Skip to main content

​
0.6.0 (2026-08-18)

  • Adds the experiment-workflow module for typed Agent Workflow invocation with journaled step cost/token evidence adapted to AgentInvoker.
  • Makes Agent Journal capture a first-class per-item lifecycle owned by AgentExperiment, using Agent Journal 1.7.0 and preserving variant, item, model, session, tool-use, and derived cost data.
  • Migrates runtime evaluation to Agent Judge 0.14.0 and stores normalized, Agent Experiment-owned judgments and verdicts. Existing 0.5 / Judge 0.13 results load automatically; re-saving writes the normalized format with the documented lossy boundary for obsolete metadata.
  • Updates to Agent Client 0.26.0, Agent Workflow 0.10.0, and Claude Code SDK 1.4.0.
  • Gives standalone consumers Jackson 2.21.6 and Jackson 3.1.6 safety floors without requiring a BOM; public 25-, 26-, and 57-JAR module closures have zero findings against the validated scan snapshot.
  • Publishes signed binary, source, and Javadoc artifacts plus one aggregate CycloneDX 1.6 JSON SBOM.
  • Requires Java 21. The credential-free reactor passed 544 tests; paid live Claude execution was not part of the release gate.

​
0.5.0 (2026-06-06)

  • Jackson 2.21.2, member pins (sdk 1.3.0, capture 1.4.0, judge 0.12.0), release 0.5.0
  • Collect status: 2026-06-03
  • Brief: abstain-aware consensus in judge aggregation

​
0.4.0 (2026-05-29)

  • Update dependencies to released versions for 0.4.0 release
  • Relax git dirty check to warn instead of throw for non-critical paths
  • Bump dependencies to next SNAPSHOT versions

​
0.3.0 (2026-05-28)

  • Update dependencies to released versions for 0.3.0 release
  • Fix RunLogManager forcing logback on all experiment-core consumers
  • Bump dependencies to next SNAPSHOT versions

​
0.2.0 (2026-05-19)

  • Update dependencies to released versions for 0.2.0 release
  • Bump agent-judge dependency from 0.10.0-SNAPSHOT to 0.10.0
  • Step 3.4: Final consolidation — all stages complete
  • Step 3.3: Jackson polymorphic serialization for both ExecutionDetail subtypes
  • Step 3.2: Add JudgeExperiment and JudgeExperimentResult
  • Step 3.1: Add JudgeScorer, supporting types, and JudgeExecutionDetail
  • Step 3.0: Stage 3 entry — verify ExecutionDetail supports JudgeExecutionDetail
  • Step 2.3: Stage 2 consolidation — verify re-eval→compare pipeline
  • Step 2.2: Add ReEvaluator for post-hoc re-scoring without re-invocation
  • Step 2.1: Add ReEvaluationContextFactory and AgentReEvaluationContextFactory
  • Update ROADMAP.md checkboxes for Step 2.0
  • Step 2.0: Stage 2 entry — verify prerequisites, document scope adjustments
  • Migrate dependencies from org.springaicommunity to io.github.markpollack
  • Step 1.3: Stage 1 consolidation — verify boundaries, compact learnings
  • Step 1.2: Rename ExperimentRunner to AgentExperiment
  • Ignore plans scratch files (META-ROADMAP.md, status.md)
  • Step 1.1: Decouple ItemResult from InvocationResult via ExecutionDetail
  • Step 1.0: Design review — map types to shared vs agent-specific boundaries
  • Ignore entire plans/ directory in .gitignore

​
0.1.0 (2026-03-30)

  • Initial release.