feat: add DeepSeek Responses provider

[ZhengdaWang (zhengdawang443)]

Summary

• Add DeepSeek as a first-class model provider using the official Responses API.

• Add CLI and Control Plane configuration, model discovery, persistence, readiness, and connection testing.

• Reuse the existing OpenAI-compatible Responses backend without introducing another agent loop.

• Store credentials in ~/.fsq/auth/deepseek.json with atomic provider replacement.

Model Discovery

• Preserve the canonical deepseek-flash alias.

• Accept versioned Flash models from 4.1 onward.

• Accept versioned Pro models from 4.0 onward.

• Apply no upper version bound.

• Exclude specialized or suffixed variants without using a fixed model allowlist.

Runtime Behavior

• Use the fixed [https://api.deepseek.com/](https://api.deepseek.com/%60) endpoint.

• Revalidate exact model membership before saving.

• Do not probe image capabilities or require image support.

• Do not retry image failures, substitute models, or fall back to another provider.

• Keep tool continuation bounded by the existing agent turn limit.

UI

• Add a dedicated DeepSeek Settings form.

• Keep discovery, save, cancellation, and recovery state isolated from other providers.

• Add DeepSeek to Home summaries and connection result dialogs.

• Preserve existing keyboard, focus, responsive, and secret-handling behavior.

Verification

• pytest -q: 3025 passed, 1 skipped

• ruff check .: passed

• npm test: 286 passed

• npm run typecheck: passed

• npm run build: passed

• uv build: wheel and source distribution passed

• Browser checks passed at 1440×900, 1024×768, and 390×844

• Full SPEC implementation audit and repair-only verification audit passed

Notes

A live DeepSeek credential smoke test was intentionally not run. No dependencies were added or changed.