@bastani/pi-ai 0.9.20-alpha.2 → 0.9.20-alpha.4

package/dist/api/openai-responses-shared.d.ts

@@ -1,5 +1,5 @@
 import type { Tool as OpenAITool, ResponseCreateParamsStreaming, ResponseInput, ResponseStreamEvent } from "openai/resources/responses/responses.js";
-import type { Api, AssistantMessage, Context, Model, Tool, Usage } from "../types.ts";
+import type { Api, AssistantMessage, Model, Tool, TranscriptContext, Usage } from "../types.ts";
 import type { AssistantMessageEventStream } from "../utils/event-stream.ts";
 /**
  * The service tier this request should carry.
@@ -40,17 +40,19 @@ export interface OpenAIResponsesStreamOptions {
 export interface ConvertResponsesMessagesOptions {
     includeSystemPrompt?: boolean;
     grammarToolInputProperties?: ReadonlyMap<string, string>;
-    deferredTools?: ReadonlyMap<string, Tool>;
-    deferredToolsMode?: "additional-tools" | "tool-search";
+    /** Whether later system messages are sent in place; otherwise they are folded into the leading prompt. */
+    supportsMidConvoSystemMessages?: boolean;
+    supportsAdditionalTools?: boolean;
+    supportsToolSearch?: boolean;
     toolOptions?: ConvertResponsesToolsOptions;
 }
 export interface ConvertResponsesToolsOptions {
     strict?: boolean | null;
     supportsStrictMode?: boolean;
     supportsOpenAIGrammarTools?: boolean;
-    deferLoading?: boolean;
+    toolSearchResult?: boolean;
 }
-export declare function convertResponsesMessages<TApi extends Api>(model: Model<TApi>, context: Context, allowedToolCallProviders: ReadonlySet<string>, options?: ConvertResponsesMessagesOptions): ResponseInput;
+export declare function convertResponsesMessages<TApi extends Api>(model: Model<TApi>, context: TranscriptContext, allowedToolCallProviders: ReadonlySet<string>, options?: ConvertResponsesMessagesOptions): ResponseInput;
 export declare function convertResponsesTools(tools: readonly Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[];
 export declare function processResponsesStream<TApi extends Api>(openaiStream: AsyncIterable<ResponseStreamEvent>, output: AssistantMessage, stream: AssistantMessageEventStream, model: Model<TApi>, options?: OpenAIResponsesStreamOptions): Promise<void>;
 //# sourceMappingURL=openai-responses-shared.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"openai-responses-shared.d.ts","sourceRoot":"","sources":["../../src/api/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAClB,6BAA6B,EAC7B,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;AA2E5E;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,2BAA2B,CAC1C,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,WAAW,CAAC,EACpC,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,GAC3E,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAE3D;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,+BAA+B,CAC9C,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,WAAW,GAAG,IAAI,GAAG,UAAU,CAAC,EACxD,OAAO,EAAE,OAAO,GACd,IAAI,CA2BN;AAED,MAAM,WAAW,4BAA4B;IAC5C,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,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,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,aAAa,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;IAC1C,iBAAiB,CAAC,EAAE,kBAAkB,GAAG,aAAa,CAAC;IACvD,WAAW,CAAC,EAAE,4BAA4B,CAAC;CAC3C;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,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,CAkNf;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAqClH;AAoCD,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,CAkUf","sourcesContent":["import type OpenAI from \"openai\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\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.ts\";\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.ts\";\nimport type { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport {\n\tappendGrammarToolInputJsonDelta,\n\ttype GrammarToolInputJsonBuffer,\n\tgetGrammarToolInput,\n\tgetJsonSchemaToolParameters,\n\tresolveGrammarConstrainedSampling,\n\tresolveJsonSchemaStrictSampling,\n} from \"./constrained-sampling.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\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\ntype ToolResultOutputContent = Array<ResponseInputText | ResponseInputImage>;\n\nfunction convertToolResultOutput<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontent: readonly (TextContent | ImageContent)[],\n): string | ToolResultOutputContent {\n\tconst textResult = content\n\t\t.filter((c): c is TextContent => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\tconst images = content.filter((c): c is ImageContent => c.type === \"image\");\n\tconst hasText = textResult.length > 0;\n\n\tif (images.length === 0 || !model.input.includes(\"image\")) {\n\t\treturn sanitizeSurrogates(hasText ? textResult : images.length > 0 ? \"(see attached image)\" : \"(no tool output)\");\n\t}\n\n\tconst output: ToolResultOutputContent = [];\n\tif (hasText) {\n\t\toutput.push({ type: \"input_text\", text: sanitizeSurrogates(textResult) });\n\t}\n\tfor (const image of images) {\n\t\toutput.push({\n\t\t\ttype: \"input_image\",\n\t\t\tdetail: \"auto\",\n\t\t\timage_url: `data:${image.mimeType};base64,${image.data}`,\n\t\t});\n\t}\n\treturn output;\n}\n\n/**\n * The service tier this request should carry.\n *\n * The **presence** of `fastRoute` is the discriminator, and its declared tier is the value. Fast\n * versus normal is model identity, so a caller who wants a different tier selects the normal sibling\n * rather than downgrading the fast one; letting a per-request option win would let a model selected,\n * recorded, persisted, and billed as `-fast` route as an ordinary request — and, because the Codex\n * routing identity keys on the final payload's tier, silently drop `originator: codex_cli_rs` too.\n *\n * A route that declares **no** tier means *no tier*, not \"fall through to the caller\". A GitHub\n * Copilot fast variant is exactly that case: it sends its own advertised `-fast` model ID and must\n * carry no OpenAI service-tier field at all, whatever the caller asks for.\n *\n * A normal model has no route, so an explicit `serviceTier` still applies there exactly as it did\n * before fast variants existed.\n */\nexport function resolveRequestedServiceTier(\n\tmodel: Pick<Model<Api>, \"fastRoute\">,\n\toptionsServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): ResponseCreateParamsStreaming[\"service_tier\"] | undefined {\n\treturn model.fastRoute ? model.fastRoute.serviceTier : optionsServiceTier;\n}\n\n/**\n * Reject a payload hook that rewrote a field the model's route owns.\n *\n * A fast variant is a distinct selected, recorded, persisted, and billed identity, and the objective\n * requires it to send its route's upstream model ID and tier. A hook that changes either would send a\n * different model — or an ordinary-tier request — under the `-fast` identity Atomic recorded, and the\n * Codex routing identity would faithfully advertise whatever the hook left behind.\n *\n * Following the repository's precedent for discarded caller instructions, this rejects loudly rather\n * than silently restoring the route's values, so a misconfigured hook surfaces instead of hiding. A\n * hook may still rewrite every other field, and a model without a route keeps unrestricted freedom.\n */\nexport function assertPayloadPreservesFastRoute(\n\tmodel: Pick<Model<Api>, \"fastRoute\" | \"id\" | \"provider\">,\n\tpayload: unknown,\n): void {\n\tconst fastRoute = model.fastRoute;\n\tif (!fastRoute) return;\n\tif (typeof payload !