DeepSeek 对话接口封装

Wraps DeepSeek's chat/completions endpoint in the same subprocess-invocation
shape as the other ruflo-*-harness skills (see ruflo-metaharness's
harness-score for the reference). No library import on ruflo's boot path.

When to use

• You want a cheap, fast completion from a non-reasoning model.
• The task fits in a single request/response — no multi-turn context.
• You want the raw content back as JSON so downstream tooling can consume it.
• For reasoning-heavy tasks (proofs, multi-step planning, hard debugging),

use deepseek-reason instead — same plugin, different model.

Data handling

The prompt and any --system text are sent over the network to api.deepseek.com, a third party. Do not put secrets, credentials, personal data or proprietary code you are not allowed to share into them. The API key is read from DEEPSEEK_API_KEY in your shell only; never paste it into a prompt, a file or a command line.

Algorithm

Implementation: scripts/chat.mjs.

1. Read DEEPSEEK_API_KEY from env. If missing, emit

{ status: 'degraded', reason: 'DEEPSEEK_API_KEY is not set', ... }
and exit 0 (ADR-150-style graceful degradation).

1. POST to https://api.deepseek.com/v1/chat/completions with

{ model, messages, temperature?, max_tokens? }. 60s hard timeout.

1. Extract choices[0].message.content and usage counters.
2. --format table prints only the content (for piping); --format json

(default) returns the full envelope.

1. --alert-on-error exits 1 on any degraded/error status (CI-friendly).

Example

node plugins/ruflo-deepseek-harness/scripts/chat.mjs \
  --prompt "In one sentence: what is HNSW?" \
  --temperature 0.2

Sample output:

{
  "status": "ok",
  "model": "deepseek-chat",
  "content": "HNSW is a graph-based approximate nearest neighbor …",
  "finishReason": "stop",
  "usage": { "promptTokens": 12, "completionTokens": 34, "totalTokens": 46 }
}