Architecture teardown of DeerFlow's source code

[NeuZhou]

I did a source-level analysis of DeerFlow's architecture. A few findings:

• The 14-layer middleware chain is well-designed for extensibility, but ordering constraints are implicit and fragile (ClarificationMiddleware must be last, learned the hard way per code comments).
• Found an orphan tool call bug in the LangGraph integration -- 93 lines that handle a crash affecting any LangGraph-based agent. If you're using LangGraph elsewhere, worth checking.
• Hash-based loop detection with kill@5 threshold is a practical solution, though it won't catch semantically similar but syntactically different loops.
• No RBAC on the coordinator/worker separation -- any worker can call any tool.

Full writeup with architecture diagrams: https://github.com/NeuZhou/awesome-ai-anatomy/tree/main/deer-flow

Happy to hear if I misread anything in the codebase.