Tools like `ncdu` and `gdu` already present a compressed, human-friendly view of the filesystem tree. A human just looks at the TUI and presses ↑ ↓ → ← to drill down, decide what matters, and ignore the rest.
Most current AI workflows instead export giant JSON dumps, then build an entire parsing + chunking + filtering layer on top of them. This burns context, adds complexity, and doesn’t match the real usage model.
So the question is: Why not let an LLM simply *operate the TUI* the same way a human does?
Just read the visible screen text, choose an action, and repeat. The TUI is already an optimized abstraction layer.
This fits the same class of interaction as modern CLI-AI agents like *Claude Code* or *OpenCode CLI*, but instead of orchestrating commands, the model would literally navigate an interactive interface (ncdu/gdu) step-by-step.
Questions:
* Does this interaction model make sense, or is there some fundamental flaw I'm missing? * Is anyone aware of existing OSS (beyond general agents like Claude Code / OpenCode CLI) that specifically lets an LLM “drive” ncdu/gdu or similar TUIs directly?
0 comments