Mythosia.AI.Abstractions - Release Notes

v4.2.0

Added

  • AIModels.Anthropic.ClaudeSonnet5_5 identifies claude-sonnet-5-5. The additive ClaudeThinkingMode enum exposes Auto, Adaptive and BetweenTools; existing reasoning/display, request-feature, capability and Run contracts are reused. Mythosia.AI 8.2.0 supplies model validation, explicit up-front thinking controls and signed-history preservation. See Sonnet 5.5 configuration and migration.
  • AIModels.OpenAI.Gpt6_1Sol identifies the published gpt-6.1-sol model. Existing Gpt6Reasoning, Run, request-feature, speed and capability contracts are reused. Mythosia.AI 8.2.0 supplies provider validation and execution.

Compatibility

  • Sonnet 5.5 defaults to adaptive/high. Common None maps to high-effort between_tools; it suppresses up-front thinking while retaining possible tool progress. Minimal is unsupported, and native BetweenTools accepts only Auto/Low/Medium/High. These are model-specific semantics; existing model identifiers and enum numeric values remain unchanged.
  • Existing required interface members, constructors, enum values, model identifiers and service defaults remain unchanged.
  • GPT-6.1 Sol requires Low, Medium, High, XHigh or Max reasoning; Auto selects Medium. None and Minimal are unsupported for this model. GPT-6 Sol and Luna retain their existing None behavior. See model selection and migration.

v4.1.0

This additive contracts release pairs with Mythosia.AI 8.1.0. Existing required interface members, constructors and enum numeric values remain unchanged.

Added

  • AIModels.DeepSeek.V4Pro identifies the supported text-only deepseek-v4-pro. The Mythosia.AI 8.1.0 core supplies model capabilities and optional Responses execution; Flash remains the default and retired constants keep their original values and warning-only obsolete annotations.

  • AIModels.xAI.Grok4_7 identifies grok-4.7 and reuses existing GrokReasoning, request-feature, speed, capability and Run contracts. The Mythosia.AI 8.1.0 core supplies reasoning validation, request wiring and processing observations; no required interface member or existing enum numeric value changes. See Grok 4.7.

  • AIModels.OpenAI.Gpt6Sol and Gpt6Luna identify gpt-6-sol and gpt-6-luna. Additive Gpt6Reasoning.None retains earlier enum numeric values; it is valid for Sol/Luna and rejected for Astra. Existing Run, reasoning, speed and request-feature contracts are reused without new required interface members. The Mythosia.AI 8.1.0 build supplies validation and execution. See model selection and settings.

  • InferenceSpeed defines ProviderDefault, Standard and Fast without changing model or reasoning effort. AIRequestFeatures.Speed and WithSpeed request extensions reuse optional request-feature contracts. Immutable AIProcessingInfo and AIRunResult.Processing preserve per-attempt requested/applied modes; unknown provider reporting stays nullable. AIModelCapabilities.StandardSpeed, .FastSpeed and .GetSpeedSupport(...) describe local tri-state support independently from live account access. Optional IAIProcessingInfoService and GetLastProcessing() expose diagnostics through IAIService references without adding required members. The Mythosia.AI 8.1.0 implementation supplies execution and observations.

  • AIModels.Anthropic.ClaudeOpus5_5 identifies claude-opus-5-5. Existing ClaudeReasoningEffort, ClaudeThinkingDisplay, request-feature and Run contracts are reused; no required interface member or existing model default changes. The Mythosia.AI 8.1.0 implementation supplies model-specific validation and preserved-thinking behavior. See usage and migration.

Fixed

  • Recognize xAI's explicit [input_too_large] context-window error, retaining both token counts in ContextLengthExceededException and streaming metadata. Unrelated input-size errors, quota errors and server failures still cannot trigger conversation compaction. Verified against a real rejected Grok request.

v4.0.0

This coordinated major release changes public contracts. See the v8 migration guide before upgrading the package family.

