Skip to main content

A learning game that turns the code AI wrote for you into a galaxy you light up by understanding it.

Project description

Codemble — an open lapis ensō whose amber star systems light up

Codemble

Turn AI-built code into a galaxy you actually understand.

Codemble is a local-first learning game for projects built with Claude Code, Codex, and other coding agents. It maps real parser evidence into a 3D galaxy and a flat architecture map, then lights each region only after you prove you understand it.

Your project · Your key · Your machine · No invented structure

Latest release CI status Python 3.11 or newer Maps Python, JavaScript, TypeScript, Go, Java, Rust, and C# projects Apache 2.0 license

Quick start · Learning loop · Documentation · Test Codemble

Codemble at galaxy level on a first run: star systems parsed from real source, each one named by its file path and wearing its import-community colour, with the import routes drawn around the systems charted so far and an as-yet unlit Home, plus language focus buttons, a Key disclosure, a notice that two files could not be read — all under tests/ — and a prompt to study codemble.cli next

Galaxy level. Every system is one module, named and coloured from the first frame; size is lines of code, brightness is how many distinct structures call it. The routes between systems fill in as you explore. Files the parser could not read stay visible and say so.

[!IMPORTANT] Codemble is in its Phase 1 tester release. It maps Python, JavaScript, TypeScript, Go, Java, Rust, C#, and mixed projects in one parser-proven galaxy, installable straight from PyPI with an in-app project picker. The technical release is complete; unaided learner runs are the evidence still being collected. Try the ten-minute tester loop.

Quick start

Two steps. The first is once per machine; the second is how you run Codemble from then on.

Step Command
1 · Install uv — the runner that fetches Codemble on demand brew install uv
2 · Chart your project — nothing to install, nothing left behind uvx codemble
Installing uv without Homebrew
# macOS · Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Already have pipx? You can skip uv entirely: pipx install codemble, then run codemble. Plain pip install codemble works too. uv is the recommended path because uvx runs the current release without adding anything to your system Python.

Codemble opens your browser — pick your project folder there. To skip the picker, pass a path: codemble ./your-ai-built-project.

The wheel already contains the web app, so Node.js is not required. No API key is needed for the galaxy, the map, the structural summary, the impact lists, source viewer, language Lens, checks, lighting, or saved progress. Add your own Anthropic or OpenAI key only if you want grounded prose explanations:

export ANTHROPIC_API_KEY=sk-ant-...   # or OPENAI_API_KEY=sk-...

Prefer to send nothing anywhere? Point Codemble at a local Ollama instead — same grounding validation, loopback only, never automatic:

ollama pull gemma4:12b && export CODEMBLE_PROVIDER=ollama

Installation, configuration, and troubleshooting →

How the learning loop works

Step What Codemble does What you gain
1. Chart Parses your project without running its code or package scripts A deterministic map made from source evidence
2. Navigate Two layers over one graph: a 3D galaxy on scripted camera rails, and a flat map of architecture and workflow. Flying to a system charts it and keeps its routes drawn Orientation without getting lost in free flight
3. Study Shows the real source, exact line numbers, what a change here reaches, neighbors, and parser-detected language idioms Context tied to code you can inspect
4. Prove Generates and scores checks from the graph—never from the model A region lights only when understanding is demonstrated
5. Return Saves progress locally; changing one file re-dims only its module A living map that stays honest as the project changes

Charting and lighting are kept apart on purpose. Exploring a system draws its routes and counts it on the star chart; only passing that region's checks turns it amber. Having been somewhere and understanding it are different claims.

No XP. No streaks. No leaderboard. The visible reward is the useful one: more of your own code becomes a sky you understand.

What it looks like

A single star system, codemble.server.app, its 31 parser-proven structures as lit worlds in the system's own colour family, each with a procedural surface and a rim atmosphere, on labelled Layer 1 and Layer 2 call guides with the call edges between them

System level. Solid guides mark call layers established by certain calls; a dashed outer guide means Codemble found no proven call path.

The study panel for create_app, showing kind, span, 53 callers, parser-proven resolution, and a structural summary marked no model needed

Study. Everything on this panel except the narration comes from the parser — and this one has no model configured at all.

Codemble's staged loading screen mapping a large project, with five named stages — finding source files, reading each file, connecting imports and calls, building graph-only checks, placing your galaxy — a live count reading 13 of 900 files, and a cancel button

Large projects (up to roughly 1,000 source files) parse in the background with visible staged progress — a real file count while reading, named steps while resolving — so a big project never looks like a frozen tab. Cancel any time to pick another.

Two layers over one graph

Codemble draws the same parser evidence two ways, switchable in the header. The map cannot show you a relationship the galaxy does not have — both layouts are computed in the graph layer and served as data.

