Open Discovery 0.5.0 — Program Evolution

Open Discovery can now improve a supplied baseline program against an external evaluator while preserving correctness and reproducible evidence.

Highlights:

• New Evolve Program skill and evaluator-driven optimization workflow.
• Immutable baseline and evaluator locking before candidate search.
• Isolated candidate lineage, external measurement, and clean winner reconstruction.
• Comparable run records with separate blind and open-book tracks.
• Local benchmarks for exact image histograms and economic dispatch.
• A pinned frontier benchmark based on NetworkX bounded-cycle search, with independent correctness checks, two adversarial graph families, and an ordinary-case regression guard.
• The broader autonomous initiative architecture, scientific reviewer, research communicator, and GitHub-ready artifact workflow developed since 0.4.1.

The harness remains local-first and prompt-driven. It does not publish, spend, or use outside compute without explicit authority.

Open Discovery 0.4.1 — Background Feature Testing

Open Discovery 0.4.1 makes Feature Tester explicitly non-blocking.

Highlights:

• every feature test runs in a separate background Codex task
• Luna/xhigh is the default tester runtime, with a recorded fallback when unavailable
• development and research conversations remain available during testing
• the Lab CEO directly supervises tester status, resources, reports, and cleanup
• tester tasks remain separate from research-direction explorer pools
• test-owned processes and artifacts must be reconciled before completion

Open Discovery 0.4.0 — Agent Research Workflows

Open Discovery 0.4.0 adds three experimental research workflows: the Hierarchy of Agents Research Model, Paper Implementer, and Feature Tester.

Highlights:

• experimental CEO -> direction leader -> explorer architecture for continuous parallel discovery
• three explorers per direction as a flexible starting default, not a permanent quota
• Luna/xhigh as the preferred runtime with practical fallback to another capable agent
• faithful paper-to-code reproduction before adaptation or optimization
• isolated feature testing with evidence-backed reports and owned-artifact cleanup
• reusable program, implementation, and feature-test templates
• issue-first contribution guidance

Local research programs, test artifacts, and user projects remain excluded from the released repository.

Open Discovery 0.3.0 — Literature Review

Open Discovery 0.3.0 adds a repo-scoped Literature Review skill that creates a project, runs one autonomous research worker, preserves source-level evidence, and delivers a verified PDF report.

Highlights:

• clear literature-review-only activation
• Luna preference with a portable fallback
• autonomous scope and search decisions without routine user interruptions
• bounded evidence checkpoints and honest access limitations
• project-local PDF tooling with installation disclosure
• topic-aware Starberry first responses

Known limitation: the long end-to-end PDF evaluation was intentionally skipped for this release.

Open Discovery 0.2.1 — Public harness cleanup

Open Discovery 0.2.1 is a public-repository cleanup release.

• Keeps the downloadable harness focused on the 45 user-facing files.
• Removes product strategy, dogfood planning, and maintainer release notes from the public tree.
• Removes the superseded general protocol so the project, literature-review, and experiment templates are the only canonical workflows.
• Expands ignore rules for local environments, logs, model files, and array artifacts.

No local research projects or experiment artifacts are included or changed.

Open Discovery 0.2.0 — Starberry

Starberry is the user-facing Open Discovery research partner. Version 0.2.0 adds isolated local projects, AI/ML, mathematics, and biology research modes, existing-project attachment, Full Auto routing, and tested project-state rules. Four independent Luna agents validated onboarding, field initialization, duplicate detection, correction handling, and project isolation before release.