@jecruz/merc-ai 0.1.4 → 0.2.0

package/dist/providers/openai-responses-shared.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-responses-shared.d.ts","sourceRoot":"","sources":["../../src/providers/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAClB,6BAA6B,EAG7B,aAAa,EAOb,mBAAmB,EAEnB,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAChB,OAAO,EAEP,KAAK,EAKL,IAAI,EAEJ,KAAK,EACL,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AAoC5E,MAAM,WAAW,4BAA4B;IAC5C,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,kBAAkB,CAAC,EAAE,CACpB,mBAAmB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,EAC9E,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KACzE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAAC;IAC/D,uBAAuB,CAAC,EAAE,CACzB,KAAK,EAAE,KAAK,EACZ,WAAW,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KAClE,IAAI,CAAC;CACV;AAED,MAAM,WAAW,+BAA+B;IAC/C,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,aAAa,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;CAC1C;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,YAAY,CAAC,EAAE,OAAO,CAAC;CACvB;AAMD,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EACxD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,wBAAwB,EAAE,WAAW,CAAC,MAAM,CAAC,EAC7C,OAAO,CAAC,EAAE,+BAA+B,GACvC,aAAa,CAmMf;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAUzG;AAMD,wBAAsB,sBAAsB,CAAC,IAAI,SAAS,GAAG,EAC5D,YAAY,EAAE,aAAa,CAAC,mBAAmB,CAAC,EAChD,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,2BAA2B,EACnC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,CAAC,EAAE,4BAA4B,GACpC,OAAO,CAAC,IAAI,CAAC,CAwQf","sourcesContent":["import type OpenAI from \"openai\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\n\tResponseFunctionCallOutputItemList,\n\tResponseFunctionToolCall,\n\tResponseInput,\n\tResponseInputContent,\n\tResponseInputImage,\n\tResponseInputItem,\n\tResponseInputText,\n\tResponseOutputMessage,\n\tResponseReasoningItem,\n\tResponseStreamEvent,\n\tResponseToolSearchOutputItemParam,\n} from \"openai/resources/responses/responses.js\";\nimport { calculateCost } from \"../models.js\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tImageContent,\n\tModel,\n\tStopReason,\n\tTextContent,\n\tTextSignatureV1,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tUsage,\n} from \"../types.js\";\nimport type { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { shortHash } from \"../utils/hash.js\";\nimport { parseStreamingJson } from \"../utils/json-parse.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport { transformMessages } from \"./transform-messages.js\";\n\n// =============================================================================\n// Utilities\n// =============================================================================\n\nfunction encodeTextSignatureV1(id: string, phase?: TextSignatureV1[\"phase\"]): string {\n\tconst payload: TextSignatureV1 = { v: 1, id };\n\tif (phase) payload.phase = phase;\n\treturn JSON.stringify(payload);\n}\n\nfunction parseTextSignature(\n\tsignature: string | undefined,\n): { id: string; phase?: TextSignatureV1[\"phase\"] } | undefined {\n\tif (!signature) return undefined;\n\tif (signature.startsWith(\"{\")) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(signature) as Partial<TextSignatureV1>;\n\t\t\tif (parsed.v === 1 && typeof parsed.id === \"string\") {\n\t\t\t\tif (parsed.phase === \"commentary\" || parsed.phase === \"final_answer\") {\n\t\t\t\t\treturn { id: parsed.id, phase: parsed.phase };\n\t\t\t\t}\n\t\t\t\treturn { id: parsed.id };\n\t\t\t}\n\t\t} catch {\n\t\t\t// Fall through to legacy plain-string handling.\n\t\t}\n\t}\n\treturn { id: signature };\n}\n\nexport interface OpenAIResponsesStreamOptions {\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\tresolveServiceTier?: (\n\t\tresponseServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t\trequestServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => ResponseCreateParamsStreaming[\"service_tier\"] | undefined;\n\tapplyServiceTierPricing?: (\n\t\tusage: Usage,\n\t\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => void;\n}\n\nexport interface ConvertResponsesMessagesOptions {\n\tincludeSystemPrompt?: boolean;\n\tdeferredTools?: ReadonlyMap<string, Tool>;\n}\n\nexport interface ConvertResponsesToolsOptions {\n\tstrict?: boolean | null;\n\tdeferLoading?: boolean;\n}\n\n// =============================================================================\n// Message conversion\n// =============================================================================\n\nexport function convertResponsesMessages<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\tallowedToolCallProviders: ReadonlySet<string>,\n\toptions?: ConvertResponsesMessagesOptions,\n): ResponseInput {\n\tconst messages: ResponseInput = [];\n\tconst loadedToolNames = new Set<string>();\n\n\tconst normalizeIdPart = (part: string): string => {\n\t\tconst sanitized = part.replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\tconst normalized = sanitized.length > 64 ? sanitized.slice(0, 64) : sanitized;\n\t\treturn normalized.replace(/_+$/, \"\");\n\t};\n\n\tconst buildForeignResponsesItemId = (itemId: string): string => {\n\t\tconst normalized = `fc_${shortHash(itemId)}`;\n\t\treturn normalized.length > 64 ? normalized.slice(0, 64) : normalized;\n\t};\n\n\tconst normalizeToolCallId = (id: string, _targetModel: Model<TApi>, source: AssistantMessage): string => {\n\t\tif (!allowedToolCallProviders.has(model.provider)) return normalizeIdPart(id);\n\t\tif (!id.includes(\"|\")) return normalizeIdPart(id);\n\t\tconst [callId, itemId] = id.split(\"|\");\n\t\tconst normalizedCallId = normalizeIdPart(callId);\n\t\tconst isForeignToolCall = source.provider !== model.provider || source.api !== model.api;\n\t\tlet normalizedItemId = isForeignToolCall ? buildForeignResponsesItemId(itemId) : normalizeIdPart(itemId);\n\t\t// OpenAI Responses API requires item id to start with \"fc\"\n\t\tif (!normalizedItemId.startsWith(\"fc_\")) {\n\t\t\tnormalizedItemId = normalizeIdPart(`fc_${normalizedItemId}`);\n\t\t}\n\t\treturn `${normalizedCallId}|${normalizedItemId}`;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);\n\n\tconst includeSystemPrompt = options?.includeSystemPrompt ?? true;\n\tif (includeSystemPrompt && context.systemPrompt) {\n\t\tconst compat = model.compat as { supportsDeveloperRole?: boolean } | undefined;\n\t\tconst role = model.reasoning && compat?.supportsDeveloperRole !== false ? \"developer\" : \"system\";\n\t\tmessages.push({\n\t\t\trole,\n\t\t\tcontent: sanitizeSurrogates(context.systemPrompt),\n\t\t});\n\t}\n\n\tlet msgIndex = 0;\n\tfor (const msg of transformedMessages) {\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tmessages.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [{ type: \"input_text\", text: sanitizeSurrogates(msg.content) }],\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ResponseInputContent[] = msg.content.map((item): ResponseInputContent => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"input_text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ResponseInputText;\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttype: \"input_image\",\n\t\t\t\t\t\tdetail: \"auto\",\n\t\t\t\t\t\timage_url: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t} satisfies ResponseInputImage;\n\t\t\t\t});\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tmessages.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\tconst output: ResponseInput = [];\n\t\t\tconst assistantMsg = msg as AssistantMessage;\n\t\t\tconst isDifferentModel =\n\t\t\t\tassistantMsg.model !== model.id &&\n\t\t\t\tassistantMsg.provider === model.provider &&\n\t\t\t\tassistantMsg.api === model.api;\n\n\t\t\tfor (const block of msg.content) {\n\t\t\t\tif (block.type === \"thinking\") {\n\t\t\t\t\tif (block.thinkingSignature) {\n\t\t\t\t\t\tconst reasoningItem = JSON.parse(block.thinkingSignature) as ResponseReasoningItem;\n\t\t\t\t\t\toutput.push(reasoningItem);\n\t\t\t\t\t}\n\t\t\t\t} else if (block.type === \"text\") {\n\t\t\t\t\tconst textBlock = block as TextContent;\n\t\t\t\t\tconst parsedSignature = parseTextSignature(textBlock.textSignature);\n\t\t\t\t\t// OpenAI requires id to be max 64 characters\n\t\t\t\t\tlet msgId = parsedSignature?.id;\n\t\t\t\t\tif (!msgId) {\n\t\t\t\t\t\tmsgId = `msg_${msgIndex}`;\n\t\t\t\t\t} else if (msgId.length > 64) {\n\t\t\t\t\t\tmsgId = `msg_${shortHash(msgId)}`;\n\t\t\t\t\t}\n\t\t\t\t\toutput.push({\n\t\t\t\t\t\ttype: \"message\",\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: [{ type: \"output_text\", text: sanitizeSurrogates(textBlock.text), annotations: [] }],\n\t\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\t\tid: msgId,\n\t\t\t\t\t\tphase: parsedSignature?.phase,\n\t\t\t\t\t} satisfies ResponseOutputMessage);\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tconst toolCall = block as ToolCall;\n\t\t\t\t\tconst [callId, itemIdRaw] = toolCall.id.split(\"|\");\n\t\t\t\t\tlet itemId: string | undefined = itemIdRaw;\n\n\t\t\t\t\t// For different-model messages, set id to undefined to avoid pairing validation.\n\t\t\t\t\t// OpenAI tracks which fc_xxx IDs were paired with rs_xxx reasoning items.