返回资讯中心
外部精选
软件工程
#开发者实时流

Shark-skills 发布:为 Claude Code 提供多智能体编排与逐步调试

来源标题:Show HN: Shark-skills Multi-agent orchestration and step-debugging for Claude

开发者 tusharsrivastava 发布 Shark-skills,为 Claude Code 提供多智能体编排和逐步调试能力。其主会话不写代码而充当操作者,负责向子智能体分派任务并审核交付,各并行通道基于同一契约开发,每次合并都经过审计。项目通过插件市场命令安装,需依赖 git 和 Node.js。

Hacker News 最新razieldecarte2026/11/17 · 今天21 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Running Claude Code like an engineering team: one operator, parallel lanes, audited merges.

Install • Why shark • Reference • Contributing

My Claude Code setup for building real features with a team of agents.

Using a single Claude Code session for a small fix works fine. Using it for a real feature usually turns into a mess of conflicting edits, silently dropped rules during merges, half-baked bug patches, and a giant commit at the end.

Throwing more agents at it doesn't fix that on its own. Parallel agents with no shared plan just make code messier, faster.

shark is what I ended up with after building a production system this way.

The main session never writes code. It becomes the operator: it briefs the subagents, checks what they hand back, and talks to you. Lanes build in parallel against one shared contract. Every merge gets audited, and every bug gets a root cause before it gets a fix.

Every rule in here exists because breaking it cost me hours. Steal them. Change them. Make them stricter for your project.

/plugin marketplace add tusharsrivastava/shark-skills /plugin install shark@shark-skills

You'll also need git, and Node.js with npx. The debugger MCP server runs through it.

git clone https://github.com/tusharsrivastava/shark-skills claude --plugin-dir ./shark-skills

Turn it on for your whole team Add this to .claude/settings.json in your repo. Everyone who opens the repo gets the plugin.

{ "extraKnownMarketplaces": { "shark-skills": { "source": { "source": "github", "repo": "tusharsrivastava/shark-skills" } } }, "enabledPlugins": { "shark@shark-skills": true }, "outputStyle": "shark:asd-ste100" }

To update, run /plugin marketplace update shark-skills and start a new session.

You only need to configure it once per project. In your repo, run:

The skills never hardcode a command. They run a key from your Commands table, so the same kit works on Python, JS/TS, Go, or anything with a test command.

Key Command Limit or note
gate:fast pnpm lint && pnpm tsc --noEmit && vitest run At most 5 min.
test <paths> vitest run <paths>

When a skill says "run gate:fast", this is the line it runs.

After /shark:init, your project file tells Claude to load /shark:operator at the start of each session. You don't need to type it every time, and your session becomes a full-fledged engineering team. Already use your own orchestration loop? Keep it in charge, and shark stays on call.

Copy AGENTS.md to your repo root and fill in the sections at the top. Already have a CLAUDE.md? Merge the sections into it. Claude Code only reads AGENTS.md when there's no CLAUDE.md, or when the claude-md-and-agents-md setting is on.

Why shark Exists

Before building shark, I built baseline, but it became too bulky for newer and improved coding models. shark skills are lean, and they scale in parallel without impacting the code quality I yearn for.

shark turns your harness into a lean, mean coding machine.

"Adding manpower to a late software project makes it later."

The Problem: You run three agents in parallel. It looks three times faster. Then you merge. Two agents changed the name of the same function. A third agent made its own version of an event. The merge takes a third of your afternoon.

The Fix: Write a contract before you split the work. One agent writes the round contract in 5 to 10 minutes. The contract gives each shared name one owner file. It lists each writer and each reader of each event. It sets the boundaries of each lane. Each lane reads the contract first. When a lane finds a shared item, it stops and reports. It does not decide alone.

"If you don't see the use of it, I certainly won't let you clear it away."