Added

  • AIModelCapabilities, ImageModelCapabilities and CapabilitySupport add immutable provider-neutral support snapshots in Mythosia.AI.Models.Capabilities. Supported, Unsupported and Unknown remain distinct. Read-only typed choice lists, nullable model identity and known limits describe available controls without asserting enabled options, account access or all valid combinations. StructuredOutput includes common prompt/repair fallback, not guaranteed native constrained decoding. No required IAIService members are added; query implementations live in the core service and request builder. See capability inspection.

  • Rich Run result (major migration): AIRun.Result changes from Task<string> to Task<AIRunResult>. The immutable snapshot exposes Text, defensively copied Usage and Citations, Provider, nullable RequestedModel and actual Model, RoundCount, FinishReason, and RawFinishReason. AIFinishReason defines Unknown, Stop, MaxTokens, ToolCalls, ContentFilter, and Other. Custom AIRun implementations must update their override and construct the new result; consumers must rebuild. Ordinary completion and StructuredStreamRun<T>.Result keep their existing result types. See migration. RequestedModel is nullable: it captures the explicit model sent, including Perplexity model overrides, and remains null when presets, profiles, or server routing send no single explicit model. It is independent from the actual final response Model.

  • Completion cancellation contract: the string and Message IAIService.GetCompletionAsync signatures append optional CancellationToken cancellationToken = default. Existing source callers retain their profile/context positions; custom implementations must update both signatures and propagate the token, and consumers must rebuild. This is an intentional major API change, not an optional capability with a wait-only fallback. The token cancels client execution without guaranteeing provider inference/billing cancellation. See migration.

  • FunctionDefinition.HandlerWithCancellation adds Func<Dictionary<string, object>, CancellationToken, Task<string>> without removing Handler; setting either replaces the same underlying handler. FunctionCallResult.IsCancelled identifies cancellation alongside IsError = true, including result snapshots. Core execution and the RAG/MCP adapters propagate these semantics without adding required IAIService members.

  • ImageQuality, ImageBackground, ImageOutputFormat, ImageResolution, ImageAspectRatio, and ImageSizeKind provide named image choices. Immutable ImageSize factories distinguish automatic sizing, exact pixels, and resolution/ratio presets.

  • AIModels.OpenAI.GptImage2_5Sunburst, GptImage2_5Flare, GptImage2_5Sunburst_260908, and GptImage2_5Flare_260908 identify the GPT Image 2.5 aliases and September 8, 2026 snapshots. Select them through the existing ImageGenerationRequest.Model / ImageEditRequest.Model; no new required interface members or image-model default changes are introduced. Provider-specific validation is implemented in Mythosia.AI. See the image guide.

  • Perplexity Agent API configuration, hosted-tool, search, and response contracts for the replacement integration. Common messages, runs, citations, and function batches are reused without adding required IAIService members.

  • AIModels.xAI.GrokImagineImage2_0 identifies grok-imagine-image-2.0 for the existing optional IImageGenerationService. Image options and provider support are described below.

  • AIModels.xAI.Grok4_6 identifies grok-4.6. GrokReasoning.XHigh is appended without changing earlier enum numeric values. Provider validation and common request-level reasoning for this model are implemented by Mythosia.AI v8.0.0; unsupported levels are rejected explicitly.

  • AIModels.Google.Gemini3_7Flash and Gemini3_8Flash identify the generally available gemini-3.7-flash and gemini-3.8-flash models. They reuse the existing GeminiThinkingLevel and common ReasoningLevel contracts; model-specific validation is provided by Mythosia.AI v8.0.0.

  • AIModels.DeepSeek.Flash identifies deepseek-flash (V4.1 Flash). DeepSeekReasoning defines native Auto, Low, High, and Max effort values; provider validation, thinking activation, user image input, reasoning events, and tool rounds are implemented by Mythosia.AI v8.0.0 through existing shared contracts.

