@nemesis-oss/ollama-sdk 1.19.0 → 1.20.0

package/dist/index.d.ts

@@ -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 PingResult, w as VisionInput } from './types-Y3u0FtdN.js';
-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-Y3u0FtdN.js';
-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-DYCxYGG-.js';
-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 NETWORK_ERROR_CODES, r as NOOP_LOGGER, s as NextFunction, t as OllamaAbortError, u as OllamaAgentMaxIterationsError, v as OllamaAgentMaxToolCallsError, w as OllamaAgentToolLoopError, x as OllamaAuthError, y as OllamaBadGatewayError, z as OllamaClientError, J as OllamaClientErrorOptions, K as OllamaErrorRequestContext, Q as OllamaErrorResponseContext, U as OllamaGenericClientError, V as OllamaIncompatibleModelError, W as OllamaMcpError, X as OllamaModelRoutingError, Y as OllamaNetworkError, Z as OllamaNotFoundError, _ as OllamaOverloadedError, $ as OllamaPaymentRequiredError, a0 as OllamaQuotaExceededError, a1 as OllamaRateLimitError, a2 as OllamaRequestTooLargeError, a3 as OllamaRequestValidationError, a4 as OllamaResponseValidationError, a5 as OllamaServerError, a6 as OllamaServerVersionUnknownError, a7 as OllamaSkillInvalidError, a8 as OllamaSkillNotFoundError, a9 as OllamaStreamError, aa as OllamaTimeoutError, ab as OllamaToolTimeoutError, ac as OllamaToolValidationError, ad as OllamaUnsupportedCapabilityError, ae as RequestContext, af as RequestLifecycleEvent, ag as ResponseContext, ah as SseEvent, ai as ThinkingEvent, aj as TokenEvent, ak as ToolCallEvent, al as composeMiddleware, am as createConsoleLogger, an as extractUsage, ao as isNetworkLevelError, ap as mapError, aq as parseSseStream } from './runtime-DYCxYGG-.js';
+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-CnsZ24tm.js';
+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 UsageGroupBy, N as UsageRange, O as UsageScope, Q as WebSearchResult, X as imageStringNeedsResolution, Y as resolveImageInput } from './types-CnsZ24tm.js';
+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-DPSrPwAa.js';
+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 NETWORK_ERROR_CODES, r as NOOP_LOGGER, s as NextFunction, t as OllamaAbortError, u as OllamaAgentMaxIterationsError, v as OllamaAgentMaxToolCallsError, w as OllamaAgentToolLoopError, x as OllamaAuthError, y as OllamaBadGatewayError, z as OllamaClientError, J as OllamaClientErrorOptions, K as OllamaCloudModelVersionError, Q as OllamaErrorRequestContext, U as OllamaErrorResponseContext, V as OllamaGenericClientError, W as OllamaIncompatibleModelError, X as OllamaMcpError, Y as OllamaModelRoutingError, Z as OllamaNetworkError, _ as OllamaNotFoundError, $ as OllamaOverloadedError, a0 as OllamaPaymentRequiredError, a1 as OllamaQuotaExceededError, a2 as OllamaRateLimitError, a3 as OllamaRequestTooLargeError, a4 as OllamaRequestValidationError, a5 as OllamaResponseValidationError, a6 as OllamaServerError, a7 as OllamaServerVersionUnknownError, a8 as OllamaSkillInvalidError, a9 as OllamaSkillNotFoundError, aa as OllamaStreamError, ab as OllamaTimeoutError, ac as OllamaToolTimeoutError, ad as OllamaToolValidationError, ae as OllamaUnsupportedCapabilityError, af as RequestContext, ag as RequestLifecycleEvent, ah as ResponseContext, ai as SseEvent, aj as ThinkingEvent, ak as TokenEvent, al as ToolCallEvent, am as composeMiddleware, an as createConsoleLogger, ao as extractUsage, ap as isNetworkLevelError, aq as mapError, ar as parseRetryAfterMs, as as parseSseStream } from './runtime-DPSrPwAa.js';
 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-CAvwqIjx.js';
