@nemesis-oss/ollama-sdk 1.11.0 → 1.12.0

package/dist/index.d.cts

@@ -1,12 +1,12 @@
-import { T as ThinkingMetadata, M as ModelResponse, R as RequestCancellationOptions, S as ShowRequestOptions, a as ShowResponse, P as PullRequestOptions, b as ProgressResponse, c as PushRequestOptions, C as CreateRequestOptions, D as DeleteRequestOptions, d as StatusResponse, e as CopyRequestOptions, f as PsResponse, V as VersionResponse, g as ToolParameters, h as ThinkValue, i as ToolCall, j as ChatRequestOptions, k as ModelOptions, l as ToolDefinition, F as FormatOption, m as ChatResponse, n as Message, G as GenerateRequestOptions, o as GenerateResponse, E as EmbedRequestOptions, p as EmbedResponse, q as EmbeddingsRequestOptions, r as EmbeddingsResponse, W as WebSearchRequestOptions, s as WebSearchResponse, t as WebFetchRequestOptions, u as WebFetchResponse, U as UsageRequestOptions, B as BalanceRequestOptions, v as VisionInput } from './types-BtQexFpT.cjs';
-export { L as ListResponse, w as Logprob, x as LogprobToken, y as ModelDetails, z as Role, A as ToolCallFunction, H as ToolFunctionDefinition, I as ToolProperty, J as UsageRange, K as UsageScope, N as WebSearchResult, O as imageStringNeedsResolution, Q as resolveImageInput } from './types-BtQexFpT.cjs';
-import { H as HttpClient, F as FetchLike, M as Middleware, L as Logger, R as RequestLifecycleHook, c as OllamaStreamEvent, A as AbortableAsyncIterable, d as OllamaStreamEventType, P as ProgressStreamResult, B as BinaryBody, O as OllamaRuntime, C as ChatStreamResult, G as GenerateStreamResult, e as RawUsageSource, T as TokenUsage } from './runtime-DTtWCxDs.cjs';
-export { D as DoneEvent, E as ErrorEvent, f as HttpBody, g as HttpClientOptions, h as HttpRequestOptions, i as LifecycleErrorEvent, j as LifecycleRetryEvent, k as LifecycleStartEvent, l as LifecycleSuccessEvent, m as LogFn, n as MessageEvent, o as MiddlewareContext, N as NANOS_PER_MS, p as NANOS_PER_SECOND, q as NOOP_LOGGER, r as NextFunction, s as OllamaAbortError, t as OllamaAgentMaxIterationsError, u as OllamaAgentMaxToolCallsError, v as OllamaAgentToolLoopError, w as OllamaAuthError, x as OllamaBadGatewayError, y as OllamaClientError, z as OllamaClientErrorOptions, J as OllamaErrorRequestContext, K as OllamaErrorResponseContext, Q as OllamaGenericClientError, U as OllamaIncompatibleModelError, V as OllamaMcpError, W as OllamaModelRoutingError, X as OllamaNetworkError, Y as OllamaNotFoundError, Z as OllamaOverloadedError, _ as OllamaQuotaExceededError, $ as OllamaRateLimitError, a0 as OllamaRequestTooLargeError, a1 as OllamaRequestValidationError, a2 as OllamaResponseValidationError, a3 as OllamaServerError, a4 as OllamaServerVersionUnknownError, a5 as OllamaSkillInvalidError, a6 as OllamaSkillNotFoundError, a7 as OllamaStreamError, a8 as OllamaTimeoutError, a9 as OllamaToolTimeoutError, aa as OllamaToolValidationError, ab as OllamaUnsupportedCapabilityError, ac as RequestContext, ad as RequestLifecycleEvent, ae as ResponseContext, af as SseEvent, ag as ThinkingEvent, ah as TokenEvent, ai as ToolCallEvent, aj as composeMiddleware, ak as createConsoleLogger, al as extractUsage, am as mapError, an as parseSseStream } from './runtime-DTtWCxDs.cjs';
+import { T as ThinkingMetadata, M as ModelResponse, R as RequestCancellationOptions, S as ShowRequestOptions, a as ShowResponse, P as PullRequestOptions, b as ProgressResponse, c as PushRequestOptions, C as CreateRequestOptions, D as DeleteRequestOptions, d as StatusResponse, e as CopyRequestOptions, f as PsResponse, V as VersionResponse, g as ToolParameters, h as ThinkValue, i as ToolCall, j as ChatRequestOptions, k as ModelOptions, l as ToolDefinition, F as FormatOption, m as ChatResponse, n as Message, G as GenerateRequestOptions, o as GenerateResponse, E as EmbedRequestOptions, p as EmbedResponse, q as EmbeddingsRequestOptions, r as EmbeddingsResponse, W as WebSearchRequestOptions, s as WebSearchResponse, t as WebFetchRequestOptions, u as WebFetchResponse, U as UsageRequestOptions, B as BalanceRequestOptions, v as PingResult, w as VisionInput } from './types-B57ZZfp3.cjs';
+export { L as ListResponse, x as Logprob, y as LogprobToken, z as ModelDetails, A as Role, H as ToolCallFunction, I as ToolFunctionDefinition, J as ToolProperty, K as UsageRange, N as UsageScope, O as WebSearchResult, Q as imageStringNeedsResolution, X as resolveImageInput } from './types-B57ZZfp3.cjs';
+import { H as HttpClient, F as FetchLike, M as Middleware, L as Logger, R as RequestLifecycleHook, c as OllamaStreamEvent, A as AbortableAsyncIterable, d as OllamaStreamEventType, P as ProgressStreamResult, B as BinaryBody, O as OllamaRuntime, C as ChatStreamResult, G as GenerateStreamResult, e as RawUsageSource, T as TokenUsage } from './runtime-DmI1cOEq.cjs';