\n\t\t\t\t\t// By omitting the id, we avoid triggering that validation (like cross-provider does).\n\t\t\t\t\tif (isDifferentModel && itemId?.startsWith(\"fc_\")) {\n\t\t\t\t\t\titemId = undefined;\n\t\t\t\t\t}\n\n\t\t\t\t\toutput.push({\n\t\t\t\t\t\ttype: \"function_call\",\n\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\targuments: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (output.length === 0) continue;\n\t\t\tmessages.push(...output);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst textResult = msg.content\n\t\t\t\t.filter((c): c is TextContent => c.type === \"text\")\n\t\t\t\t.map((c) => c.text)\n\t\t\t\t.join(\"\\n\");\n\t\t\tconst hasImages = msg.content.some((c): c is ImageContent => c.type === \"image\");\n\t\t\tconst hasText = textResult.length > 0;\n\t\t\tconst [callId] = msg.toolCallId.split(\"|\");\n\n\t\t\tlet output: string | ResponseFunctionCallOutputItemList;\n\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\tconst contentParts: ResponseFunctionCallOutputItemList = [];\n\n\t\t\t\tif (hasText) {\n\t\t\t\t\tcontentParts.push({\n\t\t\t\t\t\ttype: \"input_text\",\n\t\t\t\t\t\ttext: sanitizeSurrogates(textResult),\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tfor (const block of msg.content) {\n\t\t\t\t\tif (block.type === \"image\") {\n\t\t\t\t\t\tcontentParts.push({\n\t\t\t\t\t\t\ttype: \"input_image\",\n\t\t\t\t\t\t\tdetail: \"auto\",\n\t\t\t\t\t\t\timage_url: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\toutput = contentParts;\n\t\t\t} else {\n\t\t\t\toutput = sanitizeSurrogates(hasText ? textResult : hasImages ? \"(see attached image)\" : \"(no tool output)\");\n\t\t\t}\n\n\t\t\tmessages.push({\n\t\t\t\ttype: \"function_call_output\",\n\t\t\t\tcall_id: callId,\n\t\t\t\toutput,\n\t\t\t});\n\n\t\t\tconst deferredTools: Tool[] = [];\n\t\t\tfor (const name of msg.addedToolNames ?? []) {\n\t\t\t\tconst tool = options?.deferredTools?.get(name);\n\t\t\t\tif (!tool || loadedToolNames.has(name)) continue;\n\t\t\t\tloadedToolNames.add(name);\n\t\t\t\tdeferredTools.push(tool);\n\t\t\t}\n\t\t\tif (deferredTools.length > 0) {\n\t\t\t\tconst names = deferredTools.map((tool) => tool.name);\n\t\t\t\tconst searchCallId = `pi_tool_load_${shortHash(`${msg.toolCallId}:${names.join(\",\")}`)}`;\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"tool_search_call\",\n\t\t\t\t\tcall_id: searchCallId,\n\t\t\t\t\texecution: \"client\",\n\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\targuments: { query: names.join(\" \"), limit: names.length },\n\t\t\t\t} satisfies ResponseInputItem);\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"tool_search_output\",\n\t\t\t\t\tcall_id: searchCallId,\n\t\t\t\t\texecution: \"client\",\n\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\ttools: convertResponsesTools(deferredTools, { deferLoading: true }),\n\t\t\t\t} satisfies ResponseToolSearchOutputItemParam);\n\t\t\t}\n\t\t}\n\t\tmsgIndex++;\n\t}\n\n\treturn messages;\n}\n\n// =============================================================================\n// Tool conversion\n// =============================================================================\n\nexport function convertResponsesTools(tools: Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[] {\n\tconst strict = options?.strict === undefined ? false : options.strict;\n\treturn tools.map((tool) => ({\n\t\ttype: \"function\",\n\t\tname: tool.name,\n\t\tdescription: tool.description,\n\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\tstrict,\n\t\t...(options?.deferLoading ? { defer_loading: true } : {}),\n\t}));\n}\n\n// =============================================================================\n// Stream processing\n// =============================================================================\n\nexport async function processResponsesStream<TApi extends Api>(\n\topenaiStream: AsyncIterable<ResponseStreamEvent>,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<TApi>,\n\toptions?: OpenAIResponsesStreamOptions,\n): Promise<void> {\n\tlet currentItem: ResponseReasoningItem | ResponseOutputMessage | ResponseFunctionToolCall | null = null;\n\tlet currentBlock: ThinkingContent | TextContent | (ToolCall & { partialJson: string }) | null = null;\n\tlet terminalEventSeen = false;\n\tconst reasoningBlocksById = new Map<string, ThinkingContent>();\n\tconst blocks = output.content;\n\tconst blockIndex = () => blocks.length - 1;\n\tconst backfillReasoningSignatures = (responseOutput: ResponseReasoningItem[]): void => {\n\t\tfor (const item of responseOutput) {\n\t\t\tif (!item.encrypted_content) continue;\n\t\t\tconst block = reasoningBlocksById.get(item.id);\n\t\t\tif (!block?.thinkingSignature) continue;\n\n\t\t\tconst storedItem = JSON.parse(block.thinkingSignature) as ResponseReasoningItem;\n\t\t\tif (storedItem.encrypted_content) continue;\n\t\t\tblock.thinkingSignature = JSON.stringify({ ...storedItem, encrypted_content: item.encrypted_content });\n\t\t}\n\t};\n\n\tfor await (const event of openaiStream) {\n\t\tif (event.type === \"response.created\") {\n\t\t\toutput.responseId = event.response.id;\n\t\t} else if (event.type === \"response.output_item.added\") {\n\t\t\tconst item = event.item;\n\t\t\tif (item.type === \"reasoning\") {\n\t\t\t\tcurrentItem = item;\n\t\t\t\tcurrentBlock = { type: \"thinking\", thinking: \"\" };\n\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t} else if (item.type === \"message\") {\n\t\t\t\tcurrentItem = item;\n\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t} else if (item.type === \"function_call\") {\n\t\t\t\tcurrentItem = item;\n\t\t\t\tcurrentBlock = {\n\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\tid: `${item.call_id}|${item.id}`,\n\t\t\t\t\tname: item.name,\n\t\t\t\t\targuments: {},\n\t\t\t\t\tpartialJson: item.arguments || \"\",\n\t\t\t\t};\n\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t}\n\t\t} else if (event.type === \"response.reasoning_summary_part.added\") {\n\t\t\tif (currentItem && currentItem.type === \"reasoning\") {\n\t\t\t\tcurrentItem.summary = currentItem.summary || [];\n\t\t\t\tcurrentItem.summary.push(event.part);\n\t\t\t}\n\t\t} else if (event.type === \"response.reasoning_summary_text.delta\") {\n\t\t\tif (currentItem?.type === \"reasoning\" && currentBlock?.type === \"thinking\") {\n\t\t\t\tcurrentItem.summary = currentItem.summary || [];\n\t\t\t\tconst lastPart = currentItem.summary[currentItem.summary.length - 1];\n\t\t\t\tif (lastPart) {\n\t\t\t\t\tcurrentBlock.thinking += event.delta;\n\t\t\t\t\tlastPart.text += event.delta;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tdelta: event.delta,\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} else if (event.type === \"response.reasoning_summary_part.done\") {\n\t\t\tif (currentItem?.type === \"reasoning\" && currentBlock?.type === \"thinking\") {\n\t\t\t\tcurrentItem.summary = currentItem.summary || [];\n\t\t\t\tconst lastPart = currentItem.summary[currentItem.summary.length - 1];\n\t\t\t\tif (lastPart) {\n\t\t\t\t\tcurrentBlock.thinking += \"\\n\\n\";\n\t\t\t\t\tlastPart.text += \"\\n\\n\";\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tdelta: \"\\n\\n\",\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} else if (event.type === \"response.reasoning_text.delta\") {\n\t\t\tif (currentItem?.type === \"reasoning\" && currentBlock?.type === \"thinking\") {\n\t\t\t\tcurrentBlock.thinking += event.delta;\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\tdelta: event.delta,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (event.type === \"response.content_part.added\") {\n\t\t\tif (currentItem?.type === \"message\") {\n\t\t\t\tcurrentItem.content = currentItem.content || [];\n\t\t\t\t// Filter out ReasoningText, only accept output_text and refusal\n\t\t\t\tif (event.part.type === \"output_text\" || event.part.type === \"refusal\") {\n\t\t\t\t\tcurrentItem.content.push(event.part);\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (event.type === \"response.output_text.delta\") {\n\t\t\tif (currentItem?.type === \"message\" && currentBlock?.type === \"text\") {\n\t\t\t\tif (!currentItem.content || currentItem.content.