Only integration tests are accepted: start the server, call its public HTTP endpoints or real provider SDKs, and inspect public responses/control APIs. Test setup may build explicit fixtures, but assertions must prove observable server behavior. Source/document scans, fake clients, reflection guards and direct runtime unit tests have been removed. This inventory is documentation; tests do not read it to assert that method names appear here.
| Contract | Integration evidence |
|---|---|
| Default models, chat, embeddings, images, audio | OpenAiEndpointTests.OpenAiEndpoints_ExposeModelsAndAllModalitiesAsync |
| JSON and native SSE streaming, usage, terminal events | OpenAiEndpointTests, OpenAiSdkCompatibilityTests, OpenAiCompatibleProviderRouteTests |
| Fixed output tokenizer counts exposed on response and journal | LlmTestServerTokenCounterTests.ChatResponse_ReportsConfiguredOutputTokenUsageThroughHttpAsync |
| Reasoning and prompt-cache usage | OpenAiEndpointTests, AnthropicEndpointTests, GeminiEndpointTests, BedrockEndpointTests, ProviderApiDriftTests |
| Auth, malformed bodies, unknown models, required parameters | OpenAiEndpointTests, ProviderValidationBoundaryTests, ReviewHttpRegressionTests |
| Configure/reset/assertions through typed control client | OpenAiEndpointTests.ControlClient_ConfiguresReadsAssertionsAndResetsRuntimeAsync |
| Invalid configuration preserves the active runtime after HTTP 400 | ConfigurationBoundaryIntegrationTests |
| Per-modality auth, wrong model kind, content filter, rate limit and reset | ModalityFailureIntegrationTests |
| Stored response credential/provider isolation, count/byte capacity, store=false and reset/configure | ResponseHistoryIntegrationTests |
| Exact conversation order, scenario credentials, exhaustion and cancellation/retry | ScenarioFlowIntegrationTests |
| Parallel session queues, reset/configure, credentials and admission/disposal | SessionIsolationIntegrationTests, SessionBoundaryIntegrationTests |
| Session response history, video/fault state, cancellation and in-flight deletion | SessionStateIntegrationTests |
| Session provider matrix, official SDK JSON/SSE, native decisions, cache and journal limits | SessionProviderIntegrationTests |
| Conversation branches, turns, JSON predicates, tools, replay, exhaustion and invalid graph/configuration | ConversationHttpIntegrationTests |
| Conversation official SDK tool loops, Responses history/replay and streaming | ConversationSdkIntegrationTests |
| Conversation provider/credential/session isolation, cancellation and in-flight mutation | ConversationProviderIntegrationTests |
| Client model presets actually reach the server | LlmTestServerClientConfigurationTests |
| MEAI chat, embedding and image clients call the server | MicrosoftExtensionsAiClientTests.MicrosoftExtensionsAiClients_InvokeLlmTestServerThroughOpenAiCompatibilityEndpointsAsync |
| Official OpenAI SDK chat, Responses, tools, embeddings and modalities | OpenAiSdkCompatibilityTests, ToolFixtureClientTests |
| Azure OpenAI deployment endpoints through official SDK | AzureSdkCompatibilityTests |
| Azure/Foundry v1 endpoints through official clients | AzureV1SdkCompatibilityTests, V1AudioTranscriptionCompatibilityTests |
| Foundry inference version and auth requirements | AzureSdkCompatibilityTests, ReviewHttpRegressionTests.FoundryInference_RequiresDocumentedVersionWithoutConsumingQueueAsync |
| Anthropic Messages, SSE, tool history, cache | AnthropicEndpointTests, ToolFixtureEndpointTests |
| Gemini generation, stream, embedding, video, schema options | GeminiEndpointTests, FollowUpEndpointTests, ProviderApiDriftTests |
| Cohere/Ollama native chat and embedding | CohereEndpointTests, OllamaEndpointTests |
| Bedrock Converse/Invoke and native binary event streams | BedrockEndpointTests, BedrockSdkCompatibilityTests |
| OpenAI-compatible provider namespaces and documented operations | OpenAiCompatibleProviderRouteTests |
| Provider-neutral rate limit and content filter errors | ProviderFaultSimulationTests |
| Tool/schema operation variants, JSON and streaming, negative then successful recovery | ProviderCapabilityEvidenceTests.ClaimedToolAndSchemaCases_ExecutePositiveAndNegativeHttpEvidenceAsync |
| Tool selection, local references, invalid arguments, structured stream concatenation | ToolFixtureValidationTests |
| Parallel-tool request limits and malformed-schema/nullable/nested constraints | FollowUpEndpointTests, ToolFixtureClientTests |
| Official SDK/MEAI function invocation, tool-result continuation, model changes, caller-managed stateless history and documented include selectors | ToolFixtureClientTests, LlmTestServerResponsesContinuationTests, ResponsesCallerManagedHistoryTests |
| Video CRUD/content/delete/reset/provider isolation | ReviewVideoRegressionTests, OpenAiEndpointTests, AzureSdkCompatibilityTests |
| Native Jev/Kev/OpenAI/Clef responses, errors, images, bounds and parameter variants | DecisionEndpointTests, DecisionRuntimeTests, DecisionValidationTests, DecisionQuestionTests, DecisionInputTests, DecisionFixtureContractTests |
| Jev multipart/image/box/long-context contracts | Jev114EndpointTests, Jev114MultipartTests, TypeSafeVisionContractTests |
| Native Clef image/scalar/error envelopes and typed client | CloudflareDecisionContractTests, CloudflareDecisionErrorTests, CloudflareDecisionImageTests, CloudflareNativeClientTests |