+export { D as DoneEvent, E as ErrorEvent, f as HttpBody, g as HttpClientOptions, h as HttpRequestOptions, i as LifecycleErrorEvent, j as LifecycleRetryEvent, k as LifecycleStartEvent, l as LifecycleSuccessEvent, m as LogFn, n as MessageEvent, o as MiddlewareContext, N as NANOS_PER_MS, p as NANOS_PER_SECOND, q as NOOP_LOGGER, r as NextFunction, s as OllamaAbortError, t as OllamaAgentMaxIterationsError, u as OllamaAgentMaxToolCallsError, v as OllamaAgentToolLoopError, w as OllamaAuthError, x as OllamaBadGatewayError, y as OllamaClientError, z as OllamaClientErrorOptions, J as OllamaErrorRequestContext, K as OllamaErrorResponseContext, Q as OllamaGenericClientError, U as OllamaIncompatibleModelError, V as OllamaMcpError, W as OllamaModelRoutingError, X as OllamaNetworkError, Y as OllamaNotFoundError, Z as OllamaOverloadedError, _ as OllamaQuotaExceededError, $ as OllamaRateLimitError, a0 as OllamaRequestTooLargeError, a1 as OllamaRequestValidationError, a2 as OllamaResponseValidationError, a3 as OllamaServerError, a4 as OllamaServerVersionUnknownError, a5 as OllamaSkillInvalidError, a6 as OllamaSkillNotFoundError, a7 as OllamaStreamError, a8 as OllamaTimeoutError, a9 as OllamaToolTimeoutError, aa as OllamaToolValidationError, ab as OllamaUnsupportedCapabilityError, ac as RequestContext, ad as RequestLifecycleEvent, ae as ResponseContext, af as SseEvent, ag as ThinkingEvent, ah as TokenEvent, ai as ToolCallEvent, aj as composeMiddleware, ak as createConsoleLogger, al as extractUsage, am as mapError, an as parseSseStream } from './runtime-DmI1cOEq.cjs';
 import { z } from 'zod';
-import { R as RequestRunner, S as SystemOneChoiceAnswer, o as SystemOneNoulAnswer, p as SystemOneScoreAnswer, q as SystemOneChoiceQuestion, r as SystemOneNoulQuestion, s as SystemOneScoreQuestion, t as SystemOneRequest$1, u as SystemOneResponse$1, v as SystemOneContent, N as NativeApi, w as OpenAICompatClient, x as AnthropicCompatClient } from './anthropic-DinrABof.cjs';
-export { y as AnthropicCacheControl, z as AnthropicContentBlock, B as AnthropicImageContentBlock, C as AnthropicMessage, n as AnthropicMessageStreamEvent, A as AnthropicMessagesRequest, m as AnthropicMessagesResponse, D as AnthropicMessagesStream, E as AnthropicMetadata, F as AnthropicOutputConfig, G as AnthropicRedactedThinkingContentBlock, H as AnthropicSystem, I as AnthropicSystemTextBlock, J as AnthropicTextContentBlock, K as AnthropicThinkingConfig, L as AnthropicThinkingContentBlock, M as AnthropicTool, P as AnthropicToolChoice, Q as AnthropicToolResultContentBlock, T as AnthropicToolUseContentBlock, U as AnthropicUnknownStreamEvent, V as OllamaAnthropicContentBlock, W as OllamaAnthropicMessage, X as OllamaAnthropicMessagesRequest, Y as OllamaAnthropicSystem, Z as OllamaAnthropicSystemTextBlock, _ as OllamaAnthropicTextContentBlock, $ as OllamaAnthropicThinkingConfig, a0 as OllamaAnthropicTool, a1 as OllamaAnthropicToolResultContentBlock, a2 as OllamaOpenAIChatCompletionRequest, a3 as OllamaOpenAICompletionRequest, a4 as OllamaOpenAIEmbeddingRequest, a5 as OllamaOpenAIReasoningEffort, a6 as OllamaOpenAIResponsesRequest, a7 as OpenAIChatCompletionChoice, b as OpenAIChatCompletionChunk, a8 as OpenAIChatCompletionChunkChoice, a9 as OpenAIChatCompletionDelta, O as OpenAIChatCompletionRequest, a as OpenAIChatCompletionResponse, aa as OpenAIChatCompletionStream, ab as OpenAICompletionChoice, e as OpenAICompletionChunk, c as OpenAICompletionRequest, d as OpenAICompletionResponse, ac as OpenAICompletionStream, ad as OpenAIContentPart, ae as OpenAIEmbeddingInput, af as OpenAIEmbeddingItem, f as OpenAIEmbeddingRequest, g as OpenAIEmbeddingResponse, ag as OpenAIFunctionDefinition, ah as OpenAIImageUrlContentPart, h as OpenAIListModelsResponse, ai as OpenAIMessage, i as OpenAIModelItem, aj as OpenAIReasoningEffort, ak as OpenAIResponseFormat, al as OpenAIResponsesCompletedEvent, am as OpenAIResponsesContentPartAddedEvent, an as OpenAIResponsesContentPartDoneEvent, ao as OpenAIResponsesCreatedEvent, ap as OpenAIResponsesError, aq as OpenAIResponsesEventBase, ar as OpenAIResponsesFailedEvent, as as OpenAIResponsesFunctionCallArgumentsDeltaEvent, at as OpenAIResponsesFunctionCallArgumentsDoneEvent, au as OpenAIResponsesInProgressEvent, av as OpenAIResponsesIncompleteDetails, aw as OpenAIResponsesIncompleteEvent, ax as OpenAIResponsesOutputContent, ay as OpenAIResponsesOutputFunctionCall, az as OpenAIResponsesOutputItem, aA as OpenAIResponsesOutputItemAddedEvent, aB as OpenAIResponsesOutputItemDoneEvent, aC as OpenAIResponsesOutputMessage, aD as OpenAIResponsesOutputReasoning, aE as OpenAIResponsesOutputRefusal, aF as OpenAIResponsesOutputTextContent, aG as OpenAIResponsesOutputTextDeltaEvent, aH as OpenAIResponsesOutputTextDoneEvent, aI as OpenAIResponsesQueuedEvent, aJ as OpenAIResponsesReasoningSummaryPartAddedEvent, aK as OpenAIResponsesReasoningSummaryPartDoneEvent, aL as OpenAIResponsesReasoningSummaryTextDeltaEvent, aM as OpenAIResponsesReasoningSummaryTextDoneEvent, aN as OpenAIResponsesReasoningTextDeltaEvent, aO as OpenAIResponsesReasoningTextDoneEvent, aP as OpenAIResponsesRefusalDeltaEvent, aQ as OpenAIResponsesRefusalDoneEvent, j as OpenAIResponsesRequest, k as OpenAIResponsesResponse, aR as OpenAIResponsesStatus, aS as OpenAIResponsesStream, aT as OpenAIResponsesStreamError, l as OpenAIResponsesStreamEvent, aU as OpenAIResponsesUsage, aV as OpenAIStreamOptions, aW as OpenAITextContentPart, aX as OpenAITool, aY as OpenAIToolCall, aZ as OpenAIToolCallDelta, a_ as OpenAIUsage, a$ as SystemOneConfidence, b0 as SystemOneProbabilities } from './anthropic-DinrABof.cjs';
-import { T as ToolHandler, a as Tool, A as AnyTool, b as ToolRegistry, c as ToolExecutionResult } from './compose-lt-Qz3lI.cjs';
-export { d as AppliedSkillResult, e as ApplySkillOptions, P as ParsedFrontmatter, S as Skill, f as SkillFrontmatter, g as SkillSummary, h as ToolExecutionContext, i as ToolExecutionFailure, j as ToolExecutionSuccess, k as ToolRegistryOptions, l as applySkill, p as parseFrontmatter } from './compose-lt-Qz3lI.cjs';
