Spent some time diffing Anthropic’s public Claude Code docs against a leaked source, and the most interesting part isn’t a single hidden feature, it’s how much bigger the product feels in code than in docs.
A few things jumped out:
The command surface is wider than the docs suggest. There are real command paths for things like a hidden /advisor, /ultrareview, /session, /web-setup, and deprecated-but-still-present /output-style plumbing. Some are gated, some internal, but they’re not random stubs; they point to realproduct branches.
The agent system is also deeper than the public story. Docs cover Explore, Plan, general-purpose, etc. But the source also has a built-in verification agent whose whole job is basically: don’t trust the implementation, try to break it. That’s a very revealing design choice. It suggests Anthropic is treating “verification” as a first-class agent role, not just an instruction style.
Then there’s the forked subagent model. In code, there’s a feature-gated path where omitting subagent_type can implicitly fork a worker that inherits full context, runs in the background, and can even operate in an isolated worktree. That’s a much more aggressive orchestration model than the normal “spawn a helper agent” framing most people have in their heads.
Plan mode is another surprise. It’s not just conversational planning. It’s backed by real persistent plan files, with resume/fork recovery logic and session-specific plan slugs. Same with memory: the source has a much richer taxonomy than I expected, with typed memories (user, feedback, project, reference), private vs team scope, and rules like “convert relative dates to absolute dates before saving.”
And maybe the biggest signal: there’s a hidden server-side advisor tool wired into the app. The model can apparently call out to a stronger reviewer model that sees the whole conversation/tool history and critiques the approach. That is a very different picture from “just a CLI wrapper around one model.”
My takeaway: Claude Code’s public docs describe the product you can use today, but the source reveals the product Anthropic is actually building toward: a much more agentic system with verification, delegation, persistent planning, structured memory, remote execution, and internal reviewer models all treated as core primitives.
It feels less like “a coding CLI” and more like the shell of a full multi-agent operating environment.
The future is bright!