Layer What it is When it helps
Galaxy 3D, camera on rails through galaxy → system → study Orientation, and the shape of the whole project
Map · Architecture Modules as boxes, grouped by folder, layered by import distance from Home Seeing how the project fits together
Map · Workflow The call tree from your entrypoint, depth by depth Seeing what runs first

The Map is plain SVG, so it still works on a machine that cannot draw WebGL. Click a box and it offers both halves of a step — Read the source opens that module's real source, lens notes and relationships without leaving the layer, and Prove understanding starts its checks. Escape steps back a level, as in the Galaxy. On a compact screen the Map opens at readable 100% around Home or the selected target; Fit is an explicit whole-diagram overview. Zoom and pan survive Map refreshes and layer switches instead of snapping back after a passed check.

In Easy mode these surfaces carry plainer labels — the layer is Diagram and the tabs are How it fits together and What runs first. Same views, same evidence; only the wording follows the audience.

Open a structure, read what the parser knows first

The study panel builds itself outward from the most certain evidence: a structural summary written from parser facts alone — no key, no network, no model — then Impact, two lists of what a change here would reach and what this depends on, then grounded narration if you configured a provider, then every connection into and out of the structure with its direction, its certainty, and a file:line you can click, then the real source and the language Lens notes.

Impact rows are clickable, cited, and carry the depth at which each one was reached, traced up to three hops. A chain that passes through a relationship the parser could not prove is labelled possible for its whole length. Because it is traced from the parser alone, it is there with no key configured at all; Expert mode puts it first, and Easy mode states the same two lists in plain words below the write-up.

The explanation itself answers in at most three sentences and leads with what the structure is for. The line-by-line walkthrough is one click away behind a closed disclosure rather than the first thing you meet. Easy mode may reach for an everyday comparison; Expert mode may not.

Sections other than the narration never involve a model at all.

Easy or Expert

A header toggle changes how Codemble talks to you and how much it puts on screen: plain language, larger type, the Map by default, and a hint chip naming the nearest unlit region to Home — counted in import hops over the graph, not chosen by a model, and broken by parser-proven structure count when several sit the same distance away. The hint opens that system, then becomes an instruction rather than repeating a button the layer is already showing you. It waits until the first-run choices are done before it appears at all, and it never changes graph truth, coordinates, progress, or how a check is scored.

Codemble asks which audience you are once — the question is about you, not about the project — and each project still keeps its own mode.

You can also switch project or change Home without leaving the app.

Read the galaxy

In the galaxy In your project
A star system One source module
A planet A function or class
The Home system The selected parser-ranked entrypoint
A route or edge An import or call. A relationship the parser could not prove is drawn dashed and labeled possible, on every layer
Size Lines of code
Brightness and glow How many distinct structures call it
Colour family Import community — the eight largest groups of mutually-importing modules each get one of eight traditional Japanese hues; smaller groups stay neutral rather than share a colour
Nebula tint Language, at galaxy level
Orbit guide Solid = call layer from certain calls; dashed = no proven call path
Drifting particles A call the parser proved; a possible call stays still
Routes drawn around a system You have been there. Flying to a system charts it and its routes stay drawn
Dim → lit Not yet proven → understood

Every system is drawn, coloured and named from the first frame, whether or not you have reached it. What fills in as you explore is the web of import routes between systems — the thing that makes a large project unreadable when it is all drawn at once — and the frame the camera opens on. Hover a star and it also reports how many structures use it and how many it uses.

Understanding owns the top of the brightness range: the unlit ramp stops below the amber a lit star uses, so a busy module you have not proven can never outshine one you have — every community hue is lightness-capped beneath it, and the amber band is excluded from the community wheel entirely. Pass a region's checks and that system plays a short amber "nebula dawn" — after the light is already saved, so the animation marks a fact rather than delivering one. On the flat Map, modules with no import route from Home fold into a counted shelf (Show them draws every one), so test scaffolding never buries the connected core it cannot reach.

All seven supported languages share the same graph contract, alone or mixed in one project. Language focus changes only what you are looking at; it never changes coordinates, progress, or parser truth.

Honest by construction

Codemble is built for learners who may not yet be able to spot a confident mistake. Accuracy therefore outranks spectacle:

  • Structure, entrypoints, concepts, imports, and calls come from parsers.
  • Every explanation points to a real file:line and may name only supplied identifiers and relationships.
  • Language Lens notes appear only where a parser detected the construct.
  • Check answers come from the graph, never an LLM.
  • Approximate relationships stay visibly uncertain — a distinct colour and no drifting particles in the 3D galaxy, a genuinely dashed line in the 2D map, and the legend swatch follows whichever layer is on screen.
  • A relationship is hedged where the evidence runs out, not by default. A receiver constructed at the call site, or a method inherited from an in-project base class, resolves to one declaration and is drawn as proven; a receiver known only by a type annotation resolves but stays possible, because a subclass may override it.
  • Provider output that fails grounding validation is withheld instead of being softened into a guess — and a network failure, a rejected request and a timeout each say what actually happened rather than borrowing that wording.
  • Charting a system records that you went there. It never claims you understood it; only a passed check does that.