+import { R as RequestRunner, S as SystemOneChoiceAnswer, o as SystemOneNoulAnswer, p as SystemOneScoreAnswer, q as SystemOneChoiceQuestion, r as SystemOneNoulQuestion, s as SystemOneScoreQuestion, t as SystemOneRequest$1, u as SystemOneResponse$1, v as SystemOneContent, N as NativeApi, w as OpenAICompatClient, x as AnthropicCompatClient } from './anthropic-Did0lQxP.cjs';
+export { y as AnthropicCacheControl, z as AnthropicContentBlock, B as AnthropicImageContentBlock, C as AnthropicMessage, n as AnthropicMessageStreamEvent, A as AnthropicMessagesRequest, m as AnthropicMessagesResponse, D as AnthropicMessagesStream, E as AnthropicMetadata, F as AnthropicOutputConfig, G as AnthropicRedactedThinkingContentBlock, H as AnthropicSystem, I as AnthropicSystemTextBlock, J as AnthropicTextContentBlock, K as AnthropicThinkingConfig, L as AnthropicThinkingContentBlock, M as AnthropicTool, P as AnthropicToolChoice, Q as AnthropicToolResultContentBlock, T as AnthropicToolUseContentBlock, U as AnthropicUnknownStreamEvent, V as OllamaAnthropicContentBlock, W as OllamaAnthropicMessage, X as OllamaAnthropicMessagesRequest, Y as OllamaAnthropicSystem, Z as OllamaAnthropicSystemTextBlock, _ as OllamaAnthropicTextContentBlock, $ as OllamaAnthropicThinkingConfig, a0 as OllamaAnthropicTool, a1 as OllamaAnthropicToolResultContentBlock, a2 as OllamaOpenAIChatCompletionRequest, a3 as OllamaOpenAICompletionRequest, a4 as OllamaOpenAIEmbeddingRequest, a5 as OllamaOpenAIReasoningEffort, a6 as OllamaOpenAIResponsesRequest, a7 as OpenAIChatCompletionChoice, b as OpenAIChatCompletionChunk, a8 as OpenAIChatCompletionChunkChoice, a9 as OpenAIChatCompletionDelta, O as OpenAIChatCompletionRequest, a as OpenAIChatCompletionResponse, aa as OpenAIChatCompletionStream, ab as OpenAICompletionChoice, e as OpenAICompletionChunk, c as OpenAICompletionRequest, d as OpenAICompletionResponse, ac as OpenAICompletionStream, ad as OpenAIContentPart, ae as OpenAIEmbeddingInput, af as OpenAIEmbeddingItem, f as OpenAIEmbeddingRequest, g as OpenAIEmbeddingResponse, ag as OpenAIFunctionDefinition, ah as OpenAIImageUrlContentPart, h as OpenAIListModelsResponse, ai as OpenAIMessage, i as OpenAIModelItem, aj as OpenAIReasoningEffort, ak as OpenAIResponseFormat, al as OpenAIResponsesCompletedEvent, am as OpenAIResponsesContentPartAddedEvent, an as OpenAIResponsesContentPartDoneEvent, ao as OpenAIResponsesCreatedEvent, ap as OpenAIResponsesError, aq as OpenAIResponsesEventBase, ar as OpenAIResponsesFailedEvent, as as OpenAIResponsesFunctionCallArgumentsDeltaEvent, at as OpenAIResponsesFunctionCallArgumentsDoneEvent, au as OpenAIResponsesInProgressEvent, av as OpenAIResponsesIncompleteDetails, aw as OpenAIResponsesIncompleteEvent, ax as OpenAIResponsesOutputContent, ay as OpenAIResponsesOutputFunctionCall, az as OpenAIResponsesOutputItem, aA as OpenAIResponsesOutputItemAddedEvent, aB as OpenAIResponsesOutputItemDoneEvent, aC as OpenAIResponsesOutputMessage, aD as OpenAIResponsesOutputReasoning, aE as OpenAIResponsesOutputRefusal, aF as OpenAIResponsesOutputTextContent, aG as OpenAIResponsesOutputTextDeltaEvent, aH as OpenAIResponsesOutputTextDoneEvent, aI as OpenAIResponsesQueuedEvent, aJ as OpenAIResponsesReasoningSummaryPartAddedEvent, aK as OpenAIResponsesReasoningSummaryPartDoneEvent, aL as OpenAIResponsesReasoningSummaryTextDeltaEvent, aM as OpenAIResponsesReasoningSummaryTextDoneEvent, aN as OpenAIResponsesReasoningTextDeltaEvent, aO as OpenAIResponsesReasoningTextDoneEvent, aP as OpenAIResponsesRefusalDeltaEvent, aQ as OpenAIResponsesRefusalDoneEvent, j as OpenAIResponsesRequest, k as OpenAIResponsesResponse, aR as OpenAIResponsesStatus, aS as OpenAIResponsesStream, aT as OpenAIResponsesStreamError, l as OpenAIResponsesStreamEvent, aU as OpenAIResponsesUsage, aV as OpenAIStreamOptions, aW as OpenAITextContentPart, aX as OpenAITool, aY as OpenAIToolCall, aZ as OpenAIToolCallDelta, a_ as OpenAIUsage, a$ as SystemOneConfidence, b0 as SystemOneProbabilities } from './anthropic-Did0lQxP.cjs';
+import { T as ToolHandler, a as Tool, A as AnyTool, b as ToolRegistry, c as ToolExecutionResult } from './compose-DhvoUrWa.cjs';
+export { d as AppliedSkillResult, e as ApplySkillOptions, P as ParsedFrontmatter, S as Skill, f as SkillFrontmatter, g as SkillSummary, h as ToolExecutionContext, i as ToolExecutionFailure, j as ToolExecutionSuccess, k as ToolRegistryOptions, l as applySkill, p as parseFrontmatter } from './compose-DhvoUrWa.cjs';
 import { M as McpClientLike, a as McpToolDescriptor, b as McpTask, c as McpCallToolResult } from './types-CI-gM2u5.cjs';
 export { d as McpContentBlock, e as McpCreateTaskResult, f as McpElicitationFormRequest, g as McpElicitationHandlers, h as McpElicitationRequest, i as McpElicitationResult, j as McpElicitationUrlRequest, k as McpIcon, l as McpInputRequiredResult, m as McpListToolsParams, n as McpListToolsResult, o as McpRequestOptions, p as McpServerCapabilities, q as McpTaskMethod, r as McpTaskRequest, s as McpTaskStatus, t as McpToolAnnotations, u as McpToolCallResult, v as McpToolExecution } from './types-CI-gM2u5.cjs';
 import * as Otel from '@opentelemetry/api';
@@ -290,6 +290,15 @@ interface DiscoveredContextLength {
  * integer `num_ctx` line is present.