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tconst lastPart = currentItem.content[currentItem.content.length - 1];\n\t\t\t\tif (lastPart?.type === \"output_text\") {\n\t\t\t\t\tcurrentBlock.text += event.delta;\n\t\t\t\t\tlastPart.text += event.delta;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tdelta: event.delta,\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} else if (event.type === \"response.refusal.delta\") {\n\t\t\tif (currentItem?.type === \"message\" && currentBlock?.type === \"text\") {\n\t\t\t\tif (!currentItem.content || currentItem.content.length === 0) {\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tconst lastPart = currentItem.content[currentItem.content.length - 1];\n\t\t\t\tif (lastPart?.type === \"refusal\") {\n\t\t\t\t\tcurrentBlock.text += event.delta;\n\t\t\t\t\tlastPart.refusal += event.delta;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tdelta: event.delta,\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} else if (event.type === \"response.function_call_arguments.delta\") {\n\t\t\tif (currentItem?.type === \"function_call\" && currentBlock?.type === \"toolCall\") {\n\t\t\t\tcurrentBlock.partialJson += event.delta;\n\t\t\t\tcurrentBlock.arguments = parseStreamingJson(currentBlock.partialJson);\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\tdelta: event.delta,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (event.type === \"response.function_call_arguments.done\") {\n\t\t\tif (currentItem?.type === \"function_call\" && currentBlock?.type === \"toolCall\") {\n\t\t\t\tconst previousPartialJson = currentBlock.partialJson;\n\t\t\t\tcurrentBlock.partialJson = event.arguments;\n\t\t\t\tcurrentBlock.arguments = parseStreamingJson(currentBlock.partialJson);\n\n\t\t\t\tif (event.arguments.startsWith(previousPartialJson)) {\n\t\t\t\t\tconst delta = event.arguments.slice(previousPartialJson.length);\n\t\t\t\t\tif (delta.length > 0) {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\tpartial: output,\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} else if (event.type === \"response.output_item.done\") {\n\t\t\tconst item = event.item;\n\n\t\t\tif (item.type === \"reasoning\" && currentBlock?.type === \"thinking\") {\n\t\t\t\tconst summaryText = item.summary?.map((s) => s.text).join(\"\\n\\n\") || \"\";\n\t\t\t\tconst contentText = item.content?.map((c) => c.text).join(\"\\n\\n\") || \"\";\n\t\t\t\tcurrentBlock.thinking = summaryText || contentText || currentBlock.thinking;\n\t\t\t\tcurrentBlock.thinkingSignature = JSON.stringify(item);\n\t\t\t\treasoningBlocksById.set(item.id, currentBlock);\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t\tcurrentBlock = null;\n\t\t\t} else if (item.type === \"message\" && currentBlock?.type === \"text\") {\n\t\t\t\tcurrentBlock.text = item.content.map((c) => (c.type === \"output_text\" ? c.text : c.refusal)).join(\"\");\n\t\t\t\tcurrentBlock.textSignature = encodeTextSignatureV1(item.id, item.phase ?? undefined);\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t\tcurrentBlock = null;\n\t\t\t} else if (item.type === \"function_call\") {\n\t\t\t\tconst args =\n\t\t\t\t\tcurrentBlock?.type === \"toolCall\" && currentBlock.partialJson\n\t\t\t\t\t\t? parseStreamingJson(currentBlock.partialJson)\n\t\t\t\t\t\t: parseStreamingJson(item.arguments || \"{}\");\n\n\t\t\t\tlet toolCall: ToolCall;\n\t\t\t\tif (currentBlock?.type === \"toolCall\") {\n\t\t\t\t\t// Finalize in-place and strip the scratch buffer so replay only\n\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\tcurrentBlock.arguments = args;\n\t\t\t\t\tdelete (currentBlock as { partialJson?: string }).partialJson;\n\t\t\t\t\ttoolCall = currentBlock;\n\t\t\t\t} else {\n\t\t\t\t\ttoolCall = {\n\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\tid: `${item.call_id}|${item.id}`,\n\t\t\t\t\t\tname: item.name,\n\t\t\t\t\t\targuments: args,\n\t\t\t\t\t};\n\t\t\t\t}\n\n\t\t\t\tcurrentBlock = null;\n\t\t\t\tstream.push({ type: \"toolcall_end\", contentIndex: blockIndex(), toolCall, partial: output });\n\t\t\t}\n\t\t} else if (event.type === \"response.completed\" || event.type === \"response.incomplete\") {\n\t\t\tterminalEventSeen = true;\n\t\t\tconst response = event.response;\n\t\t\tbackfillReasoningSignatures(\n\t\t\t\tresponse.output?.filter((item): item is ResponseReasoningItem => item.type === \"reasoning\") ?? [],\n\t\t\t);\n\t\t\tif (response?.id) {\n\t\t\t\toutput.responseId = response.id;\n\t\t\t}\n\t\t\tif (response?.usage) {\n\t\t\t\tconst cachedTokens = response.usage.input_tokens_details?.cached_tokens || 0;\n\t\t\t\toutput.usage = {\n\t\t\t\t\t// OpenAI includes cached tokens in input_tokens, so subtract to get non-cached input\n\t\t\t\t\tinput: (response.usage.input_tokens || 0) - cachedTokens,\n\t\t\t\t\toutput: response.usage.output_tokens || 0,\n\t\t\t\t\tcacheRead: cachedTokens,\n\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\ttotalTokens: response.usage.total_tokens || 0,\n\t\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t\t};\n\t\t\t}\n\t\t\tcalculateCost(model, output.usage);\n\t\t\tif (options?.applyServiceTierPricing) {\n\t\t\t\tconst serviceTier = options.resolveServiceTier\n\t\t\t\t\t? options.resolveServiceTier(response?.service_tier, options.serviceTier)\n\t\t\t\t\t: (response?.service_tier ?? options.serviceTier);\n\t\t\t\toptions.applyServiceTierPricing(output.usage, serviceTier);\n\t\t\t}\n\t\t\t// Map status to stop reason\n\t\t\toutput.stopReason = mapStopReason(response?.status);\n\t\t\tif (output.content.some((b) => b.type === \"toolCall\") && output.stopReason === \"stop\") {\n\t\t\t\toutput.stopReason = \"toolUse\";\n\t\t\t}\n\t\t} else if (event.type === \"error\") {\n\t\t\tthrow new Error(`Error Code ${event.code}: ${event.message}` || \"Unknown error\");\n\t\t} else if (event.type === \"response.failed\") {\n\t\t\tterminalEventSeen = true;\n\t\t\tconst error = event.response?.error;\n\t\t\tconst details = event.response?.incomplete_details;\n\t\t\tconst msg = error\n\t\t\t\t? `${error.code || \"unknown\"}: ${error.message || \"no message\"}`\n\t\t\t\t: details?.reason\n\t\t\t\t\t? `incomplete: ${details.reason}`\n\t\t\t\t\t: \"Unknown error (no error details in response)\";\n\t\t\tthrow new Error(msg);\n\t\t}\n\t}\n\n\tif (!terminalEventSeen) {\n\t\tthrow new Error(\"OpenAI Responses stream ended before a terminal response event\");\n\t}\n}\n\nfunction mapStopReason(status: OpenAI.Responses.ResponseStatus | undefined): StopReason {\n\tif (!status) return \"stop\";\n\tswitch (status) {\n\t\tcase \"completed\":\n\t\t\treturn \"stop\";\n\t\tcase \"incomplete\":\n\t\t\treturn \"length\";\n\t\tcase \"failed\":\n\t\tcase \"cancelled\":\n\t\t\treturn \"error\";\n\t\t// These two are wonky ...\n\t\tcase \"in_progress\":\n\t\tcase \"queued\":\n\t\t\treturn \"stop\";\n\t\tdefault: {\n\t\t\tconst _exhaustive: never = status;\n\t\t\tthrow new Error(`Unhandled stop reason: ${_exhaustive}`);\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"openai-responses-shared.d.ts","sourceRoot":"","sources":["../../src/providers/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAClB,6BAA6B,EAE7B,aAAa,EAQb,mBAAmB,EAEnB,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAChB,OAAO,EAEP,KAAK,EAKL,IAAI,EAEJ,KAAK,EACL,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AAqC5E,MAAM,WAAW,4BAA4B;IAC5C,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,kBAAkB,CAAC,EAAE,CACpB,mBAAmB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,EAC9E,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KACzE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAAC;IAC/D,uBAAuB,CAAC,EAAE,CACzB,KAAK,EAAE,KAAK,EACZ,WAAW,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KAClE,IAAI,CAAC;CACV;AAED,MAAM,WAAW,+BAA+B;IAC/C,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,aAAa,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;CAC1C;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,YAAY,CAAC,EAAE,OAAO,CAAC;CACvB;AAMD,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EACxD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,wBAAwB,EAAE,WAAW,CAAC,MAAM,CAAC,EAC7C,OAAO,CAAC,EAAE,+BAA+B,GACvC,aAAa,CAoLf;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAUzG;AAeD,wBAAsB,sBAAsB,CAAC,IAAI,SAAS,GAAG,EAC5D,YAAY,EAAE,aAAa,CAAC,mBAAmB,CAAC,EAChD,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,2BAA2B,EACnC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,CAAC,EAAE,4BAA4B,GACpC,OAAO,CAAC,IAAI,CAAC,CAoOf","sourcesContent":["import type OpenAI from \"openai\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\n\tResponseFunctionCallOutputItemList,\n\tResponseInput,\n\tResponseInputContent,\n\tResponseInputImage,\n\tResponseInputItem,\n\tResponseInputText,\n\tResponseOutputItem,\n\tResponseOutputMessage,\n\tResponseReasoningItem,\n\tResponseStreamEvent,\n\tResponseToolSearchOutputItemParam,\n} from \"openai/resources/responses/responses.js\";\nimport { calculateCost } from \"../models.js\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tImageContent,\n\tModel,\n\tStopReason,\n\tTextContent,\n\tTextSignatureV1,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tUsage,\n} from \"../types.js\";\nimport type { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { shortHash } from \"../utils/hash.js\";\nimport { finalizeToolCallArguments, parseStreamingJson } from \"../utils/json-parse.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport { createProviderToolCallIdNormalizer } from \"../utils/tool-call-id.js\";\nimport { transformMessages } from \"./transform-messages.js\";\n\n// =============================================================================\n// Utilities\n// =============================================================================\n\nfunction encodeTextSignatureV1(id: string, phase?: TextSignatureV1[\"phase\"]): string {\n\tconst payload: TextSignatureV1 = { v: 1, id };\n\tif (phase) payload.phase = phase;\n\treturn JSON.stringify(payload);\n}\n\nfunction parseTextSignature(\n\tsignature: string | undefined,\n): { id: string; phase?: TextSignatureV1[\"phase\"] } | undefined {\n\tif (!signature) return undefined;\n\tif (signature.startsWith(\"{\")) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(signature) as Partial<TextSignatureV1>;\n\t\t\tif (parsed.v === 1 && typeof parsed.id === \"string\") {\n\t\t\t\tif (parsed.phase === \"commentary\" || parsed.phase === \"final_answer\") {\n\t\t\t\t\treturn { id: parsed.id, phase: parsed.phase };\n\t\t\t\t}\n\t\t\t\treturn { id: parsed.id };\n\t\t\t}\n\t\t} catch {\n\t\t\t// Fall through to legacy plain-string handling.