Mission#

Verify DeepSeek API works end-to-end through OpenClaw, both as a one-shot inference and as a real agent turn that calls tools.

Steps#

Test 1 — One-shot CLI inference#

openclaw infer model run --model deepseek-chat --prompt "Reply with exactly: deepseek-api-ok"
openclaw infer model run --model deepseek-reasoner --prompt "Reply with exactly: reasoner-ok"

Both must return their expected token in under 10 seconds. If either fails with auth error, key is wrong; rotate or recheck .env.

Test 2 — Real agent turn with a multi-step tool chain#

Run the same Christmas-joke smoke we ran for local DeepSeek:

openclaw agent --agent chris --model deepseek-chat \
  --message "Send Maxime a one-line Christmas joke on WhatsApp via the relay protocol. Tag: [notify|chris]. Session key: agent:main:whatsapp:direct:+6580860085. Confirm delivery." \
  --timeout 120 --json

Watch the JSON output. The agent must:

• Actually CALL the sessions_send tool (not describe what it would do)
• Receive a successful response from the relay
• Output a short summary mentioning the tool call ran

Test 3 — Robustness#

Repeat Test 2 three times. If even one fails to deliver, log the failure mode in this task body before deciding whether to promote DeepSeek as Chris's default.

Acceptance criteria#

• Test 1 passes for both models
• Test 2 delivers a joke to Maxime's WhatsApp without the "described instead of called" failure
• 3-of-3 success rate on Test 3
• Failure modes (if any) documented in task body