opencode-copilot-auth

Package on npm: https://www.npmjs.com/package/@zhzy0077/opencode-copilot-cli-auth

A fork of anomalyco/opencode-copilot-auth that adds native 1M token context for Claude models by routing requests through the Anthropic Messages API (/v1/messages) instead of the OpenAI-compatible /chat/completions endpoint.

How to use

Add the plugin to your opencode config:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "@zhzy0077/[email protected]"
  ]
}

Then start opencode and log in to the github-copilot provider using the device flow.

For local development before publishing, you can load the file directly:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "file:///absolute/path/to/index.mjs"
  ]
}

Note: if the file path contains opencode-copilot-auth, current opencode builds may skip loading it due to a hardcoded plugin-name filter. Use a path that does not contain that substring.

What this fork adds over upstream

1M context for Claude models

The Copilot API exposes a native Anthropic Messages endpoint at /v1/messages that accepts up to 1,000,000 tokens — far beyond the 168K hard cap on /chat/completions. This fork automatically routes all Claude model requests to that endpoint and translates the request/response format transparently.

The Copilot API metadata reports max_context_window_tokens: 200000 for most Claude models. This fork overrides those values to 1,000,000 for any model that supports /v1/messages, so opencode sends the correct window size.

Empirically verified (needle-in-haystack tests):

• claude-opus-4.7: 350K tokens ✓, 1M tokens ✓, 1.44M tokens ✗ (confirmed 1M ceiling)
• claude-sonnet-4.6: 302K tokens ✓ (secret retrieved correctly)

More models (20 vs upstream's baseline)

Uses the VSCode OAuth client (Iv1.b507a08c87ecfe98) and api.githubcopilot.com, which exposes 20 picker models including gemini-3.1-pro-preview, gemini-3-flash-preview, and gpt-4o that the enterprise endpoint omits.

Persistent token caching

The short-lived Copilot token (exchanged via /copilot_internal/v2/token) is written back to the opencode auth store with a 5-minute expiry buffer. When you run multiple opencode instances simultaneously, only the first one fetches a new token — the rest reuse the persisted one.

Model limits exposed to opencode

| Model | Context | Input (prompt) | Output | |----------------------|-----------:|---------------:|-------:| | claude-opus-4.7 | 1,000,000 | 1,000,000 | 32,000 | | claude-opus-4.6 | 1,000,000 | 1,000,000 | 32,000 | | claude-opus-4.6-1m | 1,000,000 | 936,000 | 64,000 | | claude-sonnet-4.6 | 1,000,000 | 1,000,000 | 32,000 | | claude-sonnet-4.5 | 1,000,000 | 1,000,000 | 32,000 | | claude-haiku-4.5 | 1,000,000 | 1,000,000 | 64,000 |

Claude thinking budget

When the thinking variant is selected, the plugin sends thinking_budget: 16000. When no variant is selected, it omits the field entirely.

Publishing

./script/publish.ts