dojoml 0.0.3

AI-powered autonomous ML research framework — agent runs experiments against frozen evaluation contracts on your existing pipelines.

🥋 Dojo — An AI-powered autonomous ML research framework.

Run controlled, reproducible ML experiments on your existing pipelines and build a memory of what actually works.

---

---

What is Dojo?

You define a domain — a research area pointing at your data with a fixed evaluation contract. An AI agent runs experiments inside that contract: writing training code, calling frozen load_data and evaluate tools, logging metrics, and recording findings as durable knowledge atoms.

Domain (you define)
  ├── Task            — the contract: load_data + evaluate (frozen, AI-generated at setup)
  ├── Workspace       — your repo / pipeline (local path or git url)
  └── Experiments     — agent-created, many per domain
        └── Knowledge atoms — linked across experiments, accumulating over time

The agent owns the training code. The framework owns evaluation. That separation is what makes the metrics trustworthy run-over-run, and what makes it safe to leave the agent unsupervised.

Inspired by Karpathy's autoresearch — prepare.py is frozen, train.py is fair game, program.md is what the human iterates on. Dojo generalises that pattern to any well-defined ML problem class.

---

Current Status

⚠️ Proof of Concept — under active development. Open source. Single-tenant, local-first, by design.

• Agent: Claude Agent SDK (uses your local claude CLI auth — no API key needed for runs)
• Compute: Local only (in-process / subprocess) — your data stays on your machine
• Storage: Local JSON files in .dojo/
• Tracking: File-based or MLflow (sits on top of an MLflow you already run)
• Tasks supported: RegressionTask (more types to come once regression is solid)

---

Prerequisites

• Python 3.13+
• uv
• just
• The claude CLI logged in (Claude Code) — Dojo shells out to it; no ANTHROPIC_API_KEY needed
• Node.js 18+ (only if you want the web UI)

just dev                     # install backend + frontend deps

Getting Started — California Housing in 4 commands

The CLI is a peer of the HTTP API, not a thin wrapper around it. The whole happy path runs in-process — no server needed.

mkdir housing && cd housing

# 1. Scaffold the domain (creates .dojo/, the regression Task, and PROGRAM.md)
dojo init --name housing --task-type regression --non-interactive

# 2. Describe the dataset, target, and what success looks like
$EDITOR PROGRAM.md

# 3. AI generates load_data + evaluate from PROGRAM.md, verifies them against
#    the regression contract, and freezes the task. Re-run after edits.
dojo task setup

# 4. Run the agent — events stream live to your terminal
dojo run --max-turns 30

If the AI keeps generating the wrong adapters (verification failures on real-world pipelines, e.g. unusual pandas multi-indexes, custom dataset APIs, or wrapping an existing evaluator), use Opus 4.7 for tool generation instead of the default Sonnet:

DOJO_AGENT__TOOL_GENERATION_MODEL=claude-opus-4-7 dojo task setup

Opus is slower (~30–60s vs 15–30s) but noticeably better at translating a messy PROGRAM.md into correct load_data / evaluate modules. Set it permanently in .dojo/config.yaml under agent.tool_generation_model if you want it as the default.

A reasonable starter PROGRAM.md for California housing:

## Goal
Predict California median house value (regression). Minimise RMSE on a 20% held-out test split.

## Dataset
Use `sklearn.datasets.fetch_california_housing(return_X_y=True)`.
Features and target both come back as numpy arrays — no column names needed.
https://scikit-learn.org/stable/modules/generated/sklearn.datasets.fetch_california_housing.html

## Target
Median house value (in $100,000s) for census blocks in California.

## Success
Beat a linear baseline. Try at least one tree-based model. Avoid overfitting.

What happens under the hood:

• dojo init writes .dojo/config.yaml, creates the domain + regression task with expected_metrics = [rmse, r2, mae], scaffolds PROGRAM.md, and sets current_domain_id.
• dojo task setup reads PROGRAM.md, asks the AI to generate load_data + evaluate, runs each tool in a sandbox against its ToolContract, and freezes the task. Verification failures tell you which tool failed and why — fix PROGRAM.md (or the tool code) and re-run.
• dojo run starts the agent in-process. The agent writes training code; load_data and evaluate stay frozen. The metric dict from evaluate is the only source of truth — complete_experiment rejects metric keys outside the contract, so the agent can't smuggle in custom numbers.

