Plinth

The current generation of AI coding tools, including the one I use every day, has reached a local maximum. They have optimized the chat-with-your-code paradigm to its limit. They are still stateless and gullible. They rebuild their understanding of a codebase from scratch on every prompt. They optimize for satisfying the user rather than for delivering correct code.

Plinth is the AI IDE that earns the right to write code. The architecture is plan-first, glass-box, stateful. The model cannot claim a task is done. It cannot modify files outside the approved plan. It cannot rewrite the test that asserts the bug is fixed. It must submit its work to an adversarial verification engine that compiles the code, runs the tests, grounds syntax against live documentation, and only then accepts the change.

Plinth is not a tool for beginners learning to code. It is a deterministic, verifiable, stateful engineering collaborator for senior engineers and large teams working on large codebases.

The architecture is five non-negotiable pieces.

A three-layer codebase brain that combines static intelligence (real-time AST parsing, dependency graphs, symbol-level impact analysis), persistent memory (cross-session retention of design decisions and failed approaches), and temporal knowledge (git history mining that predicts failure vectors before code is written).

An execution sandbox and adversarial verification engine that the model cannot bypass. A plan-first workflow that produces a deterministic graph of intended changes for the developer to approve before any code is written. A glass-box decision tree UI that replaces the chat window with an inspectable thought process, where the developer can correct an assumption mid-flight without starting over.

An enterprise air-gapped mode that runs entirely on self-hosted models for security-conscious environments. Each piece exists to remove a different way the current generation lies.