Skip to content

COLEMAN 4.0 — Make every CI budget decision inspectable

CI budgets force a choice: which tests run now, which wait, and what evidence supports that decision? COLEMAN 4.0 makes those choices easier to execute, inspect and reproduce while keeping test-case prioritization and selection at the center of the project.

This release adds a shared, bounded application API and optional local MCP tools, backed by independent mathematical checks. They build on the recent 3.x foundations for versioned results, protected evaluation, finite coverage selection and scenario-based resource allocation.

New in 4.0

  • One workflow across library, CLI and tools. Discover approved configurations, submit bounded CI work, inspect status, cancel, retrieve results and export portable evidence. Native COLEMAN services perform the calculation and execution.
  • Independent checks before confident claims. Small exact reference cases check feasibility, objective values, bounds and ties without calling the production solver. Separate opt-in performance observations record their environment, resource use and limitations.
  • Results with a contract. Versioned handles and canonical artifacts preserve explicit units, meanings, source identity and provenance. Zero, missing evidence and unavailable results remain distinct.
  • Bounded execution with clear outcomes. Queue, work, output, retention and worker resource limits have documented behavior. Cancellation, expiry and interrupted writes cannot masquerade as completed jobs. A feasible incumbent remains explicitly different from a proved optimum.
  • Optional tools, a service-free base. The base wheel needs no MCP SDK, model, credentials or remote service. Configure the optional local stdio adapter once; supported clients can then use its tools directly.

Try a bounded CI decision

With Python 3.14 or later on Linux or macOS, install this repository's GitHub release wheel in an activated virtual environment:

python -m pip install 'https://github.com/jacksonpradolima/coleman/releases/download/v4.0.0/coleman-4.0.0-py3-none-any.whl#sha256=80736c9eb2dc8e16bd1db3c6b30e48b0a3160803b2b35ff86a99fd598541a27c'
coleman application --request '{"schema_version":1,"request_id":"first-ci-decision","spec_id":"weighted-overlap-3","parameters":{"budget":3}}'

The packaged synthetic fixture selects A+C within a 3-second modeled budget and returns the complete typed result. It demonstrates a checkable decision; it is not a measurement of production CI savings. native-ci-smoke-v1 exercises the actual TCPCI runner on two synthetic builds.

For MCP, install the same wheel with its optional extra and configure coleman-mcp in a local client:

python -m pip install 'coleman[mcp] @ https://github.com/jacksonpradolima/coleman/releases/download/v4.0.0/coleman-4.0.0-py3-none-any.whl#sha256=80736c9eb2dc8e16bd1db3c6b30e48b0a3160803b2b35ff86a99fd598541a27c'

Use this repository's assets: the bare PyPI name belongs to a different project. Release assets include signed distributions and an SBOM. The application guide includes complete library examples, tool operations, errors and limits.

Compatibility and evidence

Existing coleman run, sweeps, RunSpec workflows and native result/checkpoint interfaces remain available. Existing Parquet fields keep their meanings; old artifacts are not silently reinterpreted. New application bundles use schema version 1 and reject unknown future versions. Portable experiment provenance explicitly declares its redactions and source digest while leaving native result/specification bytes intact.

The release is validated with the full repository suite, native CI workflows, independent reference checks and noneditable installed-wheel tests outside the checkout. Optional MCP lifecycle tests use the real SDK over stdio, offline and without an LLM. Dependency and code-scanning findings are checked before publication; advisory results are time-dependent.

Worker memory and output sampling are local guards with possible transient overshoot, not a hostile-code sandbox. Hosted deployment and durable application-job recovery are outside this release. Export/import retains completed evidence; native checkpoint continuation remains a separate interface.

COLEMAN can report insufficient evidence, use a conservative fallback and preserve negative or inconclusive outcomes. Seeds support reproducibility; they do not replace independent observations. No universal algorithm superiority or production wall-clock savings are claimed.

Explore the release

Release integrity correction

The initial SBOM contained only the distribution-directory metadata. It was replaced with a verified inventory of the published wheel's distribution metadata, including COLEMAN 4.0.0 and the wheel SHA-256. Dependencies are not bundled in the wheel; their all-platform lock audit remains separate. Wheel, source archive, signatures and the version tag were unchanged. Installation examples now pin the published wheel digest.