RubricLoop v0.1.0a2

AgentEval is now RubricLoop, an open-source evaluation project from 人在回路.

• Updated repository, bilingual documentation, branding, and live demo.
• New rubricloop command and RUBRICLOOP_* model settings.
• The agenteval command, AGENTEVAL_* settings, Python imports, and report schema remain compatible. New environment-variable names take precedence.
• The package distribution is now rubricloop. Remove the previous agenteval-traces distribution before installing this package into an existing environment.

Validation: 8 offline tests, package build, fresh installed-wheel demo, legacy CLI alias, and environment-variable fallback/precedence checks. This is a naming and packaging update; automatic evaluator repair remains a roadmap item.

Live demo · Migration notes

AgentEval v0.1.0a1

First standalone public release of AgentEval.

• Evaluate text conversations or cleaned trace trees with trusted code rules and configurable model judges.
• Preserve full trace IDs and distinguish PASS, FAIL, N/A and evaluation errors.
• Open self-contained HTML reports or consume structured JSON.
• Run the bundled offline demo without an API key.

Validation: 8 offline tests, source/wheel builds, and an installed-wheel demo in a separate environment. Model integration tests use a mock provider; this release makes no claim about judge accuracy. Automatic evaluator repair and regression acceptance remain roadmap items.

Python 3.10+. MIT. Install the attached wheel or follow the repository quick start. Package distribution name: agenteval-traces; command: agenteval.