CodeDojo is a deliberate-practice tool for developers in the AI era. Hunt injected bugs in real codebases. Rebuild real features from git history. Train against code that actually matters — from the command line.
New developers reach feature-shipping competence without earning the mental model of the codebase. When the AI is wrong, there's nothing to fall back on.
Experienced developers spend the day reviewing AI output. Pattern-matching skills atrophy from skimming generated diffs instead of authoring carefully.
CodeDojo is deliberate practice that reverses both — active recall for newcomers, adversarial sharpening for experts.
Both modes run against real code with a validator-gated Socratic AI coach — and work on Go, Python, JavaScript, TypeScript and Rust repositories, using each language's normal test command.
CodeDojo injects one plausible bug into a real repository — a flipped boundary, a dropped error check, a weakened equality. You inspect the code, run tests, ask for limited Socratic hints, then submit the file, line range, and diagnosis. Partial credit for proximity.
CodeDojo reverts a real feature commit from the repo's history and hands you a stripped-down feature description. Rebuild the behavior until the original tests pass — then it grades your work against the reference diff and approach quality.
Test results and mutation logs are computed facts. Reviewer diagnosis grading extracts concrete evidence deterministically; the LLM coach is bounded to a cached mechanism judgment and newcomer approach quality.
An actual recorded reviewer session, played back from CodeDojo's own JSON export. Step through the injected bug, the test run, the coach hint, the diagnosis, and the graded reveal — no install, no backend.
Two thin mode layers on a shared spine of interfaces. Reviewer and newcomer share infrastructure but keep their domain logic separate.
Starts an isolated session, runs commands, reads and writes files, reports diffs. Docker preferred, local fallback.
Returns Socratic hints and grading feedback. Output is validated — a stricter prompt is retried if a response leaks code.
SQLite persistence for sessions, events, scores, mutation logs, streaks, and cached commit-history scans.
A state machine — Created → Running → Submitted → Graded → Closed — with hint budget and streak tracking.
A go-git wrapper: clone, open-local, detect language from standard project markers. Never mutates the original path.
AST and text operators — boundary, conditional, error-drop, equality-weakening — each with a full original/mutated log.
CodeDojo runs build and test commands from repositories that may be unfamiliar or intentionally broken. Treat every practice repo as untrusted code — two driver modes share one interface.
CodeDojo tries Docker first and falls back to the local driver with a warning when Docker is unreachable. Full detail in the security model.
Prerequisites: Go 1.23+ and Git. Docker is optional but recommended for stronger sandboxing. Build the CLI once, then start a kata.
# clone and buildgit clone \ https://github.com/dhruv1794/CodeDojocd CodeDojomake build # first-run wizard:# pick coach backend + defaults./bin/codedojo init # optional: Docker sandbox imagesmake images
./bin/codedojo review \ --repo testdata/sample-go-repo \ --difficulty 2 --budget 3 # inside the REPLtestscat calculator/calculator.gohint nudgesubmit calculator.go:13 \ "boundary comparison"
./bin/codedojo learn \ --repo testdata/sample-go-repo \ --difficulty 2 --budget 3 # inside the REPLtestswrite calculator/calculator.godiffsubmit
codedojo serve launches the complete experience — repo browser, file editor, test runner, hints, submission — at localhost:8080. There is no hosted version, by design: CodeDojo clones repositories and runs their test suites on your machine. Keeping it local means your code never leaves your computer, and there's no account or cloud service in the loop.
# launch the full app on localhost./bin/codedojo serve \ --repo testdata/sample-go-repo \ --port 8080 # then open http://localhost:8080# repo browser, recent repos,# breadcrumbs — no account, no cloud ./bin/codedojo status # past sessions./bin/codedojo stats # aggregate stats
No install needed to see it in action — the recorded session above plays a full kata in your browser.
The MVP is feature-complete across both modes and five languages. Architecture and design notes live in the repository.
Full detail in the architecture notes.
MIT licensed. No account, no hosted service — your code stays local.