== \"object\" || payload === null || Array.isArray(payload)) {\n\t\tthrow new Error(\n\t\t\t`A request hook changed payload for fast model \"${model.provider}/${model.id}\" to a value that cannot carry ` +\n\t\t\t\t`the route's upstream model and service tier. A fast model variant must send its route's upstream model and ` +\n\t\t\t\t`service tier, because it is recorded, persisted, and billed under its own identity. Select the normal model ` +\n\t\t\t\t`\"${model.provider}/${fastRoute.baseModelId}\" if the request needs different routing, or return an object that ` +\n\t\t\t\t`preserves those fields.`,\n\t\t);\n\t}\n\tconst record = payload as Record<string, unknown>;\n\tconst conflicts: string[] = [];\n\tif (record.model !== fastRoute.upstreamModelId) {\n\t\tconflicts.push(`model (expected \"${fastRoute.upstreamModelId}\", got ${JSON.stringify(record.model)})`);\n\t}\n\tif (record.service_tier !== fastRoute.serviceTier) {\n\t\tconflicts.push(`service_tier (expected ${JSON.stringify(fastRoute.serviceTier)}, got ${JSON.stringify(record.service_tier)})`);\n\t}\n\tif (conflicts.length === 0) return;\n\tthrow new Error(\n\t\t`A request hook changed ${conflicts.join(\" and \")} for fast model \"${model.provider}/${model.id}\". ` +\n\t\t\t`A fast model variant must send its route's upstream model and service tier, because it is ` +\n\t\t\t`recorded, persisted, and billed under its own identity. Select the normal model \"${model.provider}/${fastRoute.baseModelId}\" ` +\n\t\t\t`if the request needs different routing, or stop rewriting those fields.`,\n\t);\n}\n\nexport interface OpenAIResponsesStreamOptions {\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\n\tresolveServiceTier?: (\n\t\tresponseServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t\trequestServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => ResponseCreateParamsStreaming[\"service_tier\"] | undefined;\n\tapplyServiceTierPricing?: (\n\t\tusage: Usage,\n\t\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\t) => void;\n}\n\nexport interface ConvertResponsesMessagesOptions {\n\tincludeSystemPrompt?: boolean;\n\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\n\tdeferredTools?: ReadonlyMap<string, Tool>;\n\tdeferredToolsMode?: \"additional-tools\" | \"tool-search\";\n\ttoolOptions?: ConvertResponsesToolsOptions;\n}\n\nexport interface ConvertResponsesToolsOptions {\n\tstrict?: boolean | null;\n\tsupportsStrictMode?: boolean;\n\tsupportsOpenAIGrammarTools?: boolean;\n\tdeferLoading?: boolean;\n}\n\n// =============================================================================\n// Message conversion\n// =============================================================================\n\nexport function convertResponsesMessages<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\tallowedToolCallProviders: ReadonlySet<string>,\n\toptions?: ConvertResponsesMessagesOptions,\n): ResponseInput {\n\tconst messages: ResponseInput = [];\n\tconst loadedToolNames = new Set<string>();\n\n\tconst normalizeIdPart = (part: string): string => {\n\t\tconst sanitized = part.replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\tconst normalized = sanitized.length > 64 ? sanitized.slice(0, 64) : sanitized;\n\t\treturn normalized.replace(/_+$/, \"\");\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 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\tconst customInputProperty = options?.grammarToolInputProperties?.get(toolCall.name);\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\t// When replaying custom-tool calls as a function_call, also drop non-fc_* ids such as\n\t\t\t\t\t// ctc_* custom-tool ids because function_call item ids must be fc_*.\n\t\t\t\t\tif (\n\t\t\t\t\t\t(isDifferentModel && itemId?.startsWith(\"fc_\")) ||\n\t\t\t\t\t\t(customInputProperty === undefined && !itemId?.startsWith(\"fc_\"))\n\t\t\t\t\t) {\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\tif (customInputProperty !== undefined) {\n\t\t\t\t\t\toutput.push({\n\t\t\t\t\t\t\ttype: \"custom_tool_call\",\n\t\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\t\tinput: sanitizeSurrogates(\n\t\t\t\t\t\t\t\tgetGrammarToolInput(toolCall.name, toolCall.arguments, customInputProperty),\n\t\t\t\t\t\t\t),\n\t\t\t\t\t\t\t...(canReplayNamespace && toolCall.namespace !== undefined\n\t\t\t\t\t\t\t\t? { namespace: toolCall.namespace }\n\t\t\t\t\t\t\t\t: {}),\n\t\t\t\t\t\t} satisfies ResponseOutputItem);\n\t\t\t\t\t} else {\n\t\t\t\t\t\toutput.push({\n\t\t\t\t\t\t\ttype: \"function_call\",\n\t\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\t\targuments: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t\t...(canReplayNamespace && toolCall.namespace !== undefined\n\t\t\t\t\t\t\t\t? { namespace: toolCall.namespace }\n\t\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\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 [callId] = msg.toolCallId.split(\"|\");\n\t\t\tconst output = convertToolResultOutput(model, msg.content);\n\n\t\t\tif (options?.grammarToolInputProperties?.has(msg.toolName)) {\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"custom_tool_call_output\",\n\t\t\t\t\tcall_id: callId,\n\t\t\t\t\toutput,\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"function_call_output\",\n\t\t\t\t\tcall_id: callId,\n\t\t\t\t\toutput,\n\t\t\t\t});\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 && options?.deferredToolsMode === \"additional-tools\") {\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"additional_tools\",\n\t\t\t\t\trole: \"developer\",\n\t\t\t\t\ttools: convertResponsesTools(deferredTools, options.toolOptions),\n\t\t\t\t} satisfies ResponseInputItem);\n\t\t\t} else if (deferredTools.length > 0 && options?.deferredToolsMode === \"tool-search\") {\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, {\n\t\t\t\t\t\t...options.toolOptions,\n\t\t\t\t\t\tdeferLoading: true,\n\t\t\t\t\t}),\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: readonly Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[] {\n\tconst defaultStrict = options?.strict === undefined ? false : options.strict;\n\tconst supportsStrictMode = options?.supportsStrictMode ?? true;\n\tconst supportsOpenAIGrammarTools = options?.supportsOpenAIGrammarTools ?? false;\n\n\treturn tools.map((tool) => {\n\t\tconst grammar = resolveGrammarConstrainedSampling(tool, supportsOpenAIGrammarTools);\n\t\tif (grammar) {\n\t\t\treturn {\n\t\t\t\ttype: \"custom\",\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tformat: {\n\t\t\t\t\ttype: \"grammar\",\n\t\t\t\t\tsyntax: grammar.format,\n\t\t\t\t\tdefinition: grammar.definition,\n\t\t\t\t},\n\t\t\t\t...(options?.deferLoading ? { defer_loading: true } : {}),\n\t\t\t} satisfies OpenAITool;\n\t\t}\n\n\t\tconst constrainedStrict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);\n\t\tconst strict = constrainedStrict ?? defaultStrict;\n\t\tconst functionTool: Omit<Extract<OpenAITool, { type: \"function\" }>, \"strict\"> & {\n\t\t\tstrict?: Extract<OpenAITool, { type: \"function\" }>[\"strict\"];\n\t\t} = {\n\t\t\ttype: \"function\",\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: getJsonSchemaToolParameters(tool, strict === true) as Record<string, unknown>,\n\t\t\t...