  • AIModels.Anthropic.ClaudeFable5_1 and ClaudeMythos5_1 expose the new Claude model IDs. Mythos 5.1 requires limited-access approval.

  • ClaudeThinkingDisplay.Updates represents readable progress updates without reasoning summaries.

  • ClaudeThinkingPrefixMismatchBehavior.Error and .DropBlock describe the provider's invalid-prefix handling.

  • ClaudeInputTransformation carries the provider's transformation type, request-local path, reason, and optional response/model identity.

Changed

  • Snapshots preserve the exact standard ReadOnlyCollection<T> / ReadOnlyDictionary<TKey, TValue> wrapper type inside typed containers, including supported backing collections, cycles and shared references. Standard non-generic Hashtable and SortedList copies retain their key comparers. No JSON dependency or public signature change is added.

  • Request/message snapshots retain multidimensional and nonzero-bound array shape, shared references and cycles, and the comparers of standard Dictionary<,>, SortedDictionary<,> and SortedList<,> containers. default(JsonElement) / Undefined remains valid snapshot data; document-owned JSON values remain detached. Core usage aggregation reports Int32 overflow instead of returning wrapped counts.

  • Message and function-argument snapshots detach System.Text.Json.JsonElement and JsonNode values from caller-owned documents and mutable nodes without adding a JSON package dependency. Unknown custom reference types retain their existing immutability responsibility.

  • TokenUsage.TotalTokens documents preservation of an explicitly reported provider total; Run output sequences permit only one enumeration. Public member signatures remain unchanged by these audit fixes.

  • AIModels.Perplexity.Sonar now identifies perplexity/sonar for the Agent API. Legacy Sonar model selections and Sonar-specific response contracts are removed as an intentional breaking migration ahead of the provider's 2026-09-27 endpoint retirement. See the migration guide.

  • DeepSeek V4Flash / Chat / Reasoner and OpenAI Gpt5, Gpt5Mini, Gpt5Nano, Gpt5Pro, O3, and O3Pro constants carry warning-only [Obsolete] annotations. Their names and original wire values are preserved; no replacement ID is silently substituted by the library. The provider temporarily routes deepseek-v4-flash to V4.1 Flash; use Flash explicitly in new code.

Compatibility

  • Breaking image-option migration: string Quality, Background, and OutputFormat become enums; Size becomes ImageSize; the separate request AspectRatio property is removed. OutputFormat now defaults to ImageOutputFormat.Auto. Update generation and edit callers using the migration guide; method signatures, result types, and image-model defaults remain. Provider adapters validate enum values, sizing modes, and format selectors before HTTP. No string compatibility conversion, image codec, or provider SDK is added.
  • Existing common enum values are retained. The completion cancellation parameter changes the IAIService contract and requires custom implementation updates and consumer rebuilds. The Perplexity-specific removals and changed Sonar wire ID require migration. Other obsolete model references produce warnings, which callers treating warnings as errors may need to resolve.
  • These controls are implemented by Mythosia.AI v8.0.0; the contracts package remains independent of provider SDKs and the full core implementation.

v3.1.0

Added

  • GPT-6 Astra contracts โ€” AIModels.OpenAI.Gpt6Astra, Gpt6Reasoning with Max, and Gpt6ReasoningMode for Standard/Pro configuration.
  • Optional run capability โ€” IAIRunService, AIRun, and IAIService extension bridges expose one execution's result, output events, cancellation, asynchronous disposal, and supported steering. Startup reuses the existing StreamOptions and AIRequestContext types.
  • Optional request features โ€” IAIRequestFeatureService, AIRequestFeatures, common reasoning levels and cache-preservation options, native web-search options, and provider-bound file-search store descriptors.
  • Fluent request configuration โ€” WithReasoning, WithWebSearch, and WithFileSearch retain the concrete service type while delegating to the optional request-feature capability; GetLastCitations reads its source references.
  • Citations โ€” AICitation, StreamingContent.Citation, StreamingContentType.Citation, and AIRun.Citations carry provider-supplied web and file references with response/content-local offsets.
  • Asynchronous function metadata โ€” FunctionDefinition.AllowAsync, AiFunctionAttribute.AllowAsync, and FunctionCall.IsAsync distinguish an opted-in function declaration from a provider's actual asynchronous invocation. Function-call snapshots preserve IsAsync.

