Codex Subscription Provider for Copilot

An unofficial Visual Studio Code extension that connects GitHub Copilot Chat to local Codex models using your ChatGPT subscription. Codex handles reasoning while Copilot executes tools and workspace edits, without requiring an OpenAI API key.

[!IMPORTANT] This is an independent, unofficial project. It is not affiliated with or endorsed by OpenAI, GitHub, or Microsoft.

简体中文

How It Works

GitHub Copilot Agent
  -> conversation, images, selected model, and available tools
Codex Subscription Provider
  -> local Codex runtime and ChatGPT/Codex subscription
Codex
  -> text response or structured tool-call request
Provider
  -> response returned to Copilot Agent
GitHub Copilot Agent
  -> executes VS Code tools, MCP calls, and workspace edits

The provider does not register or execute SAP, MCP, terminal, or file-editing tools itself. It only forwards the tools supplied by Copilot for the current request. Copilot remains responsible for tool execution, confirmation, and virtual-file changes.

Features

• Explicit Codex model selection in the Copilot model picker
• Configurable model reasoning effort
• Image and screenshot input
• Structured forwarding and required-call enforcement for Copilot tools
• Action evidence checks for read, write, activation, and test workflows
• One-time setup for health checks, performance, and tool policy
• ChatGPT subscription authentication with API-key authentication rejected
• Automatic discovery of the local Codex runtime
• Read-only, isolated Codex working directory with network and MCP disabled
• Runtime, model-catalog, ABAP FS tool, and last-request diagnostics
• No extension-specific analytics or telemetry

Requirements

• Visual Studio Code 1.109.0 or later
• GitHub Copilot Chat in Agent mode
• Codex signed in with Sign in with ChatGPT
• For the Thin distribution, an official Codex runtime available from an explicitly configured path, the system PATH, or the official OpenAI VS Code extension

Installation

Install the Thin distribution from the Visual Studio Marketplace, or install a VSIX from GitHub Releases.

The one-time setup verifies the environment and offers a recommended Balanced performance profile and require-for-actions tool policy.

After installation:

1. Run Developer: Reload Window.
2. Complete the one-time setup or run Codex Subscription Provider: Run Health Check.
3. Open Copilot Chat and switch to Agent mode.
4. Select an explicit Codex Subscription | ... model.
5. Enable only the tools needed for the task, including ABAP FS tools when working with SAP virtual files.
6. Ask Copilot to analyze, edit, review, activate, or test your workspace.

Configuration

{
  "copilotCodexProvider.runtimeMode": "auto",
  "copilotCodexProvider.codexExecutablePath": "",
  "copilotCodexProvider.modelSource": "discovered",
  "copilotCodexProvider.enabledModels": [],
  "copilotCodexProvider.reasoningEffort": "medium",
  "copilotCodexProvider.toolUsePolicy": "require-for-actions",
  "copilotCodexProvider.missingToolBehavior": "notify-and-stop",
  "copilotCodexProvider.showOnboarding": true
}

Runtime Modes

• auto: configured path, system PATH, official OpenAI VS Code extension, then the bundled runtime when available.
• external: require an external official Codex runtime.
• bundled: require the runtime included in a Full distribution.

Models and Performance

The extension discovers subscription-visible models through the local codex app-server model catalog. Each model is exposed as a separate Copilot model-picker entry. The selected model ID is passed to Codex without a silent fallback.

Use Codex Subscription Provider: Choose Performance Profile to select:

• Fast: low reasoning
• Balanced: medium reasoning
• Deep: high reasoning

The underlying reasoningEffort setting also supports model-default, minimal, low, medium, high, xhigh, max, and ultra. Unsupported values are rejected instead of silently downgraded.

Tool Policies

• copilot-default: follow the tool mode supplied by Copilot.
• require-for-actions: require tool-result evidence for detected read, write, activation, and test actions.
• strict-abapfs: apply action enforcement only in ABAP or SAP context.

With the recommended notify-and-stop missing-tool behavior, the provider stops before calling Codex when Copilot does not supply a required capability. It then identifies the missing capability instead of allowing an ungrounded claim that an action was completed.

The provider can only enforce tools that Copilot supplies. It does not bypass Copilot or call ABAP FS/MCP tools directly.

Image Input

• Up to 20 images and 50 MiB per request
• Duplicate images removed using SHA-256
• Temporary image files deleted after each request
• Requests rejected when the selected model does not declare image support

Diagnostics

• Codex Subscription Provider: Manage
• Codex Subscription Provider: Run Health Check
• Codex Subscription Provider: Check Codex Runtime
• Codex Subscription Provider: Check ABAP FS Tools
• Codex Subscription Provider: Refresh Models
• Codex Subscription Provider: Show Available Models
• Codex Subscription Provider: Show Last Request Details
• Codex Subscription Provider: Reset Onboarding

Privacy and Security

The extension checks the local Codex ChatGPT authentication state and removes common API-key and custom OpenAI endpoint variables before starting Codex. It does not log prompts, source code, image contents, credentials, tool arguments, or tool results.

Requests are processed by the selected local Codex runtime under the applicable OpenAI terms. Copilot and any enabled MCP tools remain subject to their own terms and privacy policies.

See PRIVACY.md, SECURITY.md, and THIRD_PARTY_NOTICES.md for details.

Development

npm ci
npm run validate
npm run package:thin

The test suite covers subscription-auth validation, runtime resolution, model catalog parsing, image handling, action and tool-capability policies, structured tool-call bridging, and SDK loading.

Contributions are welcome. Read CONTRIBUTING.md before opening an issue or pull request.

Distributions

Marketplace releases use the Thin package. Full Win64 packages are distributed only through GitHub Releases.

License

The extension source code is licensed under the MIT License. OpenAI Codex components retain their original Apache-2.0 license.