\n\t\t}\n\t}\n\treturn { id: signature };\n}\n\nexport interface OpenAIResponsesStreamOptions {\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\tresolveServiceTier?: (\n\t\tresponseServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t\trequestServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => ResponseCreateParamsStreaming[\"service_tier\"] | undefined;\n\tapplyServiceTierPricing?: (\n\t\tusage: Usage,\n\t\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => void;\n}\n\nexport interface ConvertResponsesMessagesOptions {\n\tincludeSystemPrompt?: boolean;\n\tdeferredTools?: ReadonlyMap<string, Tool>;\n}\n\nexport interface ConvertResponsesToolsOptions {\n\tstrict?: boolean | null;\n\tdeferLoading?: boolean;\n}\n\n// =============================================================================\n// Message conversion\n// =============================================================================\n\nexport function convertResponsesMessages<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\tallowedToolCallProviders: ReadonlySet<string>,\n\toptions?: ConvertResponsesMessagesOptions,\n): ResponseInput {\n\tconst messages: ResponseInput = [];\n\tconst loadedToolNames = new Set<string>();\n\n\tconst normalizeToolCallId = createProviderToolCallIdNormalizer(model, allowedToolCallProviders);\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id, _model, source) =>\n\t\tnormalizeToolCallId(id, source),\n\t);\n\n\tconst includeSystemPrompt = options?.includeSystemPrompt ?? true;\n\tif (includeSystemPrompt && context.systemPrompt) {\n\t\tconst compat = model.compat as { supportsDeveloperRole?: boolean } | undefined;\n\t\tconst role = model.reasoning && compat?.supportsDeveloperRole !== false ? \"developer\" : \"system\";\n\t\tmessages.push({\n\t\t\trole,\n\t\t\tcontent: sanitizeSurrogates(context.systemPrompt),\n\t\t});\n\t}\n\n\tlet msgIndex = 0;\n\tfor (const msg of transformedMessages) {\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tmessages.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [{ type: \"input_text\", text: sanitizeSurrogates(msg.content) }],\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ResponseInputContent[] = msg.content.map((item): ResponseInputContent => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"input_text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ResponseInputText;\n\t\t\t\t\t}\n\t\t\t\t\treturn {\n\t\t\t\t\t\ttype: \"input_image\",\n\t\t\t\t\t\tdetail: \"auto\",\n\t\t\t\t\t\timage_url: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t} satisfies ResponseInputImage;\n\t\t\t\t});\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tmessages.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\tconst output: ResponseInput = [];\n\t\t\tconst assistantMsg = msg as AssistantMessage;\n\t\t\tconst isSameProviderAndApi = assistantMsg.provider === model.provider && assistantMsg.api === model.api;\n\t\t\tconst isSameModel = isSameProviderAndApi && assistantMsg.model === model.id;\n\t\t\tconst isDifferentModel = isSameProviderAndApi && assistantMsg.model !== model.id;\n\t\t\tlet textBlockIndex = 0;\n\n\t\t\tfor (const block of msg.content) {\n\t\t\t\tif (block.type === \"thinking\") {\n\t\t\t\t\tif (block.thinkingSignature) {\n\t\t\t\t\t\tconst reasoningItem = JSON.parse(block.thinkingSignature) as ResponseReasoningItem;\n\t\t\t\t\t\toutput.push(reasoningItem);\n\t\t\t\t\t}\n\t\t\t\t} else if (block.type === \"text\") {\n\t\t\t\t\tconst textBlock = block as TextContent;\n\t\t\t\t\tconst parsedSignature = parseTextSignature(textBlock.textSignature);\n\t\t\t\t\tconst fallbackMessageId =\n\t\t\t\t\t\ttextBlockIndex === 0 ? `msg_pi_${msgIndex}` : `msg_pi_${msgIndex}_${textBlockIndex}`;\n\t\t\t\t\ttextBlockIndex++;\n\t\t\t\t\t// OpenAI requires id to be max 64 characters\n\t\t\t\t\tlet msgId = parsedSignature?.id;\n\t\t\t\t\tif (!msgId) {\n\t\t\t\t\t\tmsgId = fallbackMessageId;\n\t\t\t\t\t} else if (msgId.length > 64) {\n\t\t\t\t\t\tmsgId = `msg_${shortHash(msgId)}`;\n\t\t\t\t\t}\n\t\t\t\t\toutput.push({\n\t\t\t\t\t\ttype: \"message\",\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: [{ type: \"output_text\", text: sanitizeSurrogates(textBlock.text), annotations: [] }],\n\t\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\t\tid: msgId,\n\t\t\t\t\t\tphase: parsedSignature?.phase,\n\t\t\t\t\t} satisfies ResponseOutputMessage);\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tconst toolCall = block as ToolCall;\n\t\t\t\t\tconst [callId, itemIdRaw] = toolCall.id.split(\"|\");\n\t\t\t\t\tlet itemId: string | undefined = itemIdRaw;\n\n\t\t\t\t\t// For different-model messages, set id to undefined to avoid pairing validation.\n\t\t\t\t\t// OpenAI tracks which fc_xxx IDs were paired with rs_xxx reasoning items.\n\t\t\t\t\t// By omitting the id, we avoid triggering that validation (like cross-provider does).\n\t\t\t\t\tif (isDifferentModel && itemId?.startsWith(\"fc_\")) {\n\t\t\t\t\t\titemId = undefined;\n\t\t\t\t\t}\n\n\t\t\t\t\tconst canReplayNamespace = isSameModel || options?.deferredTools?.has(toolCall.name) === true;\n\n\t\t\t\t\toutput.push({\n\t\t\t\t\t\ttype: \"function_call\",\n\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\targuments: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t...(canReplayNamespace && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (output.length === 0) continue;\n\t\t\tmessages.push(...output);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst textResult = msg.content\n\t\t\t\t.filter((c): c is TextContent => c.type === \"text\")\n\t\t\t\t.map((c) => c.text)\n\t\t\t\t.join(\"\\n\");\n\t\t\tconst hasImages = msg.content.some((c): c is ImageContent => c.type === \"image\");\n\t\t\tconst hasText = textResult.length > 0;\n\t\t\tconst [callId] = msg.toolCallId.split(\"|\");\n\n\t\t\tlet output: string | ResponseFunctionCallOutputItemList;\n\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\tconst contentParts: ResponseFunctionCallOutputItemList = [];\n\n\t\t\t\tif (hasText) {\n\t\t\t\t\tcontentParts.push({\n\t\t\t\t\t\ttype: \"input_text\",\n\t\t\t\t\t\ttext: sanitizeSurrogates(textResult),\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tfor (const block of msg.content) {\n\t\t\t\t\tif (block.type === \"image\") {\n\t\t\t\t\t\tcontentParts.push({\n\t\t\t\t\t\t\ttype: \"input_image\",\n\t\t\t\t\t\t\tdetail: \"auto\",\n\t\t\t\t\t\t\timage_url: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\toutput = contentParts;\n\t\t\t} else {\n\t\t\t\toutput = sanitizeSurrogates(hasText ? textResult : hasImages ? \"(see attached image)\" : \"(no tool output)\");\n\t\t\t}\n\n\t\t\tmessages.push({\n\t\t\t\ttype: \"function_call_output\",\n\t\t\t\tcall_id: callId,\n\t\t\t\toutput,\n\t\t\t});\n\n\t\t\tconst deferredTools: Tool[] = [];\n\t\t\tfor (const name of msg.addedToolNames ?? []) {\n\t\t\t\tconst tool = options?.deferredTools?.get(name);\n\t\t\t\tif (!tool || loadedToolNames.has(name)) continue;\n\t\t\t\tloadedToolNames.add(name);\n\t\t\t\tdeferredTools.push(tool);\n\t\t\t}\n\t\t\tif (deferredTools.length > 0) {\n\t\t\t\tconst names = deferredTools.map((tool) => tool.name);\n\t\t\t\tconst searchCallId = `pi_tool_load_${shortHash(`${msg.toolCallId}:${names.join(\",\")}`)}`;\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"tool_search_call\",\n\t\t\t\t\tcall_id: searchCallId,\n\t\t\t\t\texecution: \"client\",\n\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\targuments: { query: names.join(\" \"), limit: names.length },\n\t\t\t\t} satisfies ResponseInputItem);\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"tool_search_output\",\n\t\t\t\t\tcall_id: searchCallId,\n\t\t\t\t\texecution: \"client\",\n\t\t\t\t\tstatus: \"completed\",\n\t\t\t\t\ttools: convertResponsesTools(deferredTools, { deferLoading: true }),\n\t\t\t\t} satisfies ResponseToolSearchOutputItemParam);\n\t\t\t}\n\t\t}\n\t\tmsgIndex++;\n\t}\n\n\treturn messages;\n}\n\n// =============================================================================\n// Tool conversion\n// =============================================================================\n\nexport function convertResponsesTools(tools: Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[] {\n\tconst strict = options?.strict === undefined ? false : options.strict;\n\treturn tools.map((tool) => ({\n\t\ttype: \"function\",\n\t\tname: tool.name,\n\t\tdescription: tool.description,\n\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\tstrict,\n\t\t...(options?.deferLoading ? { defer_loading: true } : {}),\n\t}));\n}\n\n// =============================================================================\n// Stream processing\n// =============================================================================\n\ntype StreamingToolCall = ToolCall & { partialJson: string };\n\n// Responses can interleave events for several output items, so each item keeps its own slot keyed by\n// output_index instead of a single \"current\" block.