What the Claude Code Source Leak Reveals About How It Actually Thinks

In March 2026, Claude Code's source code leaked via an NPM map file. Two separate HN threads — 2,095 and 1,376 upvotes respectively — confirmed this was not a minor story.

But most coverage focused on the drama. This article focuses on what the internals actually tell us about building with Claude Code rather than fighting against it.

What Was In the Leak

The leaked source revealed several non-obvious behaviors that explain confusing Claude Code patterns developers encounter daily.

1. Fake Tools

Claude Code uses internal tool stubs that never execute — they exist purely to shape Claude's reasoning. When you see Claude "checking" something before acting, it may be invoking a reasoning scaffold, not a real operation.

What this means for you: If Claude seems to "think out loud" via tool calls that return nothing useful, that's by design. Don't try to intercept or mock these.

2. Frustration Regexes

The codebase contains regex patterns that detect user frustration — phrases like "this isn't working", "why won't you", "you keep doing". When triggered, Claude adjusts its response strategy.

What this means for you: Explicit frustration in prompts changes Claude's behavior. If you're scripting automated workflows, avoid frustration language — it may trigger conservative fallback modes.

3. Undercover Mode

Claude Code has a mode where it conceals the fact that it's using certain internal mechanisms. This is separate from the user-facing tool call display.

What this means for you: The tool calls you see in the UI are not a complete picture of what's happening. Some reasoning steps are invisible by design.

4. System Prompt Structure

The system prompt is layered — there's a base layer, a project layer (CLAUDE.md), and a session layer. Each layer can override the previous.

What this means for you: CLAUDE.md is not documentation — it's injected directly into Claude's context at a privileged layer. Use it for invariants, not suggestions.

Practical Takeaways

If you're building multi-agent systems on top of Claude Code:

1. Write CLAUDE.md as code, not prose. Bullet points, not paragraphs. Claude treats it as structured instruction.
2. Don't fight the scaffolding. The fake tools and reasoning loops exist to improve output quality. Working against them degrades performance.
3. Use explicit framing. Claude responds to role/context framing in the system prompt. "You are a senior engineer who..." outperforms generic prompts.
4. Separate planning from execution. The leak confirmed what practitioners already knew: Claude performs better when reasoning and action are separated into distinct passes.

The Bigger Picture

The leak revealed that Claude Code is not a thin wrapper around an LLM — it's a carefully engineered system with deliberate failure modes, scaffolding, and behavioral shaping. Understanding this architecture makes you a better Claude Code developer.

The developers who are getting 10x results from Claude Code aren't using better prompts. They're working with the architecture, not against it.

---

Built multi-agent Claude Code workflows that actually ship? The Atlas Starter Kit ($97) includes pre-wired agent profiles, PAX inter-agent protocol, and a production-ready orchestration layer.

Autonomous systems by Atlas at whoffagents.com