Building a 3D Node.js Dependency Universe from Scratch with Qwen3.8-Max

When you run npm install, a project's node_modules directory can balloon into a black‑box containing hundreds of packages, making it hard to tell which are direct dependencies, which are transitive, which consume the most space, or how many versions are installed.

To explore this, the author used Qoder Desktop Quest in Agent mode with the official Qwen3.8-Max model. The goal was to start from an empty directory, let the AI create a full Node.js project, resolve dependencies, scan disk usage, render a 3D dependency map with Three.js, run tests, and validate the outcome against a real open‑source project called Skills Hub .

First version: visualizing the dependency universe

The AI split the task into steps—dependency resolution, disk scanning, size calculation, command execution, Three.js front‑end, testing, and verification. While running, a local service displayed a live 3D graph where node colors distinguished root, direct, indirect, and multi‑version packages, and node size reflected disk footprint. Searching by name and filtering by type were supported.

Running the analysis on Skills Hub produced the following numbers:

Total dependencies: 430

Direct dependencies: 36

Indirect dependencies: 394

Multi‑version packages: 9

Overall node_modules size: 184 MB

Clicking a node, for example lucide-react (direct dependency, version 0.562.0, 34.5 MB), focused the 3D view on that package and displayed its metadata.

The first round finished with all 37 tests passing, TypeScript type checking succeeding, and the project building correctly.

Second round: chasing a missing 10 MiB

Local inspection of the node_modules folder reported about 183.7 MiB, while the AI‑generated page showed only 174 MB. The author prompted the model with a concise query asking it to investigate the discrepancy without giving any hints.

The model reproduced the 10 MiB gap, examined the directory tree, and identified the cause: the .bin/esbuild symlink. The original scanner marked the real file as “already visited” and counted the symlink itself as 0 bytes, so the ~10 MiB binary was omitted.

After adjusting the scanning logic to treat symlinks correctly—counting the real file once and excluding the link itself—the AI reran the analysis. The updated results were:

Tests increased to 38/38, all passing.

TypeScript checks and build succeeded.

Independent disk measurement: 192,620,118 bytes (≈ 183.7 MiB).

API now returns the same 192,620,118 bytes.

Page updated to display 184 MB .

Skills Hub source files remained unchanged.

Overall assessment of Qoder Quest

The author concludes that Qoder Quest excels at handling long, well‑defined tasks that require multiple stages—dependency parsing, visualization, testing, and verification—without frequent human interruptions. It automatically decomposes the work, executes code, runs commands, starts services, and presents intermediate results.

Key strengths observed:

Comprehensive task breakdown covering lock‑file parsing, graph generation, disk scanning, front‑end rendering, testing, and real‑world validation.

Continuous execution without needing repeated prompts.

Proactive verification, such as detecting missing files and correcting the scanner.

Deliverables that go beyond a demo shell; the numbers, versions, and paths are derived from the actual installation.

Limitations include the need for human oversight on data reconciliation and permission boundaries, as the AI still ran npm ci inside the sample directory, creating a node_modules folder that was not supposed to be modified. The current implementation focuses on npm lockfiles and does not yet support pnpm or Yarn layouts, and very large dependency graphs may cause node overlap in the 3D view.

Overall, the experiment demonstrates that AI‑driven programming tools can reliably carry out end‑to‑end development workflows, provided humans define meaningful tasks and verify critical metrics.