  */
 declare function extractParameterNumCtx(parameters?: string | undefined): number | undefined;
+/**
+ * Whether `model` is currently resident (loaded in VRAM) per a `GET /api/ps`
+ * response — exact or tag-less-prefix match, same resolution rules as Ollama
+ * itself applies to model names. See {@link matchesModelReference}.
+ */
+declare function isModelResident(psModels: readonly {
+    readonly name?: string | undefined;
+    readonly model?: string | undefined;
+}[], model: string): boolean;
 /**
  * Finds the context window Ollama actually allocated for `model`, from a
  * `GET /api/ps` response. Matching mirrors Ollama's own model resolution:
@@ -357,9 +366,133 @@ declare const DEFAULT_RETRY_CONFIG: RetryConfig;
  */
 declare function withRetry<T>(operation: (attempt: number) => Promise<T>, config?: RetryConfig, signal?: AbortSignal): Promise<T>;
 
+/**
+ * Multi-host model-affinity routing — prefer the endpoint already holding the
+ * requested model in VRAM.
+ *
+ * Ollama daemons keep a bounded set of models resident (`OLLAMA_MAX_LOADED_MODELS`).
+ * With several interchangeable endpoints serving the same models, sending a request
+ * to a host that doesn't currently hold the model forces an unload/cold-load swap
+ * on that host — the thrashing `ModelAffinityScheduler` avoids for a *single*
+ * client's task queues. This router brings the same affinity to the *multi-host*
+ * request path: given the candidate endpoints `executeWithFailover` would try
+ * anyway, it reorders them so hosts whose `GET /api/ps` shows the requested model
+ * resident come first.
+ *
+ * Design constraints (see ADR 0028):
+ *
+ * - **Zero added request latency.** Reordering is synchronous and consults only
+ *   cached `/api/ps` snapshots. A request never waits for a probe: when a
+ *   snapshot is missing or stale, the request proceeds with the unchanged
+ *   candidate order and the refresh happens in the background, warming the cache
+ *   for subsequent requests.
+ * - **Race-free with `least-connections`.** The reorder runs entirely between the
+ *   registry's synchronous `candidates()`/`filterWithCapacity()` and the
+ *   endpoint's `acquire()` — no `await` in between — so the strategy's
+ *   no-two-calls-pick-the-same-idle-endpoint guarantee is untouched. Affinity
+ *   only reorders *within* the runnable set; priority tiers and per-strategy
+ *   ordering stay intact inside the resident/non-resident groups.
+ * - **Best-effort, never blocking, never failing a request.** A probe failure is
+ *   recorded with a short backoff (`failureRetryMs`) and never cached as data;
+ *   the affected endpoint is simply treated as unknown until a probe succeeds.
+ * - **Fresh-data-only decisions.** Only snapshots within `ttlMs` reorder
+ *   requests — `/api/ps` can change at any moment (evictions, other clients),
+ *   so stale data is treated as no data. Last-known snapshots are still exposed
+ *   via `status()` for observability.
+ */
+
+/**
+ * One `/api/ps` running-model entry, as the router consumes it — name/model only.
+ * Optional to match the generated `Ps` wire shape; hand-written `ModelResponse`
+ * values (required fields) are assignable as-is.
+ */
+interface AffinityRunningModel {
+    readonly name?: string | undefined;
+    readonly model?: string | undefined;
+}
+/**
+ * Options for dynamic model-affinity routing across endpoints. Configured under
+ * `OllamaClientConfig.endpointHealth.modelAffinity`; implemented by
+ * `OllamaClient`'s request path (`executeWithFailover`), which has HTTP access —
+ * the `EndpointRegistry` itself ignores this key.
+ */
+interface ModelAffinityRoutingOptions {
+    /**
+     * How long a `GET /api/ps` snapshot stays fresh enough to reorder requests
+     * (default 30_000ms). Mirrors `ModelAffinityScheduler`'s `psCacheTtlMs`.
+     */
+    readonly ttlMs?: number | undefined;
+    /**
+     * How long a failed `/api/ps` probe suppresses retries against that endpoint
+     * (default 5_000ms) — a dead host isn't probed on every request while it's
+     * down, and comes back automatically once it recovers. Failures are never
+     * cached as data.
+     */
+    readonly failureRetryMs?: number | undefined;
+}
+/** Introspection snapshot for one endpoint's affinity cache. */
+interface ModelAffinitySnapshot {
+    readonly endpointName: string;
+    readonly baseUrl: string;
+    /** Model names reported resident by the last successful probe (may be stale). */
+    readonly loadedModels: readonly string[];
+    /** When the last successful probe completed — `undefined` if it never succeeded. */
+    readonly fetchedAt: number | undefined;
+    /** Whether the snapshot is within `ttlMs` and therefore still drives reordering. */
+    readonly fresh: boolean;
+}
+/** Test seam: injectable clock — defaults to `Date.now`. */
+type NowFn = () => number;
+/**
+ * Cache-and-reorder engine for multi-host model affinity. See the module docs for
+ * the design constraints; `OllamaClient` constructs one (only when
+ * `endpointHealth.modelAffinity` is set) and calls {@link reorder} synchronously
+ * inside its request path.
+ */
+declare class ModelAffinityRouter {
+    private readonly fetchPs;
+    private readonly ttlMs;
+    private readonly failureRetryMs;
+    private readonly now;
+    private readonly cache;
+    private readonly failedAt;
+    private readonly inflight;
+    constructor(fetchPs: (endpoint: OllamaEndpoint) => Promise<readonly AffinityRunningModel[]>, options?: ModelAffinityRoutingOptions & {
+        readonly now?: NowFn;
+    });
+    /**
+     * Synchronously reorders `candidates` so endpoints whose fresh `/api/ps`
+     * snapshot shows `model` resident come first (relative order preserved within
+     * each group). Returns the input unchanged — and kicks a background refresh —
+     * when any candidate's snapshot is missing or stale: a wrong reorder is worse
+     * than no reorder, and the first request after a cache miss simply isn't
+     * reordered. No-op without a `model` or with fewer than two candidates.