-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 OllamaOpenAIChatContentPart, a4 as OllamaOpenAIChatMessage, a5 as OllamaOpenAICompletionRequest, a6 as OllamaOpenAIEmbeddingRequest, a7 as OllamaOpenAIImageDataUrl, a8 as OllamaOpenAIReasoningEffort, a9 as OllamaOpenAIResponsesRequest, aa as OpenAIChatCompletionChoice, b as OpenAIChatCompletionChunk, ab as OpenAIChatCompletionChunkChoice, ac as OpenAIChatCompletionDelta, O as OpenAIChatCompletionRequest, a as OpenAIChatCompletionResponse, ad as OpenAIChatCompletionStream, ae as OpenAICompletionChoice, e as OpenAICompletionChunk, c as OpenAICompletionRequest, d as OpenAICompletionResponse, af as OpenAICompletionStream, ag as OpenAIContentPart, ah as OpenAIEmbeddingInput, ai as OpenAIEmbeddingItem, f as OpenAIEmbeddingRequest, g as OpenAIEmbeddingResponse, aj as OpenAIFunctionDefinition, ak as OpenAIImageUrlContentPart, h as OpenAIListModelsResponse, al as OpenAIMessage, i as OpenAIModelItem, am as OpenAIReasoningEffort, an as OpenAIResponseFormat, ao as OpenAIResponsesCompletedEvent, ap as OpenAIResponsesContentPartAddedEvent, aq as OpenAIResponsesContentPartDoneEvent, ar as OpenAIResponsesCreatedEvent, as as OpenAIResponsesError, at as OpenAIResponsesEventBase, au as OpenAIResponsesFailedEvent, av as OpenAIResponsesFunctionCallArgumentsDeltaEvent, aw as OpenAIResponsesFunctionCallArgumentsDoneEvent, ax as OpenAIResponsesInProgressEvent, ay as OpenAIResponsesIncompleteDetails, az as OpenAIResponsesIncompleteEvent, aA as OpenAIResponsesOutputContent, aB as OpenAIResponsesOutputFunctionCall, aC as OpenAIResponsesOutputItem, aD as OpenAIResponsesOutputItemAddedEvent, aE as OpenAIResponsesOutputItemDoneEvent, aF as OpenAIResponsesOutputMessage, aG as OpenAIResponsesOutputReasoning, aH as OpenAIResponsesOutputRefusal, aI as OpenAIResponsesOutputTextContent, aJ as OpenAIResponsesOutputTextDeltaEvent, aK as OpenAIResponsesOutputTextDoneEvent, aL as OpenAIResponsesQueuedEvent, aM as OpenAIResponsesReasoningSummaryPartAddedEvent, aN as OpenAIResponsesReasoningSummaryPartDoneEvent, aO as OpenAIResponsesReasoningSummaryTextDeltaEvent, aP as OpenAIResponsesReasoningSummaryTextDoneEvent, aQ as OpenAIResponsesReasoningTextDeltaEvent, aR as OpenAIResponsesReasoningTextDoneEvent, aS as OpenAIResponsesRefusalDeltaEvent, aT as OpenAIResponsesRefusalDoneEvent, j as OpenAIResponsesRequest, k as OpenAIResponsesResponse, aU as OpenAIResponsesStatus, aV as OpenAIResponsesStream, aW as OpenAIResponsesStreamError, l as OpenAIResponsesStreamEvent, aX as OpenAIResponsesUsage, aY as OpenAIStreamOptions, aZ as OpenAITextContentPart, a_ as OpenAITool, a$ as OpenAIToolCall, b0 as OpenAIToolCallDelta, b1 as OpenAIUsage, b2 as SystemOneConfidence, b3 as SystemOneProbabilities } from './anthropic-CAvwqIjx.js';
-import { T as ToolHandler, a as Tool, A as AnyTool, b as ToolRegistry, c as ToolExecutionResult } from './compose-75UFDzOO.js';
-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-75UFDzOO.js';
+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-BFOC6eI7.js';
+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 OllamaOpenAIChatContentPart, a4 as OllamaOpenAIChatMessage, a5 as OllamaOpenAICompletionRequest, a6 as OllamaOpenAIEmbeddingRequest, a7 as OllamaOpenAIImageDataUrl, a8 as OllamaOpenAIReasoningEffort, a9 as OllamaOpenAIResponsesRequest, aa as OpenAIChatCompletionChoice, b as OpenAIChatCompletionChunk, ab as OpenAIChatCompletionChunkChoice, ac as OpenAIChatCompletionDelta, O as OpenAIChatCompletionRequest, a as OpenAIChatCompletionResponse, ad as OpenAIChatCompletionStream, ae as OpenAICompletionChoice, e as OpenAICompletionChunk, c as OpenAICompletionRequest, d as OpenAICompletionResponse, af as OpenAICompletionStream, ag as OpenAIContentPart, ah as OpenAIEmbeddingInput, ai as OpenAIEmbeddingItem, f as OpenAIEmbeddingRequest, g as OpenAIEmbeddingResponse, aj as OpenAIFunctionDefinition, ak as OpenAIImageUrlContentPart, h as OpenAIListModelsResponse, al as OpenAIMessage, i as OpenAIModelItem, am as OpenAIReasoningEffort, an as OpenAIResponseFormat, ao as OpenAIResponsesCompletedEvent, ap as OpenAIResponsesContentPartAddedEvent, aq as OpenAIResponsesContentPartDoneEvent, ar as OpenAIResponsesCreatedEvent, as as OpenAIResponsesError, at as OpenAIResponsesEventBase, au as OpenAIResponsesFailedEvent, av as OpenAIResponsesFunctionCallArgumentsDeltaEvent, aw as OpenAIResponsesFunctionCallArgumentsDoneEvent, ax as OpenAIResponsesInProgressEvent, ay as OpenAIResponsesIncompleteDetails, az as OpenAIResponsesIncompleteEvent, aA as OpenAIResponsesOutputContent, aB as OpenAIResponsesOutputFunctionCall, aC as