@mupt-ai/dari-gateway 0.2.0 → 0.3.0

package/dist/providers/pi_context.js

@@ -0,0 +1,405 @@
+import { encodeOpenAIReasoningSignature } from "@mupt-ai/dari-router/protocols";
+import { isRecord } from "@mupt-ai/dari-router/policy-engine";
+import { HttpError } from "../http/errors.js";
+import { parseDataImageUrl } from "../interaction/image.js";
+import { providerToolCallId, toolWireName } from "../interaction/tool_identity.js";
+import { compatibleProviderContinuation, sameProviderIdentity, } from "../interaction/provider_continuation_state.js";
+import { isOpenAIResponsesApi } from "./model_capabilities.js";
+const UNKNOWN_SOURCE = {
+    provider: "dari",
+    api: "interaction-history",
+    model: "unknown",
+};
+export function buildContext(request, target) {
+    const systemParts = request.instructions ? [request.instructions] : [];
+    const messages = [];
+    const toolCalls = new Map();
+    const timestamp = Date.now();
+    let pending = null;
+    let sawConversationItem = false;
+    const flushAssistant = () => {
+        if (!pending)
+            return;
+        if (pending.content.length === 0)
+            pending.content.push({ type: "text", text: "" });
+        pending.stopReason = pending.content.some((part) => part.type === "toolCall") ? "toolUse" : "stop";
+        messages.push(pending);
+        pending = null;
+    };
+    const assistantFor = (source) => {
+        if (pending && !sameProviderIdentity(pending, source))
+            flushAssistant();
+        pending ??= {
+            role: "assistant",
+            content: [],
+            ...source,
+            usage: zeroUsage(),
+            stopReason: "stop",
+            timestamp,
+        };
+        return pending;
+    };
+    const assistantForReasoning = (source) => {
+        // Codex can persist a streamed tool call before reasoning recovered from
+        // the terminal response. Both items still belong to one assistant turn.
+        if (pending && sameProviderIdentity(pending, UNKNOWN_SOURCE)) {
+            pending.provider = source.provider;
+            pending.api = source.api;
+            pending.model = source.model;
+        }
+        return assistantFor(source);
+    };
+    for (const [itemIndex, item] of request.items.entries()) {
+        if (item.type === "message" && (item.role === "system" || item.role === "developer")) {
+            flushAssistant();
+            const text = readableContent(item.content);
+            if (!text)
+                continue;
+            if (!sawConversationItem) {
+                // Leading preamble is identical on every request, so folding it into
+                // the system prompt keeps the provider payload prefix stable.
+                systemParts.push(text);
+                continue;
+            }
+            // Mid-conversation system/developer notes arrive between turns (Claude
+            // Code appends a `<total_tokens>` reminder after every tool result).
+            // Folding them into the system prompt mutates the front of the provider
+            // payload, so prefix-matching prompt caches miss and re-write the whole
+            // conversation at cache-write prices. Keep them positional instead.
+            appendPositionalNote(messages, text, timestamp);
+            continue;
+        }
+        sawConversationItem = true;
+        if (item.type === "message" && item.role === "user") {
+            flushAssistant();
+            appendUserMessage(messages, userContent(item.content, item.contentShape), timestamp);
+            continue;
+        }
+        if (item.type === "tool_result") {
+            flushAssistant();
+            const call = toolCalls.get(item.callId);
+            messages.push({
+                role: "toolResult",
+                toolCallId: call?.id ?? item.callId,
+                toolName: call?.name ?? item.name ?? "tool",
+                content: toolResultContent(item.output, target),
+                isError: item.isError ?? false,
+                timestamp,
+                ...(item.loadedTools?.length ? { addedToolNames: [...item.loadedTools] } : {}),
+            });
+            continue;
+        }
+        if (item.type === "reasoning_configuration") {
+            if (!isOpenAIResponsesApi(target.api)) {
+                throw new HttpError(400, "Reasoning configuration updates require a GPT-6 Responses model.", "unsupported_reasoning_configuration", "input");
+            }
+            const assistant = assistantFor(target);
+            assistant.content.push({
+                type: "providerItem",
+                item: {
+                    type: "configuration_update",
+                    reasoning: { effort: item.effort === "off" ? "none" : item.effort },
+                },
+                name: "providerItem",
+                arguments: {},
+            });
+            continue;
+        }
+        if (item.type === "reasoning") {
+            const continuation = compatibleProviderContinuation(item.continuation, target);
+            if (continuation) {
+                const assistant = assistantForReasoning(continuation.source);
+                const firstNonReasoning = assistant.content.findIndex((content) => content.type !== "thinking");
+                assistant.content.splice(firstNonReasoning === -1 ? assistant.content.length : firstNonReasoning, 0, thinkingContent(item, continuation));
+                // Managed-effort Claude models bind each thinking block to the prefix
+                // it was generated under, which ended in an effort marker. Pi only
+                // re-inserts that marker before turns that carry their level; without
+                // it the API silently drops the replayed block.
