• uaiCopilot.enableLiveUpdates (default: true): subscribe to the live channels. This supplements periodic health-check polling; if the socket never connects, the extension keeps working with polling only (offline-first).
• uaiCopilot.liveUpdatesReconnectMs (default: 5000): base reconnect delay; backoff grows exponentially up to a 60s cap.

The WebSocket URL is derived from uaiCopilot.apiUrl (e.g. http://127.0.0.1:8081 → ws://127.0.0.1:8081/clusters). Live /analytics frames and the master/slaves cluster status (/api/cluster-status) are surfaced in the diagnostics panel.

Tooltip behavior now includes:

• consistent connectivity header semantics with status bar,
• explicit LAN and Cloud online ratios,
• per-node scope tags like [LAN] and [Cloud],
• local system indicator (🟢/🟡) derived from local endpoint health.

Phase 288 (Android API) Configuration

For deployments using the Phase 288 Android API service, configure the extension to query the enriched health endpoints:

Settings (.vscode/settings.json or VS Code settings UI):

{
  "uaiCopilot.apiUrl": "http://192.168.1.98:8080",
  "uaiCopilot.healthCheckPaths": [
    "/health",
    "/system",
    "/clusters"
  ],
  "uaiCopilot.healthCheckInterval": 30000,
  "uaiCopilot.timeout": 5000
}

The /health endpoint on phase 288 includes:

• System hostname and timestamp
• CPU usage percentage
• RAM metrics (used/total GB)
• GPU details and estimated TFLOPS (via nvidia-smi when available)
• GPU compute estimates for RTX, Tesla, and AMD models

The /system endpoint provides additional system status and metrics. The /clusters endpoint returns live cluster node state.

This ensures the extension properly displays:

• Total Compute TF values in tooltips (e.g., "29.3TF", "2.1PF")
• Per-node GPU utilization and memory
• Accurate hostname resolution (vs. IP duplication)

Diagnostics, Feedback, and Issue Reporting

The diagnostics panel includes a feedback submission form, a Cluster Status table (master/slaves from /api/cluster-status), and a Live Analytics section fed by the /analytics WebSocket channel. The extension also captures runtime warning/error context and queues issue reports for submission when the backend is reachable.

Capabilities:

• submit feedback from diagnostics or command palette,
• include/exclude diagnostics payload on feedback submission,
• capture logger warnings/errors,
• capture guarded runtime errors and process-level unhandled failures,
• redact sensitive fields (token, password, apikey, etc.) before queuing/submission.

Reports are submitted via POST {apiUrl}/extension/reports (handled by the Phase 288 backend in android_api_server.py); the redacted payload is stored server-side and is retrievable by an operator via the token-gated GET /extension/reports.

Commands

• UAI Copilot: Open (uai-copilot.open): Opens the UAI Copilot extension.
• UAI Copilot: Open UAI Chat (uai-copilot.openUaiChat): Focuses the UAI Chat sidebar (OpenAI-compatible chat completions).
• UAI Copilot: Reconnect (uai-copilot.reconnect): Forces a re-check of the backend connection.
• UAI Copilot: Diagnostics (uai-copilot.diagnostics): Opens a diagnostics panel with connection state and logs.
• UAI Copilot: Test Connection (uai-copilot.testConnection): Performs a single connection test and shows results.
• UAI Copilot: Report Issue / Send Feedback (uai-copilot.reportIssue): Submits user feedback with optional diagnostics context.

Troubleshooting

• If you see Connection Error / -1ms, verify that the backend's health endpoint is reachable:
  • http://127.0.0.1:8081/health
  • http://127.0.0.1:8081/database/health

Development

This extension is built in TypeScript and packaged with vsce.

cd modules/UAI-Extension/vscode-extension
npm install
npm run compile
npm run package
npm run publish