(options?.deferLoading ? { defer_loading: true } : {}),\n\t\t};\n\t\tif (supportsStrictMode) {\n\t\t\tfunctionTool.strict = strict;\n\t\t}\n\t\treturn functionTool as OpenAITool;\n\t});\n}\n\n// =============================================================================\n// Stream processing\n// =============================================================================\n\ntype StreamingToolCall = ToolCall & {\n\tpartialJson?: string;\n\tcustomInput?: {\n\t\tproperty: string;\n\t\tjsonBuffer: GrammarToolInputJsonBuffer;\n\t};\n};\n\nfunction getCustomToolCallInput(block: StreamingToolCall): string {\n\tconst property = block.customInput?.property;\n\tif (property === undefined) return \"\";\n\tconst value = block.arguments[property];\n\treturn typeof value === \"string\" ? value : \"\";\n}\n\nfunction appendCustomToolCallInput(block: StreamingToolCall, nextInput: string, close: boolean): string | undefined {\n\tconst customInput = block.customInput;\n\tif (!customInput) return undefined;\n\tconst delta = appendGrammarToolInputJsonDelta(customInput.jsonBuffer, customInput.property, nextInput, close);\n\tblock.arguments = { [customInput.property]: nextInput };\n\treturn delta;\n}\n\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\ntype ToolCallOutputSlot = Extract<ResponsesOutputSlot, { type: \"toolCall\" }>;\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 sawTerminalResponseEvent = false;\n\tconst outputSlots = new Map<number, ResponsesOutputSlot>();\n\tconst reasoningBlocksById = new Map<string, ThinkingContent>();\n\tconst applyMessagePhaseStopReason = (item: ResponseOutputItem): void => {\n\t\tif (item.type === \"message\" && item.phase === \"final_answer\") {\n\t\t\toutput.stopReason = \"stop\";\n\t\t}\n\t};\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 pushToolCallDelta = (slot: ToolCallOutputSlot, delta: string | undefined): void => {\n\t\tif (delta === undefined) return;\n\t\tstream.push({\n\t\t\ttype: \"toolcall_delta\",\n\t\t\tcontentIndex: slot.contentIndex,\n\t\t\tdelta,\n\t\t\tpartial: output,\n\t\t});\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\tapplyMessagePhaseStopReason(item);\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}|${item.id}`,\n\t\t\t\tname: item.name,\n\t\t\t\targuments: {},\n\t\t\t\t...(item.namespace !== undefined ? { namespace: item.namespace } : {}),\n\t\t\t\tpartialJson: item.arguments || \"\",\n\t\t\t};\n\t\t\toutput.content.push(block);\n\t\t\tconst slot = {\n\t\t\t\ttype: \"toolCall\",\n\t\t\t\tblock,\n\t\t\t\tcontentIndex: output.content.length - 1,\n\t\t\t} satisfies ResponsesOutputSlot;\n\t\t\toutputSlots.set(outputIndex, slot);\n\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: slot.contentIndex, partial: output });\n\t\t\treturn slot;\n\t\t}\n\t\tif (item.type === \"custom_tool_call\") {\n\t\t\tconst inputProperty = options?.grammarToolInputProperties?.get(item.name) ?? \"input\";\n\t\t\tconst input = item.input || \"\";\n\t\t\tconst block: StreamingToolCall = {\n\t\t\t\ttype: \"toolCall\",\n\t\t\t\tid: `${item.call_id}|${item.id}`,\n\t\t\t\tname: item.name,\n\t\t\t\targuments: { [inputProperty]: input },\n\t\t\t\t...(item.namespace !== undefined ? { namespace: item.namespace } : {}),\n\t\t\t\tcustomInput: {\n\t\t\t\t\tproperty: inputProperty,\n\t\t\t\t\tjsonBuffer: { input: \"\", started: false, closed: false },\n\t\t\t\t},\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\treturn outputSlots.get(outputIndex) ?? createSlot(outputIndex, item);\n\t};\n\t// Azure OpenAI can omit reasoning.encrypted_content from response.output_item.done\n\t// and provide it only in response.completed.response.output. Backfill the\n\t// persisted reasoning signature from the terminal response to keep store:false\n\t// multi-turn replay stateless. See https://github.com/earendil-works/pi/issues/6409.\n\tconst backfillReasoningSignatures = (responseOutput: ResponseOutputItem[]): void => {\n\t\tfor (const item of responseOutput) {\n\t\t\tif (item.type !== \"reasoning\" || !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({\n\t\t\t\t...storedItem,\n\t\t\t\tencrypted_content: item.encrypted_content,\n\t\t\t});\n\t\t}\n\t};\n\tconst finalizeResponse = (\n\t\tresponse: Extract<ResponseStreamEvent, { type: \"response.completed\" | \"response.incomplete\" }>[\"response\"],\n\t): void => {\n\t\tsawTerminalResponseEvent = true;\n\t\tbackfillReasoningSignatures(response.output ?? []);\n\t\tif (response?.id) {\n\t\t\toutput.responseId = response.id;\n\t\t}\n\t\tif (response?.usage) {\n\t\t\tconst inputDetails = response.usage.input_tokens_details as\n\t\t\t\t| { cached_tokens?: number; cache_write_tokens?: number }\n\t\t\t\t| undefined;\n\t\t\tconst cachedTokens = inputDetails?.cached_tokens || 0;\n\t\t\tconst cacheWriteTokens = inputDetails?.cache_write_tokens || 0;\n\t\t\toutput.usage = {\n\t\t\t\t// OpenAI includes cached and cache-write tokens in input_tokens, so subtract both.\n\t\t\t\tinput: Math.max(0, (response.usage.input_tokens || 0) - cachedTokens - cacheWriteTokens),\n\t\t\t\toutput: response.usage.output_tokens || 0,\n\t\t\t\tcacheRead: cachedTokens,\n\t\t\t\tcacheWrite: cacheWriteTokens,\n\t\t\t\treasoning: response.usage.output_tokens_details?.reasoning_tokens || 0,\n\t\t\t\ttotalTokens: response.usage.total_tokens || 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t}\n\t\tcalculateCost(model, output.usage);\n\t\tif (options?.applyServiceTierPricing) {\n\t\t\tconst serviceTier = options.resolveServiceTier\n\t\t\t\t? options.resolveServiceTier(response?.service_tier, options.serviceTier)\n\t\t\t\t: (response?.service_tier ?? options.serviceTier);\n\t\t\toptions.applyServiceTierPricing(output.usage, serviceTier);\n\t\t}\n\t\t// Map status to stop reason. For incomplete responses, retain the provider's\n\t\t// specific reason so max-output truncation and content filtering stay distinct.\n\t\tconst status = response?.status;\n\t\tconst incompleteDetails = response?.incomplete_details as { reason?: unknown } | null | undefined;\n\t\tconst incompleteReason = typeof incompleteDetails?.reason === \"string\" ? incompleteDetails.reason : undefined;\n\t\toutput.rawStopReason = incompleteReason ? `${status}.${incompleteReason}` : status;\n\t\tconst mappedStop = mapStopReason(status, incompleteReason);\n\t\toutput.stopReason = mappedStop.stopReason;\n\t\toutput.errorMessage = mappedStop.errorMessage;\n\t\tif (output.content.some((b) => b.type === \"toolCall\") && output.stopReason === \"stop\") {\n\t\t\toutput.stopReason = \"toolUse\";\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: \"\\n\\n\",\n\t\t\t\tpartial: output,\n\t\t\t});\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\n\t\t} else if (event.type === \"response.output_text.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({\n\t\t\t\ttype: \"text_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\n\t\t} else if (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({\n\t\t\t\ttype: \"text_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\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 || slot.block.partialJson === undefined) continue;\n\t\t\tslot.block.partialJson += event.delta;\n\t\t\tslot.block.arguments = parseStreamingJson(slot.block.partialJson);\n\t\t\tpushToolCallDelta(slot, event.delta);\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 || slot.block.partialJson === undefined) continue;\n\t\t\tconst previousPartialJson = slot.block.partialJson;\n\t\t\tslot.block.partialJson = event.arguments;\n\t\t\tslot.block.arguments = parseStreamingJson(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) pushToolCallDelta(slot, delta);\n\t\t\t}\n\t\t} else if (event.type === \"response.custom_tool_call_input.