+     */
+    reorder(candidates: readonly OllamaEndpoint[], model: string | undefined): readonly OllamaEndpoint[];
+    /**
+     * Forces a refresh of the affinity cache for the given endpoints (deduped with
+     * any in-flight refresh; bypasses the failure backoff, because this is an
+     * explicit caller request — `client.warmModelAffinity()`). Resolves once every
+     * probe has settled — successfully or not; individual failures never reject.
+     */
+    warm(endpoints: readonly OllamaEndpoint[]): Promise<void>;
+    /** Introspection snapshots for the given endpoints (defaults to cached ones). */
+    status(endpoints?: readonly OllamaEndpoint[]): readonly ModelAffinitySnapshot[];
+    /**
+     * Starts (or joins) a background refresh for `endpoint`, unless one is already
+     * in flight or the failure backoff suppresses it. Fire-and-forget: the promise
+     * is stored only for dedupe/warm-joins; callers never await it on the request
+     * path.
+     */
+    private kickRefresh;
+    /** Refreshes one endpoint's snapshot; never rejects (failures record a backoff). */
+    private refresh;
+}
+
 /**
  * Multi-endpoint registry with priority routing and circuit breaker failover.
  */
+
 interface OllamaEndpoint {
     readonly name: string;
     readonly baseUrl: string;
@@ -427,6 +560,18 @@ interface EndpointRegistryOptions {
      * any other in-flight request — it never blocks indefinitely on its own.
      */
     readonly maxConcurrentPerEndpoint?: number | undefined;
+    /**
+     * Dynamic model-affinity routing across endpoints: reorder each request's
+     * candidates so endpoints whose `GET /api/ps` shows the requested model already
+     * resident in VRAM are tried first — avoiding unload/cold-load swaps on
+     * `OLLAMA_MAX_LOADED_MODELS`-bounded daemons. Off unless set.
+     *
+     * Implemented by `OllamaClient`'s request path (`executeWithFailover` + a
+     * TTL-cached `ModelAffinityRouter`), not by this registry — the registry is a
+     * pure, HTTP-free data structure and simply ignores this key. See
+     * `src/providers/model-affinity-router.ts` and ADR 0028.
+     */
+    readonly modelAffinity?: ModelAffinityRoutingOptions | undefined;
 }
 declare class EndpointRegistry {
     private readonly endpoints;
@@ -1867,6 +2012,16 @@ declare class OllamaClient {
      * (streams, via `executeWithFailover`'s `holdUntil` release path).
      */
     private readonly activeRequests;
+    /**
+     * Dynamic multi-host model-affinity routing, active only when
+     * `config.endpointHealth.modelAffinity` is set — see
+     * `src/providers/model-affinity-router.ts` and ADR 0028. `undefined` means
+     * the request path skips reordering entirely (zero overhead, and no `/api/ps`
+     * probes are ever issued).
+     */
+    private readonly modelAffinity;
+    /** Default timeout for one-off probe requests (`ping`, affinity `/api/ps` refreshes). */
+    private static readonly PROBE_TIMEOUT_MS;
     constructor(config?: OllamaClientConfig);
     get modelsClient(): ModelsClient;
     /**
@@ -2200,7 +2355,49 @@ declare class OllamaClient {
     capabilities(model: string, signal?: AbortSignal): Promise<ModelCapabilities>;
     runtimeMode(): RuntimeMode;
     healthCheck(): Promise<EndpointHealthCheckResult[]>;
+    /**
+     * Ultra-lightweight liveness probe against the client's current best
+     * endpoint: `HEAD /` — the root ping Ollama's documentation defines for
+     * load-balancer health checks (HTTP 200; the daemon answers `GET /` with the
+     * plain-text body "Ollama is running", which a HEAD request doesn't transfer).
+     *
+     * Deliberately a *single-endpoint, single-shot* probe — no failover to other
+     * candidates, no retry — so the answer reflects one concrete host, not "some
+     * endpoint eventually answered". To audit every configured endpoint, use
+     * {@link healthCheck} (which probes all of them, and also reports the server
+     * version at the cost of the heavier `GET /api/version` round trip).
+     *
+     * Resolves — never throws — with `{ healthy: false, error }` for connection
+     * failures and timeouts; only caller-initiated cancellation (`signal`/
+     * `destroy()`) rejects, mirroring how any other aborted request behaves.
+     */
+    ping(options?: RequestCancellationOptions): Promise<PingResult>;
+    /**
+     * Affinity-cache snapshots for every configured endpoint — which models each
+     * host held in VRAM at the last successful `/api/ps` probe, when that was, and
+     * whether it's still fresh. Empty when `endpointHealth.modelAffinity` isn't
+     * configured (no probes are ever issued, so there is nothing to report).
+     */
+    modelAffinityStatus(): readonly ModelAffinitySnapshot[];
+    /**
+     * Explicitly refreshes the model-affinity cache for every configured endpoint
+     * (deduped with any in-flight background refresh). No-op — resolves
+     * immediately — when `endpointHealth.modelAffinity` isn't configured. Use it
+     * to warm the cache at startup so the very first request already routes by
+     * residency instead of waiting for the self-warming first-round refresh.
+     */
+    warmModelAffinity(): Promise<void>;
     endpointStatus(): EndpointHealth[];
+    /**
+     * Fetches one endpoint's `GET /api/ps` running-model list for the affinity
+     * router. Bypasses `executeWithFailover` deliberately: probes must not
+     * reenter the routing machinery they inform, must not touch the registry's
+     * circuit-breaker/failure bookkeeping (a dead host's probe failing must not
+     * cool it down harder than its real traffic already does), and must not fail
+     * over to a different host (the answer is per-host by definition). Failures
+     * propagate to the router, which records a short backoff and retries later.
+     */
+    private fetchEndpointRunningModels;
     /**
      * Immediately aborts every in-flight request and active stream managed by
      * this client — the clean teardown path for short-lived processes, CLI
@@ -3066,6 +3263,93 @@ declare const ATTR_OLLAMA_AGENT_MAX_ITERATIONS = "ollama.agent.max_iterations";
 declare const ATTR_OLLAMA_AGENT_ITERATION = "ollama.agent.iteration";
 declare const GEN_AI_SYSTEM_OLLAMA = "ollama";
 
+/**
+ * Standardized telemetry normalizer — raw nanosecond counters to derived metrics.
