Add Spans() & Thread() to TaskResult

[Alex Z (CLowbrow)]

Adds Spans() and Thread() methods to TaskResult, allowing scorers to access trace data from
within evals.

Changes

• TaskResult.Spans(ctx, ...SpanQueryOpt) returns typed []eval.Span from the experiment
  trace
• TaskResult.Thread(ctx) returns preprocessed thread entries via the project_default
  preprocessor
• eval.Span struct replaces raw map[string]any with typed fields
• api/objects package adds generic object fetch API used internally for span retrieval
• InvokeGlobal cleanup removes dead /v1/function/invoke fallback
• All new tests use VCR recorded cassettes

New API

  // Typed span with structured fields
  type Span struct {
        ID             string         `json:"id"`
        SpanID         string         `json:"span_id"`
        RootSpanID     string         `json:"root_span_id"`
        SpanParents    []string       `json:"span_parents"`
        SpanAttributes map[string]any `json:"span_attributes"`
        Input          any            `json:"input"`
        Output         any            `json:"output"`
        Metadata       map[string]any `json:"metadata"`
  }

  // Functional option for filtering spans
  type SpanQueryOpt func(*spansQuery)
  func WithSpanTypes(types ...string) SpanQueryOpt

  // Methods on TaskResult (available in scorers)
  func (r TaskResult[I, R]) Spans(ctx context.Context, opts ...SpanQueryOpt) ([]Span, error)
  func (r TaskResult[I, R]) Thread(ctx context.Context) ([]map[string]any, error)

Example

  scorer := eval.NewScorer("trace_aware", func(ctx context.Context, tr
  eval.TaskResult[string, string]) (eval.Scores, error) {
        // Get all spans from the trace
        allSpans, err := tr.Spans(ctx)
        if err != nil {
                return nil, err
        }

        // Filter to only custom spans
        customSpans, err := tr.Spans(ctx, eval.WithSpanTypes("custom"))
        if err != nil {
                return nil, err
        }

        // Get preprocessed conversation thread
        thread, err := tr.Thread(ctx)
        if err != nil {
                return nil, err
        }

        score := 0.0
        if len(allSpans) > 0 {
                score = 1.0
        }

        return eval.Scores{{
                Name:  "trace_aware",
                Score: score,
                Metadata: map[string]any{
                        "span_count":        len(allSpans),
                        "custom_span_count": len(customSpans),
                        "thread_count":      len(thread),
                },
        }}, nil
  })

Full working example: examples/internal/trace-scorer/main.go