Bilal Imamoglu’s Post

When Claude Code reads a 3,000-file codebase, it reads files. It does not know who owns them, which ones change together, which ones are dead, or why they were built the way they were. repowise fixes that. It indexes your codebase into four intelligence layers — dependency graph, git history, auto-generated documentation, and architectural decisions — and exposes them to Claude Code (and any MCP-compatible AI agent) through eight precisely designed tools. The result: Claude Code answers "why does auth work this way?" instead of "here is what auth.ts contains."

From Leaks Come Innovations: How the Claude Code Leak Inspired an Open-Source Multi-Agent Orchestration Framework Following the accidental leak of Claude Code's source code caused by a source map file mistakenly bundled into an npm package update on March 31, 2026 one of the smartest engineering moves emerged from the open-source community. A former product manager studied the exposed multi-agent orchestration architecture and rebuilt it as an independent, model-agnostic open-source framework. How It Was Built: The developer didn't copy the leaked code directly. He studied the architectural patterns specifically the orchestration layer and reimplemented them from scratch as a standalone framework. (Note: Whether this fully qualifies as "clean-room reimplementation" in the strict legal sense is debatable, since the developer had direct access to the leaked source.) The rebuilt architecture includes these core components: - The Coordinator: Breaks complex goals into executable tasks automatically. -Team System: Distributes workloads across specialized agents. - Message Bus: Enables real-time communication and data exchange between agents. -Task Scheduler: Resolves dependencies to ensure tasks execute in the correct logical order. The Architectural Edge (In-Process vs. Multi-Process): The developer, JackChen (@JackChen_x on X), named it "open-multi-agent." Unlike the claude-agent-sdk, which spawns a separate CLI process per agent (creating resource bottlenecks), this framework runs entirely in-process within a single Node.js runtime. Deployment Flexibility: Thanks to its lightweight, subprocess-free design, the framework can be deployed in virtually any modern environment serverless, Docker containers, or directly within CI/CD pipelines. 1. The leak was confirmed by Anthropic as a human packaging error, not a hack or intentional release. 2. The project is very new (launched days ago) with ~4,600 GitHub stars calling it "the strongest open-source framework" for multi-agent orchestration is premature given established alternatives like LangGraph, CrewAI, and AutoGen. 3. Several other projects also emerged from the same leak, including full Python rewrites and decentralized mirrors this wasn't the only notable response. 🔗 GitHub: https://lnkd.in/dmqXEzJT

This is wild, and surprisingly simple. A single mistake has exposed ~512,000 lines of Claude Code on npm. No hack. No reverse engineering. Just a source map accidentally shipped in a prod build. A security researcher spotted a 60MB .map file in a package version. That file essentially reconstructs the original code from the minified build. The outcome? • ~1,900 TypeScript files • Full structure, comments, and logic • Quickly mirrored across GitHub What’s inside is far from trivial: • ~40 tools, each with its own permission model • The actual system prompt used by Claude Code • A multi-agent orchestration setup (agents talking to each other) • An IDE bridge connecting VS Code & JetBrains to the CLI • Unreleased features like VOICE_MODE, KAIROS, ULTRAPLAN • Even an “Undercover Mode” meant to prevent leaks… found in the leak itself Before this gets overhyped: • This is not the model. • It’s the CLI client — the layer that connects to APIs and organizes tools. Still, the root cause is almost boring: • A bundler (likely Bun) generating source maps by default • *.map not excluded from .npmignore That’s it. And honestly, this could happen to any team. The real takeaway isn’t the leak. It’s the level of engineering behind modern AI products: • Persistent memory • Structured tool orchestration • Permission-aware execution • Multi-agent systems This isn’t just an API wrapper. It’s a full product architecture. #Claude_Code