Read the full correctness contract. A wrong explanation is a highest-severity bug—report it without mercy.

Local-first, with an explicit AI boundary

Stays on your machine Leaves only when you ask
Project discovery and parsing The bounded Study context sent to your configured provider
Graph, structural summary, impact lists, language Lens, and checks A request triggered only when you open Study
Local server and packaged web app Nothing in the background
Progress and explanation cache in ~/.codemble/ No accounts, telemetry, or Codemble cloud
Narration too, if you choose a local Ollama Nothing at all in that case

No model at all? Codemble remains a complete parser-backed map and learning game; only the optional prose narration is unavailable.

Boundaries that keep the map truthful

  • Supported source: Python, JavaScript/JSX, TypeScript/TSX, Go, Java, Rust, C#, and any mix of them. Python is parsed with the standard library's ast; the other six use tree-sitter. Unsupported languages stay outside the graph rather than being guessed, and Codemble states how many of their files it saw.
  • Scale: above roughly 1,000 supported source files, choose a subdirectory — the in-app picker offers the busiest scopes as buttons and accepts a typed path, or pass codemble --path ./project/subdirectory.
  • Ambiguous Home: choose a parser-ranked entrypoint in the app or pass --entrypoint NODE_ID.
  • Broken source: syntax errors remain visible; Codemble maps safe partial evidence instead of crashing or inventing the missing structure.
  • Rendering: the 3D galaxy needs WebGL. If your machine cannot draw it, the Map layer still works — it is plain SVG over the same parser evidence, not a degraded guess.

Help test the release

The most valuable contribution right now is not a feature request. It is a first run on a real AI-built project:

  1. Follow the ten-minute tester guide.
  2. Light at least one system.
  3. Report confusion verbatim—never paste private source or API keys.

Open an early-tester report →

Develop

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest && ruff check .

(cd web && npm install && npm run check)
(cd docs-site && npm install && npm run check && npm run build)

The load-bearing design and architecture contracts are documented, not implied:

Roadmap

Horizon Work
Now Collect unaided first-run evidence on the current release across supported project types
Next Level-of-detail rendering and clustering for larger repositories (the Go, Rust, Java, and C# adapters have shipped)
Later Read-only share links, new quest types, and the coordinated public launch

The public roadmap separates shipped work from planned work. Milestones move only when their acceptance evidence exists.

Acknowledgements

License

Codemble is released under the Apache License 2.0.


Built for the moment after “AI made it work” and before “I know how it works.”

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

codemble-0.14.0.tar.gz (3.1 MB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

codemble-0.14.0-py3-none-any.whl (880.5 kB view details)

Uploaded Python 3

File details

Details for the file codemble-0.14.0.tar.gz.

File metadata

  • Download URL: codemble-0.14.0.tar.gz
  • Upload date:
  • Size: 3.1 MB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for codemble-0.14.0.tar.gz
Algorithm Hash digest
SHA256 7f5ba1273f338a3770d296bc749e596fb4689a09ad81a798cd8b3fe68bc7d69e
MD5 695e04d52867937bc4e45c9211417adb
BLAKE2b-256 41e982435c5ce84784d5f3ea88363551a58db8a70c5b8ecb170e6966f050e7d8

See more details on using hashes here.

Provenance

The following attestation bundles were made for codemble-0.14.0.tar.gz:

Publisher: publish-pypi.yml on udhawan97/Codemble

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file codemble-0.14.0-py3-none-any.whl.

File metadata

  • Download URL: codemble-0.14.0-py3-none-any.whl
  • Upload date:
  • Size: 880.5 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for codemble-0.14.0-py3-none-any.whl
Algorithm Hash digest
SHA256 3d827bd9fa8c107b9b649e432941e7975a431ae1be3bc81944c12dd2f00f2f4e
MD5 80610297c954158a9753b31cd6adba9d
BLAKE2b-256 926b07b93add136a6dea9cb3e16b382567b654baf029ad40ca36bcad7a76f764

See more details on using hashes here.

Provenance

The following attestation bundles were made for codemble-0.14.0-py3-none-any.whl:

Publisher: publish-pypi.yml on udhawan97/Codemble

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page