VS Codeで動く推理ゲームエンジンを作った——AIが書いた事件が「ちゃんと解けるか」まで自動チェックするA developer built a VS Code extension that statically analyzes evidence…

Generative AI has made it trivial to produce mystery-story text on demand, but a scenario that reads well is not the same as a puzzle that can actually be solved. A developer has released a VS Code extension that tackles exactly this gap, statically analyzing the evidence dependency graphs inside AI-generated mystery cases to automatically flag circular dependencies and unsolvable dead ends. The project is notable because it treats logical solvability as something a machine can verify, rather than something a human playtester must discover by getting stuck.

The problem the tool addresses is familiar to anyone who has asked a large language model to "write an interesting mystery case." The output often looks plausible and is pleasant enough to read, yet it falls apart the moment you try to play it as a game. The developer describes the canonical failure mode: obtaining evidence A requires evidence B, but obtaining evidence B requires evidence A. It is the logical equivalent of a locked box whose only key is sealed inside the box itself. On the page, these dependencies are easy to overlook because the prose flows smoothly; in play, they produce a point where progress simply halts and the player can go no further.

To catch these situations, the extension models a case as a dependency graph. Each piece of evidence, clue, or unlockable location becomes a node, and the prerequisites needed to reach it become directed edges. Once a scenario is expressed this way, well-established graph algorithms can be applied. Detecting a circular dependency, for example, reduces to finding a cycle in a directed graph, while identifying an unreachable "dead end" amounts to checking whether a node can be reached from the player's starting state. Because this analysis is static, it inspects the structure of the scenario without requiring anyone to actually play through every branching path, which is where the approach draws its leverage.

This is a practical application of static analysis, a technique more commonly associated with software engineering. Compilers and linters routinely examine source code without running it to surface problems such as unreachable statements, undeclared variables, or cyclic module imports. The extension appears to borrow that same mindset and point it at narrative content, treating a mystery scenario as a kind of program whose "bug" is an unwinnable state. Framing AI-generated content as something that can be checked against formal constraints is an increasingly discussed idea, since generative models optimize for text that reads convincingly rather than text that satisfies hidden structural rules.

Choosing VS Code as the host is a pragmatic decision. The editor has become a default workspace for many developers and, increasingly, for people authoring structured text and data files. Its extension marketplace is large and its extension API is mature, giving a solo developer a realistic path to distribute a niche tool and integrate feedback directly into an authoring loop. In principle, a writer could draft or paste an AI-produced case into the editor and receive immediate warnings about broken dependency chains, much as a programmer sees squiggly underlines beneath a syntax error. This tight edit-and-verify cycle is one of the main reasons editor extensions are an attractive delivery format for validation tools.

The broader context is the growing effort to make generative-AI output verifiable rather than merely fluent. Techniques such as constrained decoding, schema validation for structured outputs, and automated test harnesses all share the goal of ensuring that generated artifacts meet requirements a human might otherwise have to check by hand. A mystery game is a compact, well-defined domain for this idea, because "the case is completable" is a property that can be stated precisely and tested mechanically. For game designers experimenting with AI-assisted content, that kind of guarantee is valuable, since manual playtesting for logical soundness is slow and easy to get wrong.

It is worth keeping the scope realistic. Verifying that a case is structurally solvable is a different and narrower claim than verifying that it is fair, satisfying, or genuinely clever to play, and the developer's stated focus is on logical consistency rather than narrative quality. Whether the approach generalizes cleanly to more complex scenarios with conditional or optional clues likely depends on how faithfully those relationships can be encoded as a graph. Even so, as a personal project it demonstrates a concrete, reproducible method for holding AI-generated content to a checkable standard, and it points toward a wider pattern of pairing generative tools with automated verification.