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot || !slot.block.customInput) continue;\n\t\t\tpushToolCallDelta(\n\t\t\t\tslot,\n\t\t\t\tappendCustomToolCallInput(slot.block, getCustomToolCallInput(slot.block) + event.delta, false),\n\t\t\t);\n\t\t} else if (event.type === \"response.custom_tool_call_input.done\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot || !slot.block.customInput) continue;\n\t\t\tpushToolCallDelta(slot, appendCustomToolCallInput(slot.block, event.input, true));\n\t\t} else if (event.type === \"response.output_item.done\") {\n\t\t\tconst item = event.item;\n\t\t\tapplyMessagePhaseStopReason(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 (\n\t\t\t\titem.type === \"function_call\" &&\n\t\t\t\tslot?.type === \"toolCall\" &&\n\t\t\t\tslot.block.partialJson !== undefined\n\t\t\t) {\n\t\t\t\tslot.block.arguments = parseStreamingJson(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\n\t\t\t\t// carries parsed arguments.\n\t\t\t\tdelete slot.block.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} else if (item.type === \"custom_tool_call\" && slot?.type === \"toolCall\" && slot.block.customInput) {\n\t\t\t\tpushToolCallDelta(\n\t\t\t\t\tslot,\n\t\t\t\t\tappendCustomToolCallInput(slot.block, item.input ?? getCustomToolCallInput(slot.block), true),\n\t\t\t\t);\n\t\t\t\tif (item.namespace !== undefined) slot.block.namespace = item.namespace;\n\t\t\t\tdelete slot.block.customInput;\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\tfinalizeResponse(event.response);\n\t\t} else if (event.type === \"error\") {\n\t\t\tthrow new Error(`Error Code ${event.code}: ${event.message}` || \"Unknown error\");\n\t\t} else if (event.type === \"response.failed\") {\n\t\t\tsawTerminalResponseEvent = true;\n\t\t\toutput.rawStopReason = event.response?.status;\n\t\t\tconst error = event.response?.error;\n\t\t\tconst details = event.response?.incomplete_details;\n\t\t\tconst msg = error\n\t\t\t\t? `${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 new Error(msg);\n\t\t}\n\t}\n\tif (!sawTerminalResponseEvent) {\n\t\tthrow new Error(\"OpenAI Responses stream ended before a terminal response event\");\n\t}\n}\n\nfunction mapStopReason(\n\tstatus: OpenAI.Responses.ResponseStatus | undefined,\n\tincompleteReason?: string,\n): { stopReason: StopReason; errorMessage?: string } {\n\tif (!status) return { stopReason: \"stop\" };\n\tswitch (status) {\n\t\tcase \"completed\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"incomplete\":\n\t\t\tif (incompleteReason === \"max_output_tokens\") {\n\t\t\t\treturn { stopReason: \"length\" };\n\t\t\t}\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: incompleteReason\n\t\t\t\t\t? `Response incomplete: ${incompleteReason}`\n\t\t\t\t\t: \"Response incomplete without a provider reason\",\n\t\t\t};\n\t\tcase \"failed\":\n\t\tcase \"cancelled\":\n\t\t\treturn { stopReason: \"error\" };\n\t\t// These two are wonky ...\n\t\tcase \"in_progress\":\n\t\tcase \"queued\":\n\t\t\treturn { stopReason: \"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/api/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAClB,6BAA6B,EAC7B,aAAa,EAQb,mBAAmB,EAEnB,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAEhB,KAAK,EAML,IAAI,EAEJ,iBAAiB,EACjB,KAAK,EACL,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AA6E5E;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,2BAA2B,CAC1C,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,WAAW,CAAC,EACpC,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,GAC3E,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAE3D;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,+BAA+B,CAC9C,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,WAAW,GAAG,IAAI,GAAG,UAAU,CAAC,EACxD,OAAO,EAAE,OAAO,GACd,IAAI,CA6BN;AAED,MAAM,WAAW,4BAA4B;IAC5C,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,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,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,0GAA0G;IAC1G,8BAA8B,CAAC,EAAE,OAAO,CAAC;IACzC,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,WAAW,CAAC,EAAE,4BAA4B,CAAC;CAC3C;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC3B;AAMD,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EACxD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,iBAAiB,EAC1B,wBAAwB,EAAE,WAAW,CAAC,MAAM,CAAC,EAC7C,OAAO,CAAC,EAAE,+BAA+B,GACvC,aAAa,CA6Mf;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAqClH;AAoCD,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,CAkUf","sourcesContent":["import type OpenAI from \"openai\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\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.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tImageContent,\n\tModel,\n\tStopReason,\n\tSystemMessage,\n\tTextContent,\n\tTextSignatureV1,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tTranscriptContext,\n\tUsage,\n} from \"../types.ts\";\nimport type { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { getSystemMessageText, renderSystemMessageUpdate } from \"../utils/text.ts\";\nimport { resolveTranscript, resolveTranscriptTools } from \"../utils/transcript.ts\";\nimport {\n\tappendGrammarToolInputJsonDelta,\n\ttype GrammarToolInputJsonBuffer,\n\tgetGrammarToolInput,\n\tgetJsonSchemaToolParameters,\n\tresolveGrammarConstrainedSampling,\n\tresolveJsonSchemaStrictSampling,\n} from \"./constrained-sampling.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\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\ntype ToolResultOutputContent = Array<ResponseInputText | ResponseInputImage>;\n\nfunction convertToolResultOutput<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontent: readonly (TextContent | ImageContent)[],\n): string | ToolResultOutputContent {\n\tconst textResult = content\n\t\t.filter((c): c is TextContent => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\tconst images = content.filter((c): c is ImageContent => c.type === \"image\");\n\tconst hasText = textResult.length > 0;\n\n\tif (images.length === 0 || !model.input.includes(\"image\")) {\n\t\treturn sanitizeSurrogates(hasText ? textResult : images.length > 0 ? \"(see attached image)\" : \"(no tool output)\");\n\t}\n\n\tconst output: ToolResultOutputContent = [];\n\tif (hasText) {\n\t\toutput.push({ type: \"input_text\", text: sanitizeSurrogates(textResult) });\n\t}\n\tfor (const image of images) {\n\t\toutput.push({\n\t\t\ttype: \"input_image\",\n\t\t\tdetail: \"auto\",\n\t\t\timage_url: `data:${image.mimeType};base64,${image.data}`,\n\t\t});\n\t}\n\treturn output;\n}\n\n/**\n * The service tier this request should carry.\n *\n * The **presence** of `fastRoute` is the discriminator, and its declared tier is the value. Fast\n * versus normal is model identity, so a caller who wants a different tier selects the normal sibling\n * rather than downgrading the fast one; letting a per-request option win would let a model selected,\n * recorded, persisted, and billed as `-fast` route as an ordinary request — and, because the Codex\n * routing identity keys on the final payload's tier, silently drop `originator: codex_cli_rs` too.\n *\n * A route that declares **no** tier means *no tier*, not \"fall through to the caller\". A GitHub\n * Copilot fast variant is exactly that case: it sends its own advertised `-fast` model ID and must\n * carry no OpenAI service-tier field at all, whatever the caller asks for.