Useful neighbours:

dojo task show               # current task status, tools, frozen?
dojo runs ls                 # recent runs
dojo runs show               # last run's events + cost
dojo program show            # print the live PROGRAM.md
dojo domain use <name>       # switch active domain

Stopping a run

dojo run blocks the foreground until the agent finishes. To stop it early:

• Ctrl-C in the running terminal — the canonical path. The orchestrator is interrupted, the framework asks the backend to summarise any durable findings as knowledge atoms (a small one-shot LLM call), then prints a final cost line. A second Ctrl-C aborts the cleanup immediately.
• dojo stop [run_id] from another terminal — marks the run STOPPED on disk. This does not halt an in-process foreground run (the orchestrator lives inside the other terminal's Python process); use it to recover records left RUNNING after a hard kill, or to stop server-mode runs.

Reviewing what happened

dojo experiments ls          # rank experiments by the primary metric (best first)
dojo experiments best        # show the single best experiment so far
dojo experiments show <id>   # full detail: hypothesis, metrics, code path, errors
dojo runs show               # last run's events + total cost

dojo experiments ls orders by the task's primary_metric and direction (e.g. rmse minimised), so the leader sits on top regardless of run order. The agent's training code is preserved per-experiment in the workspace as __dojo_train_<experiment_id>.py — cat it to reproduce a run by hand.

Running the server (optional)

If you want the web UI or HTTP API:

just run-stub                # stub agent (no LLM, deterministic)
just run-claude              # Claude agent (uses your local CLI auth)

Backend → http://localhost:8000 · Frontend → http://localhost:5173. The server reads the same .dojo/ your CLI commands write to, so a CLI-started run is visible in the UI and vice versa.

---

Config

Create .dojo/config.yaml in your project root:

agent:
  backend: stub        # "stub" (no LLM) or "claude"
tracking:
  backend: file        # "file" or "mlflow"

Or use environment variables:

DOJO_AGENT__BACKEND=claude
DOJO_TRACKING__BACKEND=mlflow

---

Tests

just test       # all tests
just lint       # ruff check
just format     # auto-fix lint + format

---

Project Structure

src/dojo/
  core/         # Domain, Task, Experiment, KnowledgeAtom, Workspace, state machine
  agents/       # AgentBackend ABC + Claude / Stub backends, orchestrator
  api/          # FastAPI app + routers (/domains, /experiments, /knowledge, /agent)
  cli/          # Typer CLI: init, run, task, runs, program, domain, config, start
  tools/        # Agent tools (experiments, knowledge, tracking) + AI tool generation
  runtime/      # LabEnvironment (DI), ExperimentService, ToolVerifier, program loader
  sandbox/      # LocalSandbox (subprocess); runs generated tools + agent code
  compute/      # Compute backends (LocalCompute today)
  storage/      # Local JSON adapters (domain, experiment, knowledge, run)
  tracking/     # FileTracker, MlflowTracker, NoopTracker
  config/       # pydantic-settings + YAML config
frontend/       # React 19 + Vite 7 + shadcn/ui (currently de-prioritized)
tests/          # unit, integration, e2e

---

Key API Endpoints

Method	Path	Description
POST	/domains	Create a research domain
POST	/domains/{id}/task	Attach a Task (regression today)
POST	/domains/{id}/tools/generate	AI-generate load_data / evaluate from PROGRAM.md, verify against contract
POST	/domains/{id}/task/freeze	Freeze the task — gated on every required tool's verification
POST	/domains/{id}/workspace/setup	One-time workspace prep (venv + deps)
POST	/agent/run	Start an agent run on a domain (requires a frozen task)
GET	/agent/runs/{id}/events	Live SSE event stream
GET	/experiments?domain_id=	List experiments
GET	/knowledge?domain_id=	List knowledge atoms
GET	/health	Health check

For architecture, conventions, and "how do I add X" recipes, see CLAUDE.md. For vision and the typed-Task design, see docs/MASTER_PLAN.md. For the ordered delivery punch-list, see docs/NEXT_STEPS.md. For the PyPI release process, see docs/RELEASING.md.