Changed

  • Function policy documentation distinguishes ordinary sequential or bounded-parallel batches from the separately bounded pool of native asynchronous handlers. Async results may be delivered as completed subsets with their original call IDs.
  • Input-taking IAIService.StreamAsync methods document their planned public withdrawal in the next major release; the existing signatures remain available during the minor transition.

Compatibility

  • Additive minor release: IAIRunService and IAIRequestFeatureService are optional capabilities, so existing custom IAIService implementations gain no required members. Extension calls fail explicitly when the capability is unavailable.
  • Existing StreamingContentType numeric values are preserved. AIRun.Citations provides a default implementation for custom run subclasses.
  • AllowAsync defaults to false; provider implementations without native asynchronous calls retain ordinary execution. The implementation of these new capabilities is supplied by Mythosia.AI v7.1.0; this contracts package does not acquire a dependency on it or on provider SDKs.

v3.0.0

This is the abstractions release paired with Mythosia.AI v7. Follow the v7 migration guide before upgrading.

Added

  • Claude adaptive-thinking controls โ€” ClaudeReasoningEffort represents Auto, Low, Medium, High, XHigh, and Max; ClaudeThinkingDisplay represents omitted or summarized reasoning output.

  • GPT-5.6 model contract โ€” AIModels.OpenAI now includes Gpt5_6 (the Sol alias), Gpt5_6Sol, Gpt5_6Terra, and Gpt5_6Luna. Gpt5_6Reasoning adds the model family's Max effort, and Gpt5_6ReasoningMode represents standard or pro reasoning execution without inventing a gpt-5.6-pro model ID.

  • Claude 5 model IDs โ€” AIModels.Anthropic.ClaudeOpus5 and ClaudeSonnet5 expose the generally available claude-opus-5 and claude-sonnet-5 API models.

  • Claude Mythos 5 model ID โ€” AIModels.Anthropic.ClaudeMythos5 exposes claude-mythos-5 for approved Project Glasswing customers. Mythos 5 is limited availability, not generally available.

  • IImageGenerationService is an optional, provider-neutral contract for services that generate or edit images. Its DefaultImageModel is deliberately independent from IAIService.Model, so selecting a chat model cannot accidentally select an image model.

  • Image request and result models โ€” ImageGenerationRequest, ImageEditRequest, ImageInput, GeneratedImage, and ImageGenerationResult describe multi-image generation, ordered reference images, optional masks, output controls, provider/model provenance, request IDs, and token usage without adding a provider dependency.

  • Current OpenAI image model IDs โ€” AIModels.OpenAI.GptImage2 exposes the gpt-image-2 alias and GptImage2_260421 exposes its current 2026-04-21 snapshot without adding either image model to chat-model selectors.

  • Current Gemini model IDs โ€” AIModels.Google adds Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, plus an image-specific catalogue for Gemini 3.1 Flash Image, Gemini 3.1 Flash-Lite Image, and Gemini 3 Pro Image.

  • Current xAI model IDs โ€” AIModels.xAI adds Grok 4.5, its grok-4.5-latest alias, grok-build-latest, and the current Grok 4.3 aliases grok-4.3-latest and grok-latest.

  • Current Grok reasoning contract โ€” GrokReasoning now represents Auto, None, Low, Medium, and High. Auto omits the provider parameter, Grok 4.3 supports None through High, and Grok 4.5 supports Low through High and cannot disable reasoning.

