Skip to content

MCP 0.4.0 unified workflows

MCP 0.4.0 completes Phase 4 with its unified workflows: core.compare_runs, a fail-closed core.release_check, core.incident_summary and core.export_report. MCP judges nothing itself; it composes Lens comparisons, the Evals release gate and Chaos run reports. Malformed or failed-call evidence always blocks readiness, and so does missing evidence of a required kind; all three kinds are required by default and required_evidence narrows that.

What MCP reads

  • evaluation_report must be a native Evals report (schema_version: "1.0"). It is validated against the bundled Evals 0.9.0 schema, then judged by evaluate_gate with GateConfig.model_validate(thresholds).
  • The gate's reasons and observed values and the effective thresholds are returned under the evaluation gate.

Thresholds

  • The Evals default applies when thresholds is omitted: min_pass_rate is 1.0 and unavailable judgments fail (max_judge_unavailable_scores: 0).
  • Every GateConfig field is accepted, including tag/metric floors, judge cost and agreement, and live-target timeout/error limits.
  • GateConfig ignores unknown keys, so MCP rejects them before calling the gate.

Compatibility

No Evals API change, dependency bump or release. Invalid reports block the release as an invalid gate instead of passing.

See the MCP unified workflows guide.