Pincushion Feedback — VS Code Extension

View, claim, and resolve stakeholder feedback pins from Copilot Chat.

Setup

Install the extension from the VS Code Marketplace
The Pincushion MCP server is auto-configured in your VS Code settings
Reload the window when prompted
Open Copilot Chat and type @pincushion

You can also run Pincushion: Configure MCP Server from the Command Palette at any time.

Commands

Command	What it does
`@pincushion /pins`	Show all open feedback pins grouped by page
`@pincushion /my-pins @username`	Show pins where you've been @mentioned
`@pincushion /resolve <pin-id>`	Claim and resolve a pin (omit ID to pick from a list)
`@pincushion /feedback-summary`	High-level project overview with priority order

Requirements

VS Code 1.99+
GitHub Copilot (for chat)
Node.js 22.22+ or Node.js 24 for the MCP server via npx (Node.js 18 and Node.js 20 are unsupported)

Settings

Setting	Default	Description
`pincushion.projectDir`	workspace root	Path to your project directory for the MCP server
`pincushion.autoConfigureMcp`	`true`	Auto-add the MCP server config on activation

Manual MCP config

If you prefer to configure manually, add this to your VS Code settings.json:

{
  "mcp": {
    "servers": {
      "pincushion": {
        "command": "npx",
        "args": ["pincushion-mcp", "--project-dir", "/path/to/your/project"]
      }
    }
  }
}

Links

Extension-host regression

npm run test:host compiles and runs the extension in official VS Code 1.105.1 using pinned @vscode/test-electron. The first run downloads that runtime. A separate temporary profile, extensions directory, and two-folder workspace keep the installed IDE and its settings untouched. No Copilot subscription, cloud model, Pincushion account, or live MCP server is used. Linux CI needs a GUI display (for example xvfb-run -a npm run test:host).

The declared minimum host can be tested with PINCUSHION_VSCODE_VERSION=1.99.0 npm run test:host. The harness accepts only the default and declared-minimum versions so an accidental runtime substitution cannot be reported as compatibility evidence.

On managed Macs, run with scoped process cleanup:

python3 /Users/josh/.codex/skills/mac-maintenance/scripts/run_scoped_test.py -- npm run test:host

PINCUSHION_VSCODE_CACHE selects the download cache. For offline runs, PINCUSHION_VSCODE_EXECUTABLE may point to an already downloaded executable matching the selected version. Missing runtimes, download failures, host failures, and version mismatches fail the command; they are never counted as skipped passing tests.

The real host checks workspace-scoped configuration without overwriting global servers (including canonical JSONC configuration), refuses invalid unsaved configuration, routes the chat request to the second root, invokes a registered synthetic language-model tool through vscode.lm.invokeTool, feeds its result back to the model, and rejects unoffered calls and cancelled requests. The model response is synthetic; this is host/API evidence, not a Copilot or cloud MCP end-to-end test.

Chat tool execution uses the request's explicitly selected tool references. Automatic discovery of the correct MCP server's tools is not implemented: the stable tool API does not provide server ownership metadata. Select tools from the intended project server. Project selection controls prompt context; it cannot establish ownership of an explicitly selected tool. Calls are bounded to eight model rounds and 32 unique tool calls per request.