Smoke Testing NCP Chat

Overview

Use the reusable smoke command instead of ad-hoc curl or UI clicking when a fast real-reply check is needed.

This smoke command:

• Sends one real chat message to a running NextClaw service
• Forces the request through the specified session-type and model
• Reads the returned SSE event stream
• Prints pass/fail, assistant text, terminal event, and error details
• Exits non-zero when the route does not produce a real assistant reply

When to Use

• A quick check is needed to confirm that one concrete chat route can return a real assistant reply.
• A specific session-type + model pair needs to be validated without opening the UI.
• A fast smoke is preferred over ad-hoc request assembly.

Command

pnpm smoke:ncp-chat -- --session-type native --model dashscope/qwen3-coder-next --port 18792

Quick Reference

pnpm smoke:ncp-chat -- --session-type codex --model dashscope/qwen3-coder-next --port 18792
pnpm smoke:ncp-chat -- --session-type claude --model minimax/MiniMax-M2.5 --port 18794
pnpm smoke:ncp-chat -- --session-type native --model openai/gpt-5.3-codex --base-url http://127.0.0.1:18792
pnpm smoke:ncp-chat -- --session-type codex --model dashscope/qwen3-coder-next --prompt "Reply exactly OK" --json

Success Criteria

• Exit code is 0
• Output shows Result: PASS
• Assistant Text is non-empty
• No run.error or message.failed

When --json is used, the key checks are:

• ok: true
• assistantText is non-empty
• terminalEvent is usually run.finished

Common Mistakes

• Testing the wrong port: pnpm dev start usually serves API on 18792 in this repo.
• Forgetting --session-type: the smoke should target the exact runtime under investigation.
• Treating one runtime as proof for another runtime: native, codex, and claude should be checked explicitly.