Open Multi-Agent Files

Evaluation V1 (#403–#409)

• New @open-multi-agent/core/eval and /eval/file entry points.
• Define reusable Scorer implementations and versioned EvalSet fixtures.
• Run deterministic offline evaluations with reports, stores, gates, CLI commands, and reference scorers.
• Sample and score production runs online without changing the business result.

Offline Run Viewer (#394, [#411])

• Open saved results and traces locally without a running service.
• Inspect task-level model, provider, token, and cost information through the viewer and CLI.

Per-run metadata (#396)

• Top-level run APIs accept bounded metadata that is propagated into results, traces, and checkpoint restores.

Zero-key onboarding (#414)

• create-oma-app@0.5.0 can install and run a deterministic Demo without API keys or model requests.
• Demo output clearly discloses simulated model responses and produces Markdown, JSON, and HTML reports.
• New --no-install and --no-run flags support controlled scaffolding workflows.

Examples catalog (#412)

• Added a machine-readable examples catalog, schema, and coverage validation.

Fixes (#390, [#417])

• The process backend cleans up descendant processes after the parent exits.
• v1.12.1 fixes installed oma binaries so the CLI launches correctly through npm-created symlinks. This was found during the v1.12.0 post-publish verification.

Compatibility

• Existing run modes remain compatible.
• Evaluation is disabled by default, and online evaluation does not change the business result.
• Scorer failures are recorded as scorer_error and excluded from score aggregates instead of being counted as zero.
• @open-multi-agent/otel@0.1.0 is not republished and remains compatible with core 1.12.1.
• [#397], [#400], [#401], and [#402] are behavior-preserving internal extractions, not new public APIs.

Install

:::bash
npm install @open-multi-agent/core@1.12.1
npm create oma-app@latest my-oma