bring your own Claude / Codex / Qwen / Kimi / Copilot plan

Instead of a raw model endpoint, DGC can drive a coding CLI you already pay for and are logged into — from the full-screen app, the classic inline REPL (dgc --classic), the editor/headless backend, or a one-shot dgc -p --engine run. Every route renders the vendor's stream the same way: its tool steps as cards with their diffs, its reasoning dimmed, and its answer as it arrives.

• Claude Code — your Anthropic Pro / Max plan (claude).
• Codex — your ChatGPT Plus / Pro plan (codex).
• Qwen Code — your Qwen OAuth plan (qwen).
• Kimi for Coding — your Moonshot plan (kimi); its prompt mode is auto-only, so DGC must be in auto and the vendor controls its automatically approved internal tool actions.
• GitHub Copilot CLI — your Copilot plan (copilot).

This is orchestration, not credential replay. DGC never reads, stores, refreshes, or replays the vendor's tokens and never sends vendor-private headers. Each engine authenticates through the vendor's own login command, which opens the vendor's own browser / device flow and keeps the token in its own store. When you run a turn, DGC shells out to the official binary in your workspace and streams its output into DGC's UI. DGC saves the prompt and normalized final assistant response in its own session; vendor tool and thinking events remain display-only. Authentication, model execution, and internal tools remain owned by the selected vendor CLI, while DGC wraps the turn with its own prompt, mode, session mapping, timeout, and cleanup. The vendor process is not wrapped by DGC's OS sandbox and inherits DGC's ambient environment, including unrelated credentials. Use a trusted vendor CLI and launch DGC with only the environment secrets that turn needs.

• dgc setup lists subscription engines first, with each one's detected login marker. The vendor performs the actual authentication check when its CLI launches.
• /connect claude|codex|qwen|kimi|copilot selects an engine (a direct /connect to a provider or URL turns delegation back off).
• Sign in once with the vendor's own command, e.g. claude auth login, codex login, qwen (device code), kimi login, copilot login. dgc doctor reports status.

• /model steers the vendor's own model (e.g. opus/sonnet/haiku for Claude).
• /think low|medium|high|xhigh|max sets the reasoning effort for engines that take one (Claude, Codex, Copilot); engines without an effort flag steer it via /model.
• The editor's composer controls, /model and /think, and the corresponding command-palette pickers all follow the active route. Subscription model pickers use vendor aliases when known and otherwise accept a vendor model id directly; they never query the configured native endpoint.
• In one-shot automation, --model and --think override only that delegated dgc -p --engine turn. They do not rewrite the saved native model or thinking level.