  • Gemini safety controls โ€” GeminiSafetyThreshold represents provider-default, disabled, and the supported blocking thresholds without coupling the common image or chat contracts to Google request types.

  • Typed function-call batches โ€” FunctionCallBatch, FunctionCallResult, and FunctionCallResultBatch preserve every call from one assistant response, its provider correlation data, and the matching ordered results. Message and StreamingContent expose typed function call/result fields instead of requiring consumers to reconstruct them from metadata.

  • Function execution modes โ€” FunctionExecutionMode and FunctionCallingPolicy.ExecutionMode select sequential or bounded-parallel handler execution. Sequential remains the default; MaxConcurrency limits parallel handlers while result ordering remains stable.

Changed

  • Nullable annotations match the actual contracts โ€” optional attribute overrides, function handlers, schema descriptions/defaults/enums/items, and other provider-populated values are explicitly nullable. Required names and type identifiers retain non-null defaults, eliminating false consumer warnings without changing the CLR signatures.

Removed

  • ChatBlock.RemoveFunctionMessages() โ€” removing function calls while retaining their results creates an invalid conversation. Clear or rebuild the conversation explicitly when a model switch requires a new history.
  • Retired or deprecated OpenAI model IDs โ€” Gpt4Vision, Gpt4oLatest, Gpt5ChatLatest, Gpt5_2Codex, Gpt5_250807, Gpt5Mini_250807, Gpt5Nano_250807, and Gpt4_1Nano were removed instead of being retained as obsolete aliases.
  • Retired Claude Opus snapshots โ€” AIModels.Anthropic.ClaudeOpus4_250514 (claude-opus-4-20250514) and ClaudeOpus4_1_250805 (claude-opus-4-1-20250805) were removed. Opus 4 retired on June 15, 2026, and Opus 4.1 retired on August 5, 2026. Use ClaudeOpus4_8 for the official direct replacement.
  • GrokReasoning.Off โ€” use Auto to preserve the provider default without serializing reasoning_effort, or None for an explicit non-reasoning Grok 4.3 request. None is invalid for always-reasoning Grok 4.5.

Compatibility

  • Breaking release for the removed ChatBlock member, retired or deprecated model constants, and GrokReasoning.Off. Consumers of the new image contract and typed function batches require Mythosia.AI.Abstractions v3.0.0 or later.

v2.5.0

Added

  • ContextLengthExceededException โ€” the provider rejected the request because the prompt exceeded the model's context window. Carries the window and the rejected prompt's size when the provider reported them, the HTTP status, how many compaction attempts were made, and RecoverySkipReason โ€” why recovery could not save it. StreamAsync never throws it; there the overflow arrives as an error chunk instead. The legacy callback API StreamCompletionAsync, which has no round loop, does throw it.
  • AIHttpErrorFactory โ€” translates a provider's HTTP failure into that exception, or a plain AIServiceException when the body says something else, and builds the metadata for streaming error chunks. Every provider words its errors differently, so detection belongs here rather than in the core or the consuming app: the app should only ever learn that the context overflowed, never how a particular vendor phrases it. Recognises OpenAI's context_length_exceeded code, the OpenAI/vLLM maximum context length is N tokens wording, Anthropic's prompt is too long and combined input length and \max_tokens` exceed context limitforms, and Google'sinput token count (N) exceeds โ€ฆ`.
    • Detection is deliberately narrow and gated to HTTP 400/413. A false positive is worse than a miss: it makes the core compact the conversation, which deletes messages irreversibly, in response to an unrelated failure. Rate limits and server errors are never considered.
    • BuildErrorMetadata only adds keys โ€” error and status_code are always present, so existing readers are unaffected.

Deprecated

  • ChatBlock.RemoveFunctionMessages() now names its removal version (v7.0) instead of "future versions". Dropping function messages leaves function results without their originating call, which the chat/completions wire format rejects.

