Story

Where Wingbox comes from

Wingbox started from one research question: can a coding agent keep a correct map of a codebase in its head while it explores? The research said: not reliably. So we stopped asking the model to build the map.

Sources

The papers that started it

Two arXiv papers and one blog post. The second paper is the one this project grew from: it gave us the question, the term belief map and the reason to build one without a model.

The origin

Theory of Code Space: Do Code Agents Understand Software Architecture?

Grigory Sapunov, arXiv:2603.00601, 28 February 2026

The same framework carried over to software. Agents explore a codebase under a budget of files they may open and write down their architectural beliefs as a structured belief map: module dependencies, invariants, design intent. The gap between active and passive exploration turns out to depend on the model, keeping the belief map in context helps some models and not others, and one model's beliefs collapsed while its smaller sibling stayed stable.

Theory of Code Space toolkit on GitHub

Idea

The idea we took from it

Theory of Code Space measures how well a model builds and keeps a belief map of a codebase. Reading it, we asked the opposite question: what if the agent did not have to build that map at all?

A parser can read every file of a repository, not a budgeted handful, and it does not forget what it read. It can write the same map the paper asks agents to produce, without a model, in seconds, and write it again after every change, so an old belief cannot outlive the code it described.

In March 2026 that became a short specification, citing the paper: walk the syntax trees of a Python and TypeScript codebase and write the belief map as a file the agent queries before it reads any source.

Timeline

From a script to Wingbox

  1. 4 Feb 2026

    Theory of Space on arXiv

    Active exploration, unstable maps and belief inertia, measured in physical space.

  2. 28 Feb 2026

    Theory of Code Space on arXiv

    The framework applied to code agents; the belief map as the agent's written-down picture of the architecture.

  3. 15 Mar 2026

    First specification

    A script that builds the belief map from the syntax tree, as JSON, for Python and TypeScript.

  4. 16 Mar 2026

    Edges from a language server, fewer tokens

    Call and usage edges from a language server; the JSON replaced by a compact text form, because every token of the map is a token the agent pays for.

  5. 19 Mar 2026

    S-expressions

    One fact per line, searchable with grep and queried by a script: the format Wingbox still uses for agents. Used every day as a Claude Code and Codex skill named codespaces.

  6. 20 Jun 2026

    Open source as codespaces

    The skill and its scripts published under the MIT license.

  7. Jul to Nov 2026

    Correct on real repositories

    Edges checked against the code of public repositories, TypeScript re-exports and path aliases, then more languages: ten by November.

  8. 15 Nov 2026

    Renamed Wingbox

    A new repository and a package on PyPI. The belief map became the code graph; the idea did not change.

  9. 16 Nov 2026

    The public benchmark

    An agent with and without the map on questions drawn at random in five public repositories.

Scope

What the papers say and what we say

The papers do not evaluate Wingbox, and their results are not ours. Theory of Code Space measures how well models build beliefs about a codebase; our benchmark measures what an agent gets right when it is handed a map built by a parser. The link between the two is our reading.

Thanks to Grigory Sapunov for publishing the benchmark, its code and its results in the open.