DeepSeek GitHub Copilot CLI setup: use the official Anthropic route and stop fighting the wrong provider type

DeepSeek's official Copilot CLI page is unusually explicit: use `anthropic` as the provider type. This is not a preference tip. It is the documented way to avoid a real request-shape mismatch.

The same page says the `openai` provider path triggers a `400` error because Copilot CLI's OpenAI integration does not echo `reasoning_content` back on later requests. DeepSeek's Anthropic Messages endpoint avoids that entire problem.

DeepSeek's page installs GitHub Copilot CLI with `npm install -g @github/copilot` and says Node.js 22 or later is required. Start there before you debug provider wiring.

If the binary is already installed but a teammate copied an older setup note, confirm the Node.js version and the binary path before changing environment variables. The official DeepSeek page assumes a modern Copilot CLI baseline.

DeepSeek's documented Linux and macOS setup is straightforward: set `COPILOT_PROVIDER_TYPE=anthropic`, point `COPILOT_PROVIDER_BASE_URL` at `https://api.deepseek.com/anthropic`, pass a DeepSeek API key, and set `COPILOT_MODEL` to `deepseek-v4-pro`.

The same page says `deepseek-v4-flash` is also available. Treat Pro as the documented baseline and Flash as the cheaper lane when the workload is broader, faster, or more repetitive.

DeepSeek's page adds one detail that many generic tutorials miss: because `deepseek-v4-pro` is not in Copilot CLI's built-in catalog, the docs tell users to set token ceilings explicitly.

The official values are `840000` max prompt tokens and `128000` max output tokens. Those numbers matter for correctness as much as for UX, because a custom model route is only stable when the client understands the model's bounds.

DeepSeek's docs also publish an offline mode flag: `COPILOT_OFFLINE=true`. The name can mislead teams if they do not read the note carefully.

The page says offline mode only blocks GitHub API calls. Your prompts still go to `api.deepseek.com`. That means offline mode is about reducing GitHub-side dependencies, not about local model execution or air-gapped inference.

DeepSeek's official examples start with `deepseek-v4-pro`, which is the right documented baseline for quality-sensitive coding-agent work. Use that when you are validating a new setup or need more careful reasoning.

Switch to `deepseek-v4-flash` when the Copilot CLI session is mostly repeated shell help, broad repo exploration, or lower-cost everyday coding. The important rule is to keep the provider and endpoint contract unchanged while you benchmark model choice.

If your team also uses the VS Code chat extension, compare this page with `/guides/deepseek-github-copilot-chat` because the extension flow and the CLI BYOK flow are related but not identical.

If the next issue is endpoint selection more broadly, continue with `/guides/deepseek-openai-vs-anthropic-api-routing`.

If you want another terminal-first agent setup route, compare `/guides/deepseek-deep-code-cli`, `/guides/deepseek-hermes-agent-setup`, and `/guides/reasonix-deepseek-coding-agent`.

If the task is source verification rather than setup, use `/guides/deepseek-official-x-updates` to distinguish current official docs from stale or weak social snippets.