Compatibility

  • Additive minor release; no breaking changes.

v2.4.0

Added

  • AIModels.Anthropic.ClaudeFable5 โ€” claude-fable-5, Anthropic's new top model tier above Opus (1M context window, 128K max output).

Compatibility

  • Additive minor release; no breaking changes.

v2.3.0

Added

  • Model id constants โ€” 2026 refresh (AIModels):
    • OpenAI: Gpt5_5, Gpt5_5_260423, Gpt5_5Pro, Gpt5_5Pro_260423, Gpt5Pro.
    • Anthropic: ClaudeOpus4_8, ClaudeOpus4_7.
    • Google: Gemini3_1ProPreview, Gemini3_5Flash, Gemini3_1FlashLite.
    • xAI: Grok4_3, Grok4_20Reasoning, Grok4_20NonReasoning, GrokBuild0_1.
    • Perplexity: SonarReasoningPro.
  • Gpt5_5Reasoning enum โ€” reasoning effort for GPT-5.5 (None/Low/Medium/High/XHigh).

Removed

  • Constants for retired or non-callable models (the underlying models no longer function at runtime):
    • xAI: Grok4 (grok-4-0709), Grok4_1Fast (grok-4-1-fast), Grok3 (grok-3) โ€” retired 2026-05-15; Grok4_20MultiAgent (grok-4.20-multi-agent-0309) โ€” multi-agent API, not chat completions.
    • Anthropic: ClaudeSonnet4_250514 (claude-sonnet-4-20250514) โ€” retires 2026-06-15.
    • Perplexity: SonarReasoning (sonar-reasoning) โ€” removed from the API 2025-12-15.
    • Google: Gemini3ProPreview (gemini-3-pro-preview) โ€” shut down 2026-03-09.
    • OpenAI: Gpt5_3CodexSpark (gpt-5.3-codex-spark) โ€” not generally accessible via the standard API.

Compatibility

  • Minor release. The removed constants reference models that are already retired/non-functional; code that references them by name requires a one-line update to a current model id.

v2.2.0

Added

  • Streaming diagnostics primitives (consumed by Mythosia.AI's new WithStreamDiagnostics extension and ReadSseLinesAsync helper).
    • StreamDiagnostics โ€” observability snapshot for a single SSE round. Fields: LinesRead, DataLinesProcessed, ParseFailures, AccumulatedTextLength, LastRawLine, Elapsed. Lets callers tell "stream silently ended" apart from "transport error after N chunks".
    • StreamReadException โ€” wraps the underlying read-time exception (IOException, NotSupportedException, ObjectDisposedException, etc.) and attaches a StreamDiagnostics snapshot taken at the moment of failure. InnerException preserves the original.
    • StreamDiagnosticsBuilder โ€” fluent configurator with independent OnRawLine(Action<string>) and OnComplete(Action<StreamDiagnostics>) hooks. Consumed by service.WithStreamDiagnostics(d => d.OnRawLine(...).OnComplete(...)) in Mythosia.AI.

Compatibility

  • Additive public API update.
  • No existing types or members were removed, renamed, or changed.

v2.1.0

Added

  • Round-scoped streaming usage
    • Added StreamingContentType.RoundUsage.
    • Added StreamingContent.RoundIndex and StreamingContent.IsFinalRound.
    • StreamingContent.Usage can now carry one-round usage on RoundUsage events while Completion keeps cumulative run usage.

Compatibility

  • Additive public API update.
  • No existing members were removed or renamed.

v2.0.0

Changed

  • IAIService.GetCompletionAsync โ€” 8 overloads โ†’ 2 (breaking change)

    • GetCompletionAsync(string prompt, AIRequestProfile? profile = null, AIRequestContext? context = null)
    • GetCompletionAsync(Message message, AIRequestProfile? profile = null, AIRequestContext? context = null)
    • Migration: callers using GetCompletionAsync(message, context) must switch to GetCompletionAsync(message, context: context).