OpenAIResponsesOutputItem, aD as OpenAIResponsesOutputItemAddedEvent, aE as OpenAIResponsesOutputItemDoneEvent, aF as OpenAIResponsesOutputMessage, aG as OpenAIResponsesOutputReasoning, aH as OpenAIResponsesOutputRefusal, aI as OpenAIResponsesOutputTextContent, aJ as OpenAIResponsesOutputTextDeltaEvent, aK as OpenAIResponsesOutputTextDoneEvent, aL as OpenAIResponsesQueuedEvent, aM as OpenAIResponsesReasoningSummaryPartAddedEvent, aN as OpenAIResponsesReasoningSummaryPartDoneEvent, aO as OpenAIResponsesReasoningSummaryTextDeltaEvent, aP as OpenAIResponsesReasoningSummaryTextDoneEvent, aQ as OpenAIResponsesReasoningTextDeltaEvent, aR as OpenAIResponsesReasoningTextDoneEvent, aS as OpenAIResponsesRefusalDeltaEvent, aT as OpenAIResponsesRefusalDoneEvent, j as OpenAIResponsesRequest, k as OpenAIResponsesResponse, aU as OpenAIResponsesStatus, aV as OpenAIResponsesStream, aW as OpenAIResponsesStreamError, l as OpenAIResponsesStreamEvent, aX as OpenAIResponsesUsage, aY as OpenAIStreamOptions, aZ as OpenAITextContentPart, a_ as OpenAITool, a$ as OpenAIToolCall, b0 as OpenAIToolCallDelta, b1 as OpenAIUsage, b2 as SystemOneConfidence, b3 as SystemOneProbabilities } from './anthropic-BFOC6eI7.js';
+import { T as ToolHandler, a as Tool, A as AnyTool, b as ToolRegistry, c as ToolExecutionResult } from './compose-DaG83yId.js';
+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-DaG83yId.js';
 import { M as McpClientLike, a as McpToolDescriptor, b as McpTask, c as McpCallToolResult } from './types-CI-gM2u5.js';
 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.js';
 import * as Otel from '@opentelemetry/api';
@@ -72,7 +72,7 @@ interface LegacyIncludedBalance {
  * See ADR 0013 and ADR 0014 for the contract-first architecture.
  */
 
-type UsageBucket = UsageMetrics & {
+type UsageBucket = UsageSummary & {
     /** Start of the bucket in UTC, inclusive. */
     readonly from: string;
     /** End of the bucket in UTC, exclusive. */
@@ -81,6 +81,17 @@ type UsageBucket = UsageMetrics & {
     readonly partial?: boolean | undefined;
 };
 
+/**
+ * AUTO-GENERATED by `npm run contract:generate`. DO NOT EDIT.
+ * Source: contracts/ir/ollama.ir.json (canonical Ollama IR).
+ * See ADR 0013 and ADR 0014 for the contract-first architecture.
+ */
+
+type UsageGroup = UsageMetrics & {
+    /** Model name. */
+    readonly model: string;
+};
+
 /**
  * AUTO-GENERATED by `npm run contract:generate`. DO NOT EDIT.
  * Source: contracts/ir/ollama.ir.json (canonical Ollama IR).
@@ -112,16 +123,31 @@ interface UsageResponse {
     readonly scope: 'self' | 'team';
     /** Set automatically: hour for 24h, day for 7d and 30d. */
     readonly granularity: 'hour' | 'day';
+    /** Selected grouping. Present when group_by is requested. */
+    readonly group_by?: 'model' | undefined;
     /** Start of the range in UTC, inclusive. */
     readonly from: string;
     /** End of the range in UTC, exclusive. */
     readonly until: string;
-    /** Usage across the entire range. */
-    readonly totals: UsageMetrics;
+    /** Usage across the entire range, including groups in the remainder. */
+    readonly totals: UsageSummary;
     /** Usage in chronological order, including buckets with no recorded requests. */
     readonly buckets: readonly UsageBucket[];
 }
 
+/**
+ * AUTO-GENERATED by `npm run contract:generate`. DO NOT EDIT.
+ * Source: contracts/ir/ollama.ir.json (canonical Ollama IR).
+ * See ADR 0013 and ADR 0014 for the contract-first architecture.
+ */
+
+type UsageSummary = UsageMetrics & {
+    /** Present when group_by is requested. Currently returns up to 10 groups selected across the entire range, ranked by usage_usd or by request_count when the range includes legacy requests. Groups use the same order in every bucket; groups without requests in a bucket are omitted. */
+    readonly groups?: readonly UsageGroup[] | undefined;
+    /** Combined usage from groups outside the current limit. Omitted when there are no requests from additional groups. */
+    readonly remainder?: UsageMetrics | undefined;
+};
+
 /**
  * Model capability detection and runtime environment inference.