+ *
+ * Ollama reports every duration as an **integer in nanoseconds** on its generation
+ * responses (`total_duration`, `load_duration`, `prompt_eval_duration`,
+ * `eval_duration` — see the official "Usage" documentation), plus token counters
+ * (`prompt_eval_count`, `prompt_eval_cached_count`, `eval_count`). Consumers that
+ * want milliseconds, tokens/second, or a cache-hit ratio otherwise each reimplement
+ * the same division-and-rounding math — this module is that math, once, with the
+ * formulas the official docs spell out:
+ *
+ * - `tokensPerSecond` = `eval_count / (eval_duration / 1e9)` — generation throughput.
+ * - `promptTokensPerSecond` = `prompt_eval_count / (prompt_eval_duration / 1e9)` —
+ *   prompt ingestion (pre-fill) throughput.
+ * - `cacheHitRatio` = `prompt_eval_cached_count / (cached + evaluated)` — see the
+ *   note on {@link FormattedTelemetry.cacheHitRatio} for why the denominator is the
+ *   *sum*, not `prompt_eval_count` alone.
+ *
+ * Pure and dependency-free: {@link formatTelemetry} never fetches, never throws, and
+ * treats every missing counter as `0` — partial responses (stream final events,
+ * older servers without `prompt_eval_cached_count`) format cleanly instead of
+ * producing `NaN`/`Infinity`.
+ *
+ * The input is structural: any response object carrying these optional fields —
+ * `GenerateResponse`, `ChatResponse`, the final `GenerateStreamEvent`,
+ * `EmbedResponse` — can be passed directly, no picking/plucking required.
+ */
+/**
+ * The raw nanosecond/token counters Ollama attaches to generation responses. Every
+ * field is optional because servers report a subset: streaming final events may omit
+ * load counters, older builds predate `prompt_eval_cached_count`, and embeddings
+ * carry no `eval_count`. Accepts full response objects structurally.
+ */
+interface RawOllamaMetrics {
+    /** Total end-to-end request latency, nanoseconds. */
+    readonly total_duration?: number | undefined;
+    /** Model cold-load and weight-transfer time into VRAM, nanoseconds. */
+    readonly load_duration?: number | undefined;
+    /** Prompt tokens freshly evaluated this request (not served from the KV cache). */
+    readonly prompt_eval_count?: number | undefined;
+    /** Prompt tokens served from the KV prompt cache (prefix reuse). */
+    readonly prompt_eval_cached_count?: number | undefined;
+    /** Prompt ingestion / pre-fill processing time, nanoseconds. */
+    readonly prompt_eval_duration?: number | undefined;
+    /** Tokens generated (the response proper). */
+    readonly eval_count?: number | undefined;
+    /** Token generation time, nanoseconds. */
+    readonly eval_duration?: number | undefined;
+}
+/** The normalized, human-scale view of one response's {@link RawOllamaMetrics}. */
+interface FormattedTelemetry {
+    /** `total_duration` in milliseconds, 2 decimals. */
+    readonly totalLatencyMs: number;
+    /** `load_duration` in milliseconds, 2 decimals — 0 when the model was already resident. */
+    readonly modelLoadMs: number;
+    /** `prompt_eval_duration` in milliseconds, 2 decimals. */
+    readonly promptEvalMs: number;
+    /** `eval_duration` in milliseconds, 2 decimals. */
+    readonly generationMs: number;
+    /** `eval_count / eval_duration` in tokens/second, 1 decimal — 0 when either is absent. */
+    readonly tokensPerSecond: number;
+    /** `prompt_eval_count / prompt_eval_duration` in tokens/second, 1 decimal. */
+    readonly promptTokensPerSecond: number;
+    /**
+     * Share of prompt tokens served from the KV cache:
+     * `cached / (cached + evaluated)`, 3 decimals.
+     *
+     * The denominator is the **sum**, not `prompt_eval_count` alone, because Ollama
+     * reports `prompt_eval_count: 0` on a *full* cache hit — the same house semantics
+     * as `ConversationSession`'s `cacheStats.hitRate` (see `src/conversation.ts`). A
+     * `cached / prompt_eval_count` formula would report a perfect cache as `0`.
+     */
+    readonly cacheHitRatio: number;
+}
+/**
+ * Derives the standardized metrics summary from one response's raw nanosecond
+ * counters. Pure — never throws, never fetches; every missing counter reads as `0`.
+ *
+ * ```ts
+ * const res = await client.chat({ model: 'llama3.2', messages });
+ * const t = formatTelemetry(res);
+ * console.log(t.tokensPerSecond, 'tok/s |', t.totalLatencyMs, 'ms |',
+ *   `${Math.round(t.cacheHitRatio * 100)}% prompt cache hit`);
+ * ```
+ */
+declare function formatTelemetry(raw: RawOllamaMetrics): FormattedTelemetry;
+
 /**
  * Cross-runtime helpers with no Node.js API dependency, so they stay safe to import from
  * the core client bundle (see ADR 0006 — Edge Runtime CI).
