LACAN-1

Deepseek Supervisor — AI skill for Claude Code

AI community

"You should know" for Claude Code on third-party models such as DeepSeek: a side pass reviews the work every few steps and tells the model what it notices.

How to install Deepseek Supervisor

This entry records only its repository, not the path inside it, so there is no exact command to give. Open LACAN-1/deepseek-supervisor and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Deepseek Supervisor does

"You should know" for Claude Code on third-party models such as DeepSeek: a side pass reviews the work every few steps and tells the model what it notices.

Alternatives in AI

Alibabacloud Ack MCP Server — Alibaba Cloud's ack-mcp-server unifies container operations capabilities, enabling AI assistants and third-par 118 ★
Open Sheet — The spreadsheet framework built for agents — write the model as React, export a live .xlsx 113 ★
SwiftAgent — A type-safe, declarative AI agent framework for Swift 100 ★

README

deepseek-supervisor

**A supervisor for Claude Code running a weaker model such as DeepSeek: what the model ran is held against what it says, a second, stronger model runs its claims and reviews its changes, and what does not hold goes back to the model before you get the answer.**

[中文说明](README.zh-CN.md)

Why

Claude Code ships a side agent, "You should know", that watches Claude at work and surfaces what it notices. It is hidden whenever `ANTHROPIC_BASE_URL` points away from Anthropic: on Claude Code 2.1.290 that one variable alone switches it off. So if you run Claude Code against DeepSeek or any other Anthropic-compatible endpoint, nothing watches the work.

A weaker model needs that more, not less. It goes wrong in a few ways again and again: "all tests pass", said about code changed since the run that passed, or about a run that failed; the same failing command run a third time with blind edits in between; an edit retried against text the file no longer holds; a test bent to fit the bug; a "don't touch config.py" forgotten; a crash on an input the request covers, that no test reaches.

An earlier version of this plugin reviewed the transcript every 6 steps, as "You should know" does. Its items were opinions, and the model could talk its way past them with one quoted line. Measured on DeepSeek, it made no difference. So this one works from evidence: what the session's own tool calls show, and what a command prints. It was built for DeepSeek, but nothing in it is DeepSeek-specific: it works with any Anthropic-compatible endpoint.

What it does

Three checks, each where its finding is read at once.

1. After every batch of tool calls: rules on the record

Code reads the session's own tool calls (what ran, in what order, how it ended) and holds them to the rules below. No model, no cost, any machine, and it cannot make evidence up. What it finds goes to the model **with the tool results of that batch**, before it picks its next step.

| Rule | Raised