\n *\n * A normal model has no route, so an explicit `serviceTier` still applies there exactly as it did\n * before fast variants existed.\n */\nexport function resolveRequestedServiceTier(\n\tmodel: Pick<Model<Api>, \"fastRoute\">,\n\toptionsServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): ResponseCreateParamsStreaming[\"service_tier\"] | undefined {\n\treturn model.fastRoute ? model.fastRoute.serviceTier : optionsServiceTier;\n}\n\n/**\n * Reject a payload hook that rewrote a field the model's route owns.\n *\n * A fast variant is a distinct selected, recorded, persisted, and billed identity, and the objective\n * requires it to send its route's upstream model ID and tier. A hook that changes either would send a\n * different model — or an ordinary-tier request — under the `-fast` identity Atomic recorded, and the\n * Codex routing identity would faithfully advertise whatever the hook left behind.\n *\n * Following the repository's precedent for discarded caller instructions, this rejects loudly rather\n * than silently restoring the route's values, so a misconfigured hook surfaces instead of hiding. A\n * hook may still rewrite every other field, and a model without a route keeps unrestricted freedom.\n */\nexport function assertPayloadPreservesFastRoute(\n\tmodel: Pick<Model<Api>, \"fastRoute\" | \"id\" | \"provider\">,\n\tpayload: unknown,\n): void {\n\tconst fastRoute = model.fastRoute;\n\tif (!fastRoute) return;\n\tif (typeof payload !== \"object\" || payload === null || Array.isArray(payload)) {\n\t\tthrow new Error(\n\t\t\t`A request hook changed payload for fast model \"${model.provider}/${model.id}\" to a value that cannot carry ` +\n\t\t\t\t`the route's upstream model and service tier. A fast model variant must send its route's upstream model and ` +\n\t\t\t\t`service tier, because it is recorded, persisted, and billed under its own identity. Select the normal model ` +\n\t\t\t\t`\"${model.provider}/${fastRoute.baseModelId}\" if the request needs different routing, or return an object that ` +\n\t\t\t\t`preserves those fields.`,\n\t\t);\n\t}\n\tconst record = payload as Record<string, unknown>;\n\tconst conflicts: string[] = [];\n\tif (record.model !== fastRoute.upstreamModelId) {\n\t\tconflicts.push(`model (expected \"${fastRoute.upstreamModelId}\", got ${JSON.stringify(record.model)})`);\n\t}\n\tif (record.service_tier !== fastRoute.serviceTier) {\n\t\tconflicts.push(\n\t\t\t`service_tier (expected ${JSON.stringify(fastRoute.serviceTier)}, got ${JSON.stringify(record.service_tier)})`,\n\t\t);\n\t}\n\tif (conflicts.length === 0) return;\n\tthrow new Error(\n\t\t`A request hook changed ${conflicts.join(\" and \")} for fast model \"${model.provider}/${model.id}\". ` +\n\t\t\t`A fast model variant must send its route's upstream model and service tier, because it is ` +\n\t\t\t`recorded, persisted, and billed under its own identity. Select the normal model \"${model.provider}/${fastRoute.baseModelId}\" ` +\n\t\t\t`if the request needs different routing, or stop rewriting those fields.`,\n\t);\n}\n\nexport interface OpenAIResponsesStreamOptions {\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\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\tgrammarToolInputProperties?: ReadonlyMap<string, string>;\n\t/** Whether later system messages are sent in place; otherwise they are folded into the leading prompt. */\n\tsupportsMidConvoSystemMessages?: boolean;\n\tsupportsAdditionalTools?: boolean;\n\tsupportsToolSearch?: boolean;\n\ttoolOptions?: ConvertResponsesToolsOptions;\n}\n\nexport interface ConvertResponsesToolsOptions {\n\tstrict?: boolean | null;\n\tsupportsStrictMode?: boolean;\n\tsupportsOpenAIGrammarTools?: boolean;\n\ttoolSearchResult?: boolean;\n}\n\n// =============================================================================\n// Message conversion\n// =============================================================================\n\nexport function convertResponsesMessages<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: TranscriptContext,\n\tallowedToolCallProviders: ReadonlySet<string>,\n\toptions?: ConvertResponsesMessagesOptions,\n): ResponseInput {\n\tconst normalizedContext = resolveTranscript(context, options?.supportsMidConvoSystemMessages);\n\tconst messages: ResponseInput = [];\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(normalizedContext.messages, model, normalizeToolCallId);\n\tconst transcriptTools = resolveTranscriptTools(\n\t\tnormalizedContext.messages,\n\t\t(options?.supportsAdditionalTools ?? false) || (options?.supportsToolSearch ?? false),\n\t);\n\tconst appendSystemToolAdditions = (message: SystemMessage, seed: string): void => {\n\t\tconst tools = transcriptTools.anchorsAdditions ? (message.toolsAdded ?? []) : [];\n\t\tif (tools.length === 0) return;\n\t\tif (options?.supportsAdditionalTools) {\n\t\t\tmessages.push({\n\t\t\t\ttype: \"additional_tools\",\n\t\t\t\trole: \"developer\",\n\t\t\t\ttools: convertResponsesTools(tools, options.toolOptions),\n\t\t\t} satisfies ResponseInputItem);\n\t\t\treturn;\n\t\t}\n\t\tif (!options?.supportsToolSearch) return;\n\t\tconst names = tools.map((tool) => tool.name);\n\t\tconst callId = `pi_tool_load_${shortHash(`${seed}:${names.join(\",\")}`)}`;\n\t\tmessages.push({\n\t\t\ttype: \"tool_search_call\",\n\t\t\tcall_id: callId,\n\t\t\texecution: \"client\",\n\t\t\tstatus: \"completed\",\n\t\t\targuments: { query: names.join(\" \"), limit: names.length },\n\t\t} satisfies ResponseInputItem);\n\t\tmessages.push({\n\t\t\ttype: \"tool_search_output\",\n\t\t\tcall_id: callId,\n\t\t\texecution: \"client\",\n\t\t\tstatus: \"completed\",\n\t\t\ttools: convertResponsesTools(tools, { ...options.toolOptions, toolSearchResult: true }),\n\t\t} satisfies ResponseToolSearchOutputItemParam);\n\t};\n\tconst includeInitialSystemMessage = options?.includeSystemPrompt ?? true;\n\tconst compat = model.compat as { supportsDeveloperRole?: boolean } | undefined;\n\tconst instructionRole = model.reasoning && compat?.supportsDeveloperRole !== false ? \"developer\" : \"system\";\n\n\tlet msgIndex = 0;\n\tlet sourceIndex = 0;\n\tfor (const msg of transformedMessages) {\n\t\tconst isLeadingSystemMessage = sourceIndex++ === 0 && msg.role === \"system\";\n\t\tif (msg.role === \"system\") {\n\t\t\tif (!isLeadingSystemMessage) appendSystemToolAdditions(msg, `system:${msgIndex}`);\n\t\t\tif (!isLeadingSystemMessage || includeInitialSystemMessage) {\n\t\t\t\tconst text = isLeadingSystemMessage ? getSystemMessageText(msg) : renderSystemMessageUpdate(msg);\n\t\t\t\tif (text.length > 0) {\n\t\t\t\t\tmessages.push({ role: instructionRole, content: sanitizeSurrogates(text) });\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (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\tconst customInputProperty = options?.grammarToolInputProperties?.get(toolCall.name);\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\t// When replaying custom-tool calls as a function_call, also drop non-fc_* ids such as\n\t\t\t\t\t// ctc_* custom-tool ids because function_call item ids must be fc_*.\n\t\t\t\t\tif (\n\t\t\t\t\t\t(isDifferentModel && itemId?.startsWith(\"fc_\")) ||\n\t\t\t\t\t\t(customInputProperty === undefined && !itemId?.startsWith(\"fc_\"))\n\t\t\t\t\t) {\n\t\t\t\t\t\titemId = undefined;\n\t\t\t\t\t}\n\n\t\t\t\t\tif (customInputProperty !