DeepSeek API integration

Why#

Local models on the 4060 (Ollama, 7-16B class) work for simple text tasks but flake on multi-step tool chains. Anthropic Sonnet/Opus is reliable but expensive at volume. DeepSeek's hosted API (api.deepseek.com) sits in the middle: their deepseek-chat and deepseek-reasoner models are meaningfully cheaper than Anthropic, OpenAI-compatible API, and competitive on coding/reasoning benchmarks. Good fit for high-volume workers like Debbie/Chris when local fails.

What's in scope#

1. Get a DeepSeek API key from platform.deepseek.com (Maxime side).
2. Store the key securely in ~/.openclaw/.env (NOT committed to any repo).
3. Register DeepSeek as an OpenClaw provider in ~/.openclaw/openclaw.json (OpenAI-compatible).
4. Add aliases deepseek-chat and deepseek-reasoner so cron jobs and agent overrides can use friendly names.
5. Smoke test from CLI: openclaw infer model run --model deepseek-chat --prompt "say: ok".
6. Smoke test through an actual agent turn (multi-step tool chain — same Christmas joke test we did for local DeepSeek).
7. Decide which agent's default should switch (probably Chris and/or Debbie if the local one keeps failing).
8. Document everything in MEMORY.md so we can re-create from scratch.

What's out of scope (for now)#

• DeepSeek's chat-completion fine-tuning, embeddings, or other endpoints — only chat for the agent turn.
• Cost dashboards specific to DeepSeek (general cost tracking will pick it up).
• Routing every cron job to it — we'll do that case-by-case once we know it works.

Security rules#

• API key NEVER appears in:
  • openclaw.json (use ${env:DEEPSEEK_API_KEY} ref)
  • any committed file
  • log output
  • console output during testing
• If the key is leaked, rotate via platform.deepseek.com and update .env.

Tasks#

• t-053: Maxime obtains DeepSeek API key, stores in .env
• t-054: Register DeepSeek as OpenClaw provider, add aliases
• t-055: Smoke test (CLI + agent turn)
• t-056: Decide which agent(s) default to DeepSeek; update agent config + identity files
• t-057: Document setup in MEMORY.md
• t-065: Feature request — gateway per-task cost tagging. Outcome: No path exists today to attribute per-call inference cost to a task; .usage-cost-cache.json and OTel exports both lack taskId. Proposal written: feature-request-task-cost-tagging.md. Two-part fix (additive field on usage entries + new openclaw.task.id OTel attribute). Awaiting Maxime's go-ahead to file as an upstream GitHub issue.