  */
@@ -219,7 +245,10 @@ declare function listAvailableModels(http: HttpClient): Promise<ModelResponse[]>
  * Dynamic context-window discovery.
  *
  * Guessing context windows ("it's probably 2048… or 4096?") is how silent
- * truncation happens. Ollama advertises the real numbers in two places, and
+ * truncation happens — and current daemons make the guess worse, not better:
+ * the unset-`num_ctx` default is tiered by VRAM (4k / 32k / 256k), so a
+ * hardcoded assumption is now wrong in both directions. Ollama advertises
+ * the real numbers in two places, and
  * they answer *different questions*:
  *
  *   - **`GET /api/ps`** reports `context_length` for every currently-running
@@ -262,14 +291,14 @@ type ContextLengthSource =
  | 'parameters'
 /** `POST /api/show` `model_info["*.context_length"]` — native GGUF maximum. */
  | 'model-info'
-/** Nothing discoverable — Ollama's conservative unset-`num_ctx` default (2048). */
+/** Nothing discoverable — the 2048 worst-case floor for the unset-`num_ctx` window (current daemons tier by VRAM: 4k / 32k / 256k). */
  | 'fallback';
 /** Result of {@link ModelsClient.getContextLength}. */
 interface DiscoveredContextLength {
     /**
      * The effective context window to program against, resolved by precedence:
      * `running` (exact, when loaded) → `parameters` (Modelfile default) →
-     * `model-info` (native max) → `2048` fallback.
+     * `model-info` (native max) → worst-case `2048` fallback.
      */
     readonly contextLength: number;
     /** Which source produced {@link contextLength}. */
@@ -329,7 +358,9 @@ interface ContextLengthSignals {
  *      (Ollama's conservative default), so native is a *ceiling*, not an
  *      expectation.
  *   4. **Fallback** — {@link OLLAMA_FALLBACK_CONTEXT_LENGTH} (2048), the
- *      conservative server-side default, so the result is always actionable.
+ *      worst-case floor for the unset-`num_ctx` window (current daemons tier
+ *      by VRAM up to 262,144), so the result is always actionable and can
+ *      never over-estimate a real window.
  *
  * Pure — used by `ModelsClient.getContextLength()` and exported for callers
  * that already hold `/api/ps` / `/api/show` responses.
@@ -363,6 +394,13 @@ interface RetryConfig {
 declare const DEFAULT_RETRY_CONFIG: RetryConfig;
 /**
  * Executes an async operation with retries according to config.
+ *
+ * Retry pacing: a 429 whose `Retry-After` header produced
+ * `OllamaRateLimitError.retryAfterMs` sleeps exactly that long before the
+ * next attempt (server authority — see {@link retryDelayMs}); all other
+ * retryable errors sleep a full-jitter exponential backoff. The delay
+ * actually used is reported to `onRetry` either way, and remains abortable
+ * through `signal` for the whole wait.
  */
 declare function withRetry<T>(operation: (attempt: number) => Promise<T>, config?: RetryConfig, signal?: AbortSignal): Promise<T>;
 
@@ -758,9 +796,11 @@ interface OllamaClientConfig {
      * Default context window (`num_ctx`) applied to every `chat`/`generate`
      * request that doesn't set `options.num_ctx` itself. Injected as
      * `options.num_ctx` on the wire, making the effective window explicit
-     * instead of leaving it to Ollama's server default (commonly 2048–4096,
-     * model-dependent) — which **silently truncates** prompts that don't fit,
-     * with no error and no warning. See `src/context-safety.ts`.
+     * instead of leaving it to Ollama's VRAM-tiered server default (4,096
+     * tokens below 24 GiB, 32,768 at 24–48 GiB, 262,144 at 48 GiB and up;
+     * daemon-wide override via `OLLAMA_CONTEXT_LENGTH`) — which **silently
+     * truncates** prompts that don't fit, with no error and no warning. See
+     * `src/context-safety.ts`.
      */
     readonly defaultContextLength?: number;
     /**
@@ -887,7 +927,8 @@ declare class ModelsClient {
     ps(request?: RequestCancellationOptions): Promise<PsResponse>;
     /**
      * Discovers the model's real context window from the server — no more
-     * 2048/4096 guesswork. Consults, in precedence order:
+     * default-window guesswork (current daemons tier the unset-`num_ctx`
+     * default by VRAM: 4k / 32k / 256k). Consults, in precedence order:
      *
      *   1. `GET /api/ps` — the window the **running instance actually
      *      allocated** (exact, reflects the Modelfile `num_ctx` default and
@@ -899,7 +940,7 @@ declare class ModelsClient {
      *   3. `POST /api/show` `model_info["<arch>.context_length"]` — the
      *      model's **native GGUF maximum** (a ceiling to raise `num_ctx`
      *      toward, not the allocated window).