@@ -3340,4 +3624,4 @@ declare class ModelAffinityScheduler {
 /** Convenience factory — `createModelAffinityScheduler(client, opts)` reads better in setups. */
 declare function createModelAffinityScheduler(client: ModelAffinityClient | OllamaClient, options?: ModelAffinitySchedulerOptions): ModelAffinityScheduler;
 
-export { ATTR_GEN_AI_OPERATION_NAME, ATTR_GEN_AI_REQUEST_MODEL, ATTR_GEN_AI_RESPONSE_MODEL, ATTR_GEN_AI_SYSTEM, ATTR_GEN_AI_TOOL_CALL_ID, ATTR_GEN_AI_TOOL_NAME, ATTR_GEN_AI_USAGE_INPUT_TOKENS, ATTR_GEN_AI_USAGE_OUTPUT_TOKENS, ATTR_HTTP_REQUEST_METHOD, ATTR_HTTP_RESPONSE_STATUS_CODE, ATTR_OLLAMA_AGENT_ITERATION, ATTR_OLLAMA_AGENT_MAX_ITERATIONS, ATTR_OLLAMA_ENDPOINT_ATTEMPT, ATTR_OLLAMA_ENDPOINT_NAME, ATTR_SERVER_ADDRESS, ATTR_SERVER_PORT, ATTR_URL_FULL, AbortableAsyncIterable, Agent, type AgentChatClient, type AgentConfig, type AgentHooks, type AgentResult, type AgentRunInput, type AgentTurn, type AnswerFor, AnthropicCompatClient, AnyTool, type BackoffOptions, BalanceRequestOptions, type BalanceResponse, type BlobUploadResult, ChatRequestOptions, ChatResponse, ChatStreamResult, type ChoiceDecision, type CompactionOptions, type ContextCheck, type ContextCheckOptions, type ContextDiscoveryRequestOptions, type ContextLengthSignals, type ContextLengthSource, type ContextWindowSource, type ConversationSendOptions, ConversationSession, type ConversationSessionOptions, CopyRequestOptions, CreateRequestOptions, type CumulativeCacheStats, DEFAULT_BACKOFF, DEFAULT_BASE_URL, DEFAULT_CONTEXT_WARNING_THRESHOLD, DEFAULT_EMBED_BATCH_SIZE, DEFAULT_EMBED_CONCURRENCY, DEFAULT_FAILOVER_CODES, DEFAULT_FETCH_BACKOFF_CONFIG, DEFAULT_RETRY_CONFIG, DEFAULT_TIMEOUT_MS, type Decision, type DefineToolOptions, DeleteRequestOptions, type DiscoveredContextLength, type EmbedBatchOptions, type EmbedBatchResult, EmbedRequestOptions, EmbedResponse, EmbeddingsRequestOptions, EmbeddingsResponse, type EndpointHealth, type EndpointHealthCheckResult, EndpointRegistry, type EndpointRegistryOptions, FetchLike, type FetchWithBackoffConfig, FormatOption, GEN_AI_SYSTEM_OLLAMA, GenerateRequestOptions, GenerateResponse, GenerateStreamResult, type HistorySanitizeOptions, HttpClient, IMAGE_TOKEN_ESTIMATE, type IncludedBalance, KEEP_ALIVE_INDEFINITE, KEEP_ALIVE_UNLOAD, KNOWN_DONE_REASONS, type KeepAlive, type KnownDoneReason, type LegacyBalanceLimit, type LegacyIncludedBalance, type LoadMcpToolsOptions, Logger, MAX_SYSTEM_ONE_IMAGES_BYTES, MAX_SYSTEM_ONE_REQUEST_BYTES, MESSAGE_OVERHEAD_TOKENS, McpBridge, type McpBridgeOptions, McpCallToolResult, McpClientLike, McpTask, McpToolDescriptor, Message, Middleware, type ModelAffinityClient, ModelAffinityScheduler, type ModelAffinitySchedulerOptions, type ModelAffinityStats, type ModelCapabilities, ModelOptions, ModelResponse, ModelsClient, type NoulDecision, OLLAMA_CLOUD_BASE_URL, OLLAMA_FALLBACK_CONTEXT_LENGTH, OllamaClient, type OllamaClientConfig, type OllamaCloudFreeTierQuotaBudgets, type OllamaCredentialConfig, type OllamaEndpoint, OllamaStream, OllamaStreamEvent, OllamaStreamEventType, OpenAICompatClient, type OversizedEmbedInput, type ParallelToolCallBehavior, ProgressResponse, ProgressStreamResult, PsResponse, PullRequestOptions, PushRequestOptions, QuotaManager, type QuotaManagerOptions, type QuotaWindowConfig, type QuotaWindowStatus, RETRYABLE_STATUS_CODES, type RankResult, RawUsageSource, RequestCancellationOptions, RequestLifecycleHook, RequestRunner, type ResponsesCreateRequest, type ResponsesCreateResponse, ResponsesModule, type ResponsesReasoningEffort, type ResponsesStreamEvent, type ResponsesToolDefinition, type ResponsesUsage, type RetryConfig, type RuntimeMode, type ScoreDecision, type SessionCompactionOptions, type SessionCompactionResult, type SessionTurn, ShowRequestOptions, ShowResponse, type SpanAttributeValue, type SpanAttributes, StatusResponse, type SupportedSchema, type SystemOneAnswer, type SystemOneAnswers, SystemOneChoiceAnswer, SystemOneChoiceQuestion, SystemOneContent, SystemOneNoulAnswer, SystemOneNoulQuestion, type SystemOneQuestion, type SystemOneQuestions, type SystemOneRequest, type SystemOneResponse, SystemOneScoreAnswer, SystemOneScoreQuestion, type SystemOneUsage, ThinkValue, ThinkingMetadata, type TimeoutSignal, TokenUsage, Tool, ToolCall, ToolCallAccumulator, ToolDefinition, ToolExecutionResult, ToolHandler, ToolParameters, ToolRegistry, type TurnCacheStats, type UsageBucket, type UsageMetrics, UsageRequestOptions, type UsageResponse, VersionResponse, VisionInput, WebFetchRequestOptions, WebFetchResponse, WebSearchRequestOptions, WebSearchResponse, addSpanEvent, batchEmbed, calculateBackoff, canonicalToolCallSignature, checkChatContext, checkEndpointHealth, checkGenerateContext, compactConversationHistory, contextWarningMessage, createDecision, createModelAffinityScheduler, createOllamaCloudFreeTierQuota, createTimeoutSignal, defineTool, detectModelCapabilities, embedBatchOverflowMessage, encodeImage, estimateChatRequestTokens, estimateGenerateRequestTokens, estimateMessageTokens, estimateSystemOneRequestBytes, estimateTokens, extractContextLength as extractNativeContextLength, extractParameterNumCtx, fetchWithBackoff, findOversizedEmbedInputs, findRunningModelContextLength, inferRuntimeMode, isKeepAliveSugar, isKnownDoneReason, isSameToolCall, listAllMcpTools, listAvailableModels, loadMcpTools, mergeToolCall, mergeToolCallArgumentsString, mergeToolCallArrays, normalizeChatStream, normalizeGenerateStream, normalizeKeepAlive, normalizeProgressStream, parseNdjsonStream, parseStructuredOutput, registerMcpTools, resolveContextLength, resolveCredentialEndpoints, sanitizeHistoryForNextTurn, setSpanError, toDataStream, toResponse, toTextStream, withRetry, withSpan, zodToJsonSchema };
