@bastani/pi-ai 0.9.20-alpha.1 → 0.9.20-alpha.3

package/CHANGELOG.md

@@ -1,9 +1,17 @@
 # Changelog
 
-This package is a Bastani fork of `@earendil-works/pi-ai`. Upstream history at the audited Pi `main` sync point (`92d8e2d17d4f357788381c49ce2cdb3f4ed1f21c`) lives in [earendil-works/pi](https://github.com/earendil-works/pi/blob/92d8e2d17d4f357788381c49ce2cdb3f4ed1f21c/packages/ai/CHANGELOG.md).
+This package is a Bastani fork of `@earendil-works/pi-ai`. Upstream history at the audited Pi `main` sync point (`6671c604766b3670ed95f405aa7856835d0ca702`) lives in [earendil-works/pi](https://github.com/earendil-works/pi/blob/6671c604766b3670ed95f405aa7856835d0ca702/packages/ai/CHANGELOG.md).
 
 ## [Unreleased]
 
+## [0.9.20-alpha.3] - 2026-09-16
+
+### Fixed
+
+- Fixed OpenAI-compatible Responses errors to identify the actual provider instead of always labeling them as OpenAI errors ([#9298](https://github.com/earendil-works/pi/issues/9298)).
+- Fixed Amazon Bedrock one-hour cache writes being priced at the five-minute rate ([#9457](https://github.com/earendil-works/pi/issues/9457)).
+- Fixed Baseten requests to send session-affinity headers from `sessionId` for automatic prompt-cache routing ([#9629](https://github.com/earendil-works/pi/issues/9629)).
+
 ## [0.9.19] - 2026-09-13
 
 Cumulative release of the `0.9.19-alpha.2` through `0.9.19-alpha.4` prereleases. Per-change details remain in the unchanged prerelease sections below.

package/NOTICE.md

@@ -6,7 +6,7 @@ monorepo at `packages/ai` and publishes at the same version as `@bastani/atomic`
 
 - Upstream package: [`@earendil-works/pi-ai`](https://www.npmjs.com/package/@earendil-works/pi-ai)
 - Original fork point: `v0.84.2` (`914cf1472e715297caa30db4b9535d534a9eb718`)
-- Pi AI fixes and generated image catalog synced through audited upstream `main`: `earendil-works/pi@92d8e2d17d4f357788381c49ce2cdb3f4ed1f21c`
+- Pi AI fixes and generated image catalog synced through audited upstream `main`: `earendil-works/pi@6671c604766b3670ed95f405aa7856835d0ca702`
 - Catalog JSON under `src/providers/data/` is generated at build time from models.dev, matching upstream. It is not committed.
 
 Original work is Copyright (c) 2025 Mario Zechner and is licensed under the MIT License.

package/README.md

@@ -1,6 +1,6 @@
 # @bastani/pi-ai
 
-Bastani-branded fork of [`@earendil-works/pi-ai`](https://www.npmjs.com/package/@earendil-works/pi-ai) from [earendil-works/pi](https://github.com/earendil-works/pi). Originally forked at **v0.84.2** (`914cf1472e715297caa30db4b9535d534a9eb718`); upstream Pi AI fixes and the generated image catalog are synced through [`92d8e2d17d4f357788381c49ce2cdb3f4ed1f21c`](https://github.com/earendil-works/pi/commit/92d8e2d17d4f357788381c49ce2cdb3f4ed1f21c), the audited Pi `main` sync point. `@bastani/pi-ai` publishes at the same version as Atomic. `npm run build` refreshes the models.dev catalog, same as upstream.
+Bastani-branded fork of [`@earendil-works/pi-ai`](https://www.npmjs.com/package/@earendil-works/pi-ai) from [earendil-works/pi](https://github.com/earendil-works/pi). Originally forked at **v0.84.2** (`914cf1472e715297caa30db4b9535d534a9eb718`); upstream Pi AI fixes and the generated image catalog are synced through [`6671c604766b3670ed95f405aa7856835d0ca702`](https://github.com/earendil-works/pi/commit/6671c604766b3670ed95f405aa7856835d0ca702), the audited Pi `main` sync point. `@bastani/pi-ai` publishes at the same version as Atomic. `npm run build` refreshes the models.dev catalog, same as upstream.
 
 The public API is a drop-in replacement: install `@bastani/pi-ai` and import from `@bastani/pi-ai` instead of `@earendil-works/pi-ai`. See [NOTICE.md](NOTICE.md). This package lives in the Atomic monorepo and publishes from `.github/workflows/publish.yml`. The first npm version must be published by hand so trusted publishing can be attached.
 
@@ -1577,6 +1577,8 @@ Provider notes:
 
 **OpenAI Codex**: Requires a ChatGPT Plus or Pro subscription. Provides access to GPT-5.x Codex models with extended context windows and reasoning capabilities. The library automatically handles session-based prompt caching when `sessionId` is provided in stream options unless `cacheRetention` is `"none"`. You can set `transport` in stream options to `"sse"`, `"websocket"`, or `"auto"` for Codex Responses transport selection. When using WebSocket with a `sessionId` and cache retention enabled, connections are reused per session and expire after 5 minutes of inactivity.
 
+Call `cleanupSessionResources(sessionId)` when finished with a Codex session so its pooled WebSocket connection does not keep the process alive. Import it from `@bastani/pi-ai`.
+
 **Azure OpenAI (Responses)**: Uses the Responses API only. Set `AZURE_OPENAI_API_KEY` and either `AZURE_OPENAI_BASE_URL` or `AZURE_OPENAI_RESOURCE_NAME`. `AZURE_OPENAI_BASE_URL` supports both `https://<resource>.openai.azure.com` and `https://<resource>.cognitiveservices.azure.com`; root endpoints are normalized to `.../openai/v1` automatically. Use `AZURE_OPENAI_API_VERSION` (defaults to `v1`) to override the API version if needed. Deployment names are treated as model IDs by default, override with `azureDeploymentName` or `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` using comma-separated `model-id=deployment` pairs (for example `gpt-4o-mini=my-deployment,gpt-4o=prod`). Legacy deployment-based URLs are intentionally unsupported.
 
 **GitHub Copilot**: If you get "The requested model is not supported" error, enable the model manually in VS Code: open Copilot Chat, click the model selector, select the model (warning icon), and click "Enable".