| Native Kev formatting and required/optional request fields over HTTP | SystemOneNativeContractTests |
| Dashboard root HTML and request/response presentation | AdminPanelTests |
| Real cache miss/hit, reasoning subsets, explicit conversation correlation and unknown usage over HTTP | DashboardAnalyticsIntegrationTests |
| Per-role input composition (system, user, assistant, tool calls, tool results, tool schemas) reconciles with OpenAI and Anthropic wire usage and drives the dashboard token flow and context mix | DashboardAnalyticsIntegrationTests |
| Stored-response capacity rejection preserves queue/graph and does not warm or evict cache | StoredResponseAdmissionIntegrationTests |
| Cache isolation across provider namespaces and independent sessions | PromptCacheNamespaceIntegrationTests |
| Longest retained prefix is checked before FIFO eviction; cache-off mode stays inactive | PromptCacheCapacityIntegrationTests |
| Request/conversation analytics totals, repeated tool-result context, unlabeled-request isolation and first-viewport readability | scripts/package-consumer/PackageTests.cs (genuine SDK and Chromium) |
| Public request journal auth, all filter/paging parameters, validation, reset | RequestJournalIntegrationTests |
| Bounded request/response diagnostics preserve full HTTP response bytes | RequestJournalIntegrationTests.Journal_BoundsRetentionWithoutTruncatingActualProviderResponsesAsync |
| Request query filters, order, JSON/NDJSON export, retention/eviction/reset and active-request evidence | RequestQueryExportIntegrationTests |
| Counts/order/no-unexpected-calls, structured tool/result predicates and graph expectations | VerificationIntegrationTests |
| Retry group/status/count, advertised Retry-After and real monotonic wait compliance | VerificationIntegrationTests |
| Incomplete evidence after eviction/truncation/paging/pending requests cannot pass | VerificationIntegrationTests |
| Seed/fixed clock/reset/replay, provider-local ordinals, session isolation and real observations | DeterminismIntegrationTests |
| Native Anthropic/Gemini/Cohere/Ollama JSON/stream identity, embedding/image/Groq audio metadata | NativeDeterminismIntegrationTests |
| Aspire starts packaged service and exercises official SDKs through allocated endpoint | AspireIntegrationTests |
| Configuration key enforced by running Aspire server | AspireAuthorizationIntegrationTests |
| Installed NuGet assets, typed configuration, official SDK and interactive Chromium dashboard | scripts/package-consumer/PackageTests.cs, run by scripts/verify-package.py with isolated build configuration/cache |
| Actual public GitHub release and all NuGet bytes | scripts/verify-release.py, executed after publication |
| Clean public-feed installed consumer | scripts/verify-package.py --public-feed, executed after publication |
E4 local qualification passed 812/812 running-server integrations, 91.66% production line coverage and a
final canonical Release build with 0 warnings / 0 errors. All 19 packages packed successfully. The
independent clean installed NuGet consumer ran the packaged Aspire service, official SDKs, E4 typed
query/export/verification/fixed metadata and interactive Chromium acceptance (1/1).
Evidence logs: artifacts/delivery/e4-build-final.log, e4-coverage-final.log, e4-pack.log, e4-consumer.log.
E3 CI 39446187097 passed at 3aae4db.
E4 has not yet been pushed to branch CI. Local NuGet acceptance does not prove public-feed publication.
Add actual server flows for new behavior and preserve the 90% coverage gate.
Each implemented operation must be exercised with every supported parameter, positive and invalid/boundary values, supported streaming modes, auth/version handling and stateful continuation where applicable. A passing subset or this inventory alone does not establish full parameter coverage. New provider claims require official method-level contracts plus passing HTTP/SDK behavior evidence.
The OpenAI/NativeWebSearch*IntegrationTests.cs suites exercise the running server through HTTP and the official Responses SDK: typed ordered outputs, citations, native SSE lifecycle, options and schema rejection, mixed function rounds, explicit/stored continuation, cache accounting, ownership, reset/cancellation, replay and bounded diagnostics. Initial qualification: 99/99 native cases and 85/85 existing provider evidence cases. The installed-package consumer adds an independent NativePackageTests flow; container verification covers native JSON/SSE and terminal request evidence. See the exact supported subset and qualification. Ordinary function-only responses retain their existing output limits and are covered separately by the large-response regression.
The eleven OpenAI/NativeFileCode*IntegrationTests.cs files run 191 HTTP/official SDK cases for the independent gpt-4o-mini file-search and gpt-4.1 code-interpreter contracts: JSON/SSE, typed ordered citations, options and every supported filter operator, exact numeric comparisons, resource matching, include projection, continuation/replay, isolation, invalid-ID normalization, atomic configuration, cache read/miss and request export. Expanded web regression coverage is 105/105, including valid unused include selectors. The independent code-only full server gate passed 1158/1158 with 92.21% line coverage; see the bounded contract and delivery evidence.