chore(main): release LaunchDarkly.ServerSdk.Ai 0.10.0

[github-actions]

🤖 I have created a release beep boop

0.10.0 (2026-06-05)

⚠ BREAKING CHANGES

• Move Role from LaunchDarkly.Sdk.Server.Ai.DataModel into LdAiConfigTypes
• Rename LdAiConfig.ModelConfiguration to LdAiConfigTypes.ModelConfig and LdAiConfig.ModelProvider to LdAiConfigTypes.ProviderConfig
• Move shared types nested in LdAiConfig into static LdAiConfigTypes — Message, ModelConfiguration, and ModelProvider
• Enforce at-most-once tracking — each metric type (duration, tokens, feedback, success/error, time-to-first-token) records once per tracker; duplicates are dropped with a warning
• Add ResumptionToken on ILdAiConfigTracker and ILdAiClient.CreateTracker(resumptionToken, context) for cross-process tracker reconstruction with the original runId
• Add per-execution runId on all AI track event payloads for billing isolation
• Change CompletionConfig to return LdAiCompletionConfig instead of ILdAiConfigTracker — obtain a tracker via config.CreateTracker()
• Default LdAiCompletionConfigDefault.Enabled to true per AISDK spec (was false on the LdAiConfig builder in 0.9.x)
• Remove ILdAiConfigTracker.Config property — read config fields from the LdAiCompletionConfig the caller already holds
• Make LdAiConfigTracker SDK-constructed only — the public constructor is removed; obtain trackers via config.CreateTracker() or ILdAiClient.CreateTracker(resumptionToken, context)
• Convert LdAiCompletionConfig and LdAiCompletionConfigDefault from records to classes — equality is reference-based instead of structural
• Remove LaunchDarkly.Sdk.Server.Ai.DataModel namespace — delete unused AiConfig, Meta, Model, Provider, and Message JSON DTO classes
• Split unified LdAiConfig into LdAiCompletionConfig (SDK output) and LdAiCompletionConfigDefault (user input) — Builder, New(), and Disabled move to the Default type; introduce abstract LdAiConfig and LdAiConfigDefault base types for future agent/judge modes

Features

• Add MetricSummary property on ILdAiConfigTracker summarizing recorded metrics (44ff485)
• Add per-execution runId on all AI track event payloads for billing isolation (44ff485)
• Add ResumptionToken on ILdAiConfigTracker and ILdAiClient.CreateTracker(resumptionToken, context) for cross-process tracker reconstruction with the original runId (44ff485)
• Change CompletionConfig to return LdAiCompletionConfig instead of ILdAiConfigTracker — obtain a tracker via config.CreateTracker() (92f799f)
• Convert LdAiCompletionConfig and LdAiCompletionConfigDefault from records to classes — equality is reference-based instead of structural (92f799f)
• Default LdAiCompletionConfigDefault.Enabled to true per AISDK spec (was false on the LdAiConfig builder in 0.9.x) (92f799f)
• Enforce at-most-once tracking — each metric type (duration, tokens, feedback, success/error, time-to-first-token) records once per tracker; duplicates are dropped with a warning (44ff485)
• Make LdAiConfigTracker SDK-constructed only — the public constructor is removed; obtain trackers via config.CreateTracker() or ILdAiClient.CreateTracker(resumptionToken, context) (92f799f)
• Mode-mismatch detection — log a warning and return the caller's default when the flag's _ldMeta.mode does not match the requested mode (per sdk-specs#229) (92f799f)
• Move Role from LaunchDarkly.Sdk.Server.Ai.DataModel into LdAiConfigTypes (ac7fd06)
• Move shared types nested in LdAiConfig into static LdAiConfigTypes — Message, ModelConfiguration, and ModelProvider (ac7fd06)
• Non-object variation handling — log an error and return the caller's default when the variation result is not an object (92f799f)
• Per-message interpolation fallback — a malformed Mustache template keeps raw content for that message rather than discarding the entire config (92f799f)
• Remove ILdAiConfigTracker.Config property — read config fields from the LdAiCompletionConfig the caller already holds (92f799f)
• Remove LaunchDarkly.Sdk.Server.Ai.DataModel namespace — delete unused AiConfig, Meta, Model, Provider, and Message JSON DTO classes (92f799f)
• Rename LdAiConfig.ModelConfiguration to LdAiConfigTypes.ModelConfig and LdAiConfig.ModelProvider to LdAiConfigTypes.ProviderConfig (ac7fd06)
• Split unified LdAiConfig into LdAiCompletionConfig (SDK output) and LdAiCompletionConfigDefault (user input) — Builder, New(), and Disabled move to the Default type; introduce abstract LdAiConfig and LdAiConfigDefault base types for future agent/judge modes (92f799f)
• Tolerant LdValue parsing — missing or wrong-typed fields degrade to typed defaults instead of discarding the whole config (92f799f)

---

This PR was generated with Release Please. See documentation.

---

Note

High Risk
The tagged release ships extensive breaking public API and telemetry/billing semantics; this PR is low-risk mechanically but merging it publishes high consumer upgrade risk.

Overview
Release Please PR that publishes LaunchDarkly.ServerSdk.Ai 0.10.0 by bumping 0.9.4 → 0.10.0 in .release-please-manifest.json, LaunchDarkly.ServerSdk.Ai.csproj, SdkInfo.Version, and PROVENANCE.md, and adding the 0.10.0 section to CHANGELOG.md.

The changelog documents a breaking AI SDK realignment with the AISDK spec: split LdAiConfig into LdAiCompletionConfig (evaluation output) vs LdAiCompletionConfigDefault (caller input); CompletionConfig returns config, not a tracker (config.CreateTracker() / ILdAiClient.CreateTracker(resumptionToken, context)); tracker construction is SDK-only with runId, resumption tokens, at-most-once metrics, and MetricSummary; shared types move to LdAiConfigTypes and DataModel is removed; completion defaults default Enabled to true; config/records become classes (reference equality). Non-breaking behavior called out includes tolerant LdValue parsing, mode-mismatch and non-object variation fallbacks, and per-message Mustache fallback.

^{Reviewed by Cursor Bugbot for commit 82deb10. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

🤖 Created releases:

• LaunchDarkly.ServerSdk.Ai-v0.10.0

🌻