v1.0.0

Added

• Structured agent trace model with Trace, Span, ArtifactRecord,
  StateDelta, TraceSink, WithTraceSink, and DefaultTraceSink
• Trace linkage through Case.TraceID, Case.Trace, Result.TraceID, and
  JSONL trace_id result fields
• Agent metrics: TaskCompletion, ToolArgumentAccuracy, PlanAdherence,
  and StepEfficiency
• RAG metrics: ContextRecall, AnswerCorrectness, and NoiseSensitivity
• JSON scenario definitions via LoadScenarios, DecodeScenarios, and
  BindScenarioDrivers
• Static reports through compare.ReportHTML, compare.ReportMarkdown,
  compare.ReportJSON, and goeval report
• Judge calibration and pairwise summaries through compare.Calibrate,
  compare.CalibrateFile, and goeval calibrate

Changed

• Scenario steps can declare required and forbidden artifact keys directly
• Result and trace redaction now share the same WithRedactors hooks
• Case.TraceID now seeds empty structured trace IDs, and trace sinks write a
  shared trace ID once per Runner
• Repeated scenario summary rows now keep all run trace IDs under
  scenario_summary.trace_ids
• Tool-call metrics and scenario tool contracts now use trace tool-call spans
  when available, with Turns fallback for legacy cases
• Scenario trace assembly now avoids synthesizing turn-derived tool spans when
  a step result already provides trace tool-call spans
• Trace sink deduplication now marks a trace ID as written only after a
  successful sink write, allowing transient failures to retry
• Judge calibration now aggregates duplicate judge or variant rows instead of
  replacing earlier rows
• goeval report --out now rejects unknown file extensions unless --format
  is supplied explicitly
• Markdown reports now escape table cell pipes, newlines, and HTML-sensitive
  characters

v0.9.0

v0.9.0 adds the eval operations layer with goeval.json profiles, manifest prerequisites, profile-aware test execution, compare policies, reliability summaries, policy-aware summary APIs, and stable case ID matching.

v0.8.0

v0.8.0 adds scenario repeat, built-in tier filtering, per-case/per-step timeouts, scenario state passing, grouped contracts, pattern tool assertions, richer artifact/output checks, normalizers, and scenario summary JSONL rows.

v0.6.0

Full Changelog: v0.4.0...v0.6.0

v0.3.0

Added

• RunResult.Metadata in JSONL result sinks, copied from Case.Metadata by default.
• Split token counts (PromptTokens, CompletionTokens) on judge responses, results, and JSONL sink rows.
• WithCaseFilter runner option for skipping cases by metadata or custom predicates.
• authoring-go-eval-suites agent skill and Claude /eval command for designing, running, and reviewing eval suites.
• compare package for baseline-vs-current JSONL result regression diffs.
• Minimal goeval CLI with test, compare, and version commands.
• JSON dataset loader (LoadCases, LoadNamedCases, LoadDataset) for external golden cases.
• Getting Started guide covering local judges, OpenAI, metrics, JSONL results, and benchmarks.
• Ollama judge adapter (adapters/ollama) for local HTTP API scoring.

Full changelog: https://github.com/igcodinap/go-eval/blob/v0.3.0/CHANGELOG.md