The Problem: Rules hide in small places: a guard clause, a branch in a helper, a check from an old incident. A rewrite or a merge removes one of them. All the tests still pass. You find the problem in production.

The Fix: Write your invariants one time. Then check each diff against them. Do not remove a rule until you know why it is there.

David J. Agans, Debugging: The 9 Indispensable Rules

The Problem: You ask an agent to fix a bug. The agent reads the code and makes a guess. It changes the code at the symptom. Two days later, the bug occurs again with a new stack trace.

The Fix: Find the cause first. Fix it second. Two different agents do these two tasks.

shark:code-debugger runs your program in a real step-through debugger (Python, JS/TS, Go, Rust, Java, .NET, C/C++). It records the real values. It gives you three items: the root cause, a failing regression test, and a fix spec. Then shark:code-writer changes the production code.

/shark:debug-rca contains the method. You can also use it yourself.

"Make the change easy (warning: this may be hard), then make the easy change."

The Problem: An agent works for a long time. It changes the structure of the code and adds three behaviours. All of it goes into one commit of 4,000 lines. Nobody can review it. You cannot revert one part of it. git bisect cannot find the bad change.

The Fix: Change the structure first, in its own commit. Then change the behaviour, in a different commit.

/shark:commit-blocks splits the working tree into conventional commits. Each commit has one reason. A refactor and a behaviour change never share a commit. A script checks that each commit passes the gate. Thus git bisect works on your history.

"How does a project get to be a year late? ... One day at a time."

The Problem: The agent runs the slow end-to-end flow. It finds one failure and fixes it. Then it runs the flow again and finds the next failure. Each cycle takes 20 minutes. The estimate grows by hours, and nobody sees it.

The Fix: Find all the known defects first, with one static scan. Then fix them in rounds.

The Working rules section of AGENTS.md lists all the rules. Each rule names the failure that it prevents. An example: pytest with a folder and a file in that folder collects only the file.

AI made writing code cheap, but the engineering is still expensive. With shark, you can build production-grade systems with confidence: systems that are well written, well documented, and easy to debug when production is on fire. That's not vibe coding. That's real engineering. It is my decade of experience building and maintaining production-scale systems, distilled into this setup.

Session

/shark:init: Set up a project once. Brings your AGENTS.md or CLAUDE.md in line with the shark template, so every session starts with the operator, or with your own orchestrator if you keep it. Sets the shark:asd-ste100 output style for the project.

/shark:operator: Run the session. Sets the mode, writes the briefs, routes the work to agents, and checks what comes back. Loads by itself after /shark:init.

/shark:handoff: Record and prune the handoff file, so the next session picks up where this one stopped.

Build and merge

Skills

deny-secrets hook: Refuses Read, Edit, Write, Grep and Bash calls on the secret files that your ## Secrets section lists (.env, *.pem and similar by default). Example files (.env.example and similar) always stay readable, so agents still learn the setting names.

mcp-debugger: The MCP server behind shark:code-debugger. Real breakpoints, stepping and variable reads for Python, JS/TS, Go, Rust, Java, .NET and C/C++. It writes its log to .claude/mcp-debugger.log, so add that to your .gitignore.

shark:asd-ste100: An output style in Simplified Technical English. Short sentences, outcome first, observed and inferred claims labelled. Your agents' reports stop rambling.

AGENTS.md: The project template, with the working rules. /shark:init applies it for you.

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md first. The short version: no project values, no stack lock-in, and commands named by their AGENTS.md key. Test your change with claude --plugin-dir <kit folder> in a project that has a filled AGENTS.md.

Found a security problem? Follow SECURITY.md, and don't open a public issue.

MIT. The AGENTS.md template and the logo in assets/ are CC0, so you can copy them with no notice at all.

ASD-STE100 and Simplified Technical English are trademarks of ASD, Brussels. shark isn't affiliated with ASD. The debugger is @debugmcp/mcp-debugger (MIT). It runs through npx and isn't bundled.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →