feat!: extend AiMetrics, MetricSummary, and TrackMetricsOf

[mattrmc1]

BEGIN_COMMIT_OVERRIDE
feat!: ILdAiConfigTracker.TrackDuration now takes a double instead of a float — custom ILdAiConfigTracker implementations must update their TrackDuration override signature
feat: Extend AiMetrics with optional ToolCalls and DurationMs (duration override)
feat: Extend MetricSummary with accumulated ToolCalls and ResumptionToken
feat: TrackMetricsOf auto-tracks tool calls, honors DurationMs override, and aligns operation-vs-extractor failure semantics with the other SDKs
END_COMMIT_OVERRIDE

Summary

Extends AiMetrics, MetricSummary, and TrackMetricsOf to bring the .NET AI SDK to parity with the JS, Python, and Java SDKs for metric tracking.

AiMetrics — new optional fields

public AiMetrics(bool success, Usage? tokens = null,
    IReadOnlyList<string> toolCalls = null, double? durationMs = null)

ToolCalls carries tool keys invoked during the operation. DurationMs provides an explicit duration override — when set, TrackMetricsOf uses it instead of the stopwatch measurement.

MetricSummary — new optional fields

public record struct MetricSummary(
    double? DurationMs,
    Feedback? Feedback,
    Usage? Tokens,
    bool? Success,
    double? TimeToFirstTokenMs,
    IReadOnlyList<string> ToolCalls = null,      // NEW
    string ResumptionToken = null                 // NEW
);

ToolCalls accumulates keys from TrackToolCall invocations. ResumptionToken surfaces the tracker's resumption token on the summary.

TrackMetricsOf — tool call auto-tracking & duration override

TrackMetricsOf now extracts ToolCalls from AiMetrics and calls TrackToolCalls automatically. It also honors DurationMs when present, falling back to the stopwatch value when null.

The error-handling semantics now match the other SDKs: if the operation throws, duration + error are tracked and the exception propagates. If the metrics extractor throws, duration is tracked but error is NOT — the AI operation itself succeeded.

TrackToolCall — accumulates keys on summary

TrackToolCall now appends to an internal list (lock-guarded) so that Summary.ToolCalls reflects all tool invocations. The existing per-call $ld:ai:tool_call event emission is unchanged.

Breaking change

ILdAiConfigTracker.TrackDuration now takes a double instead of a float:

- public void TrackDuration(float durationMs);
+ public void TrackDuration(double durationMs);

ILdAiConfigTracker is a public interface. Callers that invoke TrackDuration(...) are unaffected — a float widens implicitly to double. Anyone who implements ILdAiConfigTracker (custom or test trackers) must update their override signature. This aligns the tracker duration type with MetricSummary.DurationMs/TimeToFirstTokenMs (already double?) and avoids the precision loss of the previous (float)sw.Elapsed.TotalMilliseconds cast.

Migration

Update any custom ILdAiConfigTracker implementation's TrackDuration signature:

Before:

public void TrackDuration(float durationMs) { /* ... */ }

After:

public void TrackDuration(double durationMs) { /* ... */ }

No changes are required for code that only calls TrackDuration, TrackMetricsOf, or uses the SDK-provided tracker.

How to test

dotnet test pkgs/sdk/server-ai/test/LaunchDarkly.ServerSdk.Ai.Tests.csproj --framework net8.0

9 new test cases covering backward compat, new field population, auto tool call tracking, duration override vs. stopwatch fallback, extractor failure semantics, and summary accumulation.

Known limitations

The only breaking change is the ILdAiConfigTracker.TrackDuration signature (float → double), which affects custom implementations of the interface — see Migration above. All other changes are additive: existing constructors and call sites compile and behave identically.

---

Note

Medium Risk
Public API break on ILdAiConfigTracker.TrackDuration affects custom implementations; TrackMetricsOf error/duration semantics changed in ways that alter emitted metrics for extractor failures.

Overview
Brings the .NET server AI SDK’s metric tracking in line with the JS, Python, and Java SDKs.

Breaking: ILdAiConfigTracker.TrackDuration now takes double instead of float (call sites widen implicitly; custom interface implementations must update). Duration is recorded without casting stopwatch results to float.

Additive: AiMetrics gains optional ToolCalls and DurationMs. MetricSummary adds accumulated ToolCalls (from TrackToolCall) and ResumptionToken. TrackMetricsOf auto-emits tool-call events from AiMetrics.ToolCalls, uses DurationMs when set, stops the timer before the extractor so slow extractors don’t inflate duration, and treats operation failures (duration + error) separately from extractor failures (duration only, no generation error). TrackToolCall appends keys under a lock for summary snapshots.

^{Reviewed by Cursor Bugbot for commit 0a17bde. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b2c7bc2. Configure here.}

[tanderson-ld]

Jason should be able to review before you merge.

[jsonbailey]

Take a look at the tool call list comment, other than that this seems okay.