\ntype ResponsesOutputSlot =\n\t| { type: \"thinking\"; block: ThinkingContent; contentIndex: number }\n\t| { type: \"text\"; block: TextContent; contentIndex: number }\n\t| { type: \"toolCall\"; block: StreamingToolCall; contentIndex: number };\n\nexport async function processResponsesStream<TApi extends Api>(\n\topenaiStream: AsyncIterable<ResponseStreamEvent>,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<TApi>,\n\toptions?: OpenAIResponsesStreamOptions,\n): Promise<void> {\n\tlet terminalEventSeen = false;\n\tconst reasoningBlocksById = new Map<string, ThinkingContent>();\n\tconst outputSlots = new Map<number, ResponsesOutputSlot>();\n\tconst getSlot = <TType extends ResponsesOutputSlot[\"type\"]>(\n\t\toutputIndex: number,\n\t\ttype: TType,\n\t): Extract<ResponsesOutputSlot, { type: TType }> | undefined => {\n\t\tconst slot = outputSlots.get(outputIndex);\n\t\treturn slot?.type === type ? (slot as Extract<ResponsesOutputSlot, { type: TType }>) : undefined;\n\t};\n\tconst createSlot = (outputIndex: number, item: ResponseOutputItem): ResponsesOutputSlot | undefined => {\n\t\tif (item.type === \"reasoning\") {\n\t\t\tconst block: ThinkingContent = { type: \"thinking\", thinking: \"\" };\n\t\t\toutput.content.push(block);\n\t\t\tconst slot = {\n\t\t\t\ttype: \"thinking\",\n\t\t\t\tblock,\n\t\t\t\tcontentIndex: output.content.length - 1,\n\t\t\t} satisfies ResponsesOutputSlot;\n\t\t\toutputSlots.set(outputIndex, slot);\n\t\t\tstream.push({ type: \"thinking_start\", contentIndex: slot.contentIndex, partial: output });\n\t\t\treturn slot;\n\t\t}\n\t\tif (item.type === \"message\") {\n\t\t\tconst block: TextContent = { type: \"text\", text: \"\" };\n\t\t\toutput.content.push(block);\n\t\t\tconst slot = { type: \"text\", block, contentIndex: output.content.length - 1 } satisfies ResponsesOutputSlot;\n\t\t\toutputSlots.set(outputIndex, slot);\n\t\t\tstream.push({ type: \"text_start\", contentIndex: slot.contentIndex, partial: output });\n\t\t\treturn slot;\n\t\t}\n\t\tif (item.type === \"function_call\") {\n\t\t\tconst block: StreamingToolCall = {\n\t\t\t\ttype: \"toolCall\",\n\t\t\t\tid: `${item.call_id}|${item.id}`,\n\t\t\t\tname: item.name,\n\t\t\t\targuments: {},\n\t\t\t\t...(item.namespace !== undefined ? { namespace: item.namespace } : {}),\n\t\t\t\tpartialJson: item.arguments || \"\",\n\t\t\t};\n\t\t\toutput.content.push(block);\n\t\t\tconst slot = {\n\t\t\t\ttype: \"toolCall\",\n\t\t\t\tblock,\n\t\t\t\tcontentIndex: output.content.length - 1,\n\t\t\t} satisfies ResponsesOutputSlot;\n\t\t\toutputSlots.set(outputIndex, slot);\n\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: slot.contentIndex, partial: output });\n\t\t\treturn slot;\n\t\t}\n\t\treturn undefined;\n\t};\n\tconst getOrCreateSlot = (outputIndex: number, item: ResponseOutputItem): ResponsesOutputSlot | undefined =>\n\t\toutputSlots.get(outputIndex) ?? createSlot(outputIndex, item);\n\tconst backfillReasoningSignatures = (responseOutput: ResponseReasoningItem[]): void => {\n\t\tfor (const item of responseOutput) {\n\t\t\tif (!item.encrypted_content) continue;\n\t\t\tconst block = reasoningBlocksById.get(item.id);\n\t\t\tif (!block?.thinkingSignature) continue;\n\n\t\t\tconst storedItem = JSON.parse(block.thinkingSignature) as ResponseReasoningItem;\n\t\t\tif (storedItem.encrypted_content) continue;\n\t\t\tblock.thinkingSignature = JSON.stringify({ ...storedItem, encrypted_content: item.encrypted_content });\n\t\t}\n\t};\n\n\tlet unfinishedToolCall: ToolCall | undefined;\n\tconst finalizeOutstandingToolCalls = (): void => {\n\t\tfor (const block of output.content) {\n\t\t\tif (block.type !== \"toolCall\") continue;\n\t\t\tconst call = block as ToolCall & { partialJson?: string };\n\t\t\tif (call.partialJson === undefined) continue;\n\t\t\tunfinishedToolCall ??= block;\n\t\t\tfinalizeToolCallArguments(call, call.partialJson);\n\t\t\tdelete call.partialJson;\n\t\t}\n\t};\n\n\tfor await (const event of openaiStream) {\n\t\tif (event.type === \"response.created\") {\n\t\t\toutput.responseId = event.response.id;\n\t\t} else if (event.type === \"response.output_item.added\") {\n\t\t\tcreateSlot(event.output_index, event.item);\n\t\t} else if (event.type === \"response.reasoning_summary_text.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"thinking\");\n\t\t\tif (!slot) continue;\n\t\t\tslot.block.thinking += event.delta;\n\t\t\tstream.push({ type: \"thinking_delta\", contentIndex: slot.contentIndex, delta: event.delta, partial: output });\n\t\t} else if (event.type === \"response.reasoning_summary_part.done\") {\n\t\t\tconst slot = getSlot(event.output_index, \"thinking\");\n\t\t\tif (!slot) continue;\n\t\t\tslot.block.thinking += \"\\n\\n\";\n\t\t\tstream.push({ type: \"thinking_delta\", contentIndex: slot.contentIndex, delta: \"\\n\\n\", partial: output });\n\t\t} else if (event.type === \"response.reasoning_text.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"thinking\");\n\t\t\tif (!slot) continue;\n\t\t\tslot.block.thinking += event.delta;\n\t\t\tstream.push({ type: \"thinking_delta\", contentIndex: slot.contentIndex, delta: event.delta, partial: output });\n\t\t} else if (event.type === \"response.output_text.delta\" || event.type === \"response.refusal.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"text\");\n\t\t\tif (!slot) continue;\n\t\t\tslot.block.text += event.delta;\n\t\t\tstream.push({ type: \"text_delta\", contentIndex: slot.contentIndex, delta: event.delta, partial: output });\n\t\t} else if (event.type === \"response.function_call_arguments.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot) continue;\n\t\t\tslot.block.partialJson += event.delta;\n\t\t\tslot.block.arguments = parseStreamingJson(slot.block.partialJson);\n\t\t\tstream.push({ type: \"toolcall_delta\", contentIndex: slot.contentIndex, delta: event.delta, partial: output });\n\t\t} else if (event.type === \"response.function_call_arguments.done\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot) continue;\n\t\t\tconst previousPartialJson = slot.block.partialJson;\n\t\t\tslot.block.partialJson = event.arguments;\n\t\t\tfinalizeToolCallArguments(slot.block, slot.block.partialJson);\n\n\t\t\tif (event.arguments.startsWith(previousPartialJson)) {\n\t\t\t\tconst delta = event.arguments.slice(previousPartialJson.length);\n\t\t\t\tif (delta.length > 0) {\n\t\t\t\t\tstream.push({ type: \"toolcall_delta\", contentIndex: slot.contentIndex, delta, partial: output });\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (event.type === \"response.output_item.done\") {\n\t\t\tconst item = event.item;\n\t\t\tconst slot = getOrCreateSlot(event.output_index, item);\n\n\t\t\tif (item.type === \"reasoning\" && slot?.type === \"thinking\") {\n\t\t\t\tconst summaryText = item.summary?.map((s) => s.text).join(\"\\n\\n\") || \"\";\n\t\t\t\tconst contentText = item.content?.map((c) => c.text).join(\"\\n\\n\") || \"\";\n\t\t\t\tslot.block.thinking = summaryText || contentText || slot.block.thinking;\n\t\t\t\tslot.block.thinkingSignature = JSON.stringify(item);\n\t\t\t\treasoningBlocksById.set(item.id, slot.block);\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\t\tcontent: slot.block.thinking,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t\toutputSlots.delete(event.output_index);\n\t\t\t} else if (item.type === \"message\" && slot?.type === \"text\") {\n\t\t\t\tslot.block.text = item.content?.map((c) => (c.type === \"output_text\" ? c.text : c.refusal)).join(\"\") || \"\";\n\t\t\t\tslot.block.textSignature = encodeTextSignatureV1(item.id, item.phase ?? undefined);\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\t\tcontent: slot.block.text,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t\toutputSlots.delete(event.output_index);\n\t\t\t} else if (item.type === \"function_call\" && slot?.type === \"toolCall\") {\n\t\t\t\t// The done item's arguments are authoritative; streamed deltas are only a fallback.\n\t\t\t\tfinalizeToolCallArguments(slot.block, item.arguments ?? slot.block.partialJson);\n\t\t\t\tif (item.namespace !== undefined) slot.block.namespace = item.namespace;\n\t\t\t\t// Finalize in-place and strip the scratch buffer so replay only carries parsed arguments.