-     *   4. Fallback — Ollama's conservative unset-`num_ctx` default (2048).
+     *   4. Fallback — the 2048 worst-case floor for the unset-`num_ctx` window.
      *
      * The result carries every signal it found, so `nativeContextLength:
      * 131072` vs `contextLength: 4096` (a 32× gap seen in the official API
@@ -1378,8 +1419,8 @@ declare function sanitizeHistoryForNextTurn(messages: readonly Message[], option
 interface SessionCompactionOptions {
     /**
      * Token budget for the compacted history. Defaults to the session's own
-     * window: `options.num_ctx` from the session config, or Ollama's
-     * conservative 2048 fallback when unset — pair with
+     * window: `options.num_ctx` from the session config, or the 2048
+     * worst-case fallback when unset — pair with
      * `client.models.getContextLength()` to program against the real window.
      */
     readonly maxEstimatedTokens?: number | undefined;
@@ -1685,6 +1726,41 @@ declare const MAX_SYSTEM_ONE_IMAGES_BYTES: number;
  * `maxRequestBytes` handling).
  */
 declare const MAX_SYSTEM_ONE_REQUEST_BYTES: number;
+/**
+ * Suffix that marks a System One cloud model (e.g. `clef:cloud`).
+ *
+ * Since the 2026-10-16 upstream cloud launch, `/v1/systemone` accepts
+ * cloud models on the model field. A local Ollama server fronts these
+ * requests to Ollama's cloud — but only from Ollama >= 0.40.5
+ * ({@link CLOUD_MODEL_MIN_OLLAMA_VERSION}); the hosted
+ * `https://ollama.com` endpoint always supports them.
+ */
+declare const CLOUD_MODEL_SUFFIX = ":cloud";
+/**
+ * Minimum local Ollama server version that can front cloud models
+ * (`clef:cloud` and friends) for `/v1/systemone`, per the upstream
+ * System One cloud-launch documentation. The blanket operation-level
+ * floor (0.35.0) stays on the IR constraint; this finer, model-dependent
+ * gate is enforced by `OllamaClient.systemOne` before the request is
+ * sent — closing a docs-vs-runtime gap where the 0.40.5 requirement was
+ * documented but never implemented (ADR 0036).
+ */
+declare const CLOUD_MODEL_MIN_OLLAMA_VERSION = "0.40.5";
+/**
+ * Whether `model` names a System One cloud model (ends with
+ * {@link CLOUD_MODEL_SUFFIX}, e.g. `clef:cloud`).
+ */
+declare function isCloudModel(model: string): boolean;
+/**
+ * Compare two dot-separated Ollama version strings numerically
+ * (e.g. `'0.40.5'` vs `'0.9.0'` → positive — 40 > 9 component-wise,
+ * NOT lexicographic). Mirrors the generated runtime's private
+ * `compareVersions` so the client-side cloud-model gate ranks versions
+ * identically to the IR-constraint gate.
+ *
+ * Returns negative when `a < b`, `0` when equal, positive when `a > b`.
+ */
+declare function compareOllamaVersions(a: string, b: string): number;
 /**
  * Discriminated union of the three question kinds. The upstream OpenAPI
  * defines this inline as `oneOf` within SystemOneRequest.questions; we
@@ -2403,9 +2479,13 @@ declare class OllamaClient {
      * launch): `model` also accepts cloud models such as `clef:cloud` —
      * cloud models require Ollama >= 0.40.5 or the hosted
      * `https://ollama.com` endpoint. Requires Ollama >= 0.35.0 locally.
-     * The runtime enforces the version constraint automatically (lazily
+     * Both version constraints are enforced client-side before the request
+     * is sent: the blanket 0.35.0 floor by the generated runtime (lazily
      * fetching `/api/version` when `enforceVersion` is `'auto'` or
-     * `'strict'`). Request size is checked client-side as a dual cap:
+     * `'strict'`), and the model-dependent 0.40.5 cloud gate by this
+     * wrapper — a cloud model through a local server older than 0.40.5
+     * throws {@link OllamaCloudModelVersionError} without a round-trip
+     * (ADR 0036). Request size is checked client-side as a dual cap:
      * text and schema (images excluded) within 64 KiB AND the whole
      * request within 32 MiB.
      *
@@ -2430,13 +2510,27 @@ declare class OllamaClient {
      * `client.native.systemOne(request)` directly.
      */
     systemOne<Q extends SystemOneQuestions = SystemOneQuestions>(request: SystemOneRequest<Q>): Promise<SystemOneResponse<Q>>;
+    /**
+     * Cloud-model version gate for {@link systemOne} — throws
+     * {@link OllamaCloudModelVersionError} when a cloud model is requested
+     * through a local server older than {@link CLOUD_MODEL_MIN_OLLAMA_VERSION}
+     * (0.40.5). No-op for non-cloud models and for cloud-mode clients
+     * (hosted `ollama.com`). The `/api/version` probe goes through the normal
+     * failover machinery and is cached per client instance (reset on failure,
+     * so a later gated call retries), so at most one extra round-trip happens
+     * across all gated calls.