Top 20 Claude Code commands every developer should know: GETTING STARTED 1. claude - Opens Claude Code in your terminal. Start thinking in tasks, not questions. 2. claude "prompt" - Starts a session with a task already loaded. Clean context from the start. 3. claude -c / claude -r "name" - Continues your last session or resumes a specific named one. 4. /clear - Resets context between tasks. The single most important habit to build. 5. /compact - Compresses conversation history without losing it entirely. Use when context is above 70%. SPEED TRICKS 6. Esc / Esc Esc - Stops Claude mid-action or rewinds to any previous checkpoint instantly. 7. !command - Runs shell commands like !git status without leaving Claude. Output lands in context. 8. git diff | claude -p "review" - Pipes anything into Claude for instant code review from your terminal. 9. -p flag - Runs Claude non-interactively. Perfect for scripts, cron jobs, and CI/CD pipelines. 10. /clear again - Listed twice because it is that important. Clean context, clean output. Every time. POWER USER 11. claude -w branch-name - Works in an isolated git branch. Your main codebase stays untouched. 12. --permission-mode auto - Stops asking permission for every action. Uses AI safety classifier instead. 13. --allowedTools - Scopes exactly what Claude can and cannot do for a specific task. 14. --max-budget-usd - Caps spending per session. Essential for pipelines with predictable costs. 15. --add-dir - Gives Claude visibility across multiple directories or repos at once. ADVANCED WORKFLOWS 16. CLAUDE.md - A markdown file at project root that loads automatically every session. Write once, follows forever. 17. Hooks - Auto-formats code every time Claude edits a file. Runs 100% of the time. 18. /install-github-app - Auto-reviews every PR you push. Set it once and forget. 19. TDD workflow - Write tests first, then implement. Produces 2 to 3x better code consistently. 20. Parallel sessions - Spawn multiple Claude agents in separate branches. Three features shipping simultaneously. Which of these do you actually use daily?

We rebuilt our MCP engine last month, so Healthie's Dev Assist now runs all tools in parallel instead of sequentially. The original Dev Assist explored the schema one step at a time, so every question required multiple round-trips before you got an answer. 2.0 runs all of that in a single parallel block, so developers can now build entire solutions with Dev Assist without blowing through their token budgets (which isn’t great for token usage leaderboards, but perfect for executing!) ✅ 64% lower token consumption per session (16K tokens down to ~2.6K on complex schema explorations) ✅ from 11 API calls to 5 ✅ 55% fewer round-trips ✅ ~5x faster responses ✅ Live test queries against the API: real response shapes, not just what the schema says a field accepts Full walkthrough with code here: https://lnkd.in/ekRtAzTV

The teams building on our infrastructure are building the products that clinicians and patients use every day. Every hour a developer spends figuring out the API is an hour not spent on the product. 2.0 ships with 64% lower token consumption and 55% fewer round-trips. For a team on a paid Claude or ChatGPT plans, that's a real cost. For a team trying to go live, that's real time back!

The numbers on Healthie's Dev Assist 2.0 are hard to ignore. 64% lower token consumption, 5x faster responses, 11 API calls down to 5. For developers building on our platform, this is a meaningful unlock — faster builds, lower costs, less friction. This is what investing in infrastructure looks like. 🚀

The Register called Claude Code Routines "mildly clever cron jobs." They're not wrong. But that's also not the point. Anthropic just shipped a feature that lets Claude Code run automations on a schedule, on GitHub events, or via API trigger - in the cloud. Your Mac can be completely off. You configure the routine once, push the settings, and it executes on Anthropic's infrastructure while you're asleep or in meetings. I've been using Claude Code as a core part of how I ship. The bottleneck was never the model's intelligence - it was the session model. Every automation required either active supervision or a machine staying awake, which meant async coding workflows were mostly theoretical for anyone without a permanently-on dev server. The HN thread has real skepticism. Pro users get five routines per day. There's no local execution path. Debugging a cloud-based automation that fails at 3am isn't like debugging a local script. These are legitimate concerns for production use. But the engineers dismissing this as "cron plus an API wrapper" are describing the mechanism, not the shift. The meaningful thing isn't the scheduling - it's that agent-driven code tasks can now run fully decoupled from a developer's presence and machine. That's a different category of tool than what we had last week. The rate limits will go up. The debugging story will improve. What won't change is that async agent workflows just moved from architecture blogs to shipped software. #ClaudeCode #AItools #WebDev #SoftwareDevelopment