\n\t\t\t\tdelete (slot.block as { partialJson?: string }).partialJson;\n\t\t\t\tstream.push({\n\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\t\ttoolCall: slot.block,\n\t\t\t\t\tpartial: output,\n\t\t\t\t});\n\t\t\t\toutputSlots.delete(event.output_index);\n\t\t\t}\n\t\t} else if (event.type === \"response.completed\" || event.type === \"response.incomplete\") {\n\t\t\tterminalEventSeen = true;\n\t\t\tfinalizeOutstandingToolCalls();\n\t\t\tconst response = event.response;\n\t\t\tbackfillReasoningSignatures(\n\t\t\t\tresponse.output?.filter((item): item is ResponseReasoningItem => item.type === \"reasoning\") ?? [],\n\t\t\t);\n\t\t\tif (response?.id) {\n\t\t\t\toutput.responseId = response.id;\n\t\t\t}\n\t\t\tif (response?.usage) {\n\t\t\t\tconst cachedTokens = response.usage.input_tokens_details?.cached_tokens || 0;\n\t\t\t\toutput.usage = {\n\t\t\t\t\t// OpenAI includes cached tokens in input_tokens, so subtract to get non-cached input\n\t\t\t\t\tinput: (response.usage.input_tokens || 0) - cachedTokens,\n\t\t\t\t\toutput: response.usage.output_tokens || 0,\n\t\t\t\t\tcacheRead: cachedTokens,\n\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\ttotalTokens: response.usage.total_tokens || 0,\n\t\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t\t};\n\t\t\t}\n\t\t\tcalculateCost(model, output.usage);\n\t\t\tif (options?.applyServiceTierPricing) {\n\t\t\t\tconst serviceTier = options.resolveServiceTier\n\t\t\t\t\t? options.resolveServiceTier(response?.service_tier, options.serviceTier)\n\t\t\t\t\t: (response?.service_tier ?? options.serviceTier);\n\t\t\t\toptions.applyServiceTierPricing(output.usage, serviceTier);\n\t\t\t}\n\t\t\t// Map status to stop reason\n\t\t\tconst status = response?.status;\n\t\t\toutput.rawStopReason = status;\n\t\t\toutput.stopReason = mapStopReason(status);\n\t\t\tif (output.content.some((b) => b.type === \"toolCall\") && output.stopReason === \"stop\") {\n\t\t\t\toutput.stopReason = \"toolUse\";\n\t\t\t}\n\t\t} else if (event.type === \"error\") {\n\t\t\tthrow Object.assign(new Error(`Error Code ${event.code}: ${event.message}` || \"Unknown error\"), {\n\t\t\t\tcode: event.code,\n\t\t\t});\n\t\t} else if (event.type === \"response.failed\") {\n\t\t\tterminalEventSeen = true;\n\t\t\toutput.rawStopReason = event.response?.status;\n\t\t\tconst error = event.response?.error;\n\t\t\tconst details = event.response?.incomplete_details;\n\t\t\tconst msg = error\n\t\t\t\t? `${error.code || \"unknown\"}: ${error.message || \"no message\"}`\n\t\t\t\t: details?.reason\n\t\t\t\t\t? `incomplete: ${details.reason}`\n\t\t\t\t\t: \"Unknown error (no error details in response)\";\n\t\t\tthrow Object.assign(new Error(msg), { code: error?.code });\n\t\t}\n\t}\n\n\tfinalizeOutstandingToolCalls();\n\tif (!terminalEventSeen) {\n\t\tthrow new Error(\"OpenAI Responses stream ended before a terminal response event\");\n\t}\n\t// Preserve the existing fail-closed guard even after stripping unfinished scratch buffers.\n\tif (output.stopReason === \"toolUse\" && unfinishedToolCall) {\n\t\tthrow new Error(\n\t\t\t`OpenAI Responses stream completed with an unfinished tool call: ${unfinishedToolCall.name} (${unfinishedToolCall.id})`,\n\t\t);\n\t}\n}\n\nfunction mapStopReason(status: OpenAI.Responses.ResponseStatus | undefined): StopReason {\n\tif (!status) return \"stop\";\n\tswitch (status) {\n\t\tcase \"completed\":\n\t\t\treturn \"stop\";\n\t\tcase \"incomplete\":\n\t\t\treturn \"length\";\n\t\tcase \"failed\":\n\t\tcase \"cancelled\":\n\t\t\treturn \"error\";\n\t\t// These two are wonky ...\n\t\tcase \"in_progress\":\n\t\tcase \"queued\":\n\t\t\treturn \"stop\";\n\t\tdefault: {\n\t\t\tconst _exhaustive: never = status;\n\t\t\tthrow new Error(`Unhandled stop reason: ${_exhaustive}`);\n\t\t}\n\t}\n}\n"]}

package/dist/providers/openai-responses-shared.js

@@ -1,7 +1,8 @@
 import { calculateCost } from "../models.js";
 import { shortHash } from "../utils/hash.js";
-import { parseStreamingJson } from "../utils/json-parse.js";
+import { finalizeToolCallArguments, parseStreamingJson } from "../utils/json-parse.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
+import { createProviderToolCallIdNormalizer } from "../utils/tool-call-id.js";
 import { transformMessages } from "./transform-messages.js";
 // =============================================================================
 // Utilities
@@ -37,31 +38,8 @@ function parseTextSignature(signature) {
 export function convertResponsesMessages(model, context, allowedToolCallProviders, options) {
     const messages = [];
     const loadedToolNames = new Set();
-    const normalizeIdPart = (part) => {
-        const sanitized = part.replace(/[^a-zA-Z0-9_-]/g, "_");
-        const normalized = sanitized.length > 64 ? sanitized.slice(0, 64) : sanitized;
-        return normalized.replace(/_+$/, "");
-    };
-    const buildForeignResponsesItemId = (itemId) => {
-        const normalized = `fc_${shortHash(itemId)}`;
-        return normalized.length > 64 ? normalized.slice(0, 64) : normalized;
-    };
-    const normalizeToolCallId = (id, _targetModel, source) => {
-        if (!allowedToolCallProviders.has(model.provider))
-            return normalizeIdPart(id);
-        if (!id.includes("|"))
-            return normalizeIdPart(id);
-        const [callId, itemId] = id.split("|");
-        const normalizedCallId = normalizeIdPart(callId);
-        const isForeignToolCall = source.provider !== model.provider || source.api !== model.api;
-        let normalizedItemId = isForeignToolCall ? buildForeignResponsesItemId(itemId) : normalizeIdPart(itemId);
-        // OpenAI Responses API requires item id to start with "fc"
-        if (!normalizedItemId.startsWith("fc_")) {
-            normalizedItemId = normalizeIdPart(`fc_${normalizedItemId}`);
-        }
-        return `${normalizedCallId}|${normalizedItemId}`;
-    };
-    const transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);
+    const normalizeToolCallId = createProviderToolCallIdNormalizer(model, allowedToolCallProviders);
+    const transformedMessages = transformMessages(context.messages, model, (id, _model, source) => normalizeToolCallId(id, source));
     const includeSystemPrompt = options?.includeSystemPrompt ?? true;
     if (includeSystemPrompt && context.systemPrompt) {
         const compat = model.compat;
@@ -105,9 +83,10 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
         else if (msg.role === "assistant") {
             const output = [];
             const assistantMsg = msg;
-            const isDifferentModel = assistantMsg.model !== model.id &&
-                assistantMsg.provider === model.provider &&
-                assistantMsg.api === model.api;
+            const isSameProviderAndApi = assistantMsg.provider === model.provider && assistantMsg.api === model.api;
+            const isSameModel = isSameProviderAndApi && assistantMsg.model === model.id;
+            const isDifferentModel = isSameProviderAndApi && assistantMsg.model !== model.id;
+            let textBlockIndex = 0;
             for (const block of msg.content) {
                 if (block.type === "thinking") {
                     if (block.thinkingSignature) {
@@ -118,10 +97,12 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                 else if (block.type === "text") {
                     const textBlock = block;
                     const parsedSignature = parseTextSignature(textBlock.textSignature);
+                    const fallbackMessageId = textBlockIndex === 0 ? `msg_pi_${msgIndex}` : `msg_pi_${msgIndex}_${textBlockIndex}`;
+                    textBlockIndex++;
                     // OpenAI requires id to be max 64 characters
                     let msgId = parsedSignature?.id;
                     if (!msgId) {
-                        msgId = `msg_${msgIndex}`;
+                        msgId = fallbackMessageId;
                     }
                     else if (msgId.length > 64) {
                         msgId = `msg_${shortHash(msgId)}`;
@@ -145,12 +126,14 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                     if (isDifferentModel && itemId?.startsWith("fc_")) {
                         itemId = undefined;
                     }
+                    const canReplayNamespace = isSameModel || options?.deferredTools?.has(toolCall.name) === true;
                     output.push({
                         type: "function_call",
                         id: itemId,
                         call_id: callId,
                         name: toolCall.name,
                         arguments: JSON.stringify(toolCall.arguments),
+                        ...(canReplayNamespace && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),
                     });
                 }
             }
@@ -239,16 +222,57 @@ export function convertResponsesTools(tools, options) {
         ...(options?.deferLoading ? { defer_loading: true } : {}),
     }));
 }
-// =============================================================================
-// Stream processing
-// =============================================================================
 export async function processResponsesStream(openaiStream, output, stream, model, options) {
-    let currentItem = null;
-    let currentBlock = null;
     let terminalEventSeen = false;
     const reasoningBlocksById = new Map();
-    const blocks = output.content;
-    const blockIndex = () => blocks.length - 1;
+    const outputSlots = new Map();
+    const getSlot = (outputIndex, type) => {
+        const slot = outputSlots.get(outputIndex);
+        return slot?.type === type ? slot : undefined;
+    };
+    const createSlot = (outputIndex, item) => {
+        if (item.type === "reasoning") {
+            const block = { type: "thinking", thinking: "" };
+            output.content.push(block);
+            const slot = {
+                type: "thinking",
+                block,
+                contentIndex: output.content.length - 1,
+            };
+            outputSlots.set(outputIndex, slot);
+            stream.push({ type: "thinking_start", contentIndex: slot.contentIndex, partial: output });
+            return slot;
+        }
+        if (item.type === "message") {
+            const block = { type: "text", text: "" };
+            output.content.push(block);