+export { ATTR_GEN_AI_OPERATION_NAME, ATTR_GEN_AI_REQUEST_MODEL, ATTR_GEN_AI_RESPONSE_MODEL, ATTR_GEN_AI_SYSTEM, ATTR_GEN_AI_TOOL_CALL_ID, ATTR_GEN_AI_TOOL_NAME, ATTR_GEN_AI_USAGE_INPUT_TOKENS, ATTR_GEN_AI_USAGE_OUTPUT_TOKENS, ATTR_HTTP_REQUEST_METHOD, ATTR_HTTP_RESPONSE_STATUS_CODE, ATTR_OLLAMA_AGENT_ITERATION, ATTR_OLLAMA_AGENT_MAX_ITERATIONS, ATTR_OLLAMA_ENDPOINT_ATTEMPT, ATTR_OLLAMA_ENDPOINT_NAME, ATTR_SERVER_ADDRESS, ATTR_SERVER_PORT, ATTR_URL_FULL, AbortableAsyncIterable, type AffinityRunningModel, Agent, type AgentChatClient, type AgentConfig, type AgentHooks, type AgentResult, type AgentRunInput, type AgentTurn, type AnswerFor, AnthropicCompatClient, AnyTool, type BackoffOptions, BalanceRequestOptions, type BalanceResponse, type BlobUploadResult, ChatRequestOptions, ChatResponse, ChatStreamResult, type ChoiceDecision, type CompactionOptions, type ContextCheck, type ContextCheckOptions, type ContextDiscoveryRequestOptions, type ContextLengthSignals, type ContextLengthSource, type ContextWindowSource, type ConversationSendOptions, ConversationSession, type ConversationSessionOptions, CopyRequestOptions, CreateRequestOptions, type CumulativeCacheStats, DEFAULT_BACKOFF, DEFAULT_BASE_URL, DEFAULT_CONTEXT_WARNING_THRESHOLD, DEFAULT_EMBED_BATCH_SIZE, DEFAULT_EMBED_CONCURRENCY, DEFAULT_FAILOVER_CODES, DEFAULT_FETCH_BACKOFF_CONFIG, DEFAULT_RETRY_CONFIG, DEFAULT_TIMEOUT_MS, type Decision, type DefineToolOptions, DeleteRequestOptions, type DiscoveredContextLength, type EmbedBatchOptions, type EmbedBatchResult, EmbedRequestOptions, EmbedResponse, EmbeddingsRequestOptions, EmbeddingsResponse, type EndpointHealth, type EndpointHealthCheckResult, EndpointRegistry, type EndpointRegistryOptions, FetchLike, type FetchWithBackoffConfig, FormatOption, type FormattedTelemetry, GEN_AI_SYSTEM_OLLAMA, GenerateRequestOptions, GenerateResponse, GenerateStreamResult, type HistorySanitizeOptions, HttpClient, IMAGE_TOKEN_ESTIMATE, type IncludedBalance, KEEP_ALIVE_INDEFINITE, KEEP_ALIVE_UNLOAD, KNOWN_DONE_REASONS, type KeepAlive, type KnownDoneReason, type LegacyBalanceLimit, type LegacyIncludedBalance, type LoadMcpToolsOptions, Logger, MAX_SYSTEM_ONE_IMAGES_BYTES, MAX_SYSTEM_ONE_REQUEST_BYTES, MESSAGE_OVERHEAD_TOKENS, McpBridge, type McpBridgeOptions, McpCallToolResult, McpClientLike, McpTask, McpToolDescriptor, Message, Middleware, type ModelAffinityClient, ModelAffinityRouter, type ModelAffinityRoutingOptions, ModelAffinityScheduler, type ModelAffinitySchedulerOptions, type ModelAffinitySnapshot, type ModelAffinityStats, type ModelCapabilities, ModelOptions, ModelResponse, ModelsClient, type NoulDecision, OLLAMA_CLOUD_BASE_URL, OLLAMA_FALLBACK_CONTEXT_LENGTH, OllamaClient, type OllamaClientConfig, type OllamaCloudFreeTierQuotaBudgets, type OllamaCredentialConfig, type OllamaEndpoint, OllamaStream, OllamaStreamEvent, OllamaStreamEventType, OpenAICompatClient, type OversizedEmbedInput, type ParallelToolCallBehavior, PingResult, ProgressResponse, ProgressStreamResult, PsResponse, PullRequestOptions, PushRequestOptions, QuotaManager, type QuotaManagerOptions, type QuotaWindowConfig, type QuotaWindowStatus, RETRYABLE_STATUS_CODES, type RankResult, type RawOllamaMetrics, RawUsageSource, RequestCancellationOptions, RequestLifecycleHook, RequestRunner, type ResponsesCreateRequest, type ResponsesCreateResponse, ResponsesModule, type ResponsesReasoningEffort, type ResponsesStreamEvent, type ResponsesToolDefinition, type ResponsesUsage, type RetryConfig, type RuntimeMode, type ScoreDecision, type SessionCompactionOptions, type SessionCompactionResult, type SessionTurn, ShowRequestOptions, ShowResponse, type SpanAttributeValue, type SpanAttributes, StatusResponse, type SupportedSchema, type SystemOneAnswer, type SystemOneAnswers, SystemOneChoiceAnswer, SystemOneChoiceQuestion, SystemOneContent, SystemOneNoulAnswer, SystemOneNoulQuestion, type SystemOneQuestion, type SystemOneQuestions, type SystemOneRequest, type SystemOneResponse, SystemOneScoreAnswer, SystemOneScoreQuestion, type SystemOneUsage, ThinkValue, ThinkingMetadata, type TimeoutSignal, TokenUsage, Tool, ToolCall, ToolCallAccumulator, ToolDefinition, ToolExecutionResult, ToolHandler, ToolParameters, ToolRegistry, type TurnCacheStats, type UsageBucket, type UsageMetrics, UsageRequestOptions, type UsageResponse, VersionResponse, VisionInput, WebFetchRequestOptions, WebFetchResponse, WebSearchRequestOptions, WebSearchResponse, addSpanEvent, batchEmbed, calculateBackoff, canonicalToolCallSignature, checkChatContext, checkEndpointHealth, checkGenerateContext, compactConversationHistory, contextWarningMessage, createDecision, createModelAffinityScheduler, createOllamaCloudFreeTierQuota, createTimeoutSignal, defineTool, detectModelCapabilities, embedBatchOverflowMessage, encodeImage, estimateChatRequestTokens, estimateGenerateRequestTokens, estimateMessageTokens, estimateSystemOneRequestBytes, estimateTokens, extractContextLength as extractNativeContextLength, extractParameterNumCtx, fetchWithBackoff, findOversizedEmbedInputs, findRunningModelContextLength, formatTelemetry, inferRuntimeMode, isKeepAliveSugar, isKnownDoneReason, isModelResident, isSameToolCall, listAllMcpTools, listAvailableModels, loadMcpTools, mergeToolCall, mergeToolCallArgumentsString, mergeToolCallArrays, normalizeChatStream, normalizeGenerateStream, normalizeKeepAlive, normalizeProgressStream, parseNdjsonStream, parseStructuredOutput, registerMcpTools, resolveContextLength, resolveCredentialEndpoints, sanitizeHistoryForNextTurn, setSpanError, toDataStream, toResponse, toTextStream, withRetry, withSpan, zodToJsonSchema };