package/dist/api/bedrock-converse-stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"bedrock-converse-stream.d.ts","sourceRoot":"","sources":["../../src/api/bedrock-converse-stream.ts"],"names":[],"mappings":"AA8BA,OAAO,KAAK,EAUX,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAEb,eAAe,EAEf,aAAa,EAIb,MAAM,aAAa,CAAC;AAmBrB,MAAM,MAAM,sBAAsB,GAAG,YAAY,GAAG,SAAS,CAAC;AAE9D,MAAM,WAAW,cAAe,SAAQ,aAAa;IACpD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IAEtE,SAAS,CAAC,EAAE,aAAa,CAAC;IAE1B,eAAe,CAAC,EAAE,eAAe,CAAC;IAElC,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;;;;;;OASG;IACH,eAAe,CAAC,EAAE,sBAAsB,CAAC;IACzC;;;sGAGkG;IAClG,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzC;;;;yGAIqG;IACrG,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAcD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,yBAAyB,EAAE,cAAc,CAwO5E,CAAC;AAwKF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,yBAAyB,EAAE,mBAAmB,CAiDvF,CAAC","sourcesContent":["import type { Agent as HttpsAgent } from \"node:https\";\nimport {\n\tBedrockRuntimeClient,\n\ttype BedrockRuntimeClientConfig,\n\tBedrockRuntimeServiceException,\n\tStopReason as BedrockStopReason,\n\ttype Tool as BedrockTool,\n\tCachePointType,\n\tCacheTTL,\n\ttype ContentBlock,\n\ttype ContentBlockDeltaEvent,\n\ttype ContentBlockStartEvent,\n\ttype ContentBlockStopEvent,\n\tConversationRole,\n\tConverseStreamCommand,\n\ttype ConverseStreamMetadataEvent,\n\tDocumentFormat,\n\tImageFormat,\n\ttype Message,\n\ttype SystemContentBlock,\n\ttype ToolChoice,\n\ttype ToolConfiguration,\n\ttype ToolResultContentBlock,\n\tToolResultStatus,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport { NodeHttpHandler } from \"@smithy/node-http-handler\";\nimport type { BuildMiddleware, DeserializeMiddleware, DocumentType, HttpResponse, MetadataBearer } from \"@smithy/types\";\nimport { HttpProxyAgent } from \"http-proxy-agent\";\nimport { HttpsProxyAgent } from \"https-proxy-agent\";\nimport { calculateCost } from \"../models.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tCacheRetention,\n\tContext,\n\tDocumentContent,\n\tImageContent,\n\tModel,\n\tProviderEnv,\n\tProviderResponse,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tThinkingLevel,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"../utils/diagnostics.ts\";\nimport { assertSupportedDocumentMimeType } from \"../utils/document-input.ts\";\nimport { normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { providerHeadersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { resolveHttpProxyUrlForTarget } from \"../utils/node-http-proxy.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from \"./constrained-sampling.ts\";\nimport {\n\tadjustMaxTokensForThinking,\n\tbuildBaseOptions,\n\tclampMaxTokensToContext,\n\tclampReasoning,\n} from \"./simple-options.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\nexport type BedrockThinkingDisplay = \"summarized\" | \"omitted\";\n\nexport interface BedrockOptions extends StreamOptions {\n\tregion?: string;\n\tprofile?: string;\n\ttoolChoice?: \"auto\" | \"any\" | \"none\" | { type: \"tool\"; name: string };\n\t/* See https://docs.aws.amazon.com/bedrock/latest/userguide/inference-reasoning.html for supported models. */\n\treasoning?: ThinkingLevel;\n\t/* Custom token budgets per thinking level. Overrides default budgets. */\n\tthinkingBudgets?: ThinkingBudgets;\n\t/* Only supported by Claude 4.x models, see https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html#claude-messages-extended-thinking-tool-use-interleaved */\n\tinterleavedThinking?: boolean;\n\t/**\n\t * Controls how Claude's thinking content is returned in responses.\n\t * - \"summarized\": Thinking blocks contain summarized thinking text (default here).\n\t * - \"omitted\": Thinking content is redacted but the signature still travels back\n\t *   for multi-turn continuity, reducing time-to-first-text-token.\n\t *\n\t * Note: Anthropic's API default for Claude Opus 4.8 and Mythos Preview is\n\t * \"omitted\". We default to \"summarized\" here to keep behavior consistent with\n\t * older Claude 4 models. Only applies to Claude models on Bedrock.\n\t */\n\tthinkingDisplay?: BedrockThinkingDisplay;\n\t/** Key-value pairs attached to the inference request for cost allocation tagging.\n\t * Keys: max 64 chars, no `aws:` prefix. Values: max 256 chars. Max 50 pairs.\n\t * Tags appear in AWS Cost Explorer split cost allocation data.\n\t * @see https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ConverseStream.html */\n\trequestMetadata?: Record<string, string>;\n\t/** Bearer token for Bedrock API key authentication.\n\t * When set, bypasses SigV4 signing and sends Authorization: Bearer <token> instead.\n\t * Requires `bedrock:CallWithBearerToken` IAM permission on the token's identity.\n\t * Set via AWS_BEARER_TOKEN_BEDROCK env var or pass directly.\n\t * @see https://docs.aws.amazon.com/service-authorization/latest/reference/list_amazonbedrock.html */\n\tbearerToken?: string;\n}\n\ntype Block = (TextContent | ThinkingContent | ToolCall) & {\n\tindex?: number;\n\tpartialJson?: string;\n\t/** Scratch buffer for encrypted reasoning deltas, joined into `thinkingSignature`. */\n\tredactedChunks?: Uint8Array[];\n};\n\nconst EMPTY_TEXT_PLACEHOLDER = \"<empty>\";\n\n/** Matches the placeholder the Anthropic API path uses for redacted thinking. */\nconst REDACTED_THINKING_PLACEHOLDER = \"[Reasoning redacted]\";\n\nexport const stream: StreamFunction<\"bedrock-converse-stream\", BedrockOptions> = (\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcontext: Context,\n\toptions: BedrockOptions = {},\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"bedrock-converse-stream\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\tconst blocks = output.content as Block[];\n\n\t\t// A profile explicitly configured through pi's auth flow (the `profile`\n\t\t// option or scoped `AWS_PROFILE` on the stored credential's env) must win\n\t\t// over ambient AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY. The SDK default\n\t\t// chain already prefers a configured profile over env keys, but only when\n\t\t// `credentials` is not set on the client config. See #6957.\n\t\tconst optionsProfile = options.profile || options.env?.AWS_PROFILE;\n\t\tconst config: BedrockRuntimeClientConfig = {\n\t\t\tprofile: optionsProfile || getProviderEnvValue(\"AWS_PROFILE\", options.env),\n\t\t};\n\t\tconst configuredRegion = getConfiguredBedrockRegion(options);\n\t\tconst hasAmbientConfiguredProfile = Boolean(getProviderEnvValue(\"AWS_PROFILE\"));\n\t\tconst endpointRegion = getStandardBedrockEndpointRegion(model.baseUrl);\n\t\tconst useExplicitEndpoint = shouldUseExplicitBedrockEndpoint(\n\t\t\tmodel.baseUrl,\n\t\t\tconfiguredRegion,\n\t\t\thasAmbientConfiguredProfile,\n\t\t);\n\n\t\t// Only pin standard AWS Bedrock runtime endpoints when no region or ambient AWS_PROFILE is configured.\n\t\t// This preserves custom endpoints (VPC/proxy) from #3402 without forcing built-in\n\t\t// catalog defaults such as us-east-1 to override AWS_REGION/AWS_PROFILE.\n\t\tif (useExplicitEndpoint) {\n\t\t\tconfig.endpoint = model.baseUrl;\n\t\t}\n\n\t\t// Resolve bearer token for Bedrock API key auth.\n\t\tconst skipAuth = getProviderEnvValue(\"AWS_BEDROCK_SKIP_AUTH\", options.env) === \"1\";\n\t\tconst bearerToken =\n\t\t\toptions.bearerToken ||\n\t\t\toptions.apiKey ||\n\t\t\tgetProviderEnvValue(\"AWS_BEARER_TOKEN_BEDROCK\", options.env) ||\n\t\t\tundefined;\n\t\tconst useBearerToken = bearerToken !== undefined && !skipAuth;\n\n\t\t// in Node.js/Bun environment only\n\t\tif (typeof process !== \"undefined\" && (process.versions?.node || process.versions?.bun)) {\n\t\t\t// Region resolution: ARN-embedded > explicit option > env vars > SDK default chain.\n\t\t\t// When the model ID is an inference profile ARN, extract the region from it.\n\t\t\t// This avoids conflicts with AWS_REGION set for other services.\n\t\t\tconst arnRegionMatch = model.id.match(/^arn:aws(?:-[a-z0-9-]+)?:bedrock:([a-z0-9-]+):/);\n\t\t\tif (arnRegionMatch) {\n\t\t\t\tconfig.region = arnRegionMatch[1];\n\t\t\t} else if (configuredRegion) {\n\t\t\t\tconfig.region = configuredRegion;\n\t\t\t} else if (endpointRegion && useExplicitEndpoint) {\n\t\t\t\tconfig.region = endpointRegion;\n\t\t\t} else if (!hasAmbientConfiguredProfile) {\n\t\t\t\tconfig.region = \"us-east-1\";\n\t\t\t}\n\n\t\t\t// Support proxies that don't need authentication\n\t\t\tif (skipAuth) {\n\t\t\t\tconfig.credentials = {\n\t\t\t\t\taccessKeyId: \"dummy-access-key\",\n\t\t\t\t\tsecretAccessKey: \"dummy-secret-key\",\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst credentials = getConfiguredBedrockCredentials(options.env);\n\t\t\tif (!skipAuth && credentials && !optionsProfile) {\n\t\t\t\tconfig.credentials = credentials;\n\t\t\t}\n\n\t\t\tconst proxyUrl = resolveHttpProxyUrlForTarget(model.baseUrl, options.env);\n\t\t\tif (proxyUrl) {\n\t\t\t\t// Bedrock runtime uses NodeHttp2Handler by default since v3.798.0, which is based\n\t\t\t\t// on `http2` module and has no support for http agent.\n\t\t\t\t// Use NodeHttpHandler to support HTTP(S) proxy agents.\n\t\t\t\tconfig.requestHandler = new NodeHttpHandler({\n\t\t\t\t\thttpAgent: new HttpProxyAgent(proxyUrl),\n\t\t\t\t\thttpsAgent: new HttpsProxyAgent(proxyUrl) as unknown as HttpsAgent,\n\t\t\t\t});\n\t\t\t} else if (getProviderEnvValue(\"AWS_BEDROCK_FORCE_HTTP1\", options.env) === \"1\") {\n\t\t\t\t// Some custom endpoints require HTTP/1.1 instead of HTTP/2\n\t\t\t\tconfig.requestHandler = new NodeHttpHandler();\n\t\t\t}\n\t\t} else {\n\t\t\t// Non-Node environment (browser): fall back to us-east-1 since\n\t\t\t// there's no config file resolution available.\n\t\t\tconfig.region =\n\t\t\t\tconfiguredRegion || (endpointRegion && useExplicitEndpoint ? endpointRegion : undefined) || \"us-east-1\";\n\t\t}\n\n\t\tif (useBearerToken) {\n\t\t\tconfig.token = { token: bearerToken };\n\t\t\tconfig.authSchemePreference = [\"httpBearerAuth\"];\n\t\t}\n\n\t\t// Kept outside the try so the catch can still correlate a mid-stream failure:\n\t\t// exceptions delivered as stream events carry no HTTP metadata of their own.\n\t\tlet responseRequestId: string | undefined;\n\n\t\ttry {\n\t\t\tconst supportsStrictMode = model.compat?.supportsStrictMode ?? false;\n\t\t\tconst client = new BedrockRuntimeClient(config);\n\t\t\tlet observedRawResponse = false;\n\t\t\tif (options.onResponse) {\n\t\t\t\taddResponseHeadersMiddleware(client, options.onResponse, model, () => {\n\t\t\t\t\tobservedRawResponse = true;\n\t\t\t\t});\n\t\t\t}\n\t\t\tconst customHeaders = providerHeadersToRecord({ ...model.headers, ...options.headers });\n\t\t\tif (customHeaders) {\n\t\t\t\taddCustomHeadersMiddleware(client, customHeaders);\n\t\t\t}\n\t\t\tconst cacheRetention = resolveCacheRetention(options.cacheRetention, options.env);\n\t\t\tconst inferenceMaxTokens = options.maxTokens ?? (isAnthropicClaudeModel(model) ? model.maxTokens : undefined);\n\t\t\tlet commandInput = {\n\t\t\t\tmodelId: model.id,\n\t\t\t\tmessages: convertMessages(context, model, cacheRetention, options.env),\n\t\t\t\tsystem: buildSystemPrompt(context.systemPrompt, model, cacheRetention, options.env),\n\t\t\t\tinferenceConfig: {\n\t\t\t\t\t...(inferenceMaxTokens !== undefined && { maxTokens: inferenceMaxTokens }),\n\t\t\t\t\t// Claude Fable 5.1 rejects non-default `temperature`, `top_p`, and `top_k` on every\n\t\t\t\t\t// request. Generated metadata marks such models `supportsTemperature: false`; every\n\t\t\t\t\t// other model keeps sending the field exactly as before.\n\t\t\t\t\t// https://platform.claude.com/docs/en/build-with-claude/thinking\n\t\t\t\t\t...(options.temperature !== undefined &&\n\t\t\t\t\t\tmodel.compat?.supportsTemperature !== false && { temperature: options.temperature }),\n\t\t\t\t},\n\t\t\t\ttoolConfig: convertToolConfig(context.tools, options.toolChoice, supportsStrictMode, model),\n\t\t\t\tadditionalModelRequestFields: buildAdditionalModelRequestFields(model, options),\n\t\t\t\t...(options.requestMetadata !== undefined && { requestMetadata: options.requestMetadata }),\n\t\t\t};\n\t\t\tconst nextCommandInput = await options?.onPayload?.(commandInput, model);\n\t\t\tif (nextCommandInput !== undefined) {\n\t\t\t\tcommandInput = nextCommandInput as typeof commandInput;\n\t\t\t}\n\t\t\tconst command = new ConverseStreamCommand(commandInput);\n\n\t\t\tconst response = await client.send(command, { abortSignal: options.signal });\n\t\t\tresponseRequestId = normalizeDiagnosticValue(response.$metadata.requestId);\n\t\t\tif (!observedRawResponse && response.$metadata.httpStatusCode !== undefined) {\n\t\t\t\tconst responseHeaders: Record<string, string> = {};\n\t\t\t\tif (response.$metadata.requestId) {\n\t\t\t\t\tresponseHeaders[\"x-amzn-requestid\"] = response.$metadata.requestId;\n\t\t\t\t}\n\t\t\t\tawait options?.onResponse?.({ status: response.$metadata.httpStatusCode, headers: responseHeaders }, model);\n\t\t\t}\n\n\t\t\tfor await (const item of response.stream!) {\n\t\t\t\tif (item.messageStart) {\n\t\t\t\t\tif (item.messageStart.role !== ConversationRole.ASSISTANT) {\n\t\t\t\t\t\tthrow new Error(\"Unexpected assistant message start but got user message start instead\");\n\t\t\t\t\t}\n\t\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t\t} else if (item.contentBlockStart) {\n\t\t\t\t\thandleContentBlockStart(item.contentBlockStart, blocks, output, stream);\n\t\t\t\t} else if (item.contentBlockDelta) {\n\t\t\t\t\thandleContentBlockDelta(item.contentBlockDelta, blocks, output, stream);\n\t\t\t\t} else if (item.contentBlockStop) {\n\t\t\t\t\thandleContentBlockStop(item.contentBlockStop, blocks, output, stream);\n\t\t\t\t} else if (item.messageStop) {\n\t\t\t\t\toutput.rawStopReason = item.messageStop.stopReason;\n\t\t\t\t\tconst { stopReason, errorMessage } = mapStopReason(item.messageStop.stopReason);\n\t\t\t\t\toutput.stopReason = stopReason;\n\t\t\t\t\tif (errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = errorMessage;\n\t\t\t\t\t}\n\t\t\t\t} else if (item.metadata) {\n\t\t\t\t\thandleMetadata(item.metadata, model, output);\n\t\t\t\t} else if (item.internalServerException) {\n\t\t\t\t\tthrow item.internalServerException;\n\t\t\t\t} else if (item.modelStreamErrorException) {\n\t\t\t\t\tthrow item.modelStreamErrorException;\n\t\t\t\t} else if (item.validationException) {\n\t\t\t\t\tthrow item.validationException;\n\t\t\t\t} else if (item.throttlingException) {\n\t\t\t\t\tthrow item.throttlingException;\n\t\t\t\t} else if (item.serviceUnavailableException) {\n\t\t\t\t\tthrow item.serviceUnavailableException;\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (options.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"pending\") {\n\t\t\t\tthrow new Error(\"Bedrock stream ended without a stop reason\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\" || output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"An unknown error occurred\");\n\t\t\t}\n\n\t\t\t// A stream can settle without stopping every block, so finalize here too.\n\t\t\tfor (const block of output.content) finalizeStreamingBlock(block as Block);\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tfinalizeStreamingBlock(block as Block);\n\t\t\t}\n\t\t\toutput.stopReason = options.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatBedrockError(error);\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tappendBedrockFailureDiagnostic(output, error, responseRequestId);\n\t\t\t}\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\n/**\n * Human-readable prefixes for Bedrock SDK exception names.\n * The downstream retry logic in agent-session matches patterns like\n * `server.?error` and `service.?unavailable`, so we preserve the legacy\n * prefix format rather than using the raw SDK exception name.\n */\nconst BEDROCK_ERROR_PREFIXES: Record<string, string> = {\n\tInternalServerException: \"Internal server error\",\n\tModelStreamErrorException: \"Model stream error\",\n\tValidationException: \"Validation error\",\n\tThrottlingException: \"Throttling error\",\n\tServiceUnavailableException: \"Service unavailable\",\n};\n\n/**\n * Some models reject the account/profile's configured Bedrock data retention mode\n * (e.g. \"data retention mode 'default' is not available for this model\"). Point\n * users at the AWS docs explaining how to configure a supported mode.\n */\nconst BEDROCK_DATA_RETENTION_DOCS_URL = \"https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html\";\n\n/**\n * Format a Bedrock error with a human-readable prefix.\n * AWS SDK exceptions (both from `client.send()` and from stream event items)\n * extend BedrockRuntimeServiceException. We map the `.name` to a stable\n * human-readable prefix so downstream consumers (retry logic, context-overflow\n * detection) can distinguish error categories via simple string matching.\n */\nfunction formatBedrockError(error: unknown): string {\n\tconst norm = normalizeProviderError(error);\n\t// Surface the raw HTTP body (with status) when the SDK did not fold it into\n\t// the message; otherwise fall back to the message. This is what stops a\n\t// gateway 403 from collapsing to `Unknown: UnknownError`.\n\tconst core =\n\t\t!norm.messageCarriesBody && norm.status !== undefined && norm.body !== undefined\n\t\t\t? `${norm.status}: ${norm.body}`\n\t\t\t: norm.message;\n\tconst dataRetentionHint = /data retention mode/i.test(core)\n\t\t? ` See ${BEDROCK_DATA_RETENTION_DOCS_URL} for supported data retention modes.`\n\t\t: \"\";\n\tif (error instanceof BedrockRuntimeServiceException) {\n\t\tconst prefix = BEDROCK_ERROR_PREFIXES[error.name] ?? error.name;\n\t\treturn `${prefix}: ${core}${dataRetentionHint}`;\n\t}\n\treturn `${core}${dataRetentionHint}`;\n}\n\ntype SdkErrorMetadata = { $metadata?: { httpStatusCode?: unknown; requestId?: unknown } };\n\n/** Over-long header values are dropped rather than truncated: a truncated request id is not a request id. */\nconst MAX_BEDROCK_DIAGNOSTIC_VALUE_CHARS = 200;\n\nfunction normalizeDiagnosticValue(value: unknown): string | undefined {\n\tif (typeof value !== \"string\") return undefined;\n\tconst trimmed = value.trim();\n\tif (trimmed.length === 0 || trimmed.length > MAX_BEDROCK_DIAGNOSTIC_VALUE_CHARS) return undefined;\n\treturn trimmed;\n}\n\n/**\n * The SDK puts the modeled code on `error.name` for service exceptions and unmodeled stream errors alike, so\n * do not narrow to `BedrockRuntimeServiceException`. Modeled Bedrock errors all end in `Exception`, unlike\n * transport names such as `TimeoutError`.\n */\nfunction extractBedrockErrorCode(error: unknown): string | undefined {\n\tif (!(error instanceof Error) || !error.name.endsWith(\"Exception\")) return undefined;\n\treturn normalizeDiagnosticValue(error.name);\n}\n\n/**\n * Structured metadata alongside `errorMessage`, which stays byte-identical because `isRetryableAssistantError`\n * matches against it. Unknown fields are omitted, never guessed: a modeled mid-stream exception reaches us as\n * a bare object literal, leaving only `fallbackRequestId`. `details` only, as the throw is not always `Error`.\n */\nfunction appendBedrockFailureDiagnostic(\n\toutput: AssistantMessage,\n\terror: unknown,\n\tfallbackRequestId: string | undefined,\n): void {\n\tconst metadata = (error as SdkErrorMetadata)?.$metadata;\n\tconst details: Record<string, unknown> = {};\n\n\tif (typeof metadata?.httpStatusCode === \"number\") details.status = metadata.httpStatusCode;\n\n\tconst errorCode = extractBedrockErrorCode(error);\n\tif (errorCode !== undefined) details.errorCode = errorCode;\n\n\tconst requestId = normalizeDiagnosticValue(metadata?.requestId) ?? fallbackRequestId;\n\tif (requestId !== undefined) details.requestId = requestId;\n\n\tif (Object.keys(details).length === 0) return;\n\n\tappendAssistantMessageDiagnostic(output, { type: \"bedrock_response_failure\", timestamp: Date.now(), details });\n}\n\n/**\n * Header keys that must never be overwritten by caller-supplied headers.\n * `host` and `x-amz-*` participate in the SigV4 canonical request; `authorization`\n * is owned by SigV4 or the bearer-token path (config.token + authSchemePreference).\n * Compared case-insensitively (caller key is lower-cased before lookup).\n */\nconst RESERVED_HEADER_EXACT = new Set([\"authorization\", \"host\"]);\n\nfunction isReservedHeader(key: string): boolean {\n\tconst lower = key.toLowerCase();\n\treturn lower.startsWith(\"x-amz-\") || RESERVED_HEADER_EXACT.has(lower);\n}\n\n/**\n * Attach caller-supplied headers to the outgoing Bedrock request via a Smithy\n * `build`-step middleware. The `build` step runs after request serialisation but\n * before SigV4 signing, so injected headers are covered by the signature. Reserved\n * SigV4 / auth headers (`x-amz-*`, `authorization`, `host`) are silently skipped;\n * all other caller headers override any existing same-named header on the request.\n */\nfunction addCustomHeadersMiddleware(client: BedrockRuntimeClient, headers: Record<string, string>): void {\n\tconst middleware: BuildMiddleware<object, MetadataBearer> = (next) => async (args) => {\n\t\tconst request = args.request;\n\t\tif (request && typeof request === \"object\" && \"headers\" in request) {\n\t\t\tconst requestHeaders = (request as { headers: Record<string, string> }).headers;\n\t\t\tfor (const [key, value] of Object.entries(headers)) {\n\t\t\t\tif (!isReservedHeader(key)) {\n\t\t\t\t\trequestHeaders[key] = value;\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t\treturn next(args);\n\t};\n\tclient.middlewareStack.add(middleware, { step: \"build\", name: \"pi-ai-custom-headers\", priority: \"low\" });\n}\n\nfunction isSmithyHttpResponse(response: unknown): response is HttpResponse {\n\tif (!response || typeof response !== \"object\") return false;\n\tconst candidate = response as Partial<HttpResponse>;\n\treturn typeof candidate.statusCode === \"number\" && !!candidate.headers && typeof candidate.headers === \"object\";\n}\n\nfunction toProviderResponse(response: unknown): ProviderResponse | undefined {\n\tif (!isSmithyHttpResponse(response)) return undefined;\n\treturn { status: response.statusCode, headers: { ...response.headers } };\n}\n\n/**\n * Bedrock's modeled `$metadata` only preserves selected HTTP metadata (for example\n * requestId), so custom gateway headers are otherwise lost before callers see\n * `onResponse`. Capture the raw Smithy HTTP response at the deserialize step,\n * after the SDK receives the response but before the event stream is consumed.\n */\nfunction addResponseHeadersMiddleware(\n\tclient: BedrockRuntimeClient,\n\tonResponse: NonNullable<BedrockOptions[\"onResponse\"]>,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tonObserved: () => void,\n): void {\n\tconst middleware: DeserializeMiddleware<object, MetadataBearer> = (next) => async (args) => {\n\t\tconst result = await next(args);\n\t\tconst providerResponse = toProviderResponse(result.response);\n\t\tif (providerResponse) {\n\t\t\tonObserved();\n\t\t\tawait onResponse(providerResponse, model);\n\t\t}\n\t\treturn result;\n\t};\n\tclient.middlewareStack.add(middleware, { step: \"deserialize\", name: \"pi-ai-response-headers\" });\n}\n\nexport const streamSimple: StreamFunction<\"bedrock-converse-stream\", SimpleStreamOptions> = (\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst base = {\n\t\t...buildBaseOptions(model, context, options, undefined),\n\t\ttoolChoice: options?.toolChoice,\n\t} satisfies BedrockOptions;\n\tif (!options?.reasoning) {\n\t\treturn stream(model, context, { ...base, reasoning: undefined } satisfies BedrockOptions);\n\t}\n\n\tif (isAnthropicClaudeModel(model)) {\n\t\tif (supportsAdaptiveThinking(model.id, model.name)) {\n\t\t\treturn stream(model, context, {\n\t\t\t\t...base,\n\t\t\t\treasoning: options.reasoning,\n\t\t\t\tthinkingBudgets: options.thinkingBudgets,\n\t\t\t} satisfies BedrockOptions);\n\t\t}\n\n\t\t// Undefined means the caller did not request an output cap; let the helper use the model cap.\n\t\t// Do not coerce to 0 here, or the thinking budget would become the entire maxTokens value.\n\t\tconst adjusted = adjustMaxTokensForThinking(\n\t\t\tbase.maxTokens,\n\t\t\tmodel.maxTokens,\n\t\t\toptions.reasoning,\n\t\t\toptions.thinkingBudgets,\n\t\t);\n\n\t\tconst maxTokens = clampMaxTokensToContext(model, context, adjusted.maxTokens);\n\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tmaxTokens,\n\t\t\treasoning: options.reasoning,\n\t\t\tthinkingBudgets: {\n\t\t\t\t...(options.thinkingBudgets || {}),\n\t\t\t\t[clampReasoning(options.reasoning)!]: Math.min(adjusted.thinkingBudget, Math.max(0, maxTokens - 1024)),\n\t\t\t},\n\t\t} satisfies BedrockOptions);\n\t}\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoning: options.reasoning,\n\t\tthinkingBudgets: options.thinkingBudgets,\n\t} satisfies BedrockOptions);\n};\n\nfunction handleContentBlockStart(\n\tevent: ContentBlockStartEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst index = event.contentBlockIndex!;\n\tconst start = event.start;\n\n\tif (start?.toolUse) {\n\t\tconst block: Block = {\n\t\t\ttype: \"toolCall\",\n\t\t\tid: start.toolUse.toolUseId || \"\",\n\t\t\tname: start.toolUse.name || \"\",\n\t\t\targuments: {},\n\t\t\tpartialJson: \"\",\n\t\t\tindex,\n\t\t};\n\t\toutput.content.push(block);\n\t\tstream.push({ type: \"toolcall_start\", contentIndex: blocks.length - 1, partial: output });\n\t}\n}\n\nfunction handleContentBlockDelta(\n\tevent: ContentBlockDeltaEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst contentBlockIndex = event.contentBlockIndex!;\n\tconst delta = event.delta;\n\tlet index = blocks.findIndex((b) => b.index === contentBlockIndex);\n\tlet block = blocks[index];\n\n\tif (delta?.text !== undefined) {\n\t\t// If no text block exists yet, create one, as `handleContentBlockStart` is not sent for text blocks\n\t\tif (!block) {\n\t\t\tconst newBlock: Block = { type: \"text\", text: \"\", index: contentBlockIndex };\n\t\t\toutput.content.push(newBlock);\n\t\t\tindex = blocks.length - 1;\n\t\t\tblock = blocks[index];\n\t\t\tstream.push({ type: \"text_start\", contentIndex: index, partial: output });\n\t\t}\n\t\tif (block.type === \"text\") {\n\t\t\tblock.text += delta.text;\n\t\t\tstream.push({ type: \"text_delta\", contentIndex: index, delta: delta.text, partial: output });\n\t\t}\n\t} else if (delta?.toolUse && block?.type === \"toolCall\") {\n\t\tblock.partialJson = (block.partialJson || \"\") + (delta.toolUse.input || \"\");\n\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\tstream.push({ type: \"toolcall_delta\", contentIndex: index, delta: delta.toolUse.input || \"\", partial: output });\n\t} else if (delta?.reasoningContent) {\n\t\tlet thinkingBlock = block;\n\t\tlet thinkingIndex = index;\n\n\t\tif (!thinkingBlock) {\n\t\t\tconst newBlock: Block = { type: \"thinking\", thinking: \"\", thinkingSignature: \"\", index: contentBlockIndex };\n\t\t\toutput.content.push(newBlock);\n\t\t\tthinkingIndex = blocks.length - 1;\n\t\t\tthinkingBlock = blocks[thinkingIndex];\n\t\t\tstream.push({ type: \"thinking_start\", contentIndex: thinkingIndex, partial: output });\n\t\t}\n\n\t\tif (thinkingBlock?.type === \"thinking\") {\n\t\t\tif (delta.reasoningContent.text) {\n\t\t\t\tthinkingBlock.thinking += delta.reasoningContent.text;\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\tcontentIndex: thinkingIndex,\n\t\t\t\t\tdelta: delta.reasoningContent.text,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t}\n\t\t\t// `thinkingSignature` holds either an Anthropic signature or an opaque redacted\n\t\t\t// payload, never both: mixing them would corrupt whichever arrived first.\n\t\t\tif (delta.reasoningContent.signature && !thinkingBlock.redacted) {\n\t\t\t\tthinkingBlock.thinkingSignature =\n\t\t\t\t\t(thinkingBlock.thinkingSignature || \"\") + delta.reasoningContent.signature;\n\t\t\t}\n\t\t\tif (delta.reasoningContent.redactedContent?.length) {\n\t\t\t\t// Encrypted reasoning from non-Anthropic models on Bedrock (e.g. OpenAI GPT-5.6).\n\t\t\t\t// The payload is opaque, so keep it verbatim in `thinkingSignature` the way the\n\t\t\t\t// Anthropic path stores redacted thinking, and replay it on the next turn.\n\t\t\t\tif (!thinkingBlock.redacted) {\n\t\t\t\t\tthinkingBlock.redacted = true;\n\t\t\t\t\tthinkingBlock.thinkingSignature = \"\";\n\t\t\t\t\tthinkingBlock.thinking += REDACTED_THINKING_PLACEHOLDER;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\tcontentIndex: thinkingIndex,\n\t\t\t\t\t\tdelta: REDACTED_THINKING_PLACEHOLDER,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tthinkingBlock.redactedChunks ??= [];\n\t\t\t\tthinkingBlock.redactedChunks.push(delta.reasoningContent.redactedContent);\n\t\t\t}\n\t\t}\n\t}\n}\n\n/**\n * Encodes buffered encrypted reasoning into `thinkingSignature` and drops the scratch\n * buffer, which must never reach a persisted message: `Uint8Array` serializes to an\n * index-keyed object roughly ten times the size of the base64 payload.\n */\nfunction flushRedactedContent(block: Block): void {\n\tif (block.type !== \"thinking\" || !block.redactedChunks) return;\n\tblock.thinkingSignature = bytesToBase64(block.redactedChunks);\n\tdelete block.redactedChunks;\n}\n\n/**\n * Strips every streaming scratch field. Runs from the terminal paths as well as\n * `contentBlockStop`, because a stream can settle without stopping each block.\n */\nfunction finalizeStreamingBlock(block: Block): void {\n\tdelete block.index;\n\t// partialJson is only a streaming scratch buffer; never persist it.\n\tdelete block.partialJson;\n\tflushRedactedContent(block);\n}\n\nfunction handleMetadata(\n\tevent: ConverseStreamMetadataEvent,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\toutput: AssistantMessage,\n): void {\n\tif (event.usage) {\n\t\toutput.usage.input = event.usage.inputTokens || 0;\n\t\toutput.usage.output = event.usage.outputTokens || 0;\n\t\toutput.usage.cacheRead = event.usage.cacheReadInputTokens || 0;\n\t\toutput.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;\n\t\toutput.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;\n\t\tcalculateCost(model, output.usage);\n\t}\n}\n\nfunction handleContentBlockStop(\n\tevent: ContentBlockStopEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst index = blocks.findIndex((b) => b.index === event.contentBlockIndex);\n\tconst block = blocks[index];\n\tif (!block) return;\n\tdelete (block as Block).index;\n\n\tswitch (block.type) {\n\t\tcase \"text\":\n\t\t\tstream.push({ type: \"text_end\", contentIndex: index, content: block.text, partial: output });\n\t\t\tbreak;\n\t\tcase \"thinking\":\n\t\t\tflushRedactedContent(block);\n\t\t\tstream.push({ type: \"thinking_end\", contentIndex: index, content: block.thinking, partial: output });\n\t\t\tbreak;\n\t\tcase \"toolCall\":\n\t\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\t\t// Finalize in-place and strip the scratch buffer so replay only\n\t\t\t// carries parsed arguments.\n\t\t\tdelete (block as Block).partialJson;\n\t\t\tstream.push({ type: \"toolcall_end\", contentIndex: index, toolCall: block, partial: output });\n\t\t\tbreak;\n\t}\n}\n\n/**\n * Check if the model supports adaptive thinking (Opus 4.6+, Sonnet 4.6).\n * Checks both model ID and model name to support application inference profiles\n * whose ARNs don't contain the model name.\n */\nfunction getModelMatchCandidates(modelId: string, modelName?: string): string[] {\n\tconst values = modelName ? [modelId, modelName] : [modelId];\n\treturn values.flatMap((value) => {\n\t\tconst lower = value.toLowerCase();\n\t\treturn [lower, lower.replace(/[\\s_.:]+/g, \"-\")];\n\t});\n}\n\nfunction supportsAdaptiveThinking(modelId: string, modelName?: string): boolean {\n\tconst candidates = getModelMatchCandidates(modelId, modelName);\n\treturn candidates.some(\n\t\t(s) =>\n\t\t\ts.includes(\"opus-4-6\") ||\n\t\t\ts.includes(\"opus-4-7\") ||\n\t\t\ts.includes(\"opus-4-8\") ||\n\t\t\ts.includes(\"opus-5\") ||\n\t\t\ts.includes(\"sonnet-4-6\") ||\n\t\t\ts.includes(\"sonnet-5\") ||\n\t\t\ts.includes(\"fable-5\"),\n\t);\n}\n\nfunction supportsNativeXhighEffort(model: Model<\"bedrock-converse-stream\">): boolean {\n\tconst candidates = getModelMatchCandidates(model.id, model.name);\n\treturn candidates.some(\n\t\t(s) =>\n\t\t\ts.includes(\"opus-4-7\") ||\n\t\t\ts.includes(\"opus-4-8\") ||\n\t\t\ts.includes(\"opus-5\") ||\n\t\t\ts.includes(\"sonnet-5\") ||\n\t\t\ts.includes(\"fable-5\"),\n\t);\n}\n\nfunction mapThinkingLevelToEffort(\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tlevel: SimpleStreamOptions[\"reasoning\"],\n): \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\" {\n\tif (level === \"xhigh\" && supportsNativeXhighEffort(model)) return \"xhigh\";\n\n\tconst mapped = level ? model.thinkingLevelMap?.[level] : undefined;\n\tif (typeof mapped === \"string\") return mapped as \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\n\tswitch (level) {\n\t\tcase \"minimal\":\n\t\tcase \"low\":\n\t\t\treturn \"low\";\n\t\tcase \"medium\":\n\t\t\treturn \"medium\";\n\t\tcase \"high\":\n\t\t\treturn \"high\";\n\t\tdefault:\n\t\t\treturn \"high\";\n\t}\n}\n\n/**\n * Resolve cache retention preference.\n * Defaults to \"short\" and uses PI_CACHE_RETENTION for backward compatibility.\n */\nfunction resolveCacheRetention(cacheRetention?: CacheRetention, env?: ProviderEnv): CacheRetention {\n\tif (cacheRetention) {\n\t\treturn cacheRetention;\n\t}\n\tif (getProviderEnvValue(\"PI_CACHE_RETENTION\", env) === \"long\") {\n\t\treturn \"long\";\n\t}\n\treturn \"short\";\n}\n\n/**\n * Check if the model is an Anthropic Claude model on Bedrock.\n * Checks both model ID and model name to support application inference profiles\n * whose ARNs don't contain the model name.\n */\nfunction isAnthropicClaudeModel(model: Model<\"bedrock-converse-stream\">): boolean {\n\tconst id = model.id.toLowerCase();\n\tconst name = model.name?.toLowerCase() ?? \"\";\n\treturn (\n\t\tid.includes(\"anthropic.claude\") ||\n\t\tid.includes(\"anthropic/claude\") ||\n\t\tname.includes(\"anthropic.claude\") ||\n\t\tname.includes(\"anthropic/claude\") ||\n\t\tname.includes(\"claude\")\n\t);\n}\n\n/**\n * Check if the model supports prompt caching.\n * Supported: Claude 3.5 Haiku, Claude 3.7 Sonnet, Claude 4.x models, Claude 5 models\n *\n * For base models and system-defined inference profiles the model ID / ARN\n * contains the model name, so we can decide locally.\n *\n * For application inference profiles (whose ARNs don't contain the model name),\n * also checks model.name which is user-controlled via models.json or registerProvider.\n * As a last resort, set AWS_BEDROCK_FORCE_CACHE=1 to enable cache points.\n * Amazon Nova models have automatic caching and don't need explicit cache points.\n */\nfunction supportsPromptCaching(model: Model<\"bedrock-converse-stream\">, env?: ProviderEnv): boolean {\n\tconst candidates = getModelMatchCandidates(model.id, model.name);\n\n\tconst hasClaudeRef = candidates.some((s) => s.includes(\"claude\"));\n\tif (!hasClaudeRef) {\n\t\t// Application inference profiles don't contain the model name in the ARN.\n\t\t// Allow users to force cache points via environment variable.\n\t\tif (getProviderEnvValue(\"AWS_BEDROCK_FORCE_CACHE\", env) === \"1\") return true;\n\t\treturn false;\n\t}\n\t// Claude 5 models (fable-5, opus-5, sonnet-5)\n\tif (candidates.some((s) => s.includes(\"fable-5\") || s.includes(\"opus-5\") || s.includes(\"sonnet-5\"))) return true;\n\t// Claude 4.x models (opus-4, sonnet-4, haiku-4)\n\tif (candidates.some((s) => s.includes(\"-4-\"))) return true;\n\t// Claude 3.7 Sonnet\n\tif (candidates.some((s) => s.includes(\"claude-3-7-sonnet\"))) return true;\n\t// Claude 3.5 Haiku\n\tif (candidates.some((s) => s.includes(\"claude-3-5-haiku\"))) return true;\n\treturn false;\n}\n\n/**\n * Check if the model supports thinking signatures in reasoningContent.\n * Only Anthropic Claude models support the signature field.\n * Other models (OpenAI, Qwen, Minimax, Moonshot, etc.) reject it with:\n * \"This model doesn't support the reasoningContent.reasoningText.signature field\"\n *\n * Checks both model ID and model name to support application inference profiles.\n */\nfunction supportsThinkingSignature(model: Model<\"bedrock-converse-stream\">): boolean {\n\treturn isAnthropicClaudeModel(model);\n}\n\nfunction buildSystemPrompt(\n\tsystemPrompt: string | undefined,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcacheRetention: CacheRetention,\n\tenv?: ProviderEnv,\n): SystemContentBlock[] | undefined {\n\tif (!systemPrompt) return undefined;\n\n\tconst blocks: SystemContentBlock[] = [{ text: sanitizeSurrogates(systemPrompt) }];\n\n\t// Add cache point for supported Claude models when caching is enabled\n\tif (cacheRetention !== \"none\" && supportsPromptCaching(model, env)) {\n\t\tblocks.push({\n\t\t\tcachePoint: { type: CachePointType.DEFAULT, ...(cacheRetention === \"long\" ? { ttl: CacheTTL.ONE_HOUR } : {}) },\n\t\t});\n\t}\n\n\treturn blocks;\n}\n\nfunction normalizeToolCallId(id: string): string {\n\tconst sanitized = id.replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\treturn sanitized.length > 64 ? sanitized.slice(0, 64) : sanitized;\n}\n\nfunction createNonBlankTextBlock(text: string): ContentBlock.TextMember | undefined {\n\tconst sanitized = sanitizeSurrogates(text);\n\treturn sanitized.trim().length === 0 ? undefined : { text: sanitized };\n}\n\nfunction createRequiredTextBlock(text: string): ContentBlock.TextMember {\n\treturn createNonBlankTextBlock(text) ?? { text: EMPTY_TEXT_PLACEHOLDER };\n}\n\nfunction sanitizeBedrockDocument(value: DocumentType): DocumentType {\n\tif (Array.isArray(value)) {\n\t\treturn value.map(sanitizeBedrockDocument);\n\t}\n\tif (value !== null && typeof value === \"object\") {\n\t\treturn Object.fromEntries(\n\t\t\tObject.entries(value)\n\t\t\t\t.filter(([key]) => key.length > 0)\n\t\t\t\t.map(([key, nestedValue]) => [key, sanitizeBedrockDocument(nestedValue)]),\n\t\t);\n\t}\n\treturn value;\n}\n\nfunction convertToolResultContent(content: (TextContent | ImageContent)[]): ToolResultContentBlock[] {\n\tconst result: ToolResultContentBlock[] = [];\n\tfor (const c of content) {\n\t\tif (c.type === \"image\") {\n\t\t\tresult.push({ image: createImageBlock(c.mimeType, c.data) });\n\t\t} else {\n\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\tif (textBlock) result.push(textBlock);\n\t\t}\n\t}\n\tif (result.length === 0) result.push({ text: EMPTY_TEXT_PLACEHOLDER });\n\treturn result;\n}\n\nfunction convertMessages(\n\tcontext: Context,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcacheRetention: CacheRetention,\n\tenv?: ProviderEnv,\n): Message[] {\n\tconst result: Message[] = [];\n\tconst transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst m = transformedMessages[i];\n\n\t\tswitch (m.role) {\n\t\t\tcase \"user\": {\n\t\t\t\tconst content: ContentBlock[] = [];\n\t\t\t\tif (typeof m.content === \"string\") {\n\t\t\t\t\tcontent.push(createRequiredTextBlock(m.content));\n\t\t\t\t} else {\n\t\t\t\t\tfor (const c of m.content) {\n\t\t\t\t\t\tswitch (c.type) {\n\t\t\t\t\t\t\tcase \"text\": {\n\t\t\t\t\t\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\t\t\t\t\t\tif (textBlock) content.push(textBlock);\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tcase \"image\":\n\t\t\t\t\t\t\t\tcontent.push({ image: createImageBlock(c.mimeType, c.data) });\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\tcase \"document\":\n\t\t\t\t\t\t\t\tcontent.push({ document: createDocumentBlock(c, content.length) });\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\tdefault:\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tif (content.length === 0) content.push({ text: EMPTY_TEXT_PLACEHOLDER });\n\t\t\t\t}\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.USER,\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"assistant\": {\n\t\t\t\t// Skip assistant messages with empty content (e.g., from aborted requests)\n\t\t\t\t// Bedrock rejects messages with empty content arrays\n\t\t\t\tif (m.content.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tconst contentBlocks: ContentBlock[] = [];\n\t\t\t\tfor (const c of m.content) {\n\t\t\t\t\tswitch (c.type) {\n\t\t\t\t\t\tcase \"text\": {\n\t\t\t\t\t\t\t// Skip empty text blocks\n\t\t\t\t\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\t\t\t\t\tif (!textBlock) continue;\n\t\t\t\t\t\t\tcontentBlocks.push(textBlock);\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tcase \"toolCall\":\n\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\ttoolUse: { toolUseId: c.id, name: c.name, input: sanitizeBedrockDocument(c.arguments) },\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\tcase \"thinking\": {\n\t\t\t\t\t\t\t// Encrypted reasoning is opaque: replay the stored payload as the\n\t\t\t\t\t\t\t// `redactedContent` member instead of lowering it to reasoning text.\n\t\t\t\t\t\t\tif (c.redacted) {\n\t\t\t\t\t\t\t\tconst redactedContent = decodeRedactedContent(c.thinkingSignature);\n\t\t\t\t\t\t\t\tif (redactedContent?.length) {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({ reasoningContent: { redactedContent } });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t// Skip empty thinking blocks\n\t\t\t\t\t\t\tconst thinking = sanitizeSurrogates(c.thinking);\n\t\t\t\t\t\t\tif (thinking.trim().length === 0) continue;\n\t\t\t\t\t\t\t// Only Anthropic models support the signature field in reasoningText.\n\t\t\t\t\t\t\t// For other models, we omit the signature to avoid errors like:\n\t\t\t\t\t\t\t// \"This model doesn't support the reasoningContent.reasoningText.signature field\"\n\t\t\t\t\t\t\tif (supportsThinkingSignature(model)) {\n\t\t\t\t\t\t\t\t// Signatures arrive after thinking deltas. If a partial or externally\n\t\t\t\t\t\t\t\t// persisted message lacks a signature, Bedrock rejects the replayed\n\t\t\t\t\t\t\t\t// reasoning block. Fall back to plain text, matching Anthropic.\n\t\t\t\t\t\t\t\tif (!c.thinkingSignature || c.thinkingSignature.trim().length === 0) {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({ text: thinking });\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\t\t\treasoningContent: {\n\t\t\t\t\t\t\t\t\t\t\treasoningText: {\n\t\t\t\t\t\t\t\t\t\t\t\ttext: thinking,\n\t\t\t\t\t\t\t\t\t\t\t\tsignature: c.thinkingSignature,\n\t\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\t\treasoningContent: {\n\t\t\t\t\t\t\t\t\t\treasoningText: { text: thinking },\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tdefault:\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\t// Skip if all content blocks were filtered out\n\t\t\t\tif (contentBlocks.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.ASSISTANT,\n\t\t\t\t\tcontent: contentBlocks,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"toolResult\": {\n\t\t\t\t// Collect all consecutive toolResult messages into a single user message\n\t\t\t\t// Bedrock requires all tool results to be in one message\n\t\t\t\tconst toolResults: ContentBlock.ToolResultMember[] = [];\n\n\t\t\t\t// Add current tool result with all content blocks combined\n\t\t\t\ttoolResults.push({\n\t\t\t\t\ttoolResult: {\n\t\t\t\t\t\ttoolUseId: m.toolCallId,\n\t\t\t\t\t\tcontent: convertToolResultContent(m.content),\n\t\t\t\t\t\tstatus: m.isError ? ToolResultStatus.ERROR : ToolResultStatus.SUCCESS,\n\t\t\t\t\t},\n\t\t\t\t});\n\n\t\t\t\t// Look ahead for consecutive toolResult messages\n\t\t\t\tlet j = i + 1;\n\t\t\t\twhile (j < transformedMessages.length && transformedMessages[j].role === \"toolResult\") {\n\t\t\t\t\tconst nextMsg = transformedMessages[j] as ToolResultMessage;\n\t\t\t\t\ttoolResults.push({\n\t\t\t\t\t\ttoolResult: {\n\t\t\t\t\t\t\ttoolUseId: nextMsg.toolCallId,\n\t\t\t\t\t\t\tcontent: convertToolResultContent(nextMsg.content),\n\t\t\t\t\t\t\tstatus: nextMsg.isError ? ToolResultStatus.ERROR : ToolResultStatus.SUCCESS,\n\t\t\t\t\t\t},\n\t\t\t\t\t});\n\t\t\t\t\tj++;\n\t\t\t\t}\n\n\t\t\t\t// Skip the messages we've already processed\n\t\t\t\ti = j - 1;\n\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.USER,\n\t\t\t\t\tcontent: toolResults,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tdefault:\n\t\t\t\tcontinue;\n\t\t}\n\t}\n\n\t// Add cache point to the last user message for supported Claude models when caching is enabled\n\tif (cacheRetention !== \"none\" && supportsPromptCaching(model, env) && result.length > 0) {\n\t\tconst lastMessage = result[result.length - 1];\n\t\tif (lastMessage.role === ConversationRole.USER && lastMessage.content) {\n\t\t\t(lastMessage.content as ContentBlock[]).push({\n\t\t\t\tcachePoint: {\n\t\t\t\t\ttype: CachePointType.DEFAULT,\n\t\t\t\t\t...(cacheRetention === \"long\" ? { ttl: CacheTTL.ONE_HOUR } : {}),\n\t\t\t\t},\n\t\t\t});\n\t\t}\n\t}\n\n\treturn result;\n}\n\nfunction convertToolConfig(\n\ttools: Tool[] | undefined,\n\ttoolChoice: BedrockOptions[\"toolChoice\"],\n\tsupportsStrictMode: boolean,\n\tmodel: Model<\"bedrock-converse-stream\">,\n): ToolConfiguration | undefined {\n\t// Validate the requested choice before any early return. A forced choice on a model that\n\t// rejects it is an error regardless of whether tools were supplied: returning `undefined` for\n\t// an empty or absent tool list would discard the caller's instruction silently, which is the\n\t// behavior this guard exists to prevent.\n\t//\n\t// Claude Fable 5.1 rejects forced tool use on every request with a 400, whichever platform\n\t// serves it. Reject rather than rewriting to `auto`, which would discard an explicit caller\n\t// instruction. Mirrors the guards in `anthropic-messages.ts` and `openai-completions.ts`.\n\t// https://platform.claude.com/docs/en/build-with-claude/thinking\n\tif (model.compat?.supportsForcedToolChoice === false) {\n\t\tconst isForced = toolChoice === \"any\" || (typeof toolChoice === \"object\" && toolChoice.type === \"tool\");\n\t\tif (isForced) {\n\t\t\tconst requestedLabel = typeof toolChoice === \"string\" ? toolChoice : `tool \"${toolChoice.name}\"`;\n\t\t\tthrow new Error(\n\t\t\t\t`Model ${model.id} does not support forced tool choice (requested: ${requestedLabel}). ` +\n\t\t\t\t\t`Use toolChoice \"auto\" with strict tool use or structured outputs instead.`,\n\t\t\t);\n\t\t}\n\t}\n\n\tif (!tools?.length) return undefined;\n\t// `none` is never a forced choice, so this return can never skip a rejection above.\n\tif (toolChoice === \"none\") return undefined;\n\n\tconst bedrockTools: BedrockTool[] = tools.map((tool) => {\n\t\tconst strict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);\n\t\treturn {\n\t\t\ttoolSpec: {\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tinputSchema: { json: getJsonSchemaToolParameters(tool, strict) as unknown as DocumentType },\n\t\t\t\t...(strict === true ? { strict: true } : {}),\n\t\t\t},\n\t\t};\n\t});\n\n\tlet bedrockToolChoice: ToolChoice | undefined;\n\tswitch (toolChoice) {\n\t\tcase \"auto\":\n\t\t\tbedrockToolChoice = { auto: {} };\n\t\t\tbreak;\n\t\tcase \"any\":\n\t\t\tbedrockToolChoice = { any: {} };\n\t\t\tbreak;\n\t\tdefault:\n\t\t\tif (toolChoice?.type === \"tool\") {\n\t\t\t\tbedrockToolChoice = { tool: { name: toolChoice.name } };\n\t\t\t}\n\t}\n\n\treturn { tools: bedrockTools, toolChoice: bedrockToolChoice };\n}\n\nfunction mapStopReason(reason: string | undefined): { stopReason: StopReason; errorMessage?: string } {\n\tswitch (reason) {\n\t\tcase BedrockStopReason.END_TURN:\n\t\tcase BedrockStopReason.STOP_SEQUENCE:\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase BedrockStopReason.MAX_TOKENS:\n\t\tcase BedrockStopReason.MODEL_CONTEXT_WINDOW_EXCEEDED:\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase BedrockStopReason.TOOL_USE:\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tdefault:\n\t\t\treturn reason\n\t\t\t\t? { stopReason: \"error\", errorMessage: `Provider stopped with: ${reason}` }\n\t\t\t\t: { stopReason: \"error\" };\n\t}\n}\n\nfunction getConfiguredBedrockRegion(options: BedrockOptions): string | undefined {\n\treturn (\n\t\toptions.region ||\n\t\tgetProviderEnvValue(\"AWS_REGION\", options.env) ||\n\t\tgetProviderEnvValue(\"AWS_DEFAULT_REGION\", options.env) ||\n\t\tundefined\n\t);\n}\n\nfunction getConfiguredBedrockCredentials(env?: ProviderEnv): BedrockRuntimeClientConfig[\"credentials\"] | undefined {\n\tconst accessKeyId = getProviderEnvValue(\"AWS_ACCESS_KEY_ID\", env);\n\tconst secretAccessKey = getProviderEnvValue(\"AWS_SECRET_ACCESS_KEY\", env);\n\tif (!accessKeyId || !secretAccessKey) {\n\t\treturn undefined;\n\t}\n\tconst sessionToken = getProviderEnvValue(\"AWS_SESSION_TOKEN\", env);\n\treturn {\n\t\taccessKeyId,\n\t\tsecretAccessKey,\n\t\t...(sessionToken ? { sessionToken } : {}),\n\t};\n}\n\nfunction getStandardBedrockEndpointRegion(baseUrl: string | undefined): string | undefined {\n\tif (!baseUrl) {\n\t\treturn undefined;\n\t}\n\n\ttry {\n\t\tconst { hostname } = new URL(baseUrl);\n\t\tconst match = hostname.toLowerCase().match(/^bedrock-runtime(?:-fips)?\\.([a-z0-9-]+)\\.amazonaws\\.com(?:\\.cn)?$/);\n\t\treturn match?.[1];\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction shouldUseExplicitBedrockEndpoint(\n\tbaseUrl: string,\n\tconfiguredRegion: string | undefined,\n\thasAmbientConfiguredProfile: boolean,\n): boolean {\n\tconst endpointRegion = getStandardBedrockEndpointRegion(baseUrl);\n\tif (!endpointRegion) {\n\t\treturn true;\n\t}\n\n\treturn !configuredRegion && !hasAmbientConfiguredProfile;\n}\n\nfunction isGovCloudBedrockTarget(model: Model<\"bedrock-converse-stream\">, options: BedrockOptions): boolean {\n\tconst region = getConfiguredBedrockRegion(options);\n\tif (region?.toLowerCase().startsWith(\"us-gov-\")) {\n\t\treturn true;\n\t}\n\n\tconst modelId = model.id.toLowerCase();\n\treturn modelId.startsWith(\"us-gov.\") || modelId.startsWith(\"arn:aws-us-gov:\");\n}\n\nfunction isOpenAiGpt6AstraBedrockModel(model: Model<\"bedrock-converse-stream\">): boolean {\n\treturn /^(?:global\\.|us\\.)?openai\\.gpt-6-astra$/.test(model.id);\n}\n\nfunction buildAdditionalModelRequestFields(\n\tmodel: Model<\"bedrock-converse-stream\">,\n\toptions: BedrockOptions,\n): Record<string, any> | undefined {\n\tif (!options.reasoning || !model.reasoning) {\n\t\treturn undefined;\n\t}\n\n\tif (isAnthropicClaudeModel(model)) {\n\t\t// GovCloud Bedrock currently rejects the Claude thinking.display field.\n\t\t// Omit it there until the GovCloud Converse schema catches up.\n\t\tconst display = isGovCloudBedrockTarget(model, options) ? undefined : (options.thinkingDisplay ?? \"summarized\");\n\t\tconst result: Record<string, any> = supportsAdaptiveThinking(model.id, model.name)\n\t\t\t? {\n\t\t\t\t\tthinking: { type: \"adaptive\", ...(display !== undefined ? { display } : {}) },\n\t\t\t\t\toutput_config: { effort: mapThinkingLevelToEffort(model, options.reasoning) },\n\t\t\t\t}\n\t\t\t: (() => {\n\t\t\t\t\tconst defaultBudgets: Record<ThinkingLevel, number> = {\n\t\t\t\t\t\tminimal: 1024,\n\t\t\t\t\t\tlow: 2048,\n\t\t\t\t\t\tmedium: 8192,\n\t\t\t\t\t\thigh: 16384,\n\t\t\t\t\t\txhigh: 16384, // Budget-based Claude clamps extended levels to high\n\t\t\t\t\t\tmax: 16384,\n\t\t\t\t\t};\n\n\t\t\t\t\t// Custom budgets only cover token-based levels through high.\n\t\t\t\t\tconst level = options.reasoning === \"xhigh\" || options.reasoning === \"max\" ? \"high\" : options.reasoning;\n\t\t\t\t\tconst budget = options.thinkingBudgets?.[level] ?? defaultBudgets[options.reasoning];\n\n\t\t\t\t\treturn {\n\t\t\t\t\t\tthinking: {\n\t\t\t\t\t\t\ttype: \"enabled\",\n\t\t\t\t\t\t\tbudget_tokens: budget,\n\t\t\t\t\t\t\t...(display !== undefined ? { display } : {}),\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t})();\n\n\t\tif (!supportsAdaptiveThinking(model.id, model.name) && (options.interleavedThinking ?? true)) {\n\t\t\tresult.anthropic_beta = [\"interleaved-thinking-2025-05-14\"];\n\t\t}\n\n\t\treturn result;\n\t}\n\n\tif (isOpenAiGpt6AstraBedrockModel(model)) {\n\t\t// Bedrock maps OpenAI Chat Completions fields that are not part of Converse's\n\t\t// inferenceConfig into additionalModelRequestFields. GPT-6-Astra uses the Chat\n\t\t// Completions spelling rather than the Responses API's nested reasoning object.\n\t\t// https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-openai.html\n\t\treturn { reasoning_effort: mapThinkingLevelToEffort(model, options.reasoning) };\n\t}\n\n\treturn undefined;\n}\n\n/**\n * Build a Bedrock `DocumentBlock`. Three things differ from the Anthropic path:\n *\n * - `name` is **required**, and AWS restricts it to alphanumerics, single whitespace runs,\n *   hyphens, parentheses, and square brackets — then warns the field \"is vulnerable to prompt\n *   injections… we recommend that you specify a neutral name.\" Any caller-supplied name is\n *   sanitized to that character set, and anything left empty falls back to a generated label.\n * - `source` takes **raw bytes**, not base64: \"If you use an Amazon Web Services SDK, you don't\n *   need to encode the bytes in base64.\" So the base64 payload is decoded here, the opposite of\n *   what the Anthropic converter does.\n * - Without citations enabled, Converse degrades to plain text extraction rather than full visual\n *   PDF understanding. That limitation is documented rather than worked around here.\n *   https://platform.claude.com/docs/en/build-with-claude/pdf-support\n *\n * `DocumentFormat` also has eight non-PDF members, but `format` is hardcoded to PDF because that\n * is the only media type `DocumentContent` declares — and the seven office/markup formats have no\n * Anthropic source variant to map onto. The block's own media type is therefore verified rather\n * than read.\n */\nfunction createDocumentBlock(block: DocumentContent, index: number) {\n\tassertSupportedDocumentMimeType(block);\n\tconst sanitized = (block.name ?? \"\")\n\t\t.replace(/[^a-zA-Z0-9\\s\\-()[\\]]/g, \" \")\n\t\t.replace(/\\s+/g, \" \")\n\t\t.trim();\n\treturn {\n\t\tformat: DocumentFormat.PDF,\n\t\tname: sanitized.length > 0 ? sanitized : `document ${index + 1}`,\n\t\tsource: { bytes: base64ToBytes(block.data) },\n\t};\n}\n\nfunction createImageBlock(mimeType: string, data: string) {\n\tlet format: ImageFormat;\n\tswitch (mimeType) {\n\t\tcase \"image/jpeg\":\n\t\tcase \"image/jpg\":\n\t\t\tformat = ImageFormat.JPEG;\n\t\t\tbreak;\n\t\tcase \"image/png\":\n\t\t\tformat = ImageFormat.PNG;\n\t\t\tbreak;\n\t\tcase \"image/gif\":\n\t\t\tformat = ImageFormat.GIF;\n\t\t\tbreak;\n\t\tcase \"image/webp\":\n\t\t\tformat = ImageFormat.WEBP;\n\t\t\tbreak;\n\t\tdefault:\n\t\t\tthrow new Error(`Unknown image type: ${mimeType}`);\n\t}\n\n\treturn { source: { bytes: base64ToBytes(data) }, format };\n}\n\nfunction base64ToBytes(data: string): Uint8Array {\n\tconst binaryString = atob(data);\n\tconst bytes = new Uint8Array(binaryString.length);\n\tfor (let i = 0; i < binaryString.length; i++) {\n\t\tbytes[i] = binaryString.charCodeAt(i);\n\t}\n\treturn bytes;\n}\n\n/**\n * Decodes a stored redacted payload. The AWS SDK hands the blob over as bytes, but a\n * persisted session carries it as base64. A hand-edited or externally produced session\n * can hold a signature that is not base64; drop that block instead of failing the\n * whole request.\n */\nfunction decodeRedactedContent(signature: string | undefined): Uint8Array | undefined {\n\tif (!signature) return undefined;\n\ttry {\n\t\treturn base64ToBytes(signature);\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction bytesToBase64(chunks: Uint8Array[]): string {\n\t// Encrypted reasoning runs to tens of KB, so build the binary string in slices\n\t// rather than one concatenation per byte. The window stays under the engine's\n\t// argument-count limit for spread calls.\n\tconst WINDOW = 0x8000;\n\tlet binary = \"\";\n\tfor (const chunk of chunks) {\n\t\tfor (let i = 0; i < chunk.length; i += WINDOW) {\n\t\t\tbinary += String.fromCharCode(...chunk.subarray(i, i + WINDOW));\n\t\t}\n\t}\n\treturn btoa(binary);\n}\n"]}
+{"version":3,"file":"bedrock-converse-stream.d.ts","sourceRoot":"","sources":["../../src/api/bedrock-converse-stream.ts"],"names":[],"mappings":"AA8BA,OAAO,KAAK,EAUX,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAEb,eAAe,EAEf,aAAa,EAIb,MAAM,aAAa,CAAC;AAmBrB,MAAM,MAAM,sBAAsB,GAAG,YAAY,GAAG,SAAS,CAAC;AAE9D,MAAM,WAAW,cAAe,SAAQ,aAAa;IACpD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IAEtE,SAAS,CAAC,EAAE,aAAa,CAAC;IAE1B,eAAe,CAAC,EAAE,eAAe,CAAC;IAElC,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;;;;;;OASG;IACH,eAAe,CAAC,EAAE,sBAAsB,CAAC;IACzC;;;sGAGkG;IAClG,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzC;;;;yGAIqG;IACrG,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAcD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,yBAAyB,EAAE,cAAc,CAwO5E,CAAC;AAwKF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,yBAAyB,EAAE,mBAAmB,CAiDvF,CAAC","sourcesContent":["import type { Agent as HttpsAgent } from \"node:https\";\nimport {\n\tBedrockRuntimeClient,\n\ttype BedrockRuntimeClientConfig,\n\tBedrockRuntimeServiceException,\n\tStopReason as BedrockStopReason,\n\ttype Tool as BedrockTool,\n\tCachePointType,\n\tCacheTTL,\n\ttype ContentBlock,\n\ttype ContentBlockDeltaEvent,\n\ttype ContentBlockStartEvent,\n\ttype ContentBlockStopEvent,\n\tConversationRole,\n\tConverseStreamCommand,\n\ttype ConverseStreamMetadataEvent,\n\tDocumentFormat,\n\tImageFormat,\n\ttype Message,\n\ttype SystemContentBlock,\n\ttype ToolChoice,\n\ttype ToolConfiguration,\n\ttype ToolResultContentBlock,\n\tToolResultStatus,\n} from \"@aws-sdk/client-bedrock-runtime\";\nimport { NodeHttpHandler } from \"@smithy/node-http-handler\";\nimport type { BuildMiddleware, DeserializeMiddleware, DocumentType, HttpResponse, MetadataBearer } from \"@smithy/types\";\nimport { HttpProxyAgent } from \"http-proxy-agent\";\nimport { HttpsProxyAgent } from \"https-proxy-agent\";\nimport { calculateCost } from \"../models.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tCacheRetention,\n\tContext,\n\tDocumentContent,\n\tImageContent,\n\tModel,\n\tProviderEnv,\n\tProviderResponse,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tThinkingLevel,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"../utils/diagnostics.ts\";\nimport { assertSupportedDocumentMimeType } from \"../utils/document-input.ts\";\nimport { normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { providerHeadersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { resolveHttpProxyUrlForTarget } from \"../utils/node-http-proxy.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from \"./constrained-sampling.ts\";\nimport {\n\tadjustMaxTokensForThinking,\n\tbuildBaseOptions,\n\tclampMaxTokensToContext,\n\tclampReasoning,\n} from \"./simple-options.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\nexport type BedrockThinkingDisplay = \"summarized\" | \"omitted\";\n\nexport interface BedrockOptions extends StreamOptions {\n\tregion?: string;\n\tprofile?: string;\n\ttoolChoice?: \"auto\" | \"any\" | \"none\" | { type: \"tool\"; name: string };\n\t/* See https://docs.aws.amazon.com/bedrock/latest/userguide/inference-reasoning.html for supported models. */\n\treasoning?: ThinkingLevel;\n\t/* Custom token budgets per thinking level. Overrides default budgets. */\n\tthinkingBudgets?: ThinkingBudgets;\n\t/* Only supported by Claude 4.x models, see https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html#claude-messages-extended-thinking-tool-use-interleaved */\n\tinterleavedThinking?: boolean;\n\t/**\n\t * Controls how Claude's thinking content is returned in responses.\n\t * - \"summarized\": Thinking blocks contain summarized thinking text (default here).\n\t * - \"omitted\": Thinking content is redacted but the signature still travels back\n\t *   for multi-turn continuity, reducing time-to-first-text-token.\n\t *\n\t * Note: Anthropic's API default for Claude Opus 4.8 and Mythos Preview is\n\t * \"omitted\". We default to \"summarized\" here to keep behavior consistent with\n\t * older Claude 4 models. Only applies to Claude models on Bedrock.\n\t */\n\tthinkingDisplay?: BedrockThinkingDisplay;\n\t/** Key-value pairs attached to the inference request for cost allocation tagging.\n\t * Keys: max 64 chars, no `aws:` prefix. Values: max 256 chars. Max 50 pairs.\n\t * Tags appear in AWS Cost Explorer split cost allocation data.\n\t * @see https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ConverseStream.html */\n\trequestMetadata?: Record<string, string>;\n\t/** Bearer token for Bedrock API key authentication.\n\t * When set, bypasses SigV4 signing and sends Authorization: Bearer <token> instead.\n\t * Requires `bedrock:CallWithBearerToken` IAM permission on the token's identity.\n\t * Set via AWS_BEARER_TOKEN_BEDROCK env var or pass directly.\n\t * @see https://docs.aws.amazon.com/service-authorization/latest/reference/list_amazonbedrock.html */\n\tbearerToken?: string;\n}\n\ntype Block = (TextContent | ThinkingContent | ToolCall) & {\n\tindex?: number;\n\tpartialJson?: string;\n\t/** Scratch buffer for encrypted reasoning deltas, joined into `thinkingSignature`. */\n\tredactedChunks?: Uint8Array[];\n};\n\nconst EMPTY_TEXT_PLACEHOLDER = \"<empty>\";\n\n/** Matches the placeholder the Anthropic API path uses for redacted thinking. */\nconst REDACTED_THINKING_PLACEHOLDER = \"[Reasoning redacted]\";\n\nexport const stream: StreamFunction<\"bedrock-converse-stream\", BedrockOptions> = (\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcontext: Context,\n\toptions: BedrockOptions = {},\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"bedrock-converse-stream\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\tconst blocks = output.content as Block[];\n\n\t\t// A profile explicitly configured through pi's auth flow (the `profile`\n\t\t// option or scoped `AWS_PROFILE` on the stored credential's env) must win\n\t\t// over ambient AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY. The SDK default\n\t\t// chain already prefers a configured profile over env keys, but only when\n\t\t// `credentials` is not set on the client config. See #6957.\n\t\tconst optionsProfile = options.profile || options.env?.AWS_PROFILE;\n\t\tconst config: BedrockRuntimeClientConfig = {\n\t\t\tprofile: optionsProfile || getProviderEnvValue(\"AWS_PROFILE\", options.env),\n\t\t};\n\t\tconst configuredRegion = getConfiguredBedrockRegion(options);\n\t\tconst hasAmbientConfiguredProfile = Boolean(getProviderEnvValue(\"AWS_PROFILE\"));\n\t\tconst endpointRegion = getStandardBedrockEndpointRegion(model.baseUrl);\n\t\tconst useExplicitEndpoint = shouldUseExplicitBedrockEndpoint(\n\t\t\tmodel.baseUrl,\n\t\t\tconfiguredRegion,\n\t\t\thasAmbientConfiguredProfile,\n\t\t);\n\n\t\t// Only pin standard AWS Bedrock runtime endpoints when no region or ambient AWS_PROFILE is configured.\n\t\t// This preserves custom endpoints (VPC/proxy) from #3402 without forcing built-in\n\t\t// catalog defaults such as us-east-1 to override AWS_REGION/AWS_PROFILE.\n\t\tif (useExplicitEndpoint) {\n\t\t\tconfig.endpoint = model.baseUrl;\n\t\t}\n\n\t\t// Resolve bearer token for Bedrock API key auth.\n\t\tconst skipAuth = getProviderEnvValue(\"AWS_BEDROCK_SKIP_AUTH\", options.env) === \"1\";\n\t\tconst bearerToken =\n\t\t\toptions.bearerToken ||\n\t\t\toptions.apiKey ||\n\t\t\tgetProviderEnvValue(\"AWS_BEARER_TOKEN_BEDROCK\", options.env) ||\n\t\t\tundefined;\n\t\tconst useBearerToken = bearerToken !== undefined && !skipAuth;\n\n\t\t// in Node.js/Bun environment only\n\t\tif (typeof process !== \"undefined\" && (process.versions?.node || process.versions?.bun)) {\n\t\t\t// Region resolution: ARN-embedded > explicit option > env vars > SDK default chain.\n\t\t\t// When the model ID is an inference profile ARN, extract the region from it.\n\t\t\t// This avoids conflicts with AWS_REGION set for other services.\n\t\t\tconst arnRegionMatch = model.id.match(/^arn:aws(?:-[a-z0-9-]+)?:bedrock:([a-z0-9-]+):/);\n\t\t\tif (arnRegionMatch) {\n\t\t\t\tconfig.region = arnRegionMatch[1];\n\t\t\t} else if (configuredRegion) {\n\t\t\t\tconfig.region = configuredRegion;\n\t\t\t} else if (endpointRegion && useExplicitEndpoint) {\n\t\t\t\tconfig.region = endpointRegion;\n\t\t\t} else if (!hasAmbientConfiguredProfile) {\n\t\t\t\tconfig.region = \"us-east-1\";\n\t\t\t}\n\n\t\t\t// Support proxies that don't need authentication\n\t\t\tif (skipAuth) {\n\t\t\t\tconfig.credentials = {\n\t\t\t\t\taccessKeyId: \"dummy-access-key\",\n\t\t\t\t\tsecretAccessKey: \"dummy-secret-key\",\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst credentials = getConfiguredBedrockCredentials(options.env);\n\t\t\tif (!skipAuth && credentials && !optionsProfile) {\n\t\t\t\tconfig.credentials = credentials;\n\t\t\t}\n\n\t\t\tconst proxyUrl = resolveHttpProxyUrlForTarget(model.baseUrl, options.env);\n\t\t\tif (proxyUrl) {\n\t\t\t\t// Bedrock runtime uses NodeHttp2Handler by default since v3.798.0, which is based\n\t\t\t\t// on `http2` module and has no support for http agent.\n\t\t\t\t// Use NodeHttpHandler to support HTTP(S) proxy agents.\n\t\t\t\tconfig.requestHandler = new NodeHttpHandler({\n\t\t\t\t\thttpAgent: new HttpProxyAgent(proxyUrl),\n\t\t\t\t\thttpsAgent: new HttpsProxyAgent(proxyUrl) as unknown as HttpsAgent,\n\t\t\t\t});\n\t\t\t} else if (getProviderEnvValue(\"AWS_BEDROCK_FORCE_HTTP1\", options.env) === \"1\") {\n\t\t\t\t// Some custom endpoints require HTTP/1.1 instead of HTTP/2\n\t\t\t\tconfig.requestHandler = new NodeHttpHandler();\n\t\t\t}\n\t\t} else {\n\t\t\t// Non-Node environment (browser): fall back to us-east-1 since\n\t\t\t// there's no config file resolution available.\n\t\t\tconfig.region =\n\t\t\t\tconfiguredRegion || (endpointRegion && useExplicitEndpoint ? endpointRegion : undefined) || \"us-east-1\";\n\t\t}\n\n\t\tif (useBearerToken) {\n\t\t\tconfig.token = { token: bearerToken };\n\t\t\tconfig.authSchemePreference = [\"httpBearerAuth\"];\n\t\t}\n\n\t\t// Kept outside the try so the catch can still correlate a mid-stream failure:\n\t\t// exceptions delivered as stream events carry no HTTP metadata of their own.\n\t\tlet responseRequestId: string | undefined;\n\n\t\ttry {\n\t\t\tconst supportsStrictMode = model.compat?.supportsStrictMode ?? false;\n\t\t\tconst client = new BedrockRuntimeClient(config);\n\t\t\tlet observedRawResponse = false;\n\t\t\tif (options.onResponse) {\n\t\t\t\taddResponseHeadersMiddleware(client, options.onResponse, model, () => {\n\t\t\t\t\tobservedRawResponse = true;\n\t\t\t\t});\n\t\t\t}\n\t\t\tconst customHeaders = providerHeadersToRecord({ ...model.headers, ...options.headers });\n\t\t\tif (customHeaders) {\n\t\t\t\taddCustomHeadersMiddleware(client, customHeaders);\n\t\t\t}\n\t\t\tconst cacheRetention = resolveCacheRetention(options.cacheRetention, options.env);\n\t\t\tconst inferenceMaxTokens = options.maxTokens ?? (isAnthropicClaudeModel(model) ? model.maxTokens : undefined);\n\t\t\tlet commandInput = {\n\t\t\t\tmodelId: model.id,\n\t\t\t\tmessages: convertMessages(context, model, cacheRetention, options.env),\n\t\t\t\tsystem: buildSystemPrompt(context.systemPrompt, model, cacheRetention, options.env),\n\t\t\t\tinferenceConfig: {\n\t\t\t\t\t...(inferenceMaxTokens !== undefined && { maxTokens: inferenceMaxTokens }),\n\t\t\t\t\t// Claude Fable 5.1 rejects non-default `temperature`, `top_p`, and `top_k` on every\n\t\t\t\t\t// request. Generated metadata marks such models `supportsTemperature: false`; every\n\t\t\t\t\t// other model keeps sending the field exactly as before.\n\t\t\t\t\t// https://platform.claude.com/docs/en/build-with-claude/thinking\n\t\t\t\t\t...(options.temperature !== undefined &&\n\t\t\t\t\t\tmodel.compat?.supportsTemperature !== false && { temperature: options.temperature }),\n\t\t\t\t},\n\t\t\t\ttoolConfig: convertToolConfig(context.tools, options.toolChoice, supportsStrictMode, model),\n\t\t\t\tadditionalModelRequestFields: buildAdditionalModelRequestFields(model, options),\n\t\t\t\t...(options.requestMetadata !== undefined && { requestMetadata: options.requestMetadata }),\n\t\t\t};\n\t\t\tconst nextCommandInput = await options?.onPayload?.(commandInput, model);\n\t\t\tif (nextCommandInput !== undefined) {\n\t\t\t\tcommandInput = nextCommandInput as typeof commandInput;\n\t\t\t}\n\t\t\tconst command = new ConverseStreamCommand(commandInput);\n\n\t\t\tconst response = await client.send(command, { abortSignal: options.signal });\n\t\t\tresponseRequestId = normalizeDiagnosticValue(response.$metadata.requestId);\n\t\t\tif (!observedRawResponse && response.$metadata.httpStatusCode !== undefined) {\n\t\t\t\tconst responseHeaders: Record<string, string> = {};\n\t\t\t\tif (response.$metadata.requestId) {\n\t\t\t\t\tresponseHeaders[\"x-amzn-requestid\"] = response.$metadata.requestId;\n\t\t\t\t}\n\t\t\t\tawait options?.onResponse?.({ status: response.$metadata.httpStatusCode, headers: responseHeaders }, model);\n\t\t\t}\n\n\t\t\tfor await (const item of response.stream!) {\n\t\t\t\tif (item.messageStart) {\n\t\t\t\t\tif (item.messageStart.role !== ConversationRole.ASSISTANT) {\n\t\t\t\t\t\tthrow new Error(\"Unexpected assistant message start but got user message start instead\");\n\t\t\t\t\t}\n\t\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t\t} else if (item.contentBlockStart) {\n\t\t\t\t\thandleContentBlockStart(item.contentBlockStart, blocks, output, stream);\n\t\t\t\t} else if (item.contentBlockDelta) {\n\t\t\t\t\thandleContentBlockDelta(item.contentBlockDelta, blocks, output, stream);\n\t\t\t\t} else if (item.contentBlockStop) {\n\t\t\t\t\thandleContentBlockStop(item.contentBlockStop, blocks, output, stream);\n\t\t\t\t} else if (item.messageStop) {\n\t\t\t\t\toutput.rawStopReason = item.messageStop.stopReason;\n\t\t\t\t\tconst { stopReason, errorMessage } = mapStopReason(item.messageStop.stopReason);\n\t\t\t\t\toutput.stopReason = stopReason;\n\t\t\t\t\tif (errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = errorMessage;\n\t\t\t\t\t}\n\t\t\t\t} else if (item.metadata) {\n\t\t\t\t\thandleMetadata(item.metadata, model, output);\n\t\t\t\t} else if (item.internalServerException) {\n\t\t\t\t\tthrow item.internalServerException;\n\t\t\t\t} else if (item.modelStreamErrorException) {\n\t\t\t\t\tthrow item.modelStreamErrorException;\n\t\t\t\t} else if (item.validationException) {\n\t\t\t\t\tthrow item.validationException;\n\t\t\t\t} else if (item.throttlingException) {\n\t\t\t\t\tthrow item.throttlingException;\n\t\t\t\t} else if (item.serviceUnavailableException) {\n\t\t\t\t\tthrow item.serviceUnavailableException;\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (options.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"pending\") {\n\t\t\t\tthrow new Error(\"Bedrock stream ended without a stop reason\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\" || output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"An unknown error occurred\");\n\t\t\t}\n\n\t\t\t// A stream can settle without stopping every block, so finalize here too.\n\t\t\tfor (const block of output.content) finalizeStreamingBlock(block as Block);\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tfinalizeStreamingBlock(block as Block);\n\t\t\t}\n\t\t\toutput.stopReason = options.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatBedrockError(error);\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tappendBedrockFailureDiagnostic(output, error, responseRequestId);\n\t\t\t}\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\n/**\n * Human-readable prefixes for Bedrock SDK exception names.\n * The downstream retry logic in agent-session matches patterns like\n * `server.?error` and `service.?unavailable`, so we preserve the legacy\n * prefix format rather than using the raw SDK exception name.\n */\nconst BEDROCK_ERROR_PREFIXES: Record<string, string> = {\n\tInternalServerException: \"Internal server error\",\n\tModelStreamErrorException: \"Model stream error\",\n\tValidationException: \"Validation error\",\n\tThrottlingException: \"Throttling error\",\n\tServiceUnavailableException: \"Service unavailable\",\n};\n\n/**\n * Some models reject the account/profile's configured Bedrock data retention mode\n * (e.g. \"data retention mode 'default' is not available for this model\"). Point\n * users at the AWS docs explaining how to configure a supported mode.\n */\nconst BEDROCK_DATA_RETENTION_DOCS_URL = \"https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html\";\n\n/**\n * Format a Bedrock error with a human-readable prefix.\n * AWS SDK exceptions (both from `client.send()` and from stream event items)\n * extend BedrockRuntimeServiceException. We map the `.name` to a stable\n * human-readable prefix so downstream consumers (retry logic, context-overflow\n * detection) can distinguish error categories via simple string matching.\n */\nfunction formatBedrockError(error: unknown): string {\n\tconst norm = normalizeProviderError(error);\n\t// Surface the raw HTTP body (with status) when the SDK did not fold it into\n\t// the message; otherwise fall back to the message. This is what stops a\n\t// gateway 403 from collapsing to `Unknown: UnknownError`.\n\tconst core =\n\t\t!norm.messageCarriesBody && norm.status !== undefined && norm.body !== undefined\n\t\t\t? `${norm.status}: ${norm.body}`\n\t\t\t: norm.message;\n\tconst dataRetentionHint = /data retention mode/i.test(core)\n\t\t? ` See ${BEDROCK_DATA_RETENTION_DOCS_URL} for supported data retention modes.`\n\t\t: \"\";\n\tif (error instanceof BedrockRuntimeServiceException) {\n\t\tconst prefix = BEDROCK_ERROR_PREFIXES[error.name] ?? error.name;\n\t\treturn `${prefix}: ${core}${dataRetentionHint}`;\n\t}\n\treturn `${core}${dataRetentionHint}`;\n}\n\ntype SdkErrorMetadata = { $metadata?: { httpStatusCode?: unknown; requestId?: unknown } };\n\n/** Over-long header values are dropped rather than truncated: a truncated request id is not a request id. */\nconst MAX_BEDROCK_DIAGNOSTIC_VALUE_CHARS = 200;\n\nfunction normalizeDiagnosticValue(value: unknown): string | undefined {\n\tif (typeof value !== \"string\") return undefined;\n\tconst trimmed = value.trim();\n\tif (trimmed.length === 0 || trimmed.length > MAX_BEDROCK_DIAGNOSTIC_VALUE_CHARS) return undefined;\n\treturn trimmed;\n}\n\n/**\n * The SDK puts the modeled code on `error.name` for service exceptions and unmodeled stream errors alike, so\n * do not narrow to `BedrockRuntimeServiceException`. Modeled Bedrock errors all end in `Exception`, unlike\n * transport names such as `TimeoutError`.\n */\nfunction extractBedrockErrorCode(error: unknown): string | undefined {\n\tif (!(error instanceof Error) || !error.name.endsWith(\"Exception\")) return undefined;\n\treturn normalizeDiagnosticValue(error.name);\n}\n\n/**\n * Structured metadata alongside `errorMessage`, which stays byte-identical because `isRetryableAssistantError`\n * matches against it. Unknown fields are omitted, never guessed: a modeled mid-stream exception reaches us as\n * a bare object literal, leaving only `fallbackRequestId`. `details` only, as the throw is not always `Error`.\n */\nfunction appendBedrockFailureDiagnostic(\n\toutput: AssistantMessage,\n\terror: unknown,\n\tfallbackRequestId: string | undefined,\n): void {\n\tconst metadata = (error as SdkErrorMetadata)?.$metadata;\n\tconst details: Record<string, unknown> = {};\n\n\tif (typeof metadata?.httpStatusCode === \"number\") details.status = metadata.httpStatusCode;\n\n\tconst errorCode = extractBedrockErrorCode(error);\n\tif (errorCode !== undefined) details.errorCode = errorCode;\n\n\tconst requestId = normalizeDiagnosticValue(metadata?.requestId) ?? fallbackRequestId;\n\tif (requestId !== undefined) details.requestId = requestId;\n\n\tif (Object.keys(details).length === 0) return;\n\n\tappendAssistantMessageDiagnostic(output, { type: \"bedrock_response_failure\", timestamp: Date.now(), details });\n}\n\n/**\n * Header keys that must never be overwritten by caller-supplied headers.\n * `host` and `x-amz-*` participate in the SigV4 canonical request; `authorization`\n * is owned by SigV4 or the bearer-token path (config.token + authSchemePreference).\n * Compared case-insensitively (caller key is lower-cased before lookup).\n */\nconst RESERVED_HEADER_EXACT = new Set([\"authorization\", \"host\"]);\n\nfunction isReservedHeader(key: string): boolean {\n\tconst lower = key.toLowerCase();\n\treturn lower.startsWith(\"x-amz-\") || RESERVED_HEADER_EXACT.has(lower);\n}\n\n/**\n * Attach caller-supplied headers to the outgoing Bedrock request via a Smithy\n * `build`-step middleware. The `build` step runs after request serialisation but\n * before SigV4 signing, so injected headers are covered by the signature. Reserved\n * SigV4 / auth headers (`x-amz-*`, `authorization`, `host`) are silently skipped;\n * all other caller headers override any existing same-named header on the request.\n */\nfunction addCustomHeadersMiddleware(client: BedrockRuntimeClient, headers: Record<string, string>): void {\n\tconst middleware: BuildMiddleware<object, MetadataBearer> = (next) => async (args) => {\n\t\tconst request = args.request;\n\t\tif (request && typeof request === \"object\" && \"headers\" in request) {\n\t\t\tconst requestHeaders = (request as { headers: Record<string, string> }).headers;\n\t\t\tfor (const [key, value] of Object.entries(headers)) {\n\t\t\t\tif (!isReservedHeader(key)) {\n\t\t\t\t\trequestHeaders[key] = value;\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t\treturn next(args);\n\t};\n\tclient.middlewareStack.add(middleware, { step: \"build\", name: \"pi-ai-custom-headers\", priority: \"low\" });\n}\n\nfunction isSmithyHttpResponse(response: unknown): response is HttpResponse {\n\tif (!response || typeof response !== \"object\") return false;\n\tconst candidate = response as Partial<HttpResponse>;\n\treturn typeof candidate.statusCode === \"number\" && !!candidate.headers && typeof candidate.headers === \"object\";\n}\n\nfunction toProviderResponse(response: unknown): ProviderResponse | undefined {\n\tif (!isSmithyHttpResponse(response)) return undefined;\n\treturn { status: response.statusCode, headers: { ...response.headers } };\n}\n\n/**\n * Bedrock's modeled `$metadata` only preserves selected HTTP metadata (for example\n * requestId), so custom gateway headers are otherwise lost before callers see\n * `onResponse`. Capture the raw Smithy HTTP response at the deserialize step,\n * after the SDK receives the response but before the event stream is consumed.\n */\nfunction addResponseHeadersMiddleware(\n\tclient: BedrockRuntimeClient,\n\tonResponse: NonNullable<BedrockOptions[\"onResponse\"]>,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tonObserved: () => void,\n): void {\n\tconst middleware: DeserializeMiddleware<object, MetadataBearer> = (next) => async (args) => {\n\t\tconst result = await next(args);\n\t\tconst providerResponse = toProviderResponse(result.response);\n\t\tif (providerResponse) {\n\t\t\tonObserved();\n\t\t\tawait onResponse(providerResponse, model);\n\t\t}\n\t\treturn result;\n\t};\n\tclient.middlewareStack.add(middleware, { step: \"deserialize\", name: \"pi-ai-response-headers\" });\n}\n\nexport const streamSimple: StreamFunction<\"bedrock-converse-stream\", SimpleStreamOptions> = (\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst base = {\n\t\t...buildBaseOptions(model, context, options, undefined),\n\t\ttoolChoice: options?.toolChoice,\n\t} satisfies BedrockOptions;\n\tif (!options?.reasoning) {\n\t\treturn stream(model, context, { ...base, reasoning: undefined } satisfies BedrockOptions);\n\t}\n\n\tif (isAnthropicClaudeModel(model)) {\n\t\tif (supportsAdaptiveThinking(model.id, model.name)) {\n\t\t\treturn stream(model, context, {\n\t\t\t\t...base,\n\t\t\t\treasoning: options.reasoning,\n\t\t\t\tthinkingBudgets: options.thinkingBudgets,\n\t\t\t} satisfies BedrockOptions);\n\t\t}\n\n\t\t// Undefined means the caller did not request an output cap; let the helper use the model cap.\n\t\t// Do not coerce to 0 here, or the thinking budget would become the entire maxTokens value.\n\t\tconst adjusted = adjustMaxTokensForThinking(\n\t\t\tbase.maxTokens,\n\t\t\tmodel.maxTokens,\n\t\t\toptions.reasoning,\n\t\t\toptions.thinkingBudgets,\n\t\t);\n\n\t\tconst maxTokens = clampMaxTokensToContext(model, context, adjusted.maxTokens);\n\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tmaxTokens,\n\t\t\treasoning: options.reasoning,\n\t\t\tthinkingBudgets: {\n\t\t\t\t...(options.thinkingBudgets || {}),\n\t\t\t\t[clampReasoning(options.reasoning)!]: Math.min(adjusted.thinkingBudget, Math.max(0, maxTokens - 1024)),\n\t\t\t},\n\t\t} satisfies BedrockOptions);\n\t}\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoning: options.reasoning,\n\t\tthinkingBudgets: options.thinkingBudgets,\n\t} satisfies BedrockOptions);\n};\n\nfunction handleContentBlockStart(\n\tevent: ContentBlockStartEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst index = event.contentBlockIndex!;\n\tconst start = event.start;\n\n\tif (start?.toolUse) {\n\t\tconst block: Block = {\n\t\t\ttype: \"toolCall\",\n\t\t\tid: start.toolUse.toolUseId || \"\",\n\t\t\tname: start.toolUse.name || \"\",\n\t\t\targuments: {},\n\t\t\tpartialJson: \"\",\n\t\t\tindex,\n\t\t};\n\t\toutput.content.push(block);\n\t\tstream.push({ type: \"toolcall_start\", contentIndex: blocks.length - 1, partial: output });\n\t}\n}\n\nfunction handleContentBlockDelta(\n\tevent: ContentBlockDeltaEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst contentBlockIndex = event.contentBlockIndex!;\n\tconst delta = event.delta;\n\tlet index = blocks.findIndex((b) => b.index === contentBlockIndex);\n\tlet block = blocks[index];\n\n\tif (delta?.text !== undefined) {\n\t\t// If no text block exists yet, create one, as `handleContentBlockStart` is not sent for text blocks\n\t\tif (!block) {\n\t\t\tconst newBlock: Block = { type: \"text\", text: \"\", index: contentBlockIndex };\n\t\t\toutput.content.push(newBlock);\n\t\t\tindex = blocks.length - 1;\n\t\t\tblock = blocks[index];\n\t\t\tstream.push({ type: \"text_start\", contentIndex: index, partial: output });\n\t\t}\n\t\tif (block.type === \"text\") {\n\t\t\tblock.text += delta.text;\n\t\t\tstream.push({ type: \"text_delta\", contentIndex: index, delta: delta.text, partial: output });\n\t\t}\n\t} else if (delta?.toolUse && block?.type === \"toolCall\") {\n\t\tblock.partialJson = (block.partialJson || \"\") + (delta.toolUse.input || \"\");\n\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\tstream.push({ type: \"toolcall_delta\", contentIndex: index, delta: delta.toolUse.input || \"\", partial: output });\n\t} else if (delta?.reasoningContent) {\n\t\tlet thinkingBlock = block;\n\t\tlet thinkingIndex = index;\n\n\t\tif (!thinkingBlock) {\n\t\t\tconst newBlock: Block = { type: \"thinking\", thinking: \"\", thinkingSignature: \"\", index: contentBlockIndex };\n\t\t\toutput.content.push(newBlock);\n\t\t\tthinkingIndex = blocks.length - 1;\n\t\t\tthinkingBlock = blocks[thinkingIndex];\n\t\t\tstream.push({ type: \"thinking_start\", contentIndex: thinkingIndex, partial: output });\n\t\t}\n\n\t\tif (thinkingBlock?.type === \"thinking\") {\n\t\t\tif (delta.reasoningContent.text) {\n\t\t\t\tthinkingBlock.thinking += delta.reasoningContent.text;\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\tcontentIndex: thinkingIndex,\n\t\t\t\t\tdelta: delta.reasoningContent.text,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t}\n\t\t\t// `thinkingSignature` holds either an Anthropic signature or an opaque redacted\n\t\t\t// payload, never both: mixing them would corrupt whichever arrived first.\n\t\t\tif (delta.reasoningContent.signature && !thinkingBlock.redacted) {\n\t\t\t\tthinkingBlock.thinkingSignature =\n\t\t\t\t\t(thinkingBlock.thinkingSignature || \"\") + delta.reasoningContent.signature;\n\t\t\t}\n\t\t\tif (delta.reasoningContent.redactedContent?.length) {\n\t\t\t\t// Encrypted reasoning from non-Anthropic models on Bedrock (e.g. OpenAI GPT-5.6).\n\t\t\t\t// The payload is opaque, so keep it verbatim in `thinkingSignature` the way the\n\t\t\t\t// Anthropic path stores redacted thinking, and replay it on the next turn.\n\t\t\t\tif (!thinkingBlock.redacted) {\n\t\t\t\t\tthinkingBlock.redacted = true;\n\t\t\t\t\tthinkingBlock.thinkingSignature = \"\";\n\t\t\t\t\tthinkingBlock.thinking += REDACTED_THINKING_PLACEHOLDER;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\tcontentIndex: thinkingIndex,\n\t\t\t\t\t\tdelta: REDACTED_THINKING_PLACEHOLDER,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tthinkingBlock.redactedChunks ??= [];\n\t\t\t\tthinkingBlock.redactedChunks.push(delta.reasoningContent.redactedContent);\n\t\t\t}\n\t\t}\n\t}\n}\n\n/**\n * Encodes buffered encrypted reasoning into `thinkingSignature` and drops the scratch\n * buffer, which must never reach a persisted message: `Uint8Array` serializes to an\n * index-keyed object roughly ten times the size of the base64 payload.\n */\nfunction flushRedactedContent(block: Block): void {\n\tif (block.type !== \"thinking\" || !block.redactedChunks) return;\n\tblock.thinkingSignature = bytesToBase64(block.redactedChunks);\n\tdelete block.redactedChunks;\n}\n\n/**\n * Strips every streaming scratch field. Runs from the terminal paths as well as\n * `contentBlockStop`, because a stream can settle without stopping each block.\n */\nfunction finalizeStreamingBlock(block: Block): void {\n\tdelete block.index;\n\t// partialJson is only a streaming scratch buffer; never persist it.\n\tdelete block.partialJson;\n\tflushRedactedContent(block);\n}\n\nfunction handleMetadata(\n\tevent: ConverseStreamMetadataEvent,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\toutput: AssistantMessage,\n): void {\n\tif (event.usage) {\n\t\toutput.usage.input = event.usage.inputTokens || 0;\n\t\toutput.usage.output = event.usage.outputTokens || 0;\n\t\toutput.usage.cacheRead = event.usage.cacheReadInputTokens || 0;\n\t\toutput.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;\n\t\toutput.usage.cacheWrite1h = event.usage.cacheDetails?.reduce(\n\t\t\t(total, detail) => total + (detail.ttl === CacheTTL.ONE_HOUR ? (detail.inputTokens ?? 0) : 0),\n\t\t\t0,\n\t\t);\n\t\toutput.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;\n\t\tcalculateCost(model, output.usage);\n\t}\n}\n\nfunction handleContentBlockStop(\n\tevent: ContentBlockStopEvent,\n\tblocks: Block[],\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n): void {\n\tconst index = blocks.findIndex((b) => b.index === event.contentBlockIndex);\n\tconst block = blocks[index];\n\tif (!block) return;\n\tdelete (block as Block).index;\n\n\tswitch (block.type) {\n\t\tcase \"text\":\n\t\t\tstream.push({ type: \"text_end\", contentIndex: index, content: block.text, partial: output });\n\t\t\tbreak;\n\t\tcase \"thinking\":\n\t\t\tflushRedactedContent(block);\n\t\t\tstream.push({ type: \"thinking_end\", contentIndex: index, content: block.thinking, partial: output });\n\t\t\tbreak;\n\t\tcase \"toolCall\":\n\t\t\tblock.arguments = parseStreamingJson(block.partialJson);\n\t\t\t// Finalize in-place and strip the scratch buffer so replay only\n\t\t\t// carries parsed arguments.\n\t\t\tdelete (block as Block).partialJson;\n\t\t\tstream.push({ type: \"toolcall_end\", contentIndex: index, toolCall: block, partial: output });\n\t\t\tbreak;\n\t}\n}\n\n/**\n * Check if the model supports adaptive thinking (Opus 4.6+, Sonnet 4.6).\n * Checks both model ID and model name to support application inference profiles\n * whose ARNs don't contain the model name.\n */\nfunction getModelMatchCandidates(modelId: string, modelName?: string): string[] {\n\tconst values = modelName ? [modelId, modelName] : [modelId];\n\treturn values.flatMap((value) => {\n\t\tconst lower = value.toLowerCase();\n\t\treturn [lower, lower.replace(/[\\s_.:]+/g, \"-\")];\n\t});\n}\n\nfunction supportsAdaptiveThinking(modelId: string, modelName?: string): boolean {\n\tconst candidates = getModelMatchCandidates(modelId, modelName);\n\treturn candidates.some(\n\t\t(s) =>\n\t\t\ts.includes(\"opus-4-6\") ||\n\t\t\ts.includes(\"opus-4-7\") ||\n\t\t\ts.includes(\"opus-4-8\") ||\n\t\t\ts.includes(\"opus-5\") ||\n\t\t\ts.includes(\"sonnet-4-6\") ||\n\t\t\ts.includes(\"sonnet-5\") ||\n\t\t\ts.includes(\"fable-5\"),\n\t);\n}\n\nfunction supportsNativeXhighEffort(model: Model<\"bedrock-converse-stream\">): boolean {\n\tconst candidates = getModelMatchCandidates(model.id, model.name);\n\treturn candidates.some(\n\t\t(s) =>\n\t\t\ts.includes(\"opus-4-7\") ||\n\t\t\ts.includes(\"opus-4-8\") ||\n\t\t\ts.includes(\"opus-5\") ||\n\t\t\ts.includes(\"sonnet-5\") ||\n\t\t\ts.includes(\"fable-5\"),\n\t);\n}\n\nfunction mapThinkingLevelToEffort(\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tlevel: SimpleStreamOptions[\"reasoning\"],\n): \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\" {\n\tif (level === \"xhigh\" && supportsNativeXhighEffort(model)) return \"xhigh\";\n\n\tconst mapped = level ? model.thinkingLevelMap?.[level] : undefined;\n\tif (typeof mapped === \"string\") return mapped as \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\n\tswitch (level) {\n\t\tcase \"minimal\":\n\t\tcase \"low\":\n\t\t\treturn \"low\";\n\t\tcase \"medium\":\n\t\t\treturn \"medium\";\n\t\tcase \"high\":\n\t\t\treturn \"high\";\n\t\tdefault:\n\t\t\treturn \"high\";\n\t}\n}\n\n/**\n * Resolve cache retention preference.\n * Defaults to \"short\" and uses PI_CACHE_RETENTION for backward compatibility.\n */\nfunction resolveCacheRetention(cacheRetention?: CacheRetention, env?: ProviderEnv): CacheRetention {\n\tif (cacheRetention) {\n\t\treturn cacheRetention;\n\t}\n\tif (getProviderEnvValue(\"PI_CACHE_RETENTION\", env) === \"long\") {\n\t\treturn \"long\";\n\t}\n\treturn \"short\";\n}\n\n/**\n * Check if the model is an Anthropic Claude model on Bedrock.\n * Checks both model ID and model name to support application inference profiles\n * whose ARNs don't contain the model name.\n */\nfunction isAnthropicClaudeModel(model: Model<\"bedrock-converse-stream\">): boolean {\n\tconst id = model.id.toLowerCase();\n\tconst name = model.name?.toLowerCase() ?? \"\";\n\treturn (\n\t\tid.includes(\"anthropic.claude\") ||\n\t\tid.includes(\"anthropic/claude\") ||\n\t\tname.includes(\"anthropic.claude\") ||\n\t\tname.includes(\"anthropic/claude\") ||\n\t\tname.includes(\"claude\")\n\t);\n}\n\n/**\n * Check if the model supports prompt caching.\n * Supported: Claude 3.5 Haiku, Claude 3.7 Sonnet, Claude 4.x models, Claude 5 models\n *\n * For base models and system-defined inference profiles the model ID / ARN\n * contains the model name, so we can decide locally.\n *\n * For application inference profiles (whose ARNs don't contain the model name),\n * also checks model.name which is user-controlled via models.json or registerProvider.\n * As a last resort, set AWS_BEDROCK_FORCE_CACHE=1 to enable cache points.\n * Amazon Nova models have automatic caching and don't need explicit cache points.\n */\nfunction supportsPromptCaching(model: Model<\"bedrock-converse-stream\">, env?: ProviderEnv): boolean {\n\tconst candidates = getModelMatchCandidates(model.id, model.name);\n\n\tconst hasClaudeRef = candidates.some((s) => s.includes(\"claude\"));\n\tif (!hasClaudeRef) {\n\t\t// Application inference profiles don't contain the model name in the ARN.\n\t\t// Allow users to force cache points via environment variable.\n\t\tif (getProviderEnvValue(\"AWS_BEDROCK_FORCE_CACHE\", env) === \"1\") return true;\n\t\treturn false;\n\t}\n\t// Claude 5 models (fable-5, opus-5, sonnet-5)\n\tif (candidates.some((s) => s.includes(\"fable-5\") || s.includes(\"opus-5\") || s.includes(\"sonnet-5\"))) return true;\n\t// Claude 4.x models (opus-4, sonnet-4, haiku-4)\n\tif (candidates.some((s) => s.includes(\"-4-\"))) return true;\n\t// Claude 3.7 Sonnet\n\tif (candidates.some((s) => s.includes(\"claude-3-7-sonnet\"))) return true;\n\t// Claude 3.5 Haiku\n\tif (candidates.some((s) => s.includes(\"claude-3-5-haiku\"))) return true;\n\treturn false;\n}\n\n/**\n * Check if the model supports thinking signatures in reasoningContent.\n * Only Anthropic Claude models support the signature field.\n * Other models (OpenAI, Qwen, Minimax, Moonshot, etc.) reject it with:\n * \"This model doesn't support the reasoningContent.reasoningText.signature field\"\n *\n * Checks both model ID and model name to support application inference profiles.\n */\nfunction supportsThinkingSignature(model: Model<\"bedrock-converse-stream\">): boolean {\n\treturn isAnthropicClaudeModel(model);\n}\n\nfunction buildSystemPrompt(\n\tsystemPrompt: string | undefined,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcacheRetention: CacheRetention,\n\tenv?: ProviderEnv,\n): SystemContentBlock[] | undefined {\n\tif (!systemPrompt) return undefined;\n\n\tconst blocks: SystemContentBlock[] = [{ text: sanitizeSurrogates(systemPrompt) }];\n\n\t// Add cache point for supported Claude models when caching is enabled\n\tif (cacheRetention !== \"none\" && supportsPromptCaching(model, env)) {\n\t\tblocks.push({\n\t\t\tcachePoint: { type: CachePointType.DEFAULT, ...(cacheRetention === \"long\" ? { ttl: CacheTTL.ONE_HOUR } : {}) },\n\t\t});\n\t}\n\n\treturn blocks;\n}\n\nfunction normalizeToolCallId(id: string): string {\n\tconst sanitized = id.replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\treturn sanitized.length > 64 ? sanitized.slice(0, 64) : sanitized;\n}\n\nfunction createNonBlankTextBlock(text: string): ContentBlock.TextMember | undefined {\n\tconst sanitized = sanitizeSurrogates(text);\n\treturn sanitized.trim().length === 0 ? undefined : { text: sanitized };\n}\n\nfunction createRequiredTextBlock(text: string): ContentBlock.TextMember {\n\treturn createNonBlankTextBlock(text) ?? { text: EMPTY_TEXT_PLACEHOLDER };\n}\n\nfunction sanitizeBedrockDocument(value: DocumentType): DocumentType {\n\tif (Array.isArray(value)) {\n\t\treturn value.map(sanitizeBedrockDocument);\n\t}\n\tif (value !== null && typeof value === \"object\") {\n\t\treturn Object.fromEntries(\n\t\t\tObject.entries(value)\n\t\t\t\t.filter(([key]) => key.length > 0)\n\t\t\t\t.map(([key, nestedValue]) => [key, sanitizeBedrockDocument(nestedValue)]),\n\t\t);\n\t}\n\treturn value;\n}\n\nfunction convertToolResultContent(content: (TextContent | ImageContent)[]): ToolResultContentBlock[] {\n\tconst result: ToolResultContentBlock[] = [];\n\tfor (const c of content) {\n\t\tif (c.type === \"image\") {\n\t\t\tresult.push({ image: createImageBlock(c.mimeType, c.data) });\n\t\t} else {\n\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\tif (textBlock) result.push(textBlock);\n\t\t}\n\t}\n\tif (result.length === 0) result.push({ text: EMPTY_TEXT_PLACEHOLDER });\n\treturn result;\n}\n\nfunction convertMessages(\n\tcontext: Context,\n\tmodel: Model<\"bedrock-converse-stream\">,\n\tcacheRetention: CacheRetention,\n\tenv?: ProviderEnv,\n): Message[] {\n\tconst result: Message[] = [];\n\tconst transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst m = transformedMessages[i];\n\n\t\tswitch (m.role) {\n\t\t\tcase \"user\": {\n\t\t\t\tconst content: ContentBlock[] = [];\n\t\t\t\tif (typeof m.content === \"string\") {\n\t\t\t\t\tcontent.push(createRequiredTextBlock(m.content));\n\t\t\t\t} else {\n\t\t\t\t\tfor (const c of m.content) {\n\t\t\t\t\t\tswitch (c.type) {\n\t\t\t\t\t\t\tcase \"text\": {\n\t\t\t\t\t\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\t\t\t\t\t\tif (textBlock) content.push(textBlock);\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tcase \"image\":\n\t\t\t\t\t\t\t\tcontent.push({ image: createImageBlock(c.mimeType, c.data) });\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\tcase \"document\":\n\t\t\t\t\t\t\t\tcontent.push({ document: createDocumentBlock(c, content.length) });\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\tdefault:\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tif (content.length === 0) content.push({ text: EMPTY_TEXT_PLACEHOLDER });\n\t\t\t\t}\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.USER,\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"assistant\": {\n\t\t\t\t// Skip assistant messages with empty content (e.g., from aborted requests)\n\t\t\t\t// Bedrock rejects messages with empty content arrays\n\t\t\t\tif (m.content.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tconst contentBlocks: ContentBlock[] = [];\n\t\t\t\tfor (const c of m.content) {\n\t\t\t\t\tswitch (c.type) {\n\t\t\t\t\t\tcase \"text\": {\n\t\t\t\t\t\t\t// Skip empty text blocks\n\t\t\t\t\t\t\tconst textBlock = createNonBlankTextBlock(c.text);\n\t\t\t\t\t\t\tif (!textBlock) continue;\n\t\t\t\t\t\t\tcontentBlocks.push(textBlock);\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tcase \"toolCall\":\n\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\ttoolUse: { toolUseId: c.id, name: c.name, input: sanitizeBedrockDocument(c.arguments) },\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\tcase \"thinking\": {\n\t\t\t\t\t\t\t// Encrypted reasoning is opaque: replay the stored payload as the\n\t\t\t\t\t\t\t// `redactedContent` member instead of lowering it to reasoning text.\n\t\t\t\t\t\t\tif (c.redacted) {\n\t\t\t\t\t\t\t\tconst redactedContent = decodeRedactedContent(c.thinkingSignature);\n\t\t\t\t\t\t\t\tif (redactedContent?.length) {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({ reasoningContent: { redactedContent } });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t// Skip empty thinking blocks\n\t\t\t\t\t\t\tconst thinking = sanitizeSurrogates(c.thinking);\n\t\t\t\t\t\t\tif (thinking.trim().length === 0) continue;\n\t\t\t\t\t\t\t// Only Anthropic models support the signature field in reasoningText.\n\t\t\t\t\t\t\t// For other models, we omit the signature to avoid errors like:\n\t\t\t\t\t\t\t// \"This model doesn't support the reasoningContent.reasoningText.signature field\"\n\t\t\t\t\t\t\tif (supportsThinkingSignature(model)) {\n\t\t\t\t\t\t\t\t// Signatures arrive after thinking deltas. If a partial or externally\n\t\t\t\t\t\t\t\t// persisted message lacks a signature, Bedrock rejects the replayed\n\t\t\t\t\t\t\t\t// reasoning block. Fall back to plain text, matching Anthropic.\n\t\t\t\t\t\t\t\tif (!c.thinkingSignature || c.thinkingSignature.trim().length === 0) {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({ text: thinking });\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\t\t\treasoningContent: {\n\t\t\t\t\t\t\t\t\t\t\treasoningText: {\n\t\t\t\t\t\t\t\t\t\t\t\ttext: thinking,\n\t\t\t\t\t\t\t\t\t\t\t\tsignature: c.thinkingSignature,\n\t\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\tcontentBlocks.push({\n\t\t\t\t\t\t\t\t\treasoningContent: {\n\t\t\t\t\t\t\t\t\t\treasoningText: { text: thinking },\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tdefault:\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\t// Skip if all content blocks were filtered out\n\t\t\t\tif (contentBlocks.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.ASSISTANT,\n\t\t\t\t\tcontent: contentBlocks,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"toolResult\": {\n\t\t\t\t// Collect all consecutive toolResult messages into a single user message\n\t\t\t\t// Bedrock requires all tool results to be in one message\n\t\t\t\tconst toolResults: ContentBlock.ToolResultMember[] = [];\n\n\t\t\t\t// Add current tool result with all content blocks combined\n\t\t\t\ttoolResults.push({\n\t\t\t\t\ttoolResult: {\n\t\t\t\t\t\ttoolUseId: m.toolCallId,\n\t\t\t\t\t\tcontent: convertToolResultContent(m.content),\n\t\t\t\t\t\tstatus: m.isError ? ToolResultStatus.ERROR : ToolResultStatus.SUCCESS,\n\t\t\t\t\t},\n\t\t\t\t});\n\n\t\t\t\t// Look ahead for consecutive toolResult messages\n\t\t\t\tlet j = i + 1;\n\t\t\t\twhile (j < transformedMessages.length && transformedMessages[j].role === \"toolResult\") {\n\t\t\t\t\tconst nextMsg = transformedMessages[j] as ToolResultMessage;\n\t\t\t\t\ttoolResults.push({\n\t\t\t\t\t\ttoolResult: {\n\t\t\t\t\t\t\ttoolUseId: nextMsg.toolCallId,\n\t\t\t\t\t\t\tcontent: convertToolResultContent(nextMsg.content),\n\t\t\t\t\t\t\tstatus: nextMsg.isError ? ToolResultStatus.ERROR : ToolResultStatus.SUCCESS,\n\t\t\t\t\t\t},\n\t\t\t\t\t});\n\t\t\t\t\tj++;\n\t\t\t\t}\n\n\t\t\t\t// Skip the messages we've already processed\n\t\t\t\ti = j - 1;\n\n\t\t\t\tresult.push({\n\t\t\t\t\trole: ConversationRole.USER,\n\t\t\t\t\tcontent: toolResults,\n\t\t\t\t});\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tdefault:\n\t\t\t\tcontinue;\n\t\t}\n\t}\n\n\t// Add cache point to the last user message for supported Claude models when caching is enabled\n\tif (cacheRetention !== \"none\" && supportsPromptCaching(model, env) && result.length > 0) {\n\t\tconst lastMessage = result[result.length - 1];\n\t\tif (lastMessage.role === ConversationRole.USER && lastMessage.content) {\n\t\t\t(lastMessage.content as ContentBlock[]).push({\n\t\t\t\tcachePoint: {\n\t\t\t\t\ttype: CachePointType.DEFAULT,\n\t\t\t\t\t...(cacheRetention === \"long\" ? { ttl: CacheTTL.ONE_HOUR } : {}),\n\t\t\t\t},\n\t\t\t});\n\t\t}\n\t}\n\n\treturn result;\n}\n\nfunction convertToolConfig(\n\ttools: Tool[] | undefined,\n\ttoolChoice: BedrockOptions[\"toolChoice\"],\n\tsupportsStrictMode: boolean,\n\tmodel: Model<\"bedrock-converse-stream\">,\n): ToolConfiguration | undefined {\n\t// Validate the requested choice before any early return. A forced choice on a model that\n\t// rejects it is an error regardless of whether tools were supplied: returning `undefined` for\n\t// an empty or absent tool list would discard the caller's instruction silently, which is the\n\t// behavior this guard exists to prevent.\n\t//\n\t// Claude Fable 5.1 rejects forced tool use on every request with a 400, whichever platform\n\t// serves it. Reject rather than rewriting to `auto`, which would discard an explicit caller\n\t// instruction. Mirrors the guards in `anthropic-messages.ts` and `openai-completions.ts`.\n\t// https://platform.claude.com/docs/en/build-with-claude/thinking\n\tif (model.compat?.supportsForcedToolChoice === false) {\n\t\tconst isForced = toolChoice === \"any\" || (typeof toolChoice === \"object\" && toolChoice.type === \"tool\");\n\t\tif (isForced) {\n\t\t\tconst requestedLabel = typeof toolChoice === \"string\" ? toolChoice : `tool \"${toolChoice.name}\"`;\n\t\t\tthrow new Error(\n\t\t\t\t`Model ${model.id} does not support forced tool choice (requested: ${requestedLabel}). ` +\n\t\t\t\t\t`Use toolChoice \"auto\" with strict tool use or structured outputs instead.`,\n\t\t\t);\n\t\t}\n\t}\n\n\tif (!tools?.length) return undefined;\n\t// `none` is never a forced choice, so this return can never skip a rejection above.\n\tif (toolChoice === \"none\") return undefined;\n\n\tconst bedrockTools: BedrockTool[] = tools.map((tool) => {\n\t\tconst strict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);\n\t\treturn {\n\t\t\ttoolSpec: {\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tinputSchema: { json: getJsonSchemaToolParameters(tool, strict) as unknown as DocumentType },\n\t\t\t\t...(strict === true ? { strict: true } : {}),\n\t\t\t},\n\t\t};\n\t});\n\n\tlet bedrockToolChoice: ToolChoice | undefined;\n\tswitch (toolChoice) {\n\t\tcase \"auto\":\n\t\t\tbedrockToolChoice = { auto: {} };\n\t\t\tbreak;\n\t\tcase \"any\":\n\t\t\tbedrockToolChoice = { any: {} };\n\t\t\tbreak;\n\t\tdefault:\n\t\t\tif (toolChoice?.type === \"tool\") {\n\t\t\t\tbedrockToolChoice = { tool: { name: toolChoice.name } };\n\t\t\t}\n\t}\n\n\treturn { tools: bedrockTools, toolChoice: bedrockToolChoice };\n}\n\nfunction mapStopReason(reason: string | undefined): { stopReason: StopReason; errorMessage?: string } {\n\tswitch (reason) {\n\t\tcase BedrockStopReason.END_TURN:\n\t\tcase BedrockStopReason.STOP_SEQUENCE:\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase BedrockStopReason.MAX_TOKENS:\n\t\tcase BedrockStopReason.MODEL_CONTEXT_WINDOW_EXCEEDED:\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase BedrockStopReason.TOOL_USE:\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tdefault:\n\t\t\treturn reason\n\t\t\t\t? { stopReason: \"error\", errorMessage: `Provider stopped with: ${reason}` }\n\t\t\t\t: { stopReason: \"error\" };\n\t}\n}\n\nfunction getConfiguredBedrockRegion(options: BedrockOptions): string | undefined {\n\treturn (\n\t\toptions.region ||\n\t\tgetProviderEnvValue(\"AWS_REGION\", options.env) ||\n\t\tgetProviderEnvValue(\"AWS_DEFAULT_REGION\", options.env) ||\n\t\tundefined\n\t);\n}\n\nfunction getConfiguredBedrockCredentials(env?: ProviderEnv): BedrockRuntimeClientConfig[\"credentials\"] | undefined {\n\tconst accessKeyId = getProviderEnvValue(\"AWS_ACCESS_KEY_ID\", env);\n\tconst secretAccessKey = getProviderEnvValue(\"AWS_SECRET_ACCESS_KEY\", env);\n\tif (!accessKeyId || !secretAccessKey) {\n\t\treturn undefined;\n\t}\n\tconst sessionToken = getProviderEnvValue(\"AWS_SESSION_TOKEN\", env);\n\treturn {\n\t\taccessKeyId,\n\t\tsecretAccessKey,\n\t\t...(sessionToken ? { sessionToken } : {}),\n\t};\n}\n\nfunction getStandardBedrockEndpointRegion(baseUrl: string | undefined): string | undefined {\n\tif (!baseUrl) {\n\t\treturn undefined;\n\t}\n\n\ttry {\n\t\tconst { hostname } = new URL(baseUrl);\n\t\tconst match = hostname.toLowerCase().match(/^bedrock-runtime(?:-fips)?\\.([a-z0-9-]+)\\.amazonaws\\.com(?:\\.cn)?$/);\n\t\treturn match?.[1];\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction shouldUseExplicitBedrockEndpoint(\n\tbaseUrl: string,\n\tconfiguredRegion: string | undefined,\n\thasAmbientConfiguredProfile: boolean,\n): boolean {\n\tconst endpointRegion = getStandardBedrockEndpointRegion(baseUrl);\n\tif (!endpointRegion) {\n\t\treturn true;\n\t}\n\n\treturn !configuredRegion && !hasAmbientConfiguredProfile;\n}\n\nfunction isGovCloudBedrockTarget(model: Model<\"bedrock-converse-stream\">, options: BedrockOptions): boolean {\n\tconst region = getConfiguredBedrockRegion(options);\n\tif (region?.toLowerCase().startsWith(\"us-gov-\")) {\n\t\treturn true;\n\t}\n\n\tconst modelId = model.id.toLowerCase();\n\treturn modelId.startsWith(\"us-gov.\") || modelId.startsWith(\"arn:aws-us-gov:\");\n}\n\nfunction isOpenAiGpt6AstraBedrockModel(model: Model<\"bedrock-converse-stream\">): boolean {\n\treturn /^(?:global\\.|us\\.)?openai\\.gpt-6-astra$/.test(model.id);\n}\n\nfunction buildAdditionalModelRequestFields(\n\tmodel: Model<\"bedrock-converse-stream\">,\n\toptions: BedrockOptions,\n): Record<string, any> | undefined {\n\tif (!options.reasoning || !model.reasoning) {\n\t\treturn undefined;\n\t}\n\n\tif (isAnthropicClaudeModel(model)) {\n\t\t// GovCloud Bedrock currently rejects the Claude thinking.display field.\n\t\t// Omit it there until the GovCloud Converse schema catches up.\n\t\tconst display = isGovCloudBedrockTarget(model, options) ? undefined : (options.thinkingDisplay ?? \"summarized\");\n\t\tconst result: Record<string, any> = supportsAdaptiveThinking(model.id, model.name)\n\t\t\t? {\n\t\t\t\t\tthinking: { type: \"adaptive\", ...(display !== undefined ? { display } : {}) },\n\t\t\t\t\toutput_config: { effort: mapThinkingLevelToEffort(model, options.reasoning) },\n\t\t\t\t}\n\t\t\t: (() => {\n\t\t\t\t\tconst defaultBudgets: Record<ThinkingLevel, number> = {\n\t\t\t\t\t\tminimal: 1024,\n\t\t\t\t\t\tlow: 2048,\n\t\t\t\t\t\tmedium: 8192,\n\t\t\t\t\t\thigh: 16384,\n\t\t\t\t\t\txhigh: 16384, // Budget-based Claude clamps extended levels to high\n\t\t\t\t\t\tmax: 16384,\n\t\t\t\t\t};\n\n\t\t\t\t\t// Custom budgets only cover token-based levels through high.\n\t\t\t\t\tconst level = options.reasoning === \"xhigh\" || options.reasoning === \"max\" ? \"high\" : options.reasoning;\n\t\t\t\t\tconst budget = options.thinkingBudgets?.[level] ?? defaultBudgets[options.reasoning];\n\n\t\t\t\t\treturn {\n\t\t\t\t\t\tthinking: {\n\t\t\t\t\t\t\ttype: \"enabled\",\n\t\t\t\t\t\t\tbudget_tokens: budget,\n\t\t\t\t\t\t\t...(display !== undefined ? { display } : {}),\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t})();\n\n\t\tif (!supportsAdaptiveThinking(model.id, model.name) && (options.interleavedThinking ?? true)) {\n\t\t\tresult.anthropic_beta = [\"interleaved-thinking-2025-05-14\"];\n\t\t}\n\n\t\treturn result;\n\t}\n\n\tif (isOpenAiGpt6AstraBedrockModel(model)) {\n\t\t// Bedrock maps OpenAI Chat Completions fields that are not part of Converse's\n\t\t// inferenceConfig into additionalModelRequestFields. GPT-6-Astra uses the Chat\n\t\t// Completions spelling rather than the Responses API's nested reasoning object.\n\t\t// https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-openai.html\n\t\treturn { reasoning_effort: mapThinkingLevelToEffort(model, options.reasoning) };\n\t}\n\n\treturn undefined;\n}\n\n/**\n * Build a Bedrock `DocumentBlock`. Three things differ from the Anthropic path:\n *\n * - `name` is **required**, and AWS restricts it to alphanumerics, single whitespace runs,\n *   hyphens, parentheses, and square brackets — then warns the field \"is vulnerable to prompt\n *   injections… we recommend that you specify a neutral name.\" Any caller-supplied name is\n *   sanitized to that character set, and anything left empty falls back to a generated label.\n * - `source` takes **raw bytes**, not base64: \"If you use an Amazon Web Services SDK, you don't\n *   need to encode the bytes in base64.\" So the base64 payload is decoded here, the opposite of\n *   what the Anthropic converter does.\n * - Without citations enabled, Converse degrades to plain text extraction rather than full visual\n *   PDF understanding. That limitation is documented rather than worked around here.\n *   https://platform.claude.com/docs/en/build-with-claude/pdf-support\n *\n * `DocumentFormat` also has eight non-PDF members, but `format` is hardcoded to PDF because that\n * is the only media type `DocumentContent` declares — and the seven office/markup formats have no\n * Anthropic source variant to map onto. The block's own media type is therefore verified rather\n * than read.\n */\nfunction createDocumentBlock(block: DocumentContent, index: number) {\n\tassertSupportedDocumentMimeType(block);\n\tconst sanitized = (block.name ?? \"\")\n\t\t.replace(/[^a-zA-Z0-9\\s\\-()[\\]]/g, \" \")\n\t\t.replace(/\\s+/g, \" \")\n\t\t.trim();\n\treturn {\n\t\tformat: DocumentFormat.PDF,\n\t\tname: sanitized.length > 0 ? sanitized : `document ${index + 1}`,\n\t\tsource: { bytes: base64ToBytes(block.data) },\n\t};\n}\n\nfunction createImageBlock(mimeType: string, data: string) {\n\tlet format: ImageFormat;\n\tswitch (mimeType) {\n\t\tcase \"image/jpeg\":\n\t\tcase \"image/jpg\":\n\t\t\tformat = ImageFormat.JPEG;\n\t\t\tbreak;\n\t\tcase \"image/png\":\n\t\t\tformat = ImageFormat.PNG;\n\t\t\tbreak;\n\t\tcase \"image/gif\":\n\t\t\tformat = ImageFormat.GIF;\n\t\t\tbreak;\n\t\tcase \"image/webp\":\n\t\t\tformat = ImageFormat.WEBP;\n\t\t\tbreak;\n\t\tdefault:\n\t\t\tthrow new Error(`Unknown image type: ${mimeType}`);\n\t}\n\n\treturn { source: { bytes: base64ToBytes(data) }, format };\n}\n\nfunction base64ToBytes(data: string): Uint8Array {\n\tconst binaryString = atob(data);\n\tconst bytes = new Uint8Array(binaryString.length);\n\tfor (let i = 0; i < binaryString.length; i++) {\n\t\tbytes[i] = binaryString.charCodeAt(i);\n\t}\n\treturn bytes;\n}\n\n/**\n * Decodes a stored redacted payload. The AWS SDK hands the blob over as bytes, but a\n * persisted session carries it as base64. A hand-edited or externally produced session\n * can hold a signature that is not base64; drop that block instead of failing the\n * whole request.\n */\nfunction decodeRedactedContent(signature: string | undefined): Uint8Array | undefined {\n\tif (!signature) return undefined;\n\ttry {\n\t\treturn base64ToBytes(signature);\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction bytesToBase64(chunks: Uint8Array[]): string {\n\t// Encrypted reasoning runs to tens of KB, so build the binary string in slices\n\t// rather than one concatenation per byte. The window stays under the engine's\n\t// argument-count limit for spread calls.\n\tconst WINDOW = 0x8000;\n\tlet binary = \"\";\n\tfor (const chunk of chunks) {\n\t\tfor (let i = 0; i < chunk.length; i += WINDOW) {\n\t\t\tbinary += String.fromCharCode(...chunk.subarray(i, i + WINDOW));\n\t\t}\n\t}\n\treturn btoa(binary);\n}\n"]}

package/dist/api/bedrock-converse-stream.js

@@ -536,6 +536,7 @@ function handleMetadata(event, model, output) {
         output.usage.output = event.usage.outputTokens || 0;
         output.usage.cacheRead = event.usage.cacheReadInputTokens || 0;
         output.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;
+        output.usage.cacheWrite1h = event.usage.cacheDetails?.reduce((total, detail) => total + (detail.ttl === CacheTTL.ONE_HOUR ? (detail.inputTokens ?? 0) : 0), 0);
         output.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;
         calculateCost(model, output.usage);
     }