The Claude Code source code leaked yesterday. I spent hours reading all 11 layers of architecture while it was up so you don't have to. Buried in the thousands of lines of code was a humbling realization: I’ve been using this tool completely wrong. And statistically, you probably are too. Most of us open it, type a prompt, wait for a response, and type another. Here is the reality: 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝗰𝗵𝗮𝘁 𝗮𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 𝘄𝗶𝘁𝗵 𝘁𝗲𝗿𝗺𝗶𝗻𝗮𝗹 𝗮𝗰𝗰𝗲𝘀𝘀. 𝗜𝘁 𝗶𝘀 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺. After digging through the repo, here are the 3 most critical insights that will immediately change how you engineer: 𝟭. 𝗬𝗼𝘂𝗿 𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 𝗶𝘀 𝗿𝗲-𝗿𝗲𝗮𝗱 𝗲𝘃𝗲𝗿𝘆 𝘀𝗶𝗻𝗴𝗹𝗲 𝘁𝘂𝗿𝗻 Most developers leave this blank or use 200 characters. You are allocated 40,000. Put your architecture decisions, naming conventions, and "never do this" rules here. This is the highest-leverage configuration in the codebase to make the AI understand your specific repo. 𝟮. 𝗙𝗶𝘃𝗲 𝗮𝗴𝗲𝗻𝘁𝘀 𝗰𝗼𝘀𝘁 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗮𝘀 𝗼𝗻𝗲 When Claude forks a subagent, it creates a byte-identical copy of the parent context. The API caches this. You can spin up 5 agents simultaneously, one for a security audit, one refactoring, one testing, and share the cache. Using it single-threaded is a massive waste of its capability. 𝟯. 𝗧𝗵𝗲𝗿𝗲 𝗮𝗿𝗲 𝟮𝟱+ 𝗵𝗶𝗱𝗱𝗲𝗻 𝗟𝗶𝗳𝗲𝗰𝘆𝗰𝗹𝗲 𝗛𝗼𝗼𝗸𝘀 You can intercept the pipeline at will. Imagine automatically attaching your latest test results or recent git diffs to every prompt without typing a single word. That is the power of the UserPromptSubmit hook. The developers getting 10x output aren't writing magically better prompts. They are configuring, parallelizing, and hooking into the architecture. Stop starting from scratch every session. Use --continue. Build your context. Have you set up your local CLAUDE.md file yet, or are you still relying on manual, zero-shot prompting? -- Post inspired by various X articles during yesterday's havoc.

4 takeaways from Claude Code leak this week Someone accidentally leaked Claude code's source code—512,000 lines of TypeScript due to a forgotten .map file on npm. Fortunately, TypeScript is my primary coding language, and here's what I did with it — and what you can take away right now. 𝟭. 𝗧𝗵𝗲 𝗯𝗼𝗿𝗶𝗻𝗴 𝘀𝘁𝘂𝗳𝗳 𝗴𝗼𝘁 𝘁𝗵𝗲𝗺 Even interns know not to commit sensitive code. DevOps teams have automatic systems to confirm this, and production builds include automation to validate. Unfortunately, this all got missed, leading to the accidental leak. 𝟮. 𝗦𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗲 𝗱𝗮𝘁𝗮 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 𝗼𝗻 𝘁𝗵𝗲 𝘀𝗲𝗿𝘃𝗲𝗿 In Development 101, we learn that sensitive data should stay on the server. Since Claude code is a client-side application, leaked feature flags revealed the application's roadmap. Lesson: Focus on software development best practices, even in the AI era. 𝟯. 𝗦𝘁𝗼𝗽 𝗿𝗲𝗹𝘆𝗶𝗻𝗴 𝗼𝗻 𝗼𝗻𝗲-𝘀𝗵𝗼𝘁 𝗽𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴 𝗳𝗼𝗿 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝘄𝗼𝗿𝗸 Claude code utilizes dozens of agents, skills, multi-step guardrails, and revalidation loops. The LLM acts as the brain, while the application ensures it remains focused on the task at hand. Lesson: Your application's 'thin' layer is more valuable than many LinkedIn gurus suggest. 𝟰. 𝗠𝗲𝗺𝗼𝗿𝘆 𝗮𝗻𝗱 𝗗𝗿𝗲𝗮𝗺𝗶𝗻𝗴 (𝗠𝗼𝘀𝘁 𝗳𝗮𝘀𝗰𝗶𝗻𝗮𝘁𝗶𝗻𝗴) Claude code features an advanced memory system that uses only the most relevant information: - It tracks multiple sessions - Consolidates knowledge even when not in use (Dreaming mode) This is truly inspirational. While I thought they were geniuses (which they are), seeing the magic trick behind the curtain makes it all seem within reach.