+            const slot = { type: "text", block, contentIndex: output.content.length - 1 };
+            outputSlots.set(outputIndex, slot);
+            stream.push({ type: "text_start", contentIndex: slot.contentIndex, partial: output });
+            return slot;
+        }
+        if (item.type === "function_call") {
+            const block = {
+                type: "toolCall",
+                id: `${item.call_id}|${item.id}`,
+                name: item.name,
+                arguments: {},
+                ...(item.namespace !== undefined ? { namespace: item.namespace } : {}),
+                partialJson: item.arguments || "",
+            };
+            output.content.push(block);
+            const slot = {
+                type: "toolCall",
+                block,
+                contentIndex: output.content.length - 1,
+            };
+            outputSlots.set(outputIndex, slot);
+            stream.push({ type: "toolcall_start", contentIndex: slot.contentIndex, partial: output });
+            return slot;
+        }
+        return undefined;
+    };
+    const getOrCreateSlot = (outputIndex, item) => outputSlots.get(outputIndex) ?? createSlot(outputIndex, item);
     const backfillReasoningSignatures = (responseOutput) => {
         for (const item of responseOutput) {
             if (!item.encrypted_content)
@@ -262,214 +286,123 @@ export async function processResponsesStream(openaiStream, output, stream, model
             block.thinkingSignature = JSON.stringify({ ...storedItem, encrypted_content: item.encrypted_content });
         }
     };
+    let unfinishedToolCall;
+    const finalizeOutstandingToolCalls = () => {
+        for (const block of output.content) {
+            if (block.type !== "toolCall")
+                continue;
+            const call = block;
+            if (call.partialJson === undefined)
+                continue;
+            unfinishedToolCall ??= block;
+            finalizeToolCallArguments(call, call.partialJson);
+            delete call.partialJson;
+        }
+    };
     for await (const event of openaiStream) {
         if (event.type === "response.created") {
             output.responseId = event.response.id;
         }
         else if (event.type === "response.output_item.added") {
-            const item = event.item;
-            if (item.type === "reasoning") {
-                currentItem = item;
-                currentBlock = { type: "thinking", thinking: "" };
-                output.content.push(currentBlock);
-                stream.push({ type: "thinking_start", contentIndex: blockIndex(), partial: output });
-            }
-            else if (item.type === "message") {
-                currentItem = item;
-                currentBlock = { type: "text", text: "" };
-                output.content.push(currentBlock);
-                stream.push({ type: "text_start", contentIndex: blockIndex(), partial: output });
-            }
-            else if (item.type === "function_call") {
-                currentItem = item;
-                currentBlock = {
-                    type: "toolCall",
-                    id: `${item.call_id}|${item.id}`,
-                    name: item.name,
-                    arguments: {},
-                    partialJson: item.arguments || "",
-                };
-                output.content.push(currentBlock);
-                stream.push({ type: "toolcall_start", contentIndex: blockIndex(), partial: output });
-            }
-        }
-        else if (event.type === "response.reasoning_summary_part.added") {
-            if (currentItem && currentItem.type === "reasoning") {
-                currentItem.summary = currentItem.summary || [];
-                currentItem.summary.push(event.part);
-            }
+            createSlot(event.output_index, event.item);
         }
         else if (event.type === "response.reasoning_summary_text.delta") {
-            if (currentItem?.type === "reasoning" && currentBlock?.type === "thinking") {
-                currentItem.summary = currentItem.summary || [];
-                const lastPart = currentItem.summary[currentItem.summary.length - 1];
-                if (lastPart) {
-                    currentBlock.thinking += event.delta;
-                    lastPart.text += event.delta;
-                    stream.push({
-                        type: "thinking_delta",
-                        contentIndex: blockIndex(),
-                        delta: event.delta,
-                        partial: output,
-                    });
-                }
-            }
+            const slot = getSlot(event.output_index, "thinking");
+            if (!slot)
+                continue;
+            slot.block.thinking += event.delta;
+            stream.push({ type: "thinking_delta", contentIndex: slot.contentIndex, delta: event.delta, partial: output });
         }
         else if (event.type === "response.reasoning_summary_part.done") {
-            if (currentItem?.type === "reasoning" && currentBlock?.type === "thinking") {
-                currentItem.summary = currentItem.summary || [];
-                const lastPart = currentItem.summary[currentItem.summary.length - 1];
-                if (lastPart) {
-                    currentBlock.thinking += "\n\n";
-                    lastPart.text += "\n\n";
-                    stream.push({
-                        type: "thinking_delta",
-                        contentIndex: blockIndex(),
-                        delta: "\n\n",
-                        partial: output,
-                    });
-                }
-            }
+            const slot = getSlot(event.output_index, "thinking");
+            if (!slot)
+                continue;
+            slot.block.thinking += "\n\n";
+            stream.push({ type: "thinking_delta", contentIndex: slot.contentIndex, delta: "\n\n", partial: output });
         }
         else if (event.type === "response.reasoning_text.delta") {
-            if (currentItem?.type === "reasoning" && currentBlock?.type === "thinking") {
-                currentBlock.thinking += event.delta;
-                stream.push({
-                    type: "thinking_delta",
-                    contentIndex: blockIndex(),
-                    delta: event.delta,
-                    partial: output,
-                });
-            }
-        }
-        else if (event.type === "response.content_part.added") {
-            if (currentItem?.type === "message") {
-                currentItem.content = currentItem.content || [];
-                // Filter out ReasoningText, only accept output_text and refusal
-                if (event.part.type === "output_text" || event.part.type === "refusal") {
-                    currentItem.content.push(event.part);
-                }
-            }
-        }
-        else if (event.type === "response.output_text.delta") {
-            if (currentItem?.type === "message" && currentBlock?.type === "text") {
-                if (!currentItem.content || currentItem.content.length === 0) {
-                    continue;
-                }
-                const lastPart = currentItem.content[currentItem.content.length - 1];
-                if (lastPart?.type === "output_text") {
-                    currentBlock.text += event.delta;
-                    lastPart.text += event.delta;
-                    stream.push({
-                        type: "text_delta",
-                        contentIndex: blockIndex(),
-                        delta: event.delta,
-                        partial: output,
-                    });
-                }
-            }
+            const slot = getSlot(event.output_index, "thinking");
+            if (!slot)
+                continue;
+            slot.block.thinking += event.delta;
+            stream.push({ type: "thinking_delta", contentIndex: slot.contentIndex, delta: event.delta, partial: output });
         }
-        else if (event.type === "response.refusal.delta") {
-            if (currentItem?.type === "message" && currentBlock?.type === "text") {
-                if (!currentItem.content || currentItem.content.length === 0) {
-                    continue;
-                }
-                const lastPart = currentItem.content[currentItem.content.length - 1];
-                if (lastPart?.type === "refusal") {
-                    currentBlock.text += event.delta;
-                    lastPart.refusal += event.delta;
-                    stream.push({
-                        type: "text_delta",
-                        contentIndex: blockIndex(),
-                        delta: event.delta,
-                        partial: output,
-                    });
-                }
-            }
+        else if (event.type === "response.output_text.delta" || event.type === "response.refusal.delta") {
+            const slot = getSlot(event.output_index, "text");
+            if (!slot)
+                continue;
+            slot.block.text += event.delta;
+            stream.push({ type: "text_delta", contentIndex: slot.contentIndex, delta: event.delta, partial: output });
         }
         else if (event.type === "response.function_call_arguments.delta") {
