R4 Platform VS Code Extension

VS Code extension for the R4 / RiVault AI Platform. Provides native editor integration with Gateway-authenticated LLM endpoints, endpoint management, job monitoring, and model browsing.

Features

• @r4 Chat Participant — Chat with R4 platform LLM models directly in VS Code Chat. Supports streaming responses and conversation history.
• Endpoint Management — Sidebar tree view to list, provision, and teardown inference endpoints.
• Job Monitoring — Sidebar tree view to list jobs, view logs, and cancel running/pending jobs.
• Model Browser — Browse the UMRS model registry with scope-aware filtering (system/project/user).
• Status Bar — Connection indicator showing Gateway health.
• Agent Window MCP Tools — Exposes R4 Platform and RAG MCP servers to Copilot agent mode and the VS Code Agents window.
• Native Model Provider — Registers R4 Platform LLM endpoints as native language models, selectable in the Copilot chat/edit/agent model picker. Supports streaming and tool calling.

Chat Commands

Setup

1. Install the extension
2. Run R4: Login to Gateway from the Command Palette
3. Enter your Gateway URL and API token
4. Ensure r4-mcp is available on your PATH, or set r4.mcp.command to the executable path

Alternatively, set r4.gatewayUrl in VS Code settings and the R4_GATEWAY_URL environment variable.

After login, VS Code discovers two MCP servers from this extension:

• R4 Platform — endpoint, job, workflow, model, and platform health tools
• R4 RAG — search_docs, list_documents, and rag_status

These servers appear in Copilot agent mode and the Agents window customizations panel. The extension passes Gateway credentials to the spawned MCP process through environment variables, not command-line arguments.

Using R4 Models in Copilot

There are two ways to drive Copilot chat / edit / agent on R4 Platform models (e.g. Kimi-K2.6, GLM-5.1, DeepSeek-V4-Pro). Both require a model with a ready endpoint — the model picker only lists models that currently have a warm/active endpoint.

Option A — Native provider (this extension)

1. Run R4: Login to Gateway.
2. Open the Copilot model picker → models appear under the R4 Platform vendor. Pick one for chat, edit, or agent mode.

Option B — Copilot BYOK via Custom Endpoint (no extension required)

This uses VS Code's built-in Custom Endpoint provider (which replaced the deprecated OpenAI Compatible provider). You do not create chatLanguageModels.json by hand — VS Code opens it for you during the flow.

1. Issue a long-lived JWT (admin): ... admin user add ... --issue-token --token-days 365 --gateway-url <gateway> --profile ai4s.
2. In VS Code, run Chat: Manage Language Models (Command Palette), or open the Chat model picker → Manage Language Models (gear icon).
3. Select Add Models → Custom Endpoint.
4. Enter a group name (e.g. R4 AI4S), a display name, and paste the JWT as the API key. Select API type Chat Completions.
5. VS Code opens chatLanguageModels.json. Paste the config below and save. The url must be the full chat-completions URL (<gateway>/gw/llm/v1/chat/completions), and each id must match the model id returned by GET <gateway>/gw/llm/v1/models.

[
  {
    "name": "R4 AI4S",
    "vendor": "customendpoint",
    "apiKey": "YOUR_R4_JWT",
    "apiType": "chat-completions",
    "models": [
      {
        "id": "Kimi-K2.6",
        "name": "Kimi-K2.6 (R4)",
        "url": "https://data0.ai.r-ccs.riken.jp/ai4s/gw/llm/v1/chat/completions",
        "toolCalling": true,
        "vision": false,
        "maxInputTokens": 253952,
        "maxOutputTokens": 8192
      },
      {
        "id": "DeepSeek-V4-Pro",
        "name": "DeepSeek-V4-Pro (R4)",
        "url": "https://data0.ai.r-ccs.riken.jp/ai4s/gw/llm/v1/chat/completions",
        "toolCalling": true,
        "vision": false,
        "maxInputTokens": 1040000,
        "maxOutputTokens": 8192
      },
      {
        "id": "GLM-5.1",
        "name": "GLM-5.1 (R4)",
        "url": "https://data0.ai.r-ccs.riken.jp/ai4s/gw/llm/v1/chat/completions",
        "toolCalling": true,
        "vision": false,
        "maxInputTokens": 194560,
        "maxOutputTokens": 8192
      }
    ]
  }
]

6. Pick the model from the chat model picker. If it does not appear, restart VS Code.

Config notes:

• toolCalling: true is required — models without it are hidden from the agent-mode picker.
• maxInputTokens should track each model's served context window (query GET <gateway>/gw/llm/v1/models and read max_model_len). Leave headroom for output. Current ai4s ceilings: DeepSeek-V4-Pro 1,048,576 (1M), Kimi-K2.6 262,144, GLM-5.1 202,752. Only DeepSeek serves a 1M window; the other two are capped at their trained context and cannot be raised to 1M.
• With Option A (native provider) the context window is derived from max_model_len automatically — no per-model token config needed.
• apiKey is sent as Authorization: Bearer <JWT>, which the Gateway expects.

Note: Both paths drive Copilot chat, edit, and agent (tool calling) only. Neither drives inline ghost-text completions, which stay on Copilot's own model.

Settings

Architecture

The extension is a thin client over Gateway HTTP APIs:

VS Code Extension → Gateway HTTP (/gw/*) → UMRS

• No direct UMRS or LiteLLM access
• API token stored in VS Code SecretStorage (never plaintext)
• Graceful offline degradation when Gateway is unreachable
• All MCP tool calls pass through Gateway policy and audit
• Agent Window integration uses VS Code's MCP provider API; it does not bypass Gateway or call UMRS/LiteLLM directly.

Development

cd vscode-extension
npm install
npm run compile        # Type-check + build
npm run watch          # Watch mode (esbuild)

Press F5 in VS Code to launch the Extension Development Host.

Packaging & Installing the .vsix

To build a distributable .vsix and install it into VS Code:

cd vscode-extension
npm install                                          # first time only

# Build the production bundle and package the .vsix.
# (uses @vscode/vsce; npx fetches it on demand — no global install needed)
npx --yes @vscode/vsce package --allow-missing-repository --skip-license
# → produces r4-platform-<version>.vsix in this folder

Install the packaged extension:

# From the command line
code --install-extension r4-platform-0.1.1.vsix

# To remove a previously installed copy first
code --uninstall-extension r4-platform.r4-platform

Or install from the VS Code UI: Extensions view → ... menu → Install from VSIX... → select the .vsix file.

After installing, reload the window (Developer: Reload Window) and run R4: Login to Gateway to authenticate.

Notes:

• The .vsix is a build artifact and is git-ignored; rebuild it from source rather than committing it.
• npm run package runs the same production build that vsce package invokes via vscode:prepublish, so type errors fail the package step.

Project Structure

src/
  extension.ts          — Entry point, wires all components
  gateway-client.ts     — HTTP client for all Gateway communication
  auth.ts               — SecretStorage-based auth management
  status-bar.ts         — Connection status indicator
  chat-participant.ts   — @r4 Chat Participant
  views/
    endpoints.ts        — Endpoint tree view
    jobs.ts             — Job tree view
    models.ts           — Model tree view
resources/
  r4-icon.svg           — Activity bar icon