deepseek-migrationlisted

# Migrating an agent setup to DeepSeek DeepSeek is OpenAI-API-compatible, and Anthropic-compatible endpoints exist for harnesses like Claude Code — so migration is mostly configuration, not rewrites. What actually breaks is rarely the endpoint; it's caching behavior, tool calling, and cost assumptions. ## Step 1 — pick the integration path | Current setup | Path | |---|---| | OpenAI SDK / any OpenAI-compatible tool | Point `base_url` to `https://api.deepseek.com`, swap the key, set model ids | | Claude Code | Use the documented Anthropic-compatible endpoint (env-var style config) | | Want a native harness | DeepSeek Harness (`dsh`): `npx @deepseek-ai/dsh web` — reads `DEEPSEEK_API_KEY` | | Editor plugins (Cline etc.) | Custom provider with DeepSeek base URL | Official per-tool guides: https://api-docs.deepseek.com/guides/coding_agents Model ids: **`deepseek-chat`** (fast, cheap, non-thinking) and **`deepseek-reasoner`** (chain-of-thought, slower, pricier). ## Step 2 — the four things that actually differ 1. **Prefix caching is automatic and prefix-sensitive.** DeepSeek bills repeated identical prompt prefixes at a steep discount — but the prefix must be *identical*. Put the stable system prompt and repo map first; never shuffle timestamps or dynamic content in front of them. A reordering silently kills your cache-hit rate. 2. **`deepseek-reasoner` thinking tokens bill as output.** A reasoner call can cost many multiples of a chat call. Route: routine edi