== undefined) {\n\t\t\t\t\t\toutput.push({\n\t\t\t\t\t\t\ttype: \"custom_tool_call\",\n\t\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\t\tinput: sanitizeSurrogates(\n\t\t\t\t\t\t\t\tgetGrammarToolInput(toolCall.name, toolCall.arguments, customInputProperty),\n\t\t\t\t\t\t\t),\n\t\t\t\t\t\t\t...(isSameModel && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),\n\t\t\t\t\t\t} satisfies ResponseOutputItem);\n\t\t\t\t\t} else {\n\t\t\t\t\t\toutput.push({\n\t\t\t\t\t\t\ttype: \"function_call\",\n\t\t\t\t\t\t\tid: itemId,\n\t\t\t\t\t\t\tcall_id: callId,\n\t\t\t\t\t\t\tname: toolCall.name,\n\t\t\t\t\t\t\targuments: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t\t...(isSameModel && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),\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\tif (output.length === 0) continue;\n\t\t\tmessages.push(...output);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst [callId] = msg.toolCallId.split(\"|\");\n\t\t\tconst output = convertToolResultOutput(model, msg.content);\n\n\t\t\tif (options?.grammarToolInputProperties?.has(msg.toolName)) {\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"custom_tool_call_output\",\n\t\t\t\t\tcall_id: callId,\n\t\t\t\t\toutput,\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tmessages.push({\n\t\t\t\t\ttype: \"function_call_output\",\n\t\t\t\t\tcall_id: callId,\n\t\t\t\t\toutput,\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t\tif (!isLeadingSystemMessage) msgIndex++;\n\t}\n\n\treturn messages;\n}\n\n// =============================================================================\n// Tool conversion\n// =============================================================================\n\nexport function convertResponsesTools(tools: readonly Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[] {\n\tconst defaultStrict = options?.strict === undefined ? false : options.strict;\n\tconst supportsStrictMode = options?.supportsStrictMode ?? true;\n\tconst supportsOpenAIGrammarTools = options?.supportsOpenAIGrammarTools ?? false;\n\n\treturn tools.map((tool) => {\n\t\tconst grammar = resolveGrammarConstrainedSampling(tool, supportsOpenAIGrammarTools);\n\t\tif (grammar) {\n\t\t\treturn {\n\t\t\t\ttype: \"custom\",\n\t\t\t\tname: tool.name,\n\t\t\t\tdescription: tool.description,\n\t\t\t\tformat: {\n\t\t\t\t\ttype: \"grammar\",\n\t\t\t\t\tsyntax: grammar.format,\n\t\t\t\t\tdefinition: grammar.definition,\n\t\t\t\t},\n\t\t\t\t...(options?.toolSearchResult ? { defer_loading: true } : {}),\n\t\t\t} satisfies OpenAITool;\n\t\t}\n\n\t\tconst constrainedStrict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);\n\t\tconst strict = constrainedStrict ?? defaultStrict;\n\t\tconst functionTool: Omit<Extract<OpenAITool, { type: \"function\" }>, \"strict\"> & {\n\t\t\tstrict?: Extract<OpenAITool, { type: \"function\" }>[\"strict\"];\n\t\t} = {\n\t\t\ttype: \"function\",\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: getJsonSchemaToolParameters(tool, strict === true) as Record<string, unknown>,\n\t\t\t...(options?.toolSearchResult ? { defer_loading: true } : {}),\n\t\t};\n\t\tif (supportsStrictMode) {\n\t\t\tfunctionTool.strict = strict;\n\t\t}\n\t\treturn functionTool as OpenAITool;\n\t});\n}\n\n// =============================================================================\n// Stream processing\n// =============================================================================\n\ntype StreamingToolCall = ToolCall & {\n\tpartialJson?: string;\n\tcustomInput?: {\n\t\tproperty: string;\n\t\tjsonBuffer: GrammarToolInputJsonBuffer;\n\t};\n};\n\nfunction getCustomToolCallInput(block: StreamingToolCall): string {\n\tconst property = block.customInput?.property;\n\tif (property === undefined) return \"\";\n\tconst value = block.arguments[property];\n\treturn typeof value === \"string\" ? value : \"\";\n}\n\nfunction appendCustomToolCallInput(block: StreamingToolCall, nextInput: string, close: boolean): string | undefined {\n\tconst customInput = block.customInput;\n\tif (!customInput) return undefined;\n\tconst delta = appendGrammarToolInputJsonDelta(customInput.jsonBuffer, customInput.property, nextInput, close);\n\tblock.arguments = { [customInput.property]: nextInput };\n\treturn delta;\n}\n\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\ntype ToolCallOutputSlot = Extract<ResponsesOutputSlot, { type: \"toolCall\" }>;\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 sawTerminalResponseEvent = false;\n\tconst outputSlots = new Map<number, ResponsesOutputSlot>();\n\tconst reasoningBlocksById = new Map<string, ThinkingContent>();\n\tconst applyMessagePhaseStopReason = (item: ResponseOutputItem): void => {\n\t\tif (item.type === \"message\" && item.phase === \"final_answer\") {\n\t\t\toutput.stopReason = \"stop\";\n\t\t}\n\t};\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 pushToolCallDelta = (slot: ToolCallOutputSlot, delta: string | undefined): void => {\n\t\tif (delta === undefined) return;\n\t\tstream.push({\n\t\t\ttype: \"toolcall_delta\",\n\t\t\tcontentIndex: slot.contentIndex,\n\t\t\tdelta,\n\t\t\tpartial: output,\n\t\t});\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\tapplyMessagePhaseStopReason(item);\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\tif (item.type === \"custom_tool_call\") {\n\t\t\tconst inputProperty = options?.grammarToolInputProperties?.get(item.name) ?? \"input\";\n\t\t\tconst input = item.input || \"\";\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: { [inputProperty]: input },\n\t\t\t\t...(item.namespace !== undefined ? { namespace: item.namespace } : {}),\n\t\t\t\tcustomInput: {\n\t\t\t\t\tproperty: inputProperty,\n\t\t\t\t\tjsonBuffer: { input: \"\", started: false, closed: false },\n\t\t\t\t},\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\treturn outputSlots.get(outputIndex) ?? createSlot(outputIndex, item);\n\t};\n\t// Azure OpenAI can omit reasoning.encrypted_content from response.output_item.done\n\t// and provide it only in response.completed.response.output. Backfill the\n\t// persisted reasoning signature from the terminal response to keep store:false\n\t// multi-turn replay stateless. See https://github.com/earendil-works/pi/issues/6409.\n\tconst backfillReasoningSignatures = (responseOutput: ResponseOutputItem[]): void => {\n\t\tfor (const item of responseOutput) {\n\t\t\tif (item.type !== \"reasoning\" || !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({\n\t\t\t\t...storedItem,\n\t\t\t\tencrypted_content: item.encrypted_content,\n\t\t\t});\n\t\t}\n\t};\n\tconst finalizeResponse = (\n\t\tresponse: Extract<ResponseStreamEvent, { type: \"response.completed\" | \"response.incomplete\" }>[\"response\"],\n\t): void => {\n\t\tsawTerminalResponseEvent = true;\n\t\tbackfillReasoningSignatures(response.output ?? []);\n\t\tif (response?.id) {\n\t\t\toutput.responseId = response.id;\n\t\t}\n\t\tif (response?.usage) {\n\t\t\tconst inputDetails = response.usage.input_tokens_details as\n\t\t\t\t| { cached_tokens?: number; cache_write_tokens?: number }\n\t\t\t\t| undefined;\n\t\t\tconst cachedTokens = inputDetails?.cached_tokens || 0;\n\t\t\tconst cacheWriteTokens = inputDetails?.cache_write_tokens || 0;\n\t\t\toutput.usage = {\n\t\t\t\t// OpenAI includes cached and cache-write tokens in input_tokens, so subtract both.