  • IAIService.StreamAsync โ€” 6 overloads โ†’ 4 (breaking change)

    • StreamAsync(string prompt, CancellationToken cancellationToken = default)
    • StreamAsync(Message message, AIRequestContext? context = null, CancellationToken cancellationToken = default)
    • StreamAsync(string prompt, StreamOptions options, CancellationToken cancellationToken = default)
    • StreamAsync(Message message, StreamOptions options, AIRequestContext? context = null, CancellationToken cancellationToken = default)
    • Migration: callers using StreamAsync(message, context, ct) must switch to StreamAsync(message, context: context, ct).

v1.1.0

Added

  • IFunctionRegisterable โ€” new interface that marks an AI service as capable of accepting function (tool) registrations at runtime.

Documentation

  • Added Relationship to Microsoft.Extensions.AI section to README โ€” a factual comparison table between IAIService and IChatClient covering state management, session handling, request parameters, function calling, streaming, conversation summarization, multimodal handling, and token usage tracking.
    • Exposes a single void AddFunction(FunctionDefinition function) method.
    • Allows extension packages (e.g., Mythosia.AI.Rag) to register tools on AIService without taking a direct dependency on the full Mythosia.AI core package.
    • AIService in Mythosia.AI v5.3.0 implements this interface via explicit interface implementation (IFunctionRegisterable.AddFunction โ†’ Functions.Add).

Purpose

Enables Mythosia.AI.Rag.WithAgenticRag<TService>() to constrain its generic parameter to where TService : IAIService, IFunctionRegisterable, keeping the RAG package dependency-free from the AI core while still being able to register the RAG search tool.

v1.0.0

Initial Release

Extracted core abstractions and shared models from Mythosia.AI into a standalone zero-dependency package.

Added

  • IAIService โ€” core interface defining the AI completion and streaming contract.

    • Completion: GetCompletionAsync overloads with string, Message, AIRequestProfile, AIRequestContext.
    • Streaming: StreamAsync overloads returning IAsyncEnumerable<string> or IAsyncEnumerable<StreamingContent>.
    • Properties: Model, Provider, SystemMessage, StatelessMode, ActivateChat.

  • Models โ€” all shared model types previously in Mythosia.AI:

    • Message, MessageContent (TextContent, ImageContent, AudioContent)
    • ChatBlock โ€” conversation container with system message and message history
    • ActorRole โ€” role enum (System, User, Assistant, Function)
    • AIRequestContext โ€” per-request context overrides
    • AIRequestProfile, AIRequestPurpose, RequestProfiles โ€” per-request parameter profiles
    • AIModels โ€” model identifier constants for all providers
    • AIProvider โ€” provider enum
    • SummaryConversationPolicy โ€” conversation summarization configuration
    • StructuredOutputPolicy โ€” structured output behavior configuration
    • RoundResult โ€” function calling round result

  • Streaming types:

    • StreamingContent, StreamingContentType โ€” streaming chunk with type discrimination
    • StreamOptions โ€” streaming behavior configuration
    • TokenUsage โ€” token count data (input, output, cached, reasoning)

  • Function calling types:

    • FunctionDefinition โ€” function schema for LLM tool use
    • FunctionCallingPolicy, FunctionCallMode โ€” function calling behavior controls

  • Attributes:

    • AiFunctionAttribute โ€” marks methods as AI-callable
    • AiParameterAttribute โ€” describes function parameters

  • Enums: ReasoningEffort, ReasoningSummary, GeminiThinkingLevel, Verbosity

  • Exceptions: AIServiceException, AgentMaxStepsExceededException

Purpose

Enables downstream libraries (e.g., Mythosia.AI.Rag) to depend on the lightweight contract package instead of the full Mythosia.AI implementation, avoiding transitive dependencies on Azure.AI.OpenAI, NJsonSchema, TiktokenSharp, etc.

Edit this page