-            if (currentItem?.type === "function_call" && currentBlock?.type === "toolCall") {
-                currentBlock.partialJson += event.delta;
-                currentBlock.arguments = parseStreamingJson(currentBlock.partialJson);
-                stream.push({
-                    type: "toolcall_delta",
-                    contentIndex: blockIndex(),
-                    delta: event.delta,
-                    partial: output,
-                });
-            }
+            const slot = getSlot(event.output_index, "toolCall");
+            if (!slot)
+                continue;
+            slot.block.partialJson += event.delta;
+            slot.block.arguments = parseStreamingJson(slot.block.partialJson);
+            stream.push({ type: "toolcall_delta", contentIndex: slot.contentIndex, delta: event.delta, partial: output });
         }
         else if (event.type === "response.function_call_arguments.done") {
-            if (currentItem?.type === "function_call" && currentBlock?.type === "toolCall") {
-                const previousPartialJson = currentBlock.partialJson;
-                currentBlock.partialJson = event.arguments;
-                currentBlock.arguments = parseStreamingJson(currentBlock.partialJson);
-                if (event.arguments.startsWith(previousPartialJson)) {
-                    const delta = event.arguments.slice(previousPartialJson.length);
-                    if (delta.length > 0) {
-                        stream.push({
-                            type: "toolcall_delta",
-                            contentIndex: blockIndex(),
-                            delta,
-                            partial: output,
-                        });
-                    }
+            const slot = getSlot(event.output_index, "toolCall");
+            if (!slot)
+                continue;
+            const previousPartialJson = slot.block.partialJson;
+            slot.block.partialJson = event.arguments;
+            finalizeToolCallArguments(slot.block, slot.block.partialJson);
+            if (event.arguments.startsWith(previousPartialJson)) {
+                const delta = event.arguments.slice(previousPartialJson.length);
+                if (delta.length > 0) {
+                    stream.push({ type: "toolcall_delta", contentIndex: slot.contentIndex, delta, partial: output });
                 }
             }
         }
         else if (event.type === "response.output_item.done") {
             const item = event.item;
-            if (item.type === "reasoning" && currentBlock?.type === "thinking") {
+            const slot = getOrCreateSlot(event.output_index, item);
+            if (item.type === "reasoning" && slot?.type === "thinking") {
                 const summaryText = item.summary?.map((s) => s.text).join("\n\n") || "";
                 const contentText = item.content?.map((c) => c.text).join("\n\n") || "";
-                currentBlock.thinking = summaryText || contentText || currentBlock.thinking;
-                currentBlock.thinkingSignature = JSON.stringify(item);
-                reasoningBlocksById.set(item.id, currentBlock);
+                slot.block.thinking = summaryText || contentText || slot.block.thinking;
+                slot.block.thinkingSignature = JSON.stringify(item);
+                reasoningBlocksById.set(item.id, slot.block);
                 stream.push({
                     type: "thinking_end",
-                    contentIndex: blockIndex(),
-                    content: currentBlock.thinking,
+                    contentIndex: slot.contentIndex,
+                    content: slot.block.thinking,
                     partial: output,
                 });
-                currentBlock = null;
+                outputSlots.delete(event.output_index);
             }
-            else if (item.type === "message" && currentBlock?.type === "text") {
-                currentBlock.text = item.content.map((c) => (c.type === "output_text" ? c.text : c.refusal)).join("");
-                currentBlock.textSignature = encodeTextSignatureV1(item.id, item.phase ?? undefined);
+            else if (item.type === "message" && slot?.type === "text") {
+                slot.block.text = item.content?.map((c) => (c.type === "output_text" ? c.text : c.refusal)).join("") || "";
+                slot.block.textSignature = encodeTextSignatureV1(item.id, item.phase ?? undefined);
                 stream.push({
                     type: "text_end",
-                    contentIndex: blockIndex(),
-                    content: currentBlock.text,
+                    contentIndex: slot.contentIndex,
+                    content: slot.block.text,
                     partial: output,
                 });
-                currentBlock = null;
+                outputSlots.delete(event.output_index);
             }
-            else if (item.type === "function_call") {
-                const args = currentBlock?.type === "toolCall" && currentBlock.partialJson
-                    ? parseStreamingJson(currentBlock.partialJson)
-                    : parseStreamingJson(item.arguments || "{}");
-                let toolCall;
-                if (currentBlock?.type === "toolCall") {
-                    // Finalize in-place and strip the scratch buffer so replay only
-                    // carries parsed arguments.
-                    currentBlock.arguments = args;
-                    delete currentBlock.partialJson;
-                    toolCall = currentBlock;
-                }
-                else {
-                    toolCall = {
-                        type: "toolCall",
-                        id: `${item.call_id}|${item.id}`,
-                        name: item.name,
-                        arguments: args,
-                    };
-                }
-                currentBlock = null;
-                stream.push({ type: "toolcall_end", contentIndex: blockIndex(), toolCall, partial: output });
+            else if (item.type === "function_call" && slot?.type === "toolCall") {
+                // The done item's arguments are authoritative; streamed deltas are only a fallback.
+                finalizeToolCallArguments(slot.block, item.arguments ?? slot.block.partialJson);
+                if (item.namespace !== undefined)
+                    slot.block.namespace = item.namespace;
+                // Finalize in-place and strip the scratch buffer so replay only carries parsed arguments.
+                delete slot.block.partialJson;
+                stream.push({
+                    type: "toolcall_end",
+                    contentIndex: slot.contentIndex,
+                    toolCall: slot.block,
+                    partial: output,
+                });
+                outputSlots.delete(event.output_index);
             }
         }
         else if (event.type === "response.completed" || event.type === "response.incomplete") {
             terminalEventSeen = true;
+            finalizeOutstandingToolCalls();
             const response = event.response;
             backfillReasoningSignatures(response.output?.filter((item) => item.type === "reasoning") ?? []);
             if (response?.id) {
@@ -495,16 +428,21 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 options.applyServiceTierPricing(output.usage, serviceTier);
             }
             // Map status to stop reason
-            output.stopReason = mapStopReason(response?.status);
+            const status = response?.status;
+            output.rawStopReason = status;
+            output.stopReason = mapStopReason(status);
             if (output.content.some((b) => b.type === "toolCall") && output.stopReason === "stop") {
                 output.stopReason = "toolUse";
             }
         }
         else if (event.type === "error") {
-            throw new Error(`Error Code ${event.code}: ${event.message}` || "Unknown error");
+            throw Object.assign(new Error(`Error Code ${event.code}: ${event.message}` || "Unknown error"), {
+                code: event.code,
+            });
         }
         else if (event.type === "response.failed") {
             terminalEventSeen = true;
+            output.rawStopReason = event.response?.status;
             const error = event.response?.error;
             const details = event.response?.incomplete_details;
             const msg = error
@@ -512,12 +450,17 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 : details?.reason
                     ? `incomplete: ${details.reason}`
                     : "Unknown error (no error details in response)";
-            throw new Error(msg);
+            throw Object.assign(new Error(msg), { code: error?.code });
         }
     }
+    finalizeOutstandingToolCalls();
     if (!terminalEventSeen) {
         throw new Error("OpenAI Responses stream ended before a terminal response event");
     }
+    // Preserve the existing fail-closed guard even after stripping unfinished scratch buffers.
+    if (output.stopReason === "toolUse" && unfinishedToolCall) {
+        throw new Error(`OpenAI Responses stream completed with an unfinished tool call: ${unfinishedToolCall.name} (${unfinishedToolCall.id})`);
+    }
 }
 function mapStopReason(status) {
     if (!status)