\n\t\t\t\tinput: Math.max(0, (response.usage.input_tokens || 0) - cachedTokens - cacheWriteTokens),\n\t\t\t\toutput: response.usage.output_tokens || 0,\n\t\t\t\tcacheRead: cachedTokens,\n\t\t\t\tcacheWrite: cacheWriteTokens,\n\t\t\t\treasoning: response.usage.output_tokens_details?.reasoning_tokens || 0,\n\t\t\t\ttotalTokens: response.usage.total_tokens || 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t}\n\t\tcalculateCost(model, output.usage);\n\t\tif (options?.applyServiceTierPricing) {\n\t\t\tconst serviceTier = options.resolveServiceTier\n\t\t\t\t? options.resolveServiceTier(response?.service_tier, options.serviceTier)\n\t\t\t\t: (response?.service_tier ?? options.serviceTier);\n\t\t\toptions.applyServiceTierPricing(output.usage, serviceTier);\n\t\t}\n\t\t// Map status to stop reason. For incomplete responses, retain the provider's\n\t\t// specific reason so max-output truncation and content filtering stay distinct.\n\t\tconst status = response?.status;\n\t\tconst incompleteDetails = response?.incomplete_details as { reason?: unknown } | null | undefined;\n\t\tconst incompleteReason = typeof incompleteDetails?.reason === \"string\" ? incompleteDetails.reason : undefined;\n\t\toutput.rawStopReason = incompleteReason ? `${status}.${incompleteReason}` : status;\n\t\tconst mappedStop = mapStopReason(status, incompleteReason);\n\t\toutput.stopReason = mappedStop.stopReason;\n\t\toutput.errorMessage = mappedStop.errorMessage;\n\t\tif (output.content.some((b) => b.type === \"toolCall\") && output.stopReason === \"stop\") {\n\t\t\toutput.stopReason = \"toolUse\";\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: \"\\n\\n\",\n\t\t\t\tpartial: output,\n\t\t\t});\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({\n\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\n\t\t} else if (event.type === \"response.output_text.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({\n\t\t\t\ttype: \"text_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\n\t\t} else if (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({\n\t\t\t\ttype: \"text_delta\",\n\t\t\t\tcontentIndex: slot.contentIndex,\n\t\t\t\tdelta: event.delta,\n\t\t\t\tpartial: output,\n\t\t\t});\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 || slot.block.partialJson === undefined) continue;\n\t\t\tslot.block.partialJson += event.delta;\n\t\t\tslot.block.arguments = parseStreamingJson(slot.block.partialJson);\n\t\t\tpushToolCallDelta(slot, event.delta);\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 || slot.block.partialJson === undefined) continue;\n\t\t\tconst previousPartialJson = slot.block.partialJson;\n\t\t\tslot.block.partialJson = event.arguments;\n\t\t\tslot.block.arguments = parseStreamingJson(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) pushToolCallDelta(slot, delta);\n\t\t\t}\n\t\t} else if (event.type === \"response.custom_tool_call_input.delta\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot || !slot.block.customInput) continue;\n\t\t\tpushToolCallDelta(\n\t\t\t\tslot,\n\t\t\t\tappendCustomToolCallInput(slot.block, getCustomToolCallInput(slot.block) + event.delta, false),\n\t\t\t);\n\t\t} else if (event.type === \"response.custom_tool_call_input.done\") {\n\t\t\tconst slot = getSlot(event.output_index, \"toolCall\");\n\t\t\tif (!slot || !slot.block.customInput) continue;\n\t\t\tpushToolCallDelta(slot, appendCustomToolCallInput(slot.block, event.input, true));\n\t\t} else if (event.type === \"response.output_item.done\") {\n\t\t\tconst item = event.item;\n\t\t\tapplyMessagePhaseStopReason(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 (\n\t\t\t\titem.type === \"function_call\" &&\n\t\t\t\tslot?.type === \"toolCall\" &&\n\t\t\t\tslot.block.partialJson !== undefined\n\t\t\t) {\n\t\t\t\tslot.block.arguments = parseStreamingJson(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\n\t\t\t\t// carries parsed arguments.\n\t\t\t\tdelete slot.block.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} else if (item.type === \"custom_tool_call\" && slot?.type === \"toolCall\" && slot.block.customInput) {\n\t\t\t\tpushToolCallDelta(\n\t\t\t\t\tslot,\n\t\t\t\t\tappendCustomToolCallInput(slot.block, item.input ?? getCustomToolCallInput(slot.block), true),\n\t\t\t\t);\n\t\t\t\tif (item.namespace !== undefined) slot.block.namespace = item.namespace;\n\t\t\t\tdelete slot.block.customInput;\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\tfinalizeResponse(event.response);\n\t\t} else if (event.type === \"error\") {\n\t\t\tthrow new Error(`Error Code ${event.code}: ${event.message}` || \"Unknown error\");\n\t\t} else if (event.type === \"response.failed\") {\n\t\t\tsawTerminalResponseEvent = true;\n\t\t\toutput.rawStopReason = event.response?.status;\n\t\t\tconst error = event.response?.error;\n\t\t\tconst details = event.response?.incomplete_details;\n\t\t\tconst msg = error\n\t\t\t\t? `${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 new Error(msg);\n\t\t}\n\t}\n\tif (!sawTerminalResponseEvent) {\n\t\tthrow new Error(\"OpenAI Responses stream ended before a terminal response event\");\n\t}\n}\n\nfunction mapStopReason(\n\tstatus: OpenAI.Responses.ResponseStatus | undefined,\n\tincompleteReason?: string,\n): { stopReason: StopReason; errorMessage?: string } {\n\tif (!status) return { stopReason: \"stop\" };\n\tswitch (status) {\n\t\tcase \"completed\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"incomplete\":\n\t\t\tif (incompleteReason === \"max_output_tokens\") {\n\t\t\t\treturn { stopReason: \"length\" };\n\t\t\t}\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: incompleteReason\n\t\t\t\t\t? `Response incomplete: ${incompleteReason}`\n\t\t\t\t\t: \"Response incomplete without a provider reason\",\n\t\t\t};\n\t\tcase \"failed\":\n\t\tcase \"cancelled\":\n\t\t\treturn { stopReason: \"error\" };\n\t\t// These two are wonky ...\n\t\tcase \"in_progress\":\n\t\tcase \"queued\":\n\t\t\treturn { stopReason: \"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/api/openai-responses-shared.js

@@ -2,6 +2,8 @@ import { calculateCost } from "../models.js";
 import { shortHash } from "../utils/hash.js";
 import { parseStreamingJson } from "../utils/json-parse.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
+import { getSystemMessageText, renderSystemMessageUpdate } from "../utils/text.js";
+import { resolveTranscript, resolveTranscriptTools } from "../utils/transcript.js";
 import { appendGrammarToolInputJsonDelta, getGrammarToolInput, getJsonSchemaToolParameters, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { transformMessages } from "./transform-messages.js";
 // =============================================================================
@@ -116,8 +118,8 @@ export function assertPayloadPreservesFastRoute(model, payload) {
 // Message conversion
 // =============================================================================
 export function convertResponsesMessages(model, context, allowedToolCallProviders, options) {
+    const normalizedContext = resolveTranscript(context, options?.supportsMidConvoSystemMessages);
     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;
@@ -142,19 +144,57 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
         }
         return `${normalizedCallId}|${normalizedItemId}`;
     };
-    const transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);
-    const includeSystemPrompt = options?.includeSystemPrompt ?? true;
-    if (includeSystemPrompt && context.systemPrompt) {
-        const compat = model.compat;
-        const role = model.reasoning && compat?.supportsDeveloperRole !== false ? "developer" : "system";