+     */
+    private assertCloudModelSupported;
+    private serverVersionCache;
     /**
      * Ollama's hosted web search tool (`POST https://ollama.com/api/web_search`) — a fixed
      * Ollama Cloud service, unrelated to `config.endpoints`/`baseUrl` and not proxied
      * through a local server. Requires an Ollama account API key: pass `apiKey` to the
      * client constructor or set `OLLAMA_API_KEY`. Applies the same default `timeoutMs` and
      * retry policy as inference calls, but never cross-endpoint-fails-over — there is only
-     * ever the one cloud endpoint to call.
+     * ever the one cloud endpoint to call. If the service answers 429 with a `Retry-After`
+     * header, the transport parses it into `OllamaRateLimitError.retryAfterMs` and the
+     * retry loop sleeps exactly that long instead of jittered backoff (see ADR 0035).
      */
     webSearch(req: WebSearchRequestOptions): Promise<WebSearchResponse>;
     /**
@@ -2451,14 +2545,24 @@ declare class OllamaClient {
      * web search, and web fetch, bucketed by hour (`range: '24h'`) or day (`'7d'`/`'30d'`).
      *
      * See {@link OllamaClient.webSearch} for the cloud-endpoint/auth/timeout/retry
-     * behavior, which applies identically here. Query parameters mirror the official
-     * endpoint: `range` (`'24h' | '7d' | '30d'`, server default `'7d'`) and `scope`
-     * (`'self' | 'team'`, server default `'self'`; team scope requires a team admin).
+     * behavior, which applies identically here. The pinned contract documents this
+     * endpoint's 429 with a `Retry-After` header ("Seconds to wait before retrying"),
+     * which the SDK parses into `OllamaRateLimitError.retryAfterMs` and honors verbatim
+     * when retrying. Query parameters mirror the official endpoint: `range`
+     * (`'24h' | '7d' | '30d'`, server default `'7d'`), `scope`
+     * (`'self' | 'team'`, server default `'self'`; team scope requires a team
+     * admin), and — since the 2026-10-22 pin (upstream 2476d5e) — `groupBy`
+     * (`group_by=model`), which makes `totals` and every bucket carry up to 10
+     * per-model `UsageGroup` entries plus a `remainder` roll-up, with the
+     * selection echoed back as `group_by` on the response.
      * Omitted options are left unset so the server applies its own defaults.
      *
      * ```ts
-     * const usage = await client.usage({ range: '24h' });
+     * const usage = await client.usage({ range: '24h', groupBy: 'model' });
      * console.log(usage.totals.request_count, usage.totals.usage_usd);
+     * for (const group of usage.totals.groups ?? []) {
+     *   console.log(group.model, group.usage_usd);
+     * }
      * for (const bucket of usage.buckets) {
      *   if (bucket.partial) continue; // current hour, still in progress
      *   // ...
@@ -2472,7 +2576,9 @@ declare class OllamaClient {
      * object (`balance_usd`/`allowance_usd`/`period`) or, on legacy plans, session/weekly
      * percentage limits; the shape is discriminated at runtime by the server response.
      * See {@link OllamaClient.webSearch} for the cloud-endpoint/auth/timeout/retry
-     * behavior, which applies identically here.
+     * behavior, which applies identically here. Like `/api/usage`, the pinned contract
+     * documents this endpoint's 429 with a `Retry-After` header, parsed into
+     * `OllamaRateLimitError.retryAfterMs` and honored verbatim when retrying (ADR 0035).
      *
      * ```ts
      * const balance = await client.balance();
@@ -2584,8 +2690,11 @@ declare function createTimeoutSignal(timeoutMs?: number, userSignal?: AbortSigna
  * fan-out embed batches, speculative-decoding pipelines) can't always
  * predict saturation ahead of time, so the SDK exposes this helper —
  * a thin wrapper around `fetch` that retries 503/429 responses with
- * the same jittered exponential backoff used by the SDK's internal
- * transport (see `./backoff.ts` and `./retry.ts`).
+ * the same retry discipline used by the SDK's internal transport (see
+ * `./backoff.ts` and `./retry.ts`): jittered exponential backoff, except
+ * that a 429 carrying `Retry-After` (RFC 9110 § 10.2.3) is honored
+ * verbatim — the server said when to retry, and sleeping less than that
+ * would only compound the rate limiting.
  *
  * The helper is intentionally standalone: it works with any
  * `fetch`-shaped function (browser `globalThis.fetch`, `undici`, an
@@ -2624,8 +2733,10 @@ interface FetchWithBackoffConfig {
      *
      * Receives the attempt index (0-based for the original request,
      * 1+ for retries), the HTTP status that triggered the retry
-     * (or `0` for network-level failures), and the calculated delay
-     * in milliseconds (already jittered).