+                if (continuation.kind !== "openai_reasoning" && continuation.effort !== undefined) {
+                    assistant.providerThinkingLevel = continuation.effort;
+                }
+            }
+            else {
+                const readable = [...item.summary, ...item.content].join("\n");
+                if (readable) {
+                    const assistant = assistantForReasoning(item.source ?? item.continuation?.source ?? UNKNOWN_SOURCE);
+                    const firstToolCall = assistant.content.findIndex((content) => content.type === "toolCall");
+                    assistant.content.splice(firstToolCall === -1 ? assistant.content.length : firstToolCall, 0, {
+                        type: "text",
+                        text: portableReasoningText(request.items, itemIndex, readable, target),
+                    });
+                }
+            }
+            continue;
+        }
+        if (item.type === "tool_call") {
+            const assistant = assistantFor(pending ? identityOf(pending) : UNKNOWN_SOURCE);
+            const id = providerToolCallId(item.callId, item.id, target.api);
+            const name = toolWireName(item.namespace, item.name);
+            assistant.content.push({
+                type: "toolCall",
+                id,
+                name,
+                arguments: toolArguments(item.kind, item.input),
+            });
+            toolCalls.set(item.callId, { id, name });
+            continue;
+        }
+        if (item.type === "hosted_tool_call") {
+            if (item.tool === "server_tool_use" || item.tool === "tool_search") {
+                // Server tool use and tool search results are Anthropic-specific.
+                // Handle them before the generic OpenAI Responses branch to avoid
+                // emitting Anthropic-shaped payloads as Responses provider items.
+                if (target.api === "anthropic-messages") {
+                    // Round-trip preservation when routing to Anthropic.
+                    if (item.tool === "server_tool_use") {
+                        const assistant = assistantFor(item.source ?? target);
+                        assistant.content.push({
+                            type: "serverToolCall",
+                            id: item.serverToolUseId ?? item.id,
+                            name: item.serverToolName ?? "tool",
+                            arguments: item.serverToolInput ?? {},
+                        });
+                    }
+                    else {
+                        const assistant = assistantFor(item.source ?? target);
+                        assistant.content.push({
+                            type: "toolSearchResult",
+                            toolUseId: item.serverToolUseId ?? item.id,
+                            content: (item.serverToolResult ?? []).map((ref) => ({
+                                type: "tool_reference",
+                                tool_name: ref.toolName,
+                            })),
+                        });
+                    }
+                }
+                else {
+                    // Emit as text for non-Anthropic providers at reduced fidelity.
+                    // Append to the current assistant turn to avoid splitting it with a
+                    // fabricated user message.
+                    const assistant = assistantFor(item.source ?? target);
+                    assistant.content.push({
+                        type: "text",
+                        text: `[Server tool ${item.serverToolName ?? item.tool}: ${JSON.stringify(item.payload)}]`,
+                    });
+                }
+            }
+            else if (isOpenAIResponsesApi(target.api)) {
+                const assistant = assistantFor(item.source ?? target);
+                assistant.content.push({
+                    type: "providerItem",
+                    item: item.payload,
+                    // Pi's generic estimator treats non-text assistant blocks like tool calls.
+                    // These compatibility fields stay in process; the Responses adapter emits only `item`.
