Python SDK for building unattended coding harnesses that turn a task into a reviewable branch or PR.
This repo is building primitives, not a chat app and not a general workflow engine. The core model is:
Task -> Blueprint -> Branch
The live source of truth is:
- AGENTS.md
examples/docs/api/- GitHub maintenance guide
- Real LLM run guide
- Roadmap & reliability plan
If one of those disagrees with older design notes, the current contract wins.
This repo has two different kinds of workflows:
- Research examples: research-faithful reference workflows that pressure-test whether CodeMinions can express patterns seen in company blogs and papers.
- Validation examples: small real-world runs that prove the SDK works with real models, real repos, and real environments.
These are not the same thing as the primitive SDK surface. Research is input. Examples are pressure tests. The SDK API is the output.
Task: structured input contractTool: typed executable capabilityNode: deterministic, agentic, judge, parallel, and loop workflow stepsBlueprint: ordered workflow definitionEnvironment:DockerEnvfor production,GitWorktreeEnvandLocalEnvfor local useMinion: runner that executes a task against a blueprint
- deterministic + agent hybrid, not pure free-form agenting
DockerEnvis the production path- human review happens at the end
- bounded retries and explicit escalation
- examples are pressure tests for API design, not the API itself
The SDK now has a v1. From here, the main job is not adding more surface area by default. The main job is:
- preserving a small primitive layer
- tightening contracts and failure semantics
- validating with real runs
- only promoting abstractions that recur across multiple workflows
- implementation in
src/codeminions/exists and is being hardened after the first v1 pass AGENTS.mddefines what counts as a primitive, helper, built-in pattern, or example-only behavior- public examples are backed by executable contract tests in
tests/test_examples_contracts.py - MCP support now lives under
src/codeminions/tools/mcp/as a package-level client subsystem - research and design references live under
design/
Run examples/validation/09_real_repo_config_resolution.py with your real model keys to validate CodeMinions:
- Populate
/Users/tamil/Developers/workflows/.envwith the Anthropic/OpenAI credentials, base URL overrides, and model aliases you need. - Start Docker locally so
DockerEnvcan launch containers (python:3.12is the default image). - Execute
uv run python examples/validation/09_real_repo_config_resolution.pyfrom the repo root. - Inspect the generated branch/diff and trace output to confirm
done()was hit and acceptance criteria satisfied.
The roadmap lays out how that workflow maps to our reliability goals.