+     * (or `0` for network-level failures), and the delay
+     * in milliseconds that will be slept (the server's `Retry-After`
+     * verbatim on a 429 that carried one; otherwise the jittered
+     * backoff sample).
      */
     readonly onRetry?: ((attempt: number, status: number, delayMs: number) => void) | undefined;
     /**
@@ -2667,11 +2778,14 @@ declare const DEFAULT_FETCH_BACKOFF_CONFIG: {
  * Wrapper around any `fetch`-shaped function that retries HTTP 429 / 503
  * responses with jittered exponential backoff (full-jitter strategy —
  * see `./backoff.ts` for the rationale and AWS Architecture Blog
- * citation).
+ * citation), honoring `Retry-After` on 429s when the server sends it.
  *
  * The wrapper:
  *   - forwards `input`/`init` unchanged to every attempt,
- *   - sleeps before each retry using the configured backoff,
+ *   - sleeps before each retry using the configured backoff — or the
+ *     server's `Retry-After` value verbatim when a 429 carries one
+ *     (no jitter, no clamping; same policy as the SDK's internal
+ *     `withRetry`, see `./retry.ts`),
  *   - retries up to `maxRetries` times (default 3), then returns the
  *     final `Response` (whatever its status) so the caller can
  *     branch on it,
@@ -4025,11 +4139,14 @@ declare function encodeImage(image: VisionInput): Promise<string>;
  * Context-window safety: client-side token estimation and pre-flight overflow
  * warnings for `/api/chat` and `/api/generate`.
  *
- * Why this exists: when `options.num_ctx` is left unset, Ollama loads the
- * model with a conservative default context window (commonly 2048–4096
- * tokens, model-dependent) and **silently truncates** any prompt that doesn't
- * fit — the model just loses the top of the conversation with no error, no
- * warning, and often no visible symptom beyond degraded answers. This module
+ * Why this exists: when `options.num_ctx` is left unset, Ollama sizes the
+ * default context window by available VRAM (4,096 tokens below 24 GiB,
+ * 32,768 at 24–48 GiB, 262,144 at 48 GiB and up; daemon-wide override via
+ * `OLLAMA_CONTEXT_LENGTH`, and cloud models run at their maximum — see the
+ * "Context length" page in the Ollama docs) and **silently truncates** any
+ * prompt that doesn't fit — the model just loses the top of the conversation
+ * with no error, no warning, and often no visible symptom beyond degraded
+ * answers. This module
  * estimates the prompt size client-side (a fast heuristic — no tokenizer
  * download, no model dependency) and surfaces a warning *before* the request
  * is sent, when the estimate approaches the effective context window.
@@ -4098,7 +4215,7 @@ type ContextWindowSource =
 'request'
 /** Client `defaultContextLength` config — injected as `num_ctx` by the pipeline. */
  | 'client-default'
-/** Neither set; Ollama's server default applies (2048–4096, model-dependent). */
+/** Neither set; Ollama's VRAM-tiered server default applies (4k / 32k / 256k). */
  | 'server-fallback';
 /** What the pre-flight check concluded about one request. */
 interface ContextCheck {
@@ -4154,9 +4271,15 @@ declare function checkChatContext(req: ChatRequestOptions, options?: ContextChec
  */
 declare function checkGenerateContext(req: GenerateRequestOptions, options?: ContextCheckOptions): ContextCheck;
 /**
- * The conservative server-side default (2048) Ollama most commonly applies
- * when `num_ctx` is unset — used to warn about *likely* truncation even when
- * no effective window is known client-side.
+ * 2048 — the worst-case floor for the window a daemon allocates when
+ * `num_ctx` is unset, used to warn about *likely* truncation even when no
+ * effective window is known client-side. Deliberately not a prediction:
+ * current daemons tier the unset default by VRAM (4,096 / 32,768 / 262,144
+ * tokens), older daemons defaulted lower, and an admin can pin any value via
+ * `OLLAMA_CONTEXT_LENGTH`. Sizing the floor at the minimum plausible window
+ * keeps this check unable to over-estimate a real window — over-estimation
+ * is the failure that silently re-enables truncation. Use
+ * `client.models.getContextLength()` when the real number matters.