+                    name: "providerItem",
+                    arguments: item.payload,
+                });
+            }
+            continue;
+        }
+        if (item.role === "assistant") {
+            const assistant = assistantFor(item.source ?? (pending ? identityOf(pending) : UNKNOWN_SOURCE));
+            for (const [contentIndex, content] of item.content.entries()) {
+                if (content.type === "text") {
+                    const signatureId = item.id ?? (item.phase ? `msg_dari_${itemIndex}_${contentIndex}` : undefined);
+                    assistant.content.push({
+                        type: "text",
+                        text: content.text,
+                        ...(content.annotations?.length ? { annotations: content.annotations } : {}),
+                        ...(signatureId ? { textSignature: textSignature(signatureId, item.phase) } : {}),
+                    });
+                }
+                else {
+                    assistant.content.push({ type: "text", text: "(assistant image omitted)" });
+                }
+            }
+        }
+    }
+    flushAssistant();
+    const context = { messages };
+    if (systemParts.length > 0)
+        context.systemPrompt = systemParts.join("\n\n");
+    const clientTools = request.tools.filter((tool) => tool.kind !== "hosted" && (tool.kind !== "server_tool" || target.api === "anthropic-messages"));
+    if (clientTools.length > 0)
+        context.tools = clientTools.map((tool) => toolToPi(tool, target.api));
+    return context;
+}
+function thinkingContent(item, continuation) {
+    if (continuation.kind === "openai_reasoning") {
+        const providerItemId = continuation.providerItemId ?? item.id;
+        return {
+            type: "thinking",
+            thinking: [...item.summary, ...item.content].join("\n"),
+            thinkingSignature: encodeOpenAIReasoningSignature({
+                ...(providerItemId ? { id: providerItemId } : {}),
+                summary: item.summary,
+                content: item.content,
+                encryptedContent: continuation.encryptedContent,
+            }),
+        };
+    }
+    if (continuation.kind === "anthropic_thinking") {
+        return {
+            type: "thinking",
+            thinking: continuation.thinking,
+            thinkingSignature: continuation.signature,
+        };
+    }
+    return {
+        type: "thinking",
+        thinking: "",
+        thinkingSignature: continuation.data,
+        redacted: true,
+    };
+}
+function userContent(content, shape) {
+    if (shape !== "parts")
+        return content.filter((part) => part.type === "text").map((part) => part.text).join("");
+    return content.map(contentToPi);
+}
+function appendPositionalNote(messages, text, timestamp) {
+    const previous = messages.at(-1);
+    if (previous?.role === "toolResult") {
+        previous.content = [...contentBlocks(previous.content), { type: "text", text }];
+        return;
+    }
+    appendUserMessage(messages, [{ type: "text", text }], timestamp);
+}
+function appendUserMessage(messages, content, timestamp) {
+    const previous = messages.at(-1);
+    if (previous?.role === "user") {
+        previous.content = [...contentBlocks(previous.content), ...contentBlocks(content)];
+        return;
+    }
+    messages.push({ role: "user", content, timestamp });
+}
+function contentBlocks(content) {
+    if (typeof content === "string")
+        return [{ type: "text", text: content }];
+    return Array.isArray(content) ? content : [];
+}
+function toolResultContent(content, target) {
+    if (content === null)
+        return [];
+    if (typeof content === "string")
+        return [{ type: "text", text: content }];
+    // Anthropic emits tool_reference parts natively via addedToolNames, so
+    // flattening them to synthetic text here would duplicate the names. Other
+    // providers have no native tool_reference representation, so keep the text
+    // fallback for them.
+    const filtered = target.api === "anthropic-messages"
+        ? content.filter((part) => part.type !== "tool_reference")
+        : content;
+    return filtered.map(contentToPi);
+}
+function contentToPi(content) {
+    if (content.type === "text")
+        return { type: "text", text: content.text };
+    if (content.type === "tool_reference")
+        return { type: "text", text: `Referenced tool: ${content.toolName}` };
+    const image = parseDataImageUrl(content.url);
+    if (!image) {
+        throw new HttpError(400, "Only base64 data URLs are supported for routed image input.", "invalid_request_error", "input.image_url");
+    }
+    return {
+        type: "image",
+        mimeType: image.mimeType,
+        data: image.data,
+        ...(content.detail !== undefined ? { detail: content.detail } : {}),
+    };
+}
+function portableReasoningText(items, itemIndex, readable, target) {
+    if (target.api !== "openai-completions")
+        return readable;
+    for (let index = itemIndex + 1; index < items.length; index += 1) {
+        const next = items[index];
+        if (next.type === "