+    const transformedMessages = transformMessages(normalizedContext.messages, model, normalizeToolCallId);
+    const transcriptTools = resolveTranscriptTools(normalizedContext.messages, (options?.supportsAdditionalTools ?? false) || (options?.supportsToolSearch ?? false));
+    const appendSystemToolAdditions = (message, seed) => {
+        const tools = transcriptTools.anchorsAdditions ? (message.toolsAdded ?? []) : [];
+        if (tools.length === 0)
+            return;
+        if (options?.supportsAdditionalTools) {
+            messages.push({
+                type: "additional_tools",
+                role: "developer",
+                tools: convertResponsesTools(tools, options.toolOptions),
+            });
+            return;
+        }
+        if (!options?.supportsToolSearch)
+            return;
+        const names = tools.map((tool) => tool.name);
+        const callId = `pi_tool_load_${shortHash(`${seed}:${names.join(",")}`)}`;
         messages.push({
-            role,
-            content: sanitizeSurrogates(context.systemPrompt),
+            type: "tool_search_call",
+            call_id: callId,
+            execution: "client",
+            status: "completed",
+            arguments: { query: names.join(" "), limit: names.length },
         });
-    }
+        messages.push({
+            type: "tool_search_output",
+            call_id: callId,
+            execution: "client",
+            status: "completed",
+            tools: convertResponsesTools(tools, { ...options.toolOptions, toolSearchResult: true }),
+        });
+    };
+    const includeInitialSystemMessage = options?.includeSystemPrompt ?? true;
+    const compat = model.compat;
+    const instructionRole = model.reasoning && compat?.supportsDeveloperRole !== false ? "developer" : "system";
     let msgIndex = 0;
+    let sourceIndex = 0;
     for (const msg of transformedMessages) {
-        if (msg.role === "user") {
+        const isLeadingSystemMessage = sourceIndex++ === 0 && msg.role === "system";
+        if (msg.role === "system") {
+            if (!isLeadingSystemMessage)
+                appendSystemToolAdditions(msg, `system:${msgIndex}`);
+            if (!isLeadingSystemMessage || includeInitialSystemMessage) {
+                const text = isLeadingSystemMessage ? getSystemMessageText(msg) : renderSystemMessageUpdate(msg);
+                if (text.length > 0) {
+                    messages.push({ role: instructionRole, content: sanitizeSurrogates(text) });
+                }
+            }
+        }
+        else if (msg.role === "user") {
             if (typeof msg.content === "string") {
                 messages.push({
                     role: "user",
@@ -233,7 +273,6 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                         (customInputProperty === undefined && !itemId?.startsWith("fc_"))) {
                         itemId = undefined;
                     }
-                    const canReplayNamespace = isSameModel || options?.deferredTools?.has(toolCall.name) === true;
                     if (customInputProperty !== undefined) {
                         output.push({
                             type: "custom_tool_call",
@@ -241,9 +280,7 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                             call_id: callId,
                             name: toolCall.name,
                             input: sanitizeSurrogates(getGrammarToolInput(toolCall.name, toolCall.arguments, customInputProperty)),
-                            ...(canReplayNamespace && toolCall.namespace !== undefined
-                                ? { namespace: toolCall.namespace }
-                                : {}),
+                            ...(isSameModel && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),
                         });
                     }
                     else {
@@ -253,9 +290,7 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                             call_id: callId,
                             name: toolCall.name,
                             arguments: JSON.stringify(toolCall.arguments),
-                            ...(canReplayNamespace && toolCall.namespace !== undefined
-                                ? { namespace: toolCall.namespace }
-                                : {}),
+                            ...(isSameModel && toolCall.namespace !== undefined ? { namespace: toolCall.namespace } : {}),
                         });
                     }
                 }
@@ -281,44 +316,9 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
                     output,
                 });
             }
-            const deferredTools = [];
-            for (const name of msg.addedToolNames ?? []) {
-                const tool = options?.deferredTools?.get(name);
-                if (!tool || loadedToolNames.has(name))
-                    continue;
-                loadedToolNames.add(name);
-                deferredTools.push(tool);
-            }
-            if (deferredTools.length > 0 && options?.deferredToolsMode === "additional-tools") {
-                messages.push({
-                    type: "additional_tools",
-                    role: "developer",
-                    tools: convertResponsesTools(deferredTools, options.toolOptions),
-                });
-            }
-            else if (deferredTools.length > 0 && options?.deferredToolsMode === "tool-search") {
-                const names = deferredTools.map((tool) => tool.name);
-                const searchCallId = `pi_tool_load_${shortHash(`${msg.toolCallId}:${names.join(",")}`)}`;
-                messages.push({
-                    type: "tool_search_call",
-                    call_id: searchCallId,
-                    execution: "client",
-                    status: "completed",
-                    arguments: { query: names.join(" "), limit: names.length },
-                });
-                messages.push({
-                    type: "tool_search_output",
-                    call_id: searchCallId,
-                    execution: "client",
-                    status: "completed",
-                    tools: convertResponsesTools(deferredTools, {
-                        ...options.toolOptions,
-                        deferLoading: true,
-                    }),
-                });
-            }
         }
-        msgIndex++;
+        if (!isLeadingSystemMessage)
+            msgIndex++;
     }
     return messages;
 }
@@ -341,7 +341,7 @@ export function convertResponsesTools(tools, options) {
                     syntax: grammar.format,
                     definition: grammar.definition,
                 },
-                ...(options?.deferLoading ? { defer_loading: true } : {}),
+                ...(options?.toolSearchResult ? { defer_loading: true } : {}),
             };
         }
         const constrainedStrict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);
@@ -351,7 +351,7 @@ export function convertResponsesTools(tools, options) {
             name: tool.name,
             description: tool.description,
             parameters: getJsonSchemaToolParameters(tool, strict === true),
-            ...(options?.deferLoading ? { defer_loading: true } : {}),
+            ...(options?.toolSearchResult ? { defer_loading: true } : {}),
         };
         if (supportsStrictMode) {
             functionTool.strict = strict;