@
🔀 Multi-provider: Claude + DeepSeek V4
Agents can now run on DeepSeek V4 Pro / Flash alongside Anthropic Claude (Opus / Sonnet / Haiku), selected per agent from the same model dropdown.
Highlights
- Provider priority: DeepSeek API key → Anthropic API key → Claude Pro/Max login. No code changes.
- DeepSeek via Anthropic-compatible endpoint (
api.deepseek.com/anthropic) — same Agent-SDK request format, tool calls, thinking mode, 1M-token context window. - Role defaults on DeepSeek: chiefs / head architect / advisors → DeepSeek V4 Pro; one-shot specialists / workers → DeepSeek V4 Flash.
- Live switch between Claude and DeepSeek mid-session; routing follows the selected model per request.
- Correct per-model cost tracking for both providers.
- In-app API keys panel (keys stored locally, never committed).
- Identity fix: agents state their real model and never misreport as Claude.
- DeepSeek treated as 1M context (compaction threshold + UI gauge).
Notes
- Claude path is unchanged: OAuth / Pro-Max / API-key all work as before.
- The local proxy never touches credentials or the OAuth refresh path.
🤖 Generated with Claude Code
@