  */
 declare const OLLAMA_FALLBACK_CONTEXT_LENGTH = 2048;
 /**
@@ -4303,4 +4426,4 @@ 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, 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 CompareDraftPerformanceOptions, type ContextCheck, type ContextCheckOptions, type ContextDiscoveryRequestOptions, type ContextLengthSignals, type ContextLengthSource, type ContextWindowSource, type ConversationSendOptions, ConversationSession, type ConversationSessionOptions, CopyRequestOptions, type CreateModelFromGgufOptions, 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 DraftBenchmarkClient, type DraftPerformanceComparison, type DraftRunStats, 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, type ProgressCallback, ProgressResponse, type ProgressSpeed, ProgressStreamResult, PsResponse, type PullProgressEvent, 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 StreamObserverCallbacks, type StreamObserverPhase, 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, WEB_FETCH_TOOL_NAME, WEB_SEARCH_TOOL_NAME, WebFetchRequestOptions, WebFetchResponse, type WebFetchToolParams, WebSearchRequestOptions, WebSearchResponse, type WebSearchToolParams, type WebToolsClient, type WebToolsOptions, addSpanEvent, batchEmbed, calculateBackoff, canonicalToolCallSignature, checkChatContext, checkEndpointHealth, checkGenerateContext, compactConversationHistory, compareDraftPerformance, computeProgressPercent, contextWarningMessage, createDecision, createModelAffinityScheduler, createOllamaCloudFreeTierQuota, createProgressSpeedTracker, createTimeoutSignal, createWebFetchTool, createWebSearchTool, createWebTools, defineTool, detectModelCapabilities, embedBatchOverflowMessage, encodeImage, estimateChatRequestTokens, estimateGenerateRequestTokens, estimateMessageTokens, estimateSystemOneRequestBytes, estimateTokens, extractContextLength as extractNativeContextLength, extractParameterNumCtx, fetchWithBackoff, findOversizedEmbedInputs, findRunningModelContextLength, formatTelemetry, formatWebFetchResult, formatWebSearchResult, inferRuntimeMode, isKeepAliveSugar, isKnownDoneReason, isModelResident, isSameToolCall, listAllMcpTools, listAvailableModels, loadMcpTools, mergeToolCall, mergeToolCallArgumentsString, mergeToolCallArrays, normalizeChatStream, normalizeGenerateStream, normalizeKeepAlive, normalizeProgressStream, observeStream, onProgress, parseNdjsonStream, parseStructuredOutput, recommendContextWindow, registerMcpTools, resolveContextLength, resolveCredentialEndpoints, sanitizeHistoryForNextTurn, setSpanError, toDataStream, toPullProgressEvent, 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, CLOUD_MODEL_MIN_OLLAMA_VERSION, CLOUD_MODEL_SUFFIX, ChatRequestOptions, ChatResponse, ChatStreamResult, type ChoiceDecision, type CompactionOptions, type CompareDraftPerformanceOptions, type ContextCheck, type ContextCheckOptions, type ContextDiscoveryRequestOptions, type ContextLengthSignals, type ContextLengthSource, type ContextWindowSource, type ConversationSendOptions, ConversationSession, type ConversationSessionOptions, CopyRequestOptions, type CreateModelFromGgufOptions, 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 DraftBenchmarkClient, type DraftPerformanceComparison, type DraftRunStats, 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, type ProgressCallback, ProgressResponse, type ProgressSpeed, ProgressStreamResult, PsResponse, type PullProgressEvent, 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 StreamObserverCallbacks, type StreamObserverPhase, 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 UsageGroup, type UsageMetrics, UsageRequestOptions, type UsageResponse, type UsageSummary, VersionResponse, VisionInput, WEB_FETCH_TOOL_NAME, WEB_SEARCH_TOOL_NAME, WebFetchRequestOptions, WebFetchResponse, type WebFetchToolParams, WebSearchRequestOptions, WebSearchResponse, type WebSearchToolParams, type WebToolsClient, type WebToolsOptions, addSpanEvent, batchEmbed, calculateBackoff, canonicalToolCallSignature, checkChatContext, checkEndpointHealth, checkGenerateContext, compactConversationHistory, compareDraftPerformance, compareOllamaVersions, computeProgressPercent, contextWarningMessage, createDecision, createModelAffinityScheduler, createOllamaCloudFreeTierQuota, createProgressSpeedTracker, createTimeoutSignal, createWebFetchTool, createWebSearchTool, createWebTools, defineTool, detectModelCapabilities, embedBatchOverflowMessage, encodeImage, estimateChatRequestTokens, estimateGenerateRequestTokens, estimateMessageTokens, estimateSystemOneRequestBytes, estimateTokens, extractContextLength as extractNativeContextLength, extractParameterNumCtx, fetchWithBackoff, findOversizedEmbedInputs, findRunningModelContextLength, formatTelemetry, formatWebFetchResult, formatWebSearchResult, inferRuntimeMode, isCloudModel, isKeepAliveSugar, isKnownDoneReason, isModelResident, isSameToolCall, listAllMcpTools, listAvailableModels, loadMcpTools, mergeToolCall, mergeToolCallArgumentsString, mergeToolCallArrays, normalizeChatStream, normalizeGenerateStream, normalizeKeepAlive, normalizeProgressStream, observeStream, onProgress, parseNdjsonStream, parseStructuredOutput, recommendContextWindow, registerMcpTools, resolveContextLength, resolveCredentialEndpoints, sanitizeHistoryForNextTurn, setSpanError, toDataStream, toPullProgressEvent, toResponse, toTextStream, withRetry, withSpan, zodToJsonSchema };