@earendil-works/pi-ai 0.82.1 → 0.84.0

package/dist/api/openai-completions.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/api/openai-completions.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,KAAK,EAOX,0BAA0B,EAI1B,MAAM,sCAAsC,CAAC;AAE9C,OAAO,KAAK,EAIX,OAAO,EAGP,KAAK,EACL,uBAAuB,EAGvB,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AA0GrB,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,8BAA8B,CAAC;IACpE,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;CAC1E;AAED,MAAM,WAAW,iCAAiC;IACjD,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACzD;AAOD,KAAK,+BAA+B,GAAG,IAAI,CAC1C,QAAQ,CAAC,uBAAuB,CAAC,EACjC,oBAAoB,GAAG,mBAAmB,CAC1C,GAAG;IACH,kBAAkB,CAAC,EAAE,uBAAuB,CAAC,oBAAoB,CAAC,CAAC;IACnE,iBAAiB,CAAC,EAAE,uBAAuB,CAAC,mBAAmB,CAAC,CAAC;CACjE,CAAC;AAmCF,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CA2ZjF,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAiBlF,CAAC;AAyWF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,+BAA+B,EACvC,OAAO,CAAC,EAAE,iCAAiC,GACzC,0BAA0B,EAAE,CA4R9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionDeveloperMessageParam,\n\tChatCompletionMessageParam,\n\tChatCompletionMessageToolCall,\n\tChatCompletionSystemMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { calculateCost, clampThinkingLevel } from \"../models.ts\";\nimport type {\n\tAssistantMessage,\n\tCacheRetention,\n\tChatTemplateKwargValue,\n\tContext,\n\tImageContent,\n\tMessage,\n\tModel,\n\tOpenAICompletionsCompat,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport {\n\tappendGrammarToolInputJsonDelta,\n\tcreateGrammarToolInputProperties,\n\ttype GrammarToolInputJsonBuffer,\n\tgetGrammarToolInput,\n\tresolveGrammarConstrainedSampling,\n\tresolveJsonSchemaStrictSampling,\n} from \"./constrained-sampling.ts\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.ts\";\nimport { clampOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasHeader(headers: ProviderHeaders | undefined, name: string): boolean {\n\tif (!headers) return false;\n\tconst expected = name.toLowerCase();\n\tfor (const [key, value] of Object.entries(headers)) {\n\t\tif (key.toLowerCase() === expected && value !== null && value.trim().length > 0) return true;\n\t}\n\treturn false;\n}\n\nfunction getClientApiKey(provider: string, apiKey: string | undefined, headers: ProviderHeaders | undefined): string {\n\tif (apiKey) return apiKey;\n\tif (hasHeader(headers, \"authorization\") || hasHeader(headers, \"cf-aig-authorization\")) return \"unused\";\n\tthrow new Error(`No API key for provider: ${provider}`);\n}\n\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nfunction getDeferredToolNames(messages: Message[]): Set<string> {\n\tconst names = new Set<string>();\n\tfor (const message of messages) {\n\t\tif (message.role === \"toolResult\") {\n\t\t\tfor (const name of message.addedToolNames ?? []) {\n\t\t\t\tnames.add(name);\n\t\t\t}\n\t\t}\n\t}\n\treturn names;\n}\n\nfunction getToolsByName(tools: Tool[] | undefined, names: Iterable<string>): Tool[] {\n\tif (!tools) return [];\n\tconst toolsByName = new Map(tools.map((tool) => [tool.name, tool]));\n\treturn Array.from(names)\n\t\t.map((name) => toolsByName.get(name))\n\t\t.filter((tool): tool is Tool => tool !== undefined);\n}\n\nfunction isTextContentBlock(block: { type: string }): block is TextContent {\n\treturn block.type === \"text\";\n}\n\nfunction isThinkingContentBlock(block: { type: string }): block is ThinkingContent {\n\treturn block.type === \"thinking\";\n}\n\nfunction isToolCallBlock(block: { type: string }): block is ToolCall {\n\treturn block.type === \"toolCall\";\n}\n\nfunction isImageContentBlock(block: { type: string }): block is ImageContent {\n\treturn block.type === \"image\";\n}\n\nfunction isEncryptedReasoningDetail(detail: unknown): detail is OpenAIEncryptedReasoningDetail {\n\tif (typeof detail !== \"object\" || detail === null) {\n\t\treturn false;\n\t}\n\tconst candidate = detail as Record<string, unknown>;\n\treturn (\n\t\tcandidate.type === \"reasoning.encrypted\" &&\n\t\ttypeof candidate.id === \"string\" &&\n\t\tcandidate.id.length > 0 &&\n\t\ttypeof candidate.data === \"string\" &&\n\t\tcandidate.data.length > 0\n\t);\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: OpenAI.Chat.Completions.ChatCompletionToolChoiceOption;\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n}\n\nexport interface ConvertCompletionsMessagesOptions {\n\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\n}\n\ninterface OpenAICompatCacheControl {\n\ttype: \"ephemeral\";\n\tttl?: string;\n}\n\ntype ResolvedOpenAICompletionsCompat = Omit<\n\tRequired<OpenAICompletionsCompat>,\n\t\"cacheControlFormat\" | \"deferredToolsMode\"\n> & {\n\tcacheControlFormat?: OpenAICompletionsCompat[\"cacheControlFormat\"];\n\tdeferredToolsMode?: OpenAICompletionsCompat[\"deferredToolsMode\"];\n};\n\ntype ResolvedChatTemplateKwargValue = string | number | boolean | null;\n\ntype ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;\n\ntype KimiToolSystemMessageParam = {\n\trole: \"system\";\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[];\n};\n\ntype OpenAIEncryptedReasoningDetail = {\n\ttype: \"reasoning.encrypted\";\n\tid: string;\n\tdata: string;\n};\n\ntype ChatCompletionTextPartWithCacheControl = ChatCompletionContentPartText & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\ntype ChatCompletionToolWithCacheControl = OpenAI.Chat.Completions.ChatCompletionTool & {\n\tcache_control?: OpenAICompatCacheControl;\n};\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\nexport const stream: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\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: model.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: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = getClientApiKey(model.provider, options?.apiKey, options?.headers);\n\t\t\tconst compat = getCompat(model);\n\t\t\tconst grammarToolInputProperties = createGrammarToolInputProperties(\n\t\t\t\tcontext.tools,\n\t\t\t\tcompat.supportsOpenAIGrammarTools,\n\t\t\t);\n\t\t\tconst cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);\n\t\t\tconst cacheSessionId = cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);\n\t\t\tlet params = buildParams(model, context, options, compat, cacheRetention, grammarToolInputProperties);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst requestOptions = {\n\t\t\t\t...(options?.signal ? { signal: options.signal } : {}),\n\t\t\t\t...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),\n\t\t\t\tmaxRetries: 0,\n\t\t\t};\n\t\t\tconst { data: openaiStream, response } = await retryProviderRequest(\n\t\t\t\t() => client.chat.completions.create(params, requestOptions).withResponse(),\n\t\t\t\t{\n\t\t\t\t\tmaxRetries: options?.maxRetries,\n\t\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t},\n\t\t\t);\n\t\t\tawait options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tinterface StreamingToolCallBlock extends ToolCall {\n\t\t\t\tpartialArgs?: string;\n\t\t\t\tcustomInput?: {\n\t\t\t\t\tproperty: string;\n\t\t\t\t\tjsonBuffer: GrammarToolInputJsonBuffer;\n\t\t\t\t};\n\t\t\t\tstreamIndex?: number;\n\t\t\t}\n\t\t\ttype StreamingBlock = TextContent | ThinkingContent | StreamingToolCallBlock;\n\t\t\ttype StreamingToolCallDelta = {\n\t\t\t\tindex?: number;\n\t\t\t\tid?: string;\n\t\t\t\ttype?: string;\n\t\t\t\tfunction?: { name?: string; arguments?: string };\n\t\t\t\tcustom?: { name?: string; input?: string };\n\t\t\t};\n\n\t\t\tlet textBlock: TextContent | null = null;\n\t\t\tlet thinkingBlock: ThinkingContent | null = null;\n\t\t\tlet hasFinishReason = false;\n\t\t\tconst toolCallBlocksByIndex = new Map<number, StreamingToolCallBlock>();\n\t\t\tconst toolCallBlocksById = new Map<string, StreamingToolCallBlock>();\n\t\t\tconst pendingReasoningDetailsByToolCallId = new Map<string, string>();\n\t\t\tconst blocks = output.content as StreamingBlock[];\n\t\t\tconst getContentIndex = (block: StreamingBlock) => blocks.indexOf(block);\n\t\t\tconst getCustomToolCallInput = (block: StreamingToolCallBlock): string => {\n\t\t\t\tconst property = block.customInput?.property;\n\t\t\t\tif (property === undefined) return \"\";\n\t\t\t\tconst value = block.arguments[property];\n\t\t\t\treturn typeof value === \"string\" ? value : \"\";\n\t\t\t};\n\t\t\tconst appendCustomToolCallInput = (\n\t\t\t\tblock: StreamingToolCallBlock,\n\t\t\t\tnextInput: string,\n\t\t\t\tclose: boolean,\n\t\t\t): string | undefined => {\n\t\t\t\tconst customInput = block.customInput;\n\t\t\t\tif (!customInput) return undefined;\n\t\t\t\tconst delta = appendGrammarToolInputJsonDelta(\n\t\t\t\t\tcustomInput.jsonBuffer,\n\t\t\t\t\tcustomInput.property,\n\t\t\t\t\tnextInput,\n\t\t\t\t\tclose,\n\t\t\t\t);\n\t\t\t\tblock.arguments = { [customInput.property]: nextInput };\n\t\t\t\treturn delta;\n\t\t\t};\n\t\t\tconst finishBlock = (block: StreamingBlock) => {\n\t\t\t\tconst contentIndex = getContentIndex(block);\n\t\t\t\tif (contentIndex === -1) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tif (block.customInput) {\n\t\t\t\t\t\tconst delta = appendCustomToolCallInput(block, getCustomToolCallInput(block), true);\n\t\t\t\t\t\tif (delta !== undefined) {\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t} else {\n\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t}\n\t\t\t\t\t// Finalize in-place and strip the scratch buffers so replay only\n\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\tdelete block.customInput;\n\t\t\t\t\tdelete block.streamIndex;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureTextBlock = () => {\n\t\t\t\tif (!textBlock) {\n\t\t\t\t\ttextBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\tblocks.push(textBlock);\n\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: getContentIndex(textBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn textBlock;\n\t\t\t};\n\t\t\tconst ensureThinkingBlock = (thinkingSignature: string) => {\n\t\t\t\tif (!thinkingBlock) {\n\t\t\t\t\tthinkingBlock = {\n\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\tthinkingSignature,\n\t\t\t\t\t};\n\t\t\t\t\tblocks.push(thinkingBlock);\n\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: getContentIndex(thinkingBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn thinkingBlock;\n\t\t\t};\n\t\t\tconst applyPendingReasoningDetail = (block: StreamingToolCallBlock) => {\n\t\t\t\tif (!block.id) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tconst pendingReasoningDetail = pendingReasoningDetailsByToolCallId.get(block.id);\n\t\t\t\tif (pendingReasoningDetail) {\n\t\t\t\t\tblock.thoughtSignature = pendingReasoningDetail;\n\t\t\t\t\tpendingReasoningDetailsByToolCallId.delete(block.id);\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureToolCallBlock = (toolCall: StreamingToolCallDelta) => {\n\t\t\t\tconst streamIndex = typeof toolCall.index === \"number\" ? toolCall.index : undefined;\n\t\t\t\tconst name = toolCall.function?.name ?? toolCall.custom?.name ?? \"\";\n\t\t\t\tlet block = streamIndex !== undefined ? toolCallBlocksByIndex.get(streamIndex) : undefined;\n\t\t\t\tif (!block && toolCall.id) {\n\t\t\t\t\tblock = toolCallBlocksById.get(toolCall.id);\n\t\t\t\t}\n\t\t\t\tif (!block) {\n\t\t\t\t\t// Note: the \"input\" fallback here should/must not be taken.  in case the LLM makes up\n\t\t\t\t\t// a tool we don't knwo about, we at least have a place to stash our stuff.\n\t\t\t\t\tconst customInputProperty = toolCall.custom\n\t\t\t\t\t\t? (grammarToolInputProperties.get(name) ?? \"input\")\n\t\t\t\t\t\t: undefined;\n\t\t\t\t\tconst hasCustomInput = customInputProperty !== undefined;\n\t\t\t\t\tblock = {\n\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\tname,\n\t\t\t\t\t\targuments: hasCustomInput ? { [customInputProperty]: \"\" } : {},\n\t\t\t\t\t\tpartialArgs: hasCustomInput ? undefined : \"\",\n\t\t\t\t\t\tcustomInput: hasCustomInput\n\t\t\t\t\t\t\t? { property: customInputProperty, jsonBuffer: { input: \"\", started: false, closed: false } }\n\t\t\t\t\t\t\t: undefined,\n\t\t\t\t\t\tstreamIndex,\n\t\t\t\t\t};\n\t\t\t\t\tif (streamIndex !== undefined) {\n\t\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t\t}\n\t\t\t\t\tif (toolCall.id) {\n\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t}\n\t\t\t\t\tblocks.push(block);\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_start\",\n\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tif (streamIndex !== undefined && block.streamIndex === undefined) {\n\t\t\t\t\tblock.streamIndex = streamIndex;\n\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t}\n\t\t\t\tif (toolCall.id) {\n\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t}\n\t\t\t\tif (!block.name && name) {\n\t\t\t\t\tblock.name = name;\n\t\t\t\t}\n\t\t\t\tif (toolCall.custom && !block.customInput) {\n\t\t\t\t\tconst customInputProperty = grammarToolInputProperties.get(block.name) ?? \"input\";\n\t\t\t\t\tblock.arguments = { [customInputProperty]: \"\" };\n\t\t\t\t\tblock.customInput = {\n\t\t\t\t\t\tproperty: customInputProperty,\n\t\t\t\t\t\tjsonBuffer: { input: \"\", started: false, closed: false },\n\t\t\t\t\t};\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t}\n\t\t\t\tapplyPendingReasoningDetail(block);\n\t\t\t\treturn block;\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (typeof chunk.model === \"string\" && chunk.model.length > 0 && chunk.model !== model.id) {\n\t\t\t\t\toutput.responseModel ||= chunk.model;\n\t\t\t\t}\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t\thasFinishReason = true;\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst block = ensureTextBlock();\n\t\t\t\t\t\tblock.text += choice.delta.content;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tconst deltaFields = choice.delta as Record<string, unknown>;\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tconst value = deltaFields[field];\n\t\t\t\t\t\tif (typeof value === \"string\" && value.length > 0) {\n\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tconst delta = deltaFields[foundReasoningField];\n\t\t\t\t\t\tif (typeof delta === \"string\" && delta.length > 0) {\n\t\t\t\t\t\t\tconst thinkingSignature =\n\t\t\t\t\t\t\t\tmodel.provider === \"opencode-go\" && foundReasoningField === \"reasoning\"\n\t\t\t\t\t\t\t\t\t? \"reasoning_content\"\n\t\t\t\t\t\t\t\t\t: foundReasoningField;\n\t\t\t\t\t\t\tconst block = ensureThinkingBlock(thinkingSignature);\n\t\t\t\t\t\t\tblock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls as StreamingToolCallDelta[]) {\n\t\t\t\t\t\t\tconst block = ensureToolCallBlock(toolCall);\n\t\t\t\t\t\t\tif (!block.id && toolCall.id) {\n\t\t\t\t\t\t\t\tblock.id = toolCall.id;\n\t\t\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tconst name = toolCall.function?.name ?? toolCall.custom?.name;\n\t\t\t\t\t\t\tif (!block.name && name) {\n\t\t\t\t\t\t\t\tblock.name = name;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.partialArgs = (block.partialArgs ?? \"\") + toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\t\t} else if (toolCall.custom?.input) {\n\t\t\t\t\t\t\t\tconst nextInput = getCustomToolCallInput(block) + toolCall.custom.input;\n\t\t\t\t\t\t\t\tdelta = appendCustomToolCallInput(block, nextInput, false) ?? \"\";\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as { reasoning_details?: unknown }).reasoning_details;\n\t\t\t\t\tif (Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (isEncryptedReasoningDetail(detail)) {\n\t\t\t\t\t\t\t\tconst serializedDetail = JSON.stringify(detail);\n\t\t\t\t\t\t\t\tconst matchingToolCall = toolCallBlocksById.get(detail.id);\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = serializedDetail;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tpendingReasoningDetailsByToolCallId.set(detail.id, serializedDetail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfor (const block of blocks) {\n\t\t\t\tfinishBlock(block);\n\t\t\t}\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 === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\t\t\tif (!hasFinishReason) {\n\t\t\t\tthrow new Error(\"Stream ended without finish_reason\");\n\t\t\t}\n\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\tdelete (block as { index?: number }).index;\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialArgs?: string }).partialArgs;\n\t\t\t\tdelete (block as { customInput?: unknown }).customInput;\n\t\t\t\tdelete (block as { streamIndex?: number }).streamIndex;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatProviderError(normalizeProviderError(error));\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\t// normalizeProviderError already stringifies the parsed body (error.error)\n\t\t\t// into errorMessage, so only append the raw metadata when it is not already\n\t\t\t// present to avoid double-printing it.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata && !output.errorMessage.includes(String(rawMetadata))) {\n\t\t\t\toutput.errorMessage += `\\n${rawMetadata}`;\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\nexport const streamSimple: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tgetClientApiKey(model.provider, options?.apiKey, options?.headers);\n\n\tconst base = buildBaseOptions(model, context, options, options?.apiKey);\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey: string,\n\toptionsHeaders?: ProviderHeaders,\n\tsessionId?: string,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n) {\n\tconst headers: ProviderHeaders = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\tif (sessionId && compat.sendSessionAffinityHeaders) {\n\t\tif (compat.sessionAffinityFormat === \"openrouter\") {\n\t\t\theaders[\"x-session-id\"] = sessionId;\n\t\t} else {\n\t\t\tif (compat.sessionAffinityFormat === \"openai\") {\n\t\t\t\theaders.session_id = sessionId;\n\t\t\t}\n\t\t\theaders[\"x-client-request-id\"] = sessionId;\n\t\t\theaders[\"x-session-affinity\"] = sessionId;\n\t\t}\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tdefaultHeaders: headers,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n\tcacheRetention: CacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env),\n\tgrammarToolInputProperties: ReadonlyMap<string, string> = createGrammarToolInputProperties(\n\t\tcontext.tools,\n\t\tcompat.supportsOpenAIGrammarTools,\n\t),\n) {\n\tconst messages = convertMessages(model, context, compat, { grammarToolInputProperties });\n\tconst cacheControl = getCompatCacheControl(compat, cacheRetention);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.id,\n\t\tmessages,\n\t\tstream: true,\n\t\tprompt_cache_key:\n\t\t\t(model.baseUrl.includes(\"api.openai.com\") && cacheRetention !== \"none\") ||\n\t\t\t(cacheRetention === \"long\" && compat.supportsLongCacheRetention)\n\t\t\t\t? clampOpenAIPromptCacheKey(options?.sessionId)\n\t\t\t\t: undefined,\n\t\tprompt_cache_retention: cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"24h\" : undefined,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tconst deferredToolNames =\n\t\tcompat.deferredToolsMode === \"kimi\" ? getDeferredToolNames(context.messages) : new Set<string>();\n\tconst activeTools = context.tools?.filter((tool) => !deferredToolNames.has(tool.name));\n\tif (activeTools && activeTools.length > 0) {\n\t\tparams.tools = convertTools(activeTools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (cacheControl) {\n\t\tapplyAnthropicCacheControl(messages, params.tools, cacheControl);\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\tconst zaiParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\tthinking?: { type: \"enabled\" | \"disabled\"; clear_thinking?: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\tzaiParams.thinking = options?.reasoningEffort ? { type: \"enabled\", clear_thinking: false } : { type: \"disabled\" };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\tconst mappedEffort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\t\tconst effort = mappedEffort === undefined ? options.reasoningEffort : mappedEffort;\n\t\t\tif (typeof effort === \"string\") {\n\t\t\t\tzaiParams.reasoning_effort = effort;\n\t\t\t}\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = {\n\t\t\tenable_thinking: !!options?.reasoningEffort,\n\t\t\tpreserve_thinking: true,\n\t\t};\n\t} else if (compat.thinkingFormat === \"chat-template\" && model.reasoning) {\n\t\tconst chatTemplateKwargs = buildChatTemplateKwargs(model, options, compat);\n\t\tif (chatTemplateKwargs) {\n\t\t\t(params as any).chat_template_kwargs = chatTemplateKwargs;\n\t\t}\n\t} else if (compat.thinkingFormat === \"deepseek\" && model.reasoning) {\n\t\tif (options?.reasoningEffort) {\n\t\t\t(params as any).thinking = { type: \"enabled\" };\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\t(params as any).thinking = { type: \"disabled\" };\n\t\t}\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort,\n\t\t\t};\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\topenRouterParams.reasoning = { effort: model.thinkingLevelMap?.off ?? \"none\" };\n\t\t}\n\t} else if (compat.thinkingFormat === \"ant-ling\" && model.reasoning && options?.reasoningEffort) {\n\t\tconst effort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\tif (typeof effort === \"string\") {\n\t\t\t(params as typeof params & { reasoning?: { effort: string } }).reasoning = { effort };\n\t\t}\n\t} else if (compat.thinkingFormat === \"together\" && model.reasoning) {\n\t\tconst togetherParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\treasoning?: { enabled: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\ttogetherParams.reasoning = { enabled: !!options?.reasoningEffort };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\ttogetherParams.reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"string-thinking\" && model.reasoning) {\n\t\tconst stringThinkingParams = params as typeof params & { thinking?: string };\n\t\tif (options?.reasoningEffort) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.off ?? \"none\";\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t} else if (!options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\tconst offValue = model.thinkingLevelMap?.off;\n\t\tif (typeof offValue === \"string\") {\n\t\t\t(params as any).reasoning_effort = offValue;\n\t\t}\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\treturn params;\n}\n\nfunction buildChatTemplateKwargs(\n\tmodel: Model<\"openai-completions\">,\n\toptions: OpenAICompletionsOptions | undefined,\n\tcompat: ResolvedOpenAICompletionsCompat,\n): Record<string, ResolvedChatTemplateKwargValue> | undefined {\n\tconst kwargs: Record<string, ResolvedChatTemplateKwargValue> = {};\n\n\tfor (const [key, value] of Object.entries(compat.chatTemplateKwargs)) {\n\t\tconst resolved = resolveChatTemplateKwargValue(model, options, value);\n\t\tif (resolved !== undefined) {\n\t\t\tkwargs[key] = resolved;\n\t\t}\n\t}\n\n\treturn Object.keys(kwargs).length > 0 ? kwargs : undefined;\n}\n\nfunction resolveChatTemplateKwargValue(\n\tmodel: Model<\"openai-completions\">,\n\toptions: OpenAICompletionsOptions | undefined,\n\tvalue: ChatTemplateKwargValue,\n): ResolvedChatTemplateKwargValue | undefined {\n\tif (typeof value !== \"object\" || value === null) {\n\t\treturn value;\n\t}\n\n\tconst reasoningEffort = options?.reasoningEffort;\n\tif (!reasoningEffort && value.omitWhenOff) {\n\t\treturn undefined;\n\t}\n\tif (value.$var === \"thinking.enabled\") {\n\t\treturn !!reasoningEffort;\n\t}\n\n\tconst mappedValue = reasoningEffort ? model.thinkingLevelMap?.[reasoningEffort] : model.thinkingLevelMap?.off;\n\treturn mappedValue === undefined ? reasoningEffort : typeof mappedValue === \"string\" ? mappedValue : undefined;\n}\n\nfunction getCompatCacheControl(\n\tcompat: ResolvedOpenAICompletionsCompat,\n\tcacheRetention: CacheRetention,\n): OpenAICompatCacheControl | undefined {\n\tif (compat.cacheControlFormat !== \"anthropic\" || cacheRetention === \"none\") {\n\t\treturn undefined;\n\t}\n\n\tconst ttl = cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"1h\" : undefined;\n\treturn { type: \"ephemeral\", ...(ttl ? { ttl } : {}) };\n}\n\nfunction applyAnthropicCacheControl(\n\tmessages: ChatCompletionMessageParam[],\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\taddCacheControlToSystemPrompt(messages, cacheControl);\n\taddCacheControlToLastTool(tools, cacheControl);\n\taddCacheControlToLastConversationMessage(messages, cacheControl);\n}\n\nfunction addCacheControlToSystemPrompt(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (const message of messages) {\n\t\tif (message.role === \"system\" || message.role === \"developer\") {\n\t\t\taddCacheControlToInstructionMessage(message, cacheControl);\n\t\t\treturn;\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastConversationMessage(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"tool\") {\n\t\t\tif (addCacheControlToMessage(message, cacheControl)) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastTool(\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tif (!tools || tools.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastTool = tools[tools.length - 1] as ChatCompletionToolWithCacheControl;\n\tlastTool.cache_control = cacheControl;\n}\n\nfunction addCacheControlToInstructionMessage(\n\tmessage: ChatCompletionInstructionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\treturn addCacheControlToTextContent(message, cacheControl);\n}\n\nfunction addCacheControlToMessage(\n\tmessage: ChatCompletionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"tool\") {\n\t\treturn addCacheControlToTextContent(message, cacheControl);\n\t}\n\treturn false;\n}\n\nfunction addCacheControlToTextContent(\n\tmessage:\n\t\t| ChatCompletionInstructionMessageParam\n\t\t| ChatCompletionAssistantMessageParam\n\t\t| ChatCompletionToolMessageParam\n\t\t| Extract<ChatCompletionMessageParam, { role: \"user\" }>,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tconst content = message.content;\n\tif (typeof content === \"string\") {\n\t\tif (content.length === 0) {\n\t\t\treturn false;\n\t\t}\n\t\tmessage.content = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: content,\n\t\t\t\tcache_control: cacheControl,\n\t\t\t},\n\t\t] as ChatCompletionTextPartWithCacheControl[];\n\t\treturn true;\n\t}\n\n\tif (!Array.isArray(content)) {\n\t\treturn false;\n\t}\n\n\tfor (let i = content.length - 1; i >= 0; i--) {\n\t\tconst part = content[i];\n\t\tif (part?.type === \"text\") {\n\t\t\tconst textPart = part as ChatCompletionTextPartWithCacheControl;\n\t\t\ttextPart.cache_control = cacheControl;\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: ResolvedOpenAICompletionsCompat,\n\toptions?: ConvertCompletionsMessagesOptions,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\t// Multiple tool calls in the same turn can share call_id but differ by item_id.\n\t\t// Preserve item-level uniqueness when replaying into Chat Completions, which\n\t\t// requires distinct tool call ids.\n\t\tif (id.includes(\"|\")) {\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\tconst separatorIndex = id.indexOf(\"|\");\n\t\t\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst combinedId = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\t\t\tif (combinedId.length <= 40) {\n\t\t\t\treturn combinedId;\n\t\t\t}\n\t\t\tconst hash = shortHash(id).slice(0, 8);\n\t\t\tconst prefix = callId.slice(0, Math.max(1, 40 - hash.length - 1));\n\t\t\treturn `${prefix}_${hash}`;\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\tif (context.systemPrompt) {\n\t\tconst useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;\n\t\tconst role = useDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role: role, content: sanitizeSurrogates(context.systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ChatCompletionContentPart[] = msg.content.map((item): ChatCompletionContentPart => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText;\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage;\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst assistantTextParts = msg.content\n\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t.filter((block) => block.text.trim().length > 0)\n\t\t\t\t.map(\n\t\t\t\t\t(block) =>\n\t\t\t\t\t\t({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t\t}) satisfies ChatCompletionContentPartText,\n\t\t\t\t);\n\t\t\tconst assistantText = assistantTextParts.map((part) => part.text).join(\"\");\n\n\t\t\tconst nonEmptyThinkingBlocks = msg.content\n\t\t\t\t.filter(isThinkingContentBlock)\n\t\t\t\t.filter((block) => block.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks\n\t\t\t\t\t\t.map((block) => sanitizeSurrogates(block.thinking))\n\t\t\t\t\t\t.join(\"\\n\\n\");\n\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }, ...assistantTextParts];\n\t\t\t\t} else {\n\t\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\t\tif (assistantText.length > 0) {\n\t\t\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tlet signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (model.provider === \"opencode-go\" && signature === \"reasoning\") {\n\t\t\t\t\t\tsignature = \"reasoning_content\";\n\t\t\t\t\t}\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t} else if (assistantText.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter(isToolCallBlock);\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc): ChatCompletionMessageToolCall => {\n\t\t\t\t\tconst customInputProperty = options?.grammarToolInputProperties?.get(tc.name);\n\t\t\t\t\tif (customInputProperty !== undefined) {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\tid: tc.id,\n\t\t\t\t\t\t\ttype: \"custom\",\n\t\t\t\t\t\t\tcustom: {\n\t\t\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\t\t\tinput: sanitizeSurrogates(getGrammarToolInput(tc.name, tc.arguments, customInputProperty)),\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t};\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\tid: tc.id,\n\t\t\t\t\t\ttype: \"function\",\n\t\t\t\t\t\tfunction: {\n\t\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t});\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (\n\t\t\t\tcompat.requiresReasoningContentOnAssistantMessages &&\n\t\t\t\tmodel.reasoning &&\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content === undefined\n\t\t\t) {\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content = \"\";\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tconst deferredToolNames = new Set<string>();\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t\t.map((block) => block.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\tconst toolResultText = hasText ? textResult : hasImages ? \"(see attached image)\" : \"(no tool output)\";\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(toolResultText),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (compat.deferredToolsMode === \"kimi\") {\n\t\t\t\t\tfor (const name of toolMsg.addedToolNames ?? []) {\n\t\t\t\t\t\tdeferredToolNames.add(name);\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (isImageContentBlock(block)) {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\n\t\t\tif (deferredToolNames.size > 0) {\n\t\t\t\tconst deferredTools = getToolsByName(context.tools, deferredToolNames);\n\t\t\t\tif (deferredTools.length > 0) {\n\t\t\t\t\tconst kimiToolMessage: KimiToolSystemMessageParam = {\n\t\t\t\t\t\trole: \"system\",\n\t\t\t\t\t\ttools: convertTools(deferredTools, compat),\n\t\t\t\t\t};\n\t\t\t\t\t// Kimi accepts a system message with tools but omits the standard content field.\n\t\t\t\t\tparams.push(kimiToolMessage as unknown as ChatCompletionMessageParam);\n\t\t\t\t}\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: ResolvedOpenAICompletionsCompat,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn tools.map((tool) => {\n\t\tconst grammar = resolveGrammarConstrainedSampling(tool, compat.supportsOpenAIGrammarTools);\n\t\tif (grammar) {\n\t\t\treturn {\n\t\t\t\ttype: \"custom\",\n\t\t\t\tcustom: {\n\t\t\t\t\tname: tool.name,\n\t\t\t\t\tdescription: tool.description,\n\t\t\t\t\tformat: {\n\t\t\t\t\t\ttype: \"grammar\",\n\t\t\t\t\t\tgrammar: {\n\t\t\t\t\t\t\tsyntax: grammar.format,\n\t\t\t\t\t\t\tdefinition: grammar.definition,\n\t\t\t\t\t\t},\n\t\t\t\t\t},\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\n\t\tconst strict = resolveJsonSchemaStrictSampling(tool, compat.supportsStrictMode !== false);\n\t\treturn {\n\t\t\ttype: \"function\",\n\t\t\tfunction: {\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tparameters: tool.parameters as Record<string, unknown>, // TypeBox already generates JSON Schema\n\t\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t\t...(compat.supportsStrictMode !== false && { strict: strict ?? false }),\n\t\t\t},\n\t\t};\n\t});\n}\n\nfunction parseChunkUsage(\n\trawUsage: {\n\t\tprompt_tokens?: number;\n\t\tcompletion_tokens?: number;\n\t\tprompt_cache_hit_tokens?: number;\n\t\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\t\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t},\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = rawUsage.prompt_tokens || 0;\n\tconst cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;\n\tconst cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;\n\n\t// Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read\n\t// tokens (hits). OpenAI does not document or emit cache_write_tokens, but\n\t// OpenRouter-compatible providers can include it as a separate write count.\n\t// OpenRouter's own provider/tests affirm the separate mapping:\n\t// https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409\n\t// Do not subtract writes from cached_tokens, otherwise spec-compliant\n\t// providers are under-reported. DS4 mirrors this contract too:\n\t// https://github.com/antirez/ds4/pull/29\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\t// OpenAI completion_tokens already includes reasoning_tokens.\n\tconst outputTokens = rawUsage.completion_tokens || 0;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\treasoning: rawUsage.completion_tokens_details?.reasoning_tokens || 0,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\treturn usage;\n}\n\nfunction mapStopReason(reason: ChatCompletionChunk.Choice[\"finish_reason\"] | string): {\n\tstopReason: StopReason;\n\terrorMessage?: string;\n} {\n\tif (reason === null) return { stopReason: \"stop\" };\n\tswitch (reason) {\n\t\tcase \"stop\":\n\t\tcase \"end\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"length\":\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase \"function_call\":\n\t\tcase \"tool_calls\":\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tcase \"content_filter\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: content_filter\" };\n\t\tcase \"network_error\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: network_error\" };\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: `Provider finish_reason: ${reason}`,\n\t\t\t};\n\t}\n}\n\n/**\n * Auto-detect compatibility settings from provider name and baseUrl.\n * Used as the base when model.compat is not set; explicit model.compat\n * entries override these detected values.\n */\nfunction detectCompat(model: Model<\"openai-completions\">): ResolvedOpenAICompletionsCompat {\n\tconst provider = model.provider;\n\tconst baseUrl = model.baseUrl;\n\n\tconst isZai =\n\t\tprovider === \"zai\" ||\n\t\tprovider === \"zai-coding-cn\" ||\n\t\tbaseUrl.includes(\"api.z.ai\") ||\n\t\tbaseUrl.includes(\"open.bigmodel.cn\");\n\tconst isTogether =\n\t\tprovider === \"together\" || baseUrl.includes(\"api.together.ai\") || baseUrl.includes(\"api.together.xyz\");\n\tconst isMoonshot = provider === \"moonshotai\" || provider === \"moonshotai-cn\" || baseUrl.includes(\"api.moonshot.\");\n\tconst isOpenRouter = provider === \"openrouter\" || baseUrl.includes(\"openrouter.ai\");\n\tconst isCloudflareWorkersAI = provider === \"cloudflare-workers-ai\" || baseUrl.includes(\"api.cloudflare.com\");\n\tconst isCloudflareAiGateway = provider === \"cloudflare-ai-gateway\" || baseUrl.includes(\"gateway.ai.cloudflare.com\");\n\tconst isNvidia = provider === \"nvidia\" || baseUrl.includes(\"integrate.api.nvidia.com\");\n\tconst isAntLing = provider === \"ant-ling\" || baseUrl.includes(\"api.ant-ling.com\");\n\n\tconst isNonStandard =\n\t\tisNvidia ||\n\t\tprovider === \"cerebras\" ||\n\t\tbaseUrl.includes(\"cerebras.ai\") ||\n\t\tprovider === \"xai\" ||\n\t\tbaseUrl.includes(\"api.x.ai\") ||\n\t\tisTogether ||\n\t\tbaseUrl.includes(\"chutes.ai\") ||\n\t\tbaseUrl.includes(\"deepseek.com\") ||\n\t\tisZai ||\n\t\tisMoonshot ||\n\t\tprovider === \"opencode\" ||\n\t\tbaseUrl.includes(\"opencode.ai\") ||\n\t\tisCloudflareWorkersAI ||\n\t\tisCloudflareAiGateway ||\n\t\tisAntLing;\n\n\tconst useMaxTokens =\n\t\tbaseUrl.includes(\"chutes.ai\") || isMoonshot || isCloudflareAiGateway || isTogether || isNvidia || isAntLing;\n\n\tconst isGrok = provider === \"xai\" || baseUrl.includes(\"api.x.ai\");\n\tconst isDeepSeek = provider === \"deepseek\" || baseUrl.includes(\"deepseek.com\");\n\tconst isOpenRouterDeveloperRoleModel =\n\t\tisOpenRouter && (model.id.startsWith(\"anthropic/\") || model.id.startsWith(\"openai/\"));\n\tconst cacheControlFormat = provider === \"openrouter\" && model.id.startsWith(\"anthropic/\") ? \"anthropic\" : undefined;\n\n\treturn {\n\t\tsupportsStore: !isNonStandard,\n\t\tsupportsDeveloperRole: isOpenRouterDeveloperRoleModel || (!isNonStandard && !isOpenRouter),\n\t\tsupportsReasoningEffort:\n\t\t\t!isGrok && !isZai && !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia && !isAntLing,\n\t\tsupportsUsageInStreaming: true,\n\t\tmaxTokensField: useMaxTokens ? \"max_tokens\" : \"max_completion_tokens\",\n\t\trequiresToolResultName: false,\n\t\trequiresAssistantAfterToolResult: false,\n\t\trequiresThinkingAsText: false,\n\t\trequiresReasoningContentOnAssistantMessages: isDeepSeek,\n\t\tthinkingFormat: isDeepSeek\n\t\t\t? \"deepseek\"\n\t\t\t: isZai\n\t\t\t\t? \"zai\"\n\t\t\t\t: isTogether\n\t\t\t\t\t? \"together\"\n\t\t\t\t\t: isAntLing\n\t\t\t\t\t\t? \"ant-ling\"\n\t\t\t\t\t\t: isOpenRouter\n\t\t\t\t\t\t\t? \"openrouter\"\n\t\t\t\t\t\t\t: \"openai\",\n\t\topenRouterRouting: {},\n\t\tvercelGatewayRouting: {},\n\t\tchatTemplateKwargs: {},\n\t\tzaiToolStream: false,\n\t\tsupportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,\n\t\tsupportsOpenAIGrammarTools: false,\n\t\tcacheControlFormat,\n\t\tsendSessionAffinityHeaders: false,\n\t\tdeferredToolsMode: undefined,\n\t\tsessionAffinityFormat: isOpenRouter ? \"openrouter\" : \"openai\",\n\t\tsupportsLongCacheRetention: !(\n\t\t\tisTogether ||\n\t\t\tisCloudflareWorkersAI ||\n\t\t\tisCloudflareAiGateway ||\n\t\t\tisNvidia ||\n\t\t\tisAntLing\n\t\t),\n\t};\n}\n\n/**\n * Get resolved compatibility settings for a model.\n * Auto-detects from provider/URL then overrides with explicit model.compat.\n */\nfunction getCompat(model: Model<\"openai-completions\">): ResolvedOpenAICompletionsCompat {\n\tconst detected = detectCompat(model);\n\tif (!model.compat) return detected;\n\n\treturn {\n\t\tsupportsStore: model.compat.supportsStore ?? detected.supportsStore,\n\t\tsupportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,\n\t\tsupportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,\n\t\tsupportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,\n\t\tmaxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,\n\t\trequiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,\n\t\trequiresAssistantAfterToolResult:\n\t\t\tmodel.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,\n\t\trequiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,\n\t\trequiresReasoningContentOnAssistantMessages:\n\t\t\tmodel.compat.requiresReasoningContentOnAssistantMessages ??\n\t\t\tdetected.requiresReasoningContentOnAssistantMessages,\n\t\tthinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,\n\t\topenRouterRouting: model.compat.openRouterRouting ?? {},\n\t\tvercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,\n\t\tchatTemplateKwargs: model.compat.chatTemplateKwargs ?? detected.chatTemplateKwargs,\n\t\tzaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,\n\t\tsupportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,\n\t\tsupportsOpenAIGrammarTools: model.compat.supportsOpenAIGrammarTools ?? detected.supportsOpenAIGrammarTools,\n\t\tcacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,\n\t\tsendSessionAffinityHeaders: model.compat.sendSessionAffinityHeaders ?? detected.sendSessionAffinityHeaders,\n\t\tdeferredToolsMode: model.compat.deferredToolsMode ?? detected.deferredToolsMode,\n\t\tsessionAffinityFormat: model.compat.sessionAffinityFormat ?? detected.sessionAffinityFormat,\n\t\tsupportsLongCacheRetention: model.compat.supportsLongCacheRetention ?? detected.supportsLongCacheRetention,\n\t};\n}\n"]}
+{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/api/openai-completions.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,KAAK,EAOX,0BAA0B,EAI1B,MAAM,sCAAsC,CAAC;AAE9C,OAAO,KAAK,EAIX,OAAO,EAGP,KAAK,EACL,uBAAuB,EAGvB,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAEb,eAAe,EAKf,MAAM,aAAa,CAAC;AA0GrB,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,8BAA8B,CAAC;IACpE,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;IAC1E,oGAAoG;IACpG,eAAe,CAAC,EAAE,eAAe,CAAC;CAClC;AAED,MAAM,WAAW,iCAAiC;IACjD,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACzD;AAOD,KAAK,+BAA+B,GAAG,IAAI,CAC1C,QAAQ,CAAC,uBAAuB,CAAC,EACjC,oBAAoB,GAAG,mBAAmB,GAAG,6BAA6B,CAC1E,GAAG;IACH,kBAAkB,CAAC,EAAE,uBAAuB,CAAC,oBAAoB,CAAC,CAAC;IACnE,iBAAiB,CAAC,EAAE,uBAAuB,CAAC,mBAAmB,CAAC,CAAC;IACjE,2BAA2B,CAAC,EAAE,uBAAuB,CAAC,6BAA6B,CAAC,CAAC;CACrF,CAAC;AAmCF,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CA8ZjF,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAkBlF,CAAC;AA4ZF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,+BAA+B,EACvC,OAAO,CAAC,EAAE,iCAAiC,GACzC,0BAA0B,EAAE,CA4R9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionDeveloperMessageParam,\n\tChatCompletionMessageParam,\n\tChatCompletionMessageToolCall,\n\tChatCompletionSystemMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { calculateCost, clampThinkingLevel } from \"../models.ts\";\nimport type {\n\tAssistantMessage,\n\tCacheRetention,\n\tChatTemplateKwargValue,\n\tContext,\n\tImageContent,\n\tMessage,\n\tModel,\n\tOpenAICompletionsCompat,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport {\n\tappendGrammarToolInputJsonDelta,\n\tcreateGrammarToolInputProperties,\n\ttype GrammarToolInputJsonBuffer,\n\tgetGrammarToolInput,\n\tresolveGrammarConstrainedSampling,\n\tresolveJsonSchemaStrictSampling,\n} from \"./constrained-sampling.ts\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.ts\";\nimport { clampOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport { buildBaseOptions, clampReasoning, MIN_ANSWER_TOKENS } from \"./simple-options.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasHeader(headers: ProviderHeaders | undefined, name: string): boolean {\n\tif (!headers) return false;\n\tconst expected = name.toLowerCase();\n\tfor (const [key, value] of Object.entries(headers)) {\n\t\tif (key.toLowerCase() === expected && value !== null && value.trim().length > 0) return true;\n\t}\n\treturn false;\n}\n\nfunction getClientApiKey(provider: string, apiKey: string | undefined, headers: ProviderHeaders | undefined): string {\n\tif (apiKey) return apiKey;\n\tif (hasHeader(headers, \"authorization\") || hasHeader(headers, \"cf-aig-authorization\")) return \"unused\";\n\tthrow new Error(`No API key for provider: ${provider}`);\n}\n\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nfunction getDeferredToolNames(messages: Message[]): Set<string> {\n\tconst names = new Set<string>();\n\tfor (const message of messages) {\n\t\tif (message.role === \"toolResult\") {\n\t\t\tfor (const name of message.addedToolNames ?? []) {\n\t\t\t\tnames.add(name);\n\t\t\t}\n\t\t}\n\t}\n\treturn names;\n}\n\nfunction getToolsByName(tools: Tool[] | undefined, names: Iterable<string>): Tool[] {\n\tif (!tools) return [];\n\tconst toolsByName = new Map(tools.map((tool) => [tool.name, tool]));\n\treturn Array.from(names)\n\t\t.map((name) => toolsByName.get(name))\n\t\t.filter((tool): tool is Tool => tool !== undefined);\n}\n\nfunction isTextContentBlock(block: { type: string }): block is TextContent {\n\treturn block.type === \"text\";\n}\n\nfunction isThinkingContentBlock(block: { type: string }): block is ThinkingContent {\n\treturn block.type === \"thinking\";\n}\n\nfunction isToolCallBlock(block: { type: string }): block is ToolCall {\n\treturn block.type === \"toolCall\";\n}\n\nfunction isImageContentBlock(block: { type: string }): block is ImageContent {\n\treturn block.type === \"image\";\n}\n\nfunction isEncryptedReasoningDetail(detail: unknown): detail is OpenAIEncryptedReasoningDetail {\n\tif (typeof detail !== \"object\" || detail === null) {\n\t\treturn false;\n\t}\n\tconst candidate = detail as Record<string, unknown>;\n\treturn (\n\t\tcandidate.type === \"reasoning.encrypted\" &&\n\t\ttypeof candidate.id === \"string\" &&\n\t\tcandidate.id.length > 0 &&\n\t\ttypeof candidate.data === \"string\" &&\n\t\tcandidate.data.length > 0\n\t);\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: OpenAI.Chat.Completions.ChatCompletionToolChoiceOption;\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\t/** Token budgets per thinking level. Only used when `compat.supportsThinkingTokenBudget` is set. */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\nexport interface ConvertCompletionsMessagesOptions {\n\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\n}\n\ninterface OpenAICompatCacheControl {\n\ttype: \"ephemeral\";\n\tttl?: string;\n}\n\ntype ResolvedOpenAICompletionsCompat = Omit<\n\tRequired<OpenAICompletionsCompat>,\n\t\"cacheControlFormat\" | \"deferredToolsMode\" | \"supportsThinkingTokenBudget\"\n> & {\n\tcacheControlFormat?: OpenAICompletionsCompat[\"cacheControlFormat\"];\n\tdeferredToolsMode?: OpenAICompletionsCompat[\"deferredToolsMode\"];\n\tsupportsThinkingTokenBudget?: OpenAICompletionsCompat[\"supportsThinkingTokenBudget\"];\n};\n\ntype ResolvedChatTemplateKwargValue = string | number | boolean | null;\n\ntype ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;\n\ntype KimiToolSystemMessageParam = {\n\trole: \"system\";\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[];\n};\n\ntype OpenAIEncryptedReasoningDetail = {\n\ttype: \"reasoning.encrypted\";\n\tid: string;\n\tdata: string;\n};\n\ntype ChatCompletionTextPartWithCacheControl = ChatCompletionContentPartText & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\ntype ChatCompletionToolWithCacheControl = OpenAI.Chat.Completions.ChatCompletionTool & {\n\tcache_control?: OpenAICompatCacheControl;\n};\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\nexport const stream: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\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: model.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\ttry {\n\t\t\tconst apiKey = getClientApiKey(model.provider, options?.apiKey, options?.headers);\n\t\t\tconst compat = getCompat(model);\n\t\t\tconst grammarToolInputProperties = createGrammarToolInputProperties(\n\t\t\t\tcontext.tools,\n\t\t\t\tcompat.supportsOpenAIGrammarTools,\n\t\t\t);\n\t\t\tconst cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);\n\t\t\tconst cacheSessionId = cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers, options?.fetch, cacheSessionId, compat);\n\t\t\tlet params = buildParams(model, context, options, compat, cacheRetention, grammarToolInputProperties);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst requestOptions = {\n\t\t\t\t...(options?.signal ? { signal: options.signal } : {}),\n\t\t\t\t...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),\n\t\t\t\tmaxRetries: 0,\n\t\t\t};\n\t\t\tconst { data: openaiStream, response } = await retryProviderRequest(\n\t\t\t\t() => client.chat.completions.create(params, requestOptions).withResponse(),\n\t\t\t\t{\n\t\t\t\t\tmaxRetries: options?.maxRetries,\n\t\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t},\n\t\t\t);\n\t\t\tawait options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tinterface StreamingToolCallBlock extends ToolCall {\n\t\t\t\tpartialArgs?: string;\n\t\t\t\tcustomInput?: {\n\t\t\t\t\tproperty: string;\n\t\t\t\t\tjsonBuffer: GrammarToolInputJsonBuffer;\n\t\t\t\t};\n\t\t\t\tstreamIndex?: number;\n\t\t\t}\n\t\t\ttype StreamingBlock = TextContent | ThinkingContent | StreamingToolCallBlock;\n\t\t\ttype StreamingToolCallDelta = {\n\t\t\t\tindex?: number;\n\t\t\t\tid?: string;\n\t\t\t\ttype?: string;\n\t\t\t\tfunction?: { name?: string; arguments?: string };\n\t\t\t\tcustom?: { name?: string; input?: string };\n\t\t\t};\n\n\t\t\tlet textBlock: TextContent | null = null;\n\t\t\tlet thinkingBlock: ThinkingContent | null = null;\n\t\t\tlet hasFinishReason = false;\n\t\t\tconst toolCallBlocksByIndex = new Map<number, StreamingToolCallBlock>();\n\t\t\tconst toolCallBlocksById = new Map<string, StreamingToolCallBlock>();\n\t\t\tconst pendingReasoningDetailsByToolCallId = new Map<string, string>();\n\t\t\tconst blocks = output.content as StreamingBlock[];\n\t\t\tconst getContentIndex = (block: StreamingBlock) => blocks.indexOf(block);\n\t\t\tconst getCustomToolCallInput = (block: StreamingToolCallBlock): string => {\n\t\t\t\tconst property = block.customInput?.property;\n\t\t\t\tif (property === undefined) return \"\";\n\t\t\t\tconst value = block.arguments[property];\n\t\t\t\treturn typeof value === \"string\" ? value : \"\";\n\t\t\t};\n\t\t\tconst appendCustomToolCallInput = (\n\t\t\t\tblock: StreamingToolCallBlock,\n\t\t\t\tnextInput: string,\n\t\t\t\tclose: boolean,\n\t\t\t): string | undefined => {\n\t\t\t\tconst customInput = block.customInput;\n\t\t\t\tif (!customInput) return undefined;\n\t\t\t\tconst delta = appendGrammarToolInputJsonDelta(\n\t\t\t\t\tcustomInput.jsonBuffer,\n\t\t\t\t\tcustomInput.property,\n\t\t\t\t\tnextInput,\n\t\t\t\t\tclose,\n\t\t\t\t);\n\t\t\t\tblock.arguments = { [customInput.property]: nextInput };\n\t\t\t\treturn delta;\n\t\t\t};\n\t\t\tconst finishBlock = (block: StreamingBlock) => {\n\t\t\t\tconst contentIndex = getContentIndex(block);\n\t\t\t\tif (contentIndex === -1) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tif (block.customInput) {\n\t\t\t\t\t\tconst delta = appendCustomToolCallInput(block, getCustomToolCallInput(block), true);\n\t\t\t\t\t\tif (delta !== undefined) {\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t} else {\n\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t}\n\t\t\t\t\t// Finalize in-place and strip the scratch buffers so replay only\n\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\tdelete block.customInput;\n\t\t\t\t\tdelete block.streamIndex;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureTextBlock = () => {\n\t\t\t\tif (!textBlock) {\n\t\t\t\t\ttextBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\tblocks.push(textBlock);\n\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: getContentIndex(textBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn textBlock;\n\t\t\t};\n\t\t\tconst ensureThinkingBlock = (thinkingSignature: string) => {\n\t\t\t\tif (!thinkingBlock) {\n\t\t\t\t\tthinkingBlock = {\n\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\tthinkingSignature,\n\t\t\t\t\t};\n\t\t\t\t\tblocks.push(thinkingBlock);\n\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: getContentIndex(thinkingBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn thinkingBlock;\n\t\t\t};\n\t\t\tconst applyPendingReasoningDetail = (block: StreamingToolCallBlock) => {\n\t\t\t\tif (!block.id) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tconst pendingReasoningDetail = pendingReasoningDetailsByToolCallId.get(block.id);\n\t\t\t\tif (pendingReasoningDetail) {\n\t\t\t\t\tblock.thoughtSignature = pendingReasoningDetail;\n\t\t\t\t\tpendingReasoningDetailsByToolCallId.delete(block.id);\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureToolCallBlock = (toolCall: StreamingToolCallDelta) => {\n\t\t\t\tconst streamIndex = typeof toolCall.index === \"number\" ? toolCall.index : undefined;\n\t\t\t\tconst name = toolCall.function?.name ?? toolCall.custom?.name ?? \"\";\n\t\t\t\tlet block = streamIndex !== undefined ? toolCallBlocksByIndex.get(streamIndex) : undefined;\n\t\t\t\tif (!block && toolCall.id) {\n\t\t\t\t\tblock = toolCallBlocksById.get(toolCall.id);\n\t\t\t\t}\n\t\t\t\tif (!block) {\n\t\t\t\t\t// Note: the \"input\" fallback here should/must not be taken.  in case the LLM makes up\n\t\t\t\t\t// a tool we don't knwo about, we at least have a place to stash our stuff.\n\t\t\t\t\tconst customInputProperty =\n\t\t\t\t\t\ttoolCall.custom && !toolCall.function ? (grammarToolInputProperties.get(name) ?? \"input\") : undefined;\n\t\t\t\t\tconst hasCustomInput = customInputProperty !== undefined;\n\t\t\t\t\tblock = {\n\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\tname,\n\t\t\t\t\t\targuments: hasCustomInput ? { [customInputProperty]: \"\" } : {},\n\t\t\t\t\t\tpartialArgs: hasCustomInput ? undefined : \"\",\n\t\t\t\t\t\tcustomInput: hasCustomInput\n\t\t\t\t\t\t\t? { property: customInputProperty, jsonBuffer: { input: \"\", started: false, closed: false } }\n\t\t\t\t\t\t\t: undefined,\n\t\t\t\t\t\tstreamIndex,\n\t\t\t\t\t};\n\t\t\t\t\tif (streamIndex !== undefined) {\n\t\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t\t}\n\t\t\t\t\tif (toolCall.id) {\n\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t}\n\t\t\t\t\tblocks.push(block);\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_start\",\n\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tif (streamIndex !== undefined && block.streamIndex === undefined) {\n\t\t\t\t\tblock.streamIndex = streamIndex;\n\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t}\n\t\t\t\tif (toolCall.id) {\n\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t}\n\t\t\t\tif (!block.name && name) {\n\t\t\t\t\tblock.name = name;\n\t\t\t\t}\n\t\t\t\tif (toolCall.custom && !toolCall.function && !block.customInput) {\n\t\t\t\t\tconst customInputProperty = grammarToolInputProperties.get(block.name) ?? \"input\";\n\t\t\t\t\tblock.arguments = { [customInputProperty]: \"\" };\n\t\t\t\t\tblock.customInput = {\n\t\t\t\t\t\tproperty: customInputProperty,\n\t\t\t\t\t\tjsonBuffer: { input: \"\", started: false, closed: false },\n\t\t\t\t\t};\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t}\n\t\t\t\tapplyPendingReasoningDetail(block);\n\t\t\t\treturn block;\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (typeof chunk.model === \"string\" && chunk.model.length > 0 && chunk.model !== model.id) {\n\t\t\t\t\toutput.responseModel ||= chunk.model;\n\t\t\t\t}\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\toutput.rawStopReason = choice.finish_reason;\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t\thasFinishReason = true;\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst block = ensureTextBlock();\n\t\t\t\t\t\tblock.text += choice.delta.content;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tconst deltaFields = choice.delta as Record<string, unknown>;\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tconst value = deltaFields[field];\n\t\t\t\t\t\tif (typeof value === \"string\" && value.length > 0) {\n\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tconst delta = deltaFields[foundReasoningField];\n\t\t\t\t\t\tif (typeof delta === \"string\" && delta.length > 0) {\n\t\t\t\t\t\t\tconst thinkingSignature =\n\t\t\t\t\t\t\t\tmodel.provider === \"opencode-go\" && foundReasoningField === \"reasoning\"\n\t\t\t\t\t\t\t\t\t? \"reasoning_content\"\n\t\t\t\t\t\t\t\t\t: foundReasoningField;\n\t\t\t\t\t\t\tconst block = ensureThinkingBlock(thinkingSignature);\n\t\t\t\t\t\t\tblock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls as StreamingToolCallDelta[]) {\n\t\t\t\t\t\t\tconst block = ensureToolCallBlock(toolCall);\n\t\t\t\t\t\t\tif (!block.id && toolCall.id) {\n\t\t\t\t\t\t\t\tblock.id = toolCall.id;\n\t\t\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tconst name = toolCall.function?.name ?? toolCall.custom?.name;\n\t\t\t\t\t\t\tif (!block.name && name) {\n\t\t\t\t\t\t\t\tblock.name = name;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.partialArgs = (block.partialArgs ?? \"\") + toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\t\t} else if (toolCall.custom?.input) {\n\t\t\t\t\t\t\t\tconst nextInput = getCustomToolCallInput(block) + toolCall.custom.input;\n\t\t\t\t\t\t\t\tdelta = appendCustomToolCallInput(block, nextInput, false) ?? \"\";\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as { reasoning_details?: unknown }).reasoning_details;\n\t\t\t\t\tif (Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (isEncryptedReasoningDetail(detail)) {\n\t\t\t\t\t\t\t\tconst serializedDetail = JSON.stringify(detail);\n\t\t\t\t\t\t\t\tconst matchingToolCall = toolCallBlocksById.get(detail.id);\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = serializedDetail;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tpendingReasoningDetailsByToolCallId.set(detail.id, serializedDetail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfor (const block of blocks) {\n\t\t\t\tfinishBlock(block);\n\t\t\t}\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 === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (!hasFinishReason && !compat.supportsFinishReason) {\n\t\t\t\toutput.stopReason = output.content.some((block) => block.type === \"toolCall\") ? \"toolUse\" : \"stop\";\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\t\t\tif ((compat.supportsFinishReason && !hasFinishReason) || output.stopReason === \"pending\") {\n\t\t\t\tthrow new Error(\"Stream ended without finish_reason\");\n\t\t\t}\n\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\tdelete (block as { index?: number }).index;\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialArgs?: string }).partialArgs;\n\t\t\t\tdelete (block as { customInput?: unknown }).customInput;\n\t\t\t\tdelete (block as { streamIndex?: number }).streamIndex;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatProviderError(normalizeProviderError(error));\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\t// normalizeProviderError already stringifies the parsed body (error.error)\n\t\t\t// into errorMessage, so only append the raw metadata when it is not already\n\t\t\t// present to avoid double-printing it.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata && !output.errorMessage.includes(String(rawMetadata))) {\n\t\t\t\toutput.errorMessage += `\\n${rawMetadata}`;\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\nexport const streamSimple: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tgetClientApiKey(model.provider, options?.apiKey, options?.headers);\n\n\tconst base = buildBaseOptions(model, context, options, options?.apiKey);\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t\tthinkingBudgets: options?.thinkingBudgets,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey: string,\n\toptionsHeaders?: ProviderHeaders,\n\tfetch?: typeof globalThis.fetch,\n\tsessionId?: string,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n) {\n\tconst headers: ProviderHeaders = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\tif (sessionId && compat.sendSessionAffinityHeaders) {\n\t\tif (compat.sessionAffinityFormat === \"openrouter\") {\n\t\t\theaders[\"x-session-id\"] = sessionId;\n\t\t} else {\n\t\t\tif (compat.sessionAffinityFormat === \"openai\") {\n\t\t\t\theaders.session_id = sessionId;\n\t\t\t}\n\t\t\theaders[\"x-client-request-id\"] = sessionId;\n\t\t\theaders[\"x-session-affinity\"] = sessionId;\n\t\t}\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tfetch,\n\t\tdefaultHeaders: headers,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n\tcacheRetention: CacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env),\n\tgrammarToolInputProperties: ReadonlyMap<string, string> = createGrammarToolInputProperties(\n\t\tcontext.tools,\n\t\tcompat.supportsOpenAIGrammarTools,\n\t),\n) {\n\tconst messages = convertMessages(model, context, compat, { grammarToolInputProperties });\n\tconst cacheControl = getCompatCacheControl(compat, cacheRetention);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.id,\n\t\tmessages,\n\t\tstream: true,\n\t\tprompt_cache_key:\n\t\t\t(model.baseUrl.includes(\"api.openai.com\") && cacheRetention !== \"none\") ||\n\t\t\t(cacheRetention === \"long\" && compat.supportsLongCacheRetention)\n\t\t\t\t? clampOpenAIPromptCacheKey(options?.sessionId)\n\t\t\t\t: undefined,\n\t\tprompt_cache_retention: cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"24h\" : undefined,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tconst deferredToolNames =\n\t\tcompat.deferredToolsMode === \"kimi\" ? getDeferredToolNames(context.messages) : new Set<string>();\n\tconst activeTools = context.tools?.filter((tool) => !deferredToolNames.has(tool.name));\n\tif (activeTools && activeTools.length > 0) {\n\t\tparams.tools = convertTools(activeTools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (cacheControl) {\n\t\tapplyAnthropicCacheControl(messages, params.tools, cacheControl);\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\tconst zaiParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\tthinking?: { type: \"enabled\" | \"disabled\"; clear_thinking?: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\tzaiParams.thinking = options?.reasoningEffort ? { type: \"enabled\", clear_thinking: false } : { type: \"disabled\" };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\tconst mappedEffort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\t\tconst effort = mappedEffort === undefined ? options.reasoningEffort : mappedEffort;\n\t\t\tif (typeof effort === \"string\") {\n\t\t\t\tzaiParams.reasoning_effort = effort;\n\t\t\t}\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\tconst effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t\tif (typeof effort === \"string\") {\n\t\t\t\t(params as any).reasoning_effort = effort;\n\t\t\t}\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = {\n\t\t\tenable_thinking: !!options?.reasoningEffort,\n\t\t\tpreserve_thinking: true,\n\t\t};\n\t} else if (compat.thinkingFormat === \"chat-template\" && model.reasoning) {\n\t\tconst chatTemplateKwargs = buildChatTemplateValues(model, options, compat.chatTemplateKwargs);\n\t\tif (chatTemplateKwargs) {\n\t\t\t(params as any).chat_template_kwargs = chatTemplateKwargs;\n\t\t}\n\t} else if (compat.thinkingFormat === \"baseten\" && model.reasoning) {\n\t\tconst basetenParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\tchat_template_args?: Record<string, ResolvedChatTemplateKwargValue>;\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\tconst chatTemplateArgs = buildChatTemplateValues(model, options, compat.chatTemplateArgs);\n\t\tif (chatTemplateArgs) {\n\t\t\tbasetenParams.chat_template_args = chatTemplateArgs;\n\t\t}\n\t\tif (compat.supportsReasoningEffort) {\n\t\t\tconst requestedEffort = options?.reasoningEffort;\n\t\t\tconst mappedEffort = requestedEffort ? model.thinkingLevelMap?.[requestedEffort] : model.thinkingLevelMap?.off;\n\t\t\tconst effort = mappedEffort === undefined ? requestedEffort : mappedEffort;\n\t\t\tif (typeof effort === \"string\") {\n\t\t\t\tbasetenParams.reasoning_effort = effort;\n\t\t\t}\n\t\t}\n\t} else if (compat.thinkingFormat === \"deepseek\" && model.reasoning) {\n\t\tif (options?.reasoningEffort) {\n\t\t\t(params as any).thinking = { type: \"enabled\" };\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\t(params as any).thinking = { type: \"disabled\" };\n\t\t}\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort,\n\t\t\t};\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\topenRouterParams.reasoning = { effort: model.thinkingLevelMap?.off ?? \"none\" };\n\t\t}\n\t} else if (compat.thinkingFormat === \"ant-ling\" && model.reasoning && options?.reasoningEffort) {\n\t\tconst effort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\tif (typeof effort === \"string\") {\n\t\t\t(params as typeof params & { reasoning?: { effort: string } }).reasoning = { effort };\n\t\t}\n\t} else if (compat.thinkingFormat === \"together\" && model.reasoning) {\n\t\tconst togetherParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\treasoning?: { enabled: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\ttogetherParams.reasoning = { enabled: !!options?.reasoningEffort };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\ttogetherParams.reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"string-thinking\" && model.reasoning) {\n\t\tconst stringThinkingParams = params as typeof params & { thinking?: string };\n\t\tif (options?.reasoningEffort) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.off ?? \"none\";\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t} else if (!options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\tconst offValue = model.thinkingLevelMap?.off;\n\t\tif (typeof offValue === \"string\") {\n\t\t\t(params as any).reasoning_effort = offValue;\n\t\t}\n\t}\n\n\t// vLLM caps reasoning with a top-level thinking_token_budget. Independent of\n\t// thinkingFormat: the same server can serve zai, qwen or chat-template models.\n\t// Reasoning and the answer share max_tokens here, so an uncapped reasoning\n\t// phase can consume the whole response and leave no answer and no tool call.\n\tif (compat.supportsThinkingTokenBudget && options?.reasoningEffort && model.reasoning) {\n\t\tconst level = clampReasoning(options.reasoningEffort)!;\n\t\tconst budgets: ThinkingBudgets = {\n\t\t\tminimal: 1024,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 16384,\n\t\t\t...options.thinkingBudgets,\n\t\t};\n\t\tconst ceiling = (params as { max_tokens?: number }).max_tokens ?? params.max_completion_tokens ?? model.maxTokens;\n\t\t// Always leave room for the answer, otherwise the budget recreates the bug it prevents.\n\t\tconst budget = Math.min(budgets[level]!, Math.max(0, ceiling - MIN_ANSWER_TOKENS));\n\t\tif (budget > 0) {\n\t\t\t(params as { thinking_token_budget?: number }).thinking_token_budget = budget;\n\t\t}\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\t// Last so custom keys override the named request fields.\n\tif (options?.samplingParams) {\n\t\tObject.assign(params, options.samplingParams);\n\t}\n\n\treturn params;\n}\n\nfunction buildChatTemplateValues(\n\tmodel: Model<\"openai-completions\">,\n\toptions: OpenAICompletionsOptions | undefined,\n\tvalues: Record<string, ChatTemplateKwargValue>,\n): Record<string, ResolvedChatTemplateKwargValue> | undefined {\n\tconst resolvedValues: Record<string, ResolvedChatTemplateKwargValue> = {};\n\n\tfor (const [key, value] of Object.entries(values)) {\n\t\tconst resolved = resolveChatTemplateKwargValue(model, options, value);\n\t\tif (resolved !== undefined) {\n\t\t\tresolvedValues[key] = resolved;\n\t\t}\n\t}\n\n\treturn Object.keys(resolvedValues).length > 0 ? resolvedValues : undefined;\n}\n\nfunction resolveChatTemplateKwargValue(\n\tmodel: Model<\"openai-completions\">,\n\toptions: OpenAICompletionsOptions | undefined,\n\tvalue: ChatTemplateKwargValue,\n): ResolvedChatTemplateKwargValue | undefined {\n\tif (typeof value !== \"object\" || value === null) {\n\t\treturn value;\n\t}\n\n\tconst reasoningEffort = options?.reasoningEffort;\n\tif (!reasoningEffort && value.omitWhenOff) {\n\t\treturn undefined;\n\t}\n\tif (value.$var === \"thinking.enabled\") {\n\t\treturn !!reasoningEffort;\n\t}\n\n\tconst mappedValue = reasoningEffort ? model.thinkingLevelMap?.[reasoningEffort] : model.thinkingLevelMap?.off;\n\treturn mappedValue === undefined ? reasoningEffort : typeof mappedValue === \"string\" ? mappedValue : undefined;\n}\n\nfunction getCompatCacheControl(\n\tcompat: ResolvedOpenAICompletionsCompat,\n\tcacheRetention: CacheRetention,\n): OpenAICompatCacheControl | undefined {\n\tif (compat.cacheControlFormat !== \"anthropic\" || cacheRetention === \"none\") {\n\t\treturn undefined;\n\t}\n\n\tconst ttl = cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"1h\" : undefined;\n\treturn { type: \"ephemeral\", ...(ttl ? { ttl } : {}) };\n}\n\nfunction applyAnthropicCacheControl(\n\tmessages: ChatCompletionMessageParam[],\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\taddCacheControlToSystemPrompt(messages, cacheControl);\n\taddCacheControlToLastTool(tools, cacheControl);\n\taddCacheControlToLastConversationMessage(messages, cacheControl);\n}\n\nfunction addCacheControlToSystemPrompt(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (const message of messages) {\n\t\tif (message.role === \"system\" || message.role === \"developer\") {\n\t\t\taddCacheControlToInstructionMessage(message, cacheControl);\n\t\t\treturn;\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastConversationMessage(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"tool\") {\n\t\t\tif (addCacheControlToMessage(message, cacheControl)) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastTool(\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tif (!tools || tools.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastTool = tools[tools.length - 1] as ChatCompletionToolWithCacheControl;\n\tlastTool.cache_control = cacheControl;\n}\n\nfunction addCacheControlToInstructionMessage(\n\tmessage: ChatCompletionInstructionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\treturn addCacheControlToTextContent(message, cacheControl);\n}\n\nfunction addCacheControlToMessage(\n\tmessage: ChatCompletionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tif (message.role === \"user\" || message.role === \"assistant\" || message.role === \"tool\") {\n\t\treturn addCacheControlToTextContent(message, cacheControl);\n\t}\n\treturn false;\n}\n\nfunction addCacheControlToTextContent(\n\tmessage:\n\t\t| ChatCompletionInstructionMessageParam\n\t\t| ChatCompletionAssistantMessageParam\n\t\t| ChatCompletionToolMessageParam\n\t\t| Extract<ChatCompletionMessageParam, { role: \"user\" }>,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tconst content = message.content;\n\tif (typeof content === \"string\") {\n\t\tif (content.length === 0) {\n\t\t\treturn false;\n\t\t}\n\t\tmessage.content = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: content,\n\t\t\t\tcache_control: cacheControl,\n\t\t\t},\n\t\t] as ChatCompletionTextPartWithCacheControl[];\n\t\treturn true;\n\t}\n\n\tif (!Array.isArray(content)) {\n\t\treturn false;\n\t}\n\n\tfor (let i = content.length - 1; i >= 0; i--) {\n\t\tconst part = content[i];\n\t\tif (part?.type === \"text\") {\n\t\t\tconst textPart = part as ChatCompletionTextPartWithCacheControl;\n\t\t\ttextPart.cache_control = cacheControl;\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: ResolvedOpenAICompletionsCompat,\n\toptions?: ConvertCompletionsMessagesOptions,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\t// Multiple tool calls in the same turn can share call_id but differ by item_id.\n\t\t// Preserve item-level uniqueness when replaying into Chat Completions, which\n\t\t// requires distinct tool call ids.\n\t\tif (id.includes(\"|\")) {\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\tconst separatorIndex = id.indexOf(\"|\");\n\t\t\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst combinedId = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\t\t\tif (combinedId.length <= 40) {\n\t\t\t\treturn combinedId;\n\t\t\t}\n\t\t\tconst hash = shortHash(id).slice(0, 8);\n\t\t\tconst prefix = callId.slice(0, Math.max(1, 40 - hash.length - 1));\n\t\t\treturn `${prefix}_${hash}`;\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\tif (context.systemPrompt) {\n\t\tconst useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;\n\t\tconst role = useDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role: role, content: sanitizeSurrogates(context.systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ChatCompletionContentPart[] = msg.content.map((item): ChatCompletionContentPart => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText;\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage;\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst assistantTextParts = msg.content\n\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t.filter((block) => block.text.trim().length > 0)\n\t\t\t\t.map(\n\t\t\t\t\t(block) =>\n\t\t\t\t\t\t({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t\t}) satisfies ChatCompletionContentPartText,\n\t\t\t\t);\n\t\t\tconst assistantText = assistantTextParts.map((part) => part.text).join(\"\");\n\n\t\t\tconst nonEmptyThinkingBlocks = msg.content\n\t\t\t\t.filter(isThinkingContentBlock)\n\t\t\t\t.filter((block) => block.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks\n\t\t\t\t\t\t.map((block) => sanitizeSurrogates(block.thinking))\n\t\t\t\t\t\t.join(\"\\n\\n\");\n\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }, ...assistantTextParts];\n\t\t\t\t} else {\n\t\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\t\tif (assistantText.length > 0) {\n\t\t\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tlet signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (model.provider === \"opencode-go\" && signature === \"reasoning\") {\n\t\t\t\t\t\tsignature = \"reasoning_content\";\n\t\t\t\t\t}\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t} else if (assistantText.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter(isToolCallBlock);\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc): ChatCompletionMessageToolCall => {\n\t\t\t\t\tconst customInputProperty = options?.grammarToolInputProperties?.get(tc.name);\n\t\t\t\t\tif (customInputProperty !== undefined) {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\tid: tc.id,\n\t\t\t\t\t\t\ttype: \"custom\",\n\t\t\t\t\t\t\tcustom: {\n\t\t\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\t\t\tinput: sanitizeSurrogates(getGrammarToolInput(tc.name, tc.arguments, customInputProperty)),\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t};\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\tid: tc.id,\n\t\t\t\t\t\ttype: \"function\",\n\t\t\t\t\t\tfunction: {\n\t\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t});\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (\n\t\t\t\tcompat.requiresReasoningContentOnAssistantMessages &&\n\t\t\t\tmodel.reasoning &&\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content === undefined\n\t\t\t) {\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content = \"\";\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tconst deferredToolNames = new Set<string>();\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t\t.map((block) => block.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\tconst toolResultText = hasText ? textResult : hasImages ? \"(see attached image)\" : \"(no tool output)\";\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(toolResultText),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (compat.deferredToolsMode === \"kimi\") {\n\t\t\t\t\tfor (const name of toolMsg.addedToolNames ?? []) {\n\t\t\t\t\t\tdeferredToolNames.add(name);\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (isImageContentBlock(block)) {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\n\t\t\tif (deferredToolNames.size > 0) {\n\t\t\t\tconst deferredTools = getToolsByName(context.tools, deferredToolNames);\n\t\t\t\tif (deferredTools.length > 0) {\n\t\t\t\t\tconst kimiToolMessage: KimiToolSystemMessageParam = {\n\t\t\t\t\t\trole: \"system\",\n\t\t\t\t\t\ttools: convertTools(deferredTools, compat),\n\t\t\t\t\t};\n\t\t\t\t\t// Kimi accepts a system message with tools but omits the standard content field.\n\t\t\t\t\tparams.push(kimiToolMessage as unknown as ChatCompletionMessageParam);\n\t\t\t\t}\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: ResolvedOpenAICompletionsCompat,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn tools.map((tool) => {\n\t\tconst grammar = resolveGrammarConstrainedSampling(tool, compat.supportsOpenAIGrammarTools);\n\t\tif (grammar) {\n\t\t\treturn {\n\t\t\t\ttype: \"custom\",\n\t\t\t\tcustom: {\n\t\t\t\t\tname: tool.name,\n\t\t\t\t\tdescription: tool.description,\n\t\t\t\t\tformat: {\n\t\t\t\t\t\ttype: \"grammar\",\n\t\t\t\t\t\tgrammar: {\n\t\t\t\t\t\t\tsyntax: grammar.format,\n\t\t\t\t\t\t\tdefinition: grammar.definition,\n\t\t\t\t\t\t},\n\t\t\t\t\t},\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\n\t\tconst strict = resolveJsonSchemaStrictSampling(tool, compat.supportsStrictMode !== false);\n\t\treturn {\n\t\t\ttype: \"function\",\n\t\t\tfunction: {\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tparameters: tool.parameters as Record<string, unknown>, // TypeBox already generates JSON Schema\n\t\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t\t...(compat.supportsStrictMode !== false && { strict: strict ?? false }),\n\t\t\t},\n\t\t};\n\t});\n}\n\nfunction parseChunkUsage(\n\trawUsage: {\n\t\tprompt_tokens?: number;\n\t\tcompletion_tokens?: number;\n\t\tprompt_cache_hit_tokens?: number;\n\t\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\t\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t},\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = rawUsage.prompt_tokens || 0;\n\tconst cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;\n\tconst cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;\n\n\t// Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read\n\t// tokens (hits). OpenAI does not document or emit cache_write_tokens, but\n\t// OpenRouter-compatible providers can include it as a separate write count.\n\t// OpenRouter's own provider/tests affirm the separate mapping:\n\t// https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409\n\t// Do not subtract writes from cached_tokens, otherwise spec-compliant\n\t// providers are under-reported. DS4 mirrors this contract too:\n\t// https://github.com/antirez/ds4/pull/29\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\t// OpenAI completion_tokens already includes reasoning_tokens.\n\tconst outputTokens = rawUsage.completion_tokens || 0;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\treasoning: rawUsage.completion_tokens_details?.reasoning_tokens || 0,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\treturn usage;\n}\n\nfunction mapStopReason(reason: ChatCompletionChunk.Choice[\"finish_reason\"] | string): {\n\tstopReason: StopReason;\n\terrorMessage?: string;\n} {\n\tif (reason === null) return { stopReason: \"stop\" };\n\tswitch (reason) {\n\t\tcase \"stop\":\n\t\tcase \"end\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"length\":\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase \"function_call\":\n\t\tcase \"tool_calls\":\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tcase \"content_filter\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: content_filter\" };\n\t\tcase \"network_error\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: network_error\" };\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: `Provider finish_reason: ${reason}`,\n\t\t\t};\n\t}\n}\n\n/**\n * Auto-detect compatibility settings from provider name and baseUrl.\n * Used as the base when model.compat is not set; explicit model.compat\n * entries override these detected values.\n */\nfunction detectCompat(model: Model<\"openai-completions\">): ResolvedOpenAICompletionsCompat {\n\tconst provider = model.provider;\n\tconst baseUrl = model.baseUrl;\n\n\tconst isZai =\n\t\tprovider === \"zai\" ||\n\t\tprovider === \"zai-coding-cn\" ||\n\t\tbaseUrl.includes(\"api.z.ai\") ||\n\t\tbaseUrl.includes(\"open.bigmodel.cn\");\n\tconst isTogether =\n\t\tprovider === \"together\" || baseUrl.includes(\"api.together.ai\") || baseUrl.includes(\"api.together.xyz\");\n\tconst isMoonshot = provider === \"moonshotai\" || provider === \"moonshotai-cn\" || baseUrl.includes(\"api.moonshot.\");\n\tconst isOpenRouter = provider === \"openrouter\" || baseUrl.includes(\"openrouter.ai\");\n\tconst isCloudflareWorkersAI = provider === \"cloudflare-workers-ai\" || baseUrl.includes(\"api.cloudflare.com\");\n\tconst isCloudflareAiGateway = provider === \"cloudflare-ai-gateway\" || baseUrl.includes(\"gateway.ai.cloudflare.com\");\n\tconst isNvidia = provider === \"nvidia\" || baseUrl.includes(\"integrate.api.nvidia.com\");\n\tconst isAntLing = provider === \"ant-ling\" || baseUrl.includes(\"api.ant-ling.com\");\n\n\tconst isNonStandard =\n\t\tisNvidia ||\n\t\tprovider === \"cerebras\" ||\n\t\tbaseUrl.includes(\"cerebras.ai\") ||\n\t\tprovider === \"xai\" ||\n\t\tbaseUrl.includes(\"api.x.ai\") ||\n\t\tisTogether ||\n\t\tbaseUrl.includes(\"chutes.ai\") ||\n\t\tbaseUrl.includes(\"deepseek.com\") ||\n\t\tisZai ||\n\t\tisMoonshot ||\n\t\tprovider === \"opencode\" ||\n\t\tbaseUrl.includes(\"opencode.ai\") ||\n\t\tisCloudflareWorkersAI ||\n\t\tisCloudflareAiGateway ||\n\t\tisAntLing;\n\n\tconst useMaxTokens =\n\t\tbaseUrl.includes(\"chutes.ai\") ||\n\t\tisMoonshot ||\n\t\tisCloudflareAiGateway ||\n\t\tisTogether ||\n\t\tisNvidia ||\n\t\tisAntLing ||\n\t\tisZai;\n\n\tconst isGrok = provider === \"xai\" || baseUrl.includes(\"api.x.ai\");\n\tconst isDeepSeek = provider === \"deepseek\" || baseUrl.includes(\"deepseek.com\");\n\tconst isOpenRouterDeveloperRoleModel =\n\t\tisOpenRouter && (model.id.startsWith(\"anthropic/\") || model.id.startsWith(\"openai/\"));\n\tconst cacheControlFormat = provider === \"openrouter\" && model.id.startsWith(\"anthropic/\") ? \"anthropic\" : undefined;\n\n\treturn {\n\t\tsupportsStore: !isNonStandard,\n\t\tsupportsDeveloperRole: isOpenRouterDeveloperRoleModel || (!isNonStandard && !isOpenRouter),\n\t\tsupportsReasoningEffort:\n\t\t\t!isGrok && !isZai && !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia && !isAntLing,\n\t\tsupportsUsageInStreaming: true,\n\t\tsupportsFinishReason: true,\n\t\tmaxTokensField: useMaxTokens ? \"max_tokens\" : \"max_completion_tokens\",\n\t\trequiresToolResultName: false,\n\t\trequiresAssistantAfterToolResult: false,\n\t\trequiresThinkingAsText: false,\n\t\trequiresReasoningContentOnAssistantMessages: isDeepSeek,\n\t\tthinkingFormat: isDeepSeek\n\t\t\t? \"deepseek\"\n\t\t\t: isZai\n\t\t\t\t? \"zai\"\n\t\t\t\t: isTogether\n\t\t\t\t\t? \"together\"\n\t\t\t\t\t: isAntLing\n\t\t\t\t\t\t? \"ant-ling\"\n\t\t\t\t\t\t: isOpenRouter\n\t\t\t\t\t\t\t? \"openrouter\"\n\t\t\t\t\t\t\t: \"openai\",\n\t\topenRouterRouting: {},\n\t\tvercelGatewayRouting: {},\n\t\tchatTemplateKwargs: {},\n\t\tchatTemplateArgs: {},\n\t\tzaiToolStream: false,\n\t\tsupportsThinkingTokenBudget: false,\n\t\tsupportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,\n\t\tsupportsOpenAIGrammarTools: false,\n\t\tcacheControlFormat,\n\t\tsendSessionAffinityHeaders: false,\n\t\tdeferredToolsMode: undefined,\n\t\tsessionAffinityFormat: isOpenRouter ? \"openrouter\" : \"openai\",\n\t\tsupportsLongCacheRetention: !(\n\t\t\tisTogether ||\n\t\t\tisCloudflareWorkersAI ||\n\t\t\tisCloudflareAiGateway ||\n\t\t\tisNvidia ||\n\t\t\tisAntLing\n\t\t),\n\t};\n}\n\n/**\n * Get resolved compatibility settings for a model.\n * Auto-detects from provider/URL then overrides with explicit model.compat.\n */\nfunction getCompat(model: Model<\"openai-completions\">): ResolvedOpenAICompletionsCompat {\n\tconst detected = detectCompat(model);\n\tif (!model.compat) return detected;\n\n\treturn {\n\t\tsupportsStore: model.compat.supportsStore ?? detected.supportsStore,\n\t\tsupportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,\n\t\tsupportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,\n\t\tsupportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,\n\t\tsupportsFinishReason: model.compat.supportsFinishReason ?? detected.supportsFinishReason,\n\t\tmaxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,\n\t\trequiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,\n\t\trequiresAssistantAfterToolResult:\n\t\t\tmodel.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,\n\t\trequiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,\n\t\trequiresReasoningContentOnAssistantMessages:\n\t\t\tmodel.compat.requiresReasoningContentOnAssistantMessages ??\n\t\t\tdetected.requiresReasoningContentOnAssistantMessages,\n\t\tthinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,\n\t\topenRouterRouting: model.compat.openRouterRouting ?? {},\n\t\tvercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,\n\t\tchatTemplateKwargs: model.compat.chatTemplateKwargs ?? detected.chatTemplateKwargs,\n\t\tchatTemplateArgs: model.compat.chatTemplateArgs ?? detected.chatTemplateArgs,\n\t\tzaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,\n\t\tsupportsThinkingTokenBudget: model.compat.supportsThinkingTokenBudget ?? detected.supportsThinkingTokenBudget,\n\t\tsupportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,\n\t\tsupportsOpenAIGrammarTools: model.compat.supportsOpenAIGrammarTools ?? detected.supportsOpenAIGrammarTools,\n\t\tcacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,\n\t\tsendSessionAffinityHeaders: model.compat.sendSessionAffinityHeaders ?? detected.sendSessionAffinityHeaders,\n\t\tdeferredToolsMode: model.compat.deferredToolsMode ?? detected.deferredToolsMode,\n\t\tsessionAffinityFormat: model.compat.sessionAffinityFormat ?? detected.sessionAffinityFormat,\n\t\tsupportsLongCacheRetention: model.compat.supportsLongCacheRetention ?? detected.supportsLongCacheRetention,\n\t};\n}\n"]}

package/dist/api/openai-completions.js

@@ -11,7 +11,7 @@ import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
 import { appendGrammarToolInputJsonDelta, createGrammarToolInputProperties, getGrammarToolInput, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
-import { buildBaseOptions } from "./simple-options.js";
+import { buildBaseOptions, clampReasoning, MIN_ANSWER_TOKENS } from "./simple-options.js";
 import { transformMessages } from "./transform-messages.js";
 /**
  * Check if conversation messages contain tool calls or tool results.
@@ -116,7 +116,7 @@ export const stream = (model, context, options) => {
                 totalTokens: 0,
                 cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
             },
-            stopReason: "stop",
+            stopReason: "pending",
             timestamp: Date.now(),
         };
         try {
@@ -125,7 +125,7 @@ export const stream = (model, context, options) => {
             const grammarToolInputProperties = createGrammarToolInputProperties(context.tools, compat.supportsOpenAIGrammarTools);
             const cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);
             const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
-            const client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);
+            const client = createClient(model, context, apiKey, options?.headers, options?.fetch, cacheSessionId, compat);
             let params = buildParams(model, context, options, compat, cacheRetention, grammarToolInputProperties);
             const nextParams = await options?.onPayload?.(params, model);
             if (nextParams !== undefined) {
@@ -255,9 +255,7 @@ export const stream = (model, context, options) => {
                 if (!block) {
                     // Note: the "input" fallback here should/must not be taken.  in case the LLM makes up
                     // a tool we don't knwo about, we at least have a place to stash our stuff.
-                    const customInputProperty = toolCall.custom
-                        ? (grammarToolInputProperties.get(name) ?? "input")
-                        : undefined;
+                    const customInputProperty = toolCall.custom && !toolCall.function ? (grammarToolInputProperties.get(name) ?? "input") : undefined;
                     const hasCustomInput = customInputProperty !== undefined;
                     block = {
                         type: "toolCall",
@@ -293,7 +291,7 @@ export const stream = (model, context, options) => {
                 if (!block.name && name) {
                     block.name = name;
                 }
-                if (toolCall.custom && !block.customInput) {
+                if (toolCall.custom && !toolCall.function && !block.customInput) {
                     const customInputProperty = grammarToolInputProperties.get(block.name) ?? "input";
                     block.arguments = { [customInputProperty]: "" };
                     block.customInput = {
@@ -326,6 +324,7 @@ export const stream = (model, context, options) => {
                     output.usage = parseChunkUsage(choice.usage, model);
                 }
                 if (choice.finish_reason) {
+                    output.rawStopReason = choice.finish_reason;
                     const finishReasonResult = mapStopReason(choice.finish_reason);
                     output.stopReason = finishReasonResult.stopReason;
                     if (finishReasonResult.errorMessage) {
@@ -431,10 +430,13 @@ export const stream = (model, context, options) => {
             if (output.stopReason === "aborted") {
                 throw new Error("Request was aborted");
             }
+            if (!hasFinishReason && !compat.supportsFinishReason) {
+                output.stopReason = output.content.some((block) => block.type === "toolCall") ? "toolUse" : "stop";
+            }
             if (output.stopReason === "error") {
                 throw new Error(output.errorMessage || "Provider returned an error stop reason");
             }
-            if (!hasFinishReason) {
+            if ((compat.supportsFinishReason && !hasFinishReason) || output.stopReason === "pending") {
                 throw new Error("Stream ended without finish_reason");
             }
             stream.push({ type: "done", reason: output.stopReason, message: output });
@@ -474,9 +476,10 @@ export const streamSimple = (model, context, options) => {
         ...base,
         reasoningEffort,
         toolChoice,
+        thinkingBudgets: options?.thinkingBudgets,
     });
 };
-function createClient(model, context, apiKey, optionsHeaders, sessionId, compat = getCompat(model)) {
+function createClient(model, context, apiKey, optionsHeaders, fetch, sessionId, compat = getCompat(model)) {
     const headers = { ...model.headers };
     if (model.provider === "github-copilot") {
         const hasImages = hasCopilotVisionInput(context.messages);
@@ -506,6 +509,7 @@ function createClient(model, context, apiKey, optionsHeaders, sessionId, compat
         apiKey,
         baseURL: model.baseUrl,
         dangerouslyAllowBrowser: true,
+        fetch,
         defaultHeaders: headers,
     });
 }
@@ -570,6 +574,12 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
     }
     else if (compat.thinkingFormat === "qwen" && model.reasoning) {
         params.enable_thinking = !!options?.reasoningEffort;
+        if (options?.reasoningEffort && compat.supportsReasoningEffort) {
+            const effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;
+            if (typeof effort === "string") {
+                params.reasoning_effort = effort;
+            }
+        }
     }
     else if (compat.thinkingFormat === "qwen-chat-template" && model.reasoning) {
         params.chat_template_kwargs = {
@@ -578,11 +588,26 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
         };
     }
     else if (compat.thinkingFormat === "chat-template" && model.reasoning) {
-        const chatTemplateKwargs = buildChatTemplateKwargs(model, options, compat);
+        const chatTemplateKwargs = buildChatTemplateValues(model, options, compat.chatTemplateKwargs);
         if (chatTemplateKwargs) {
             params.chat_template_kwargs = chatTemplateKwargs;
         }
     }
+    else if (compat.thinkingFormat === "baseten" && model.reasoning) {
+        const basetenParams = params;
+        const chatTemplateArgs = buildChatTemplateValues(model, options, compat.chatTemplateArgs);
+        if (chatTemplateArgs) {
+            basetenParams.chat_template_args = chatTemplateArgs;
+        }
+        if (compat.supportsReasoningEffort) {
+            const requestedEffort = options?.reasoningEffort;
+            const mappedEffort = requestedEffort ? model.thinkingLevelMap?.[requestedEffort] : model.thinkingLevelMap?.off;
+            const effort = mappedEffort === undefined ? requestedEffort : mappedEffort;
+            if (typeof effort === "string") {
+                basetenParams.reasoning_effort = effort;
+            }
+        }
+    }
     else if (compat.thinkingFormat === "deepseek" && model.reasoning) {
         if (options?.reasoningEffort) {
             params.thinking = { type: "enabled" };
@@ -639,6 +664,26 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
             params.reasoning_effort = offValue;
         }
     }
+    // vLLM caps reasoning with a top-level thinking_token_budget. Independent of
+    // thinkingFormat: the same server can serve zai, qwen or chat-template models.
+    // Reasoning and the answer share max_tokens here, so an uncapped reasoning
+    // phase can consume the whole response and leave no answer and no tool call.
+    if (compat.supportsThinkingTokenBudget && options?.reasoningEffort && model.reasoning) {
+        const level = clampReasoning(options.reasoningEffort);
+        const budgets = {
+            minimal: 1024,
+            low: 2048,
+            medium: 8192,
+            high: 16384,
+            ...options.thinkingBudgets,
+        };
+        const ceiling = params.max_tokens ?? params.max_completion_tokens ?? model.maxTokens;
+        // Always leave room for the answer, otherwise the budget recreates the bug it prevents.
+        const budget = Math.min(budgets[level], Math.max(0, ceiling - MIN_ANSWER_TOKENS));
+        if (budget > 0) {
+            params.thinking_token_budget = budget;
+        }
+    }
     // OpenRouter provider routing preferences
     if (model.compat?.openRouterRouting) {
         params.provider = model.compat.openRouterRouting;
@@ -655,17 +700,21 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
             params.providerOptions = { gateway: gatewayOptions };
         }
     }
+    // Last so custom keys override the named request fields.
+    if (options?.samplingParams) {
+        Object.assign(params, options.samplingParams);
+    }
     return params;
 }
-function buildChatTemplateKwargs(model, options, compat) {
-    const kwargs = {};
-    for (const [key, value] of Object.entries(compat.chatTemplateKwargs)) {
+function buildChatTemplateValues(model, options, values) {
+    const resolvedValues = {};
+    for (const [key, value] of Object.entries(values)) {
         const resolved = resolveChatTemplateKwargValue(model, options, value);
         if (resolved !== undefined) {
-            kwargs[key] = resolved;
+            resolvedValues[key] = resolved;
         }
     }
-    return Object.keys(kwargs).length > 0 ? kwargs : undefined;
+    return Object.keys(resolvedValues).length > 0 ? resolvedValues : undefined;
 }
 function resolveChatTemplateKwargValue(model, options, value) {
     if (typeof value !== "object" || value === null) {
@@ -1138,7 +1187,13 @@ function detectCompat(model) {
         isCloudflareWorkersAI ||
         isCloudflareAiGateway ||
         isAntLing;
-    const useMaxTokens = baseUrl.includes("chutes.ai") || isMoonshot || isCloudflareAiGateway || isTogether || isNvidia || isAntLing;
+    const useMaxTokens = baseUrl.includes("chutes.ai") ||
+        isMoonshot ||
+        isCloudflareAiGateway ||
+        isTogether ||
+        isNvidia ||
+        isAntLing ||
+        isZai;
     const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
     const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
     const isOpenRouterDeveloperRoleModel = isOpenRouter && (model.id.startsWith("anthropic/") || model.id.startsWith("openai/"));
@@ -1148,6 +1203,7 @@ function detectCompat(model) {
         supportsDeveloperRole: isOpenRouterDeveloperRoleModel || (!isNonStandard && !isOpenRouter),
         supportsReasoningEffort: !isGrok && !isZai && !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia && !isAntLing,
         supportsUsageInStreaming: true,
+        supportsFinishReason: true,
         maxTokensField: useMaxTokens ? "max_tokens" : "max_completion_tokens",
         requiresToolResultName: false,
         requiresAssistantAfterToolResult: false,
@@ -1167,7 +1223,9 @@ function detectCompat(model) {
         openRouterRouting: {},
         vercelGatewayRouting: {},
         chatTemplateKwargs: {},
+        chatTemplateArgs: {},
         zaiToolStream: false,
+        supportsThinkingTokenBudget: false,
         supportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,
         supportsOpenAIGrammarTools: false,
         cacheControlFormat,
@@ -1194,6 +1252,7 @@ function getCompat(model) {
         supportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,
         supportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,
         supportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,
+        supportsFinishReason: model.compat.supportsFinishReason ?? detected.supportsFinishReason,
         maxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,
         requiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,
         requiresAssistantAfterToolResult: model.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,
@@ -1204,7 +1263,9 @@ function getCompat(model) {
         openRouterRouting: model.compat.openRouterRouting ?? {},
         vercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,
         chatTemplateKwargs: model.compat.chatTemplateKwargs ?? detected.chatTemplateKwargs,
+        chatTemplateArgs: model.compat.chatTemplateArgs ?? detected.chatTemplateArgs,
         zaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,
+        supportsThinkingTokenBudget: model.compat.supportsThinkingTokenBudget ?? detected.supportsThinkingTokenBudget,
         supportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,
         supportsOpenAIGrammarTools: model.compat.supportsOpenAIGrammarTools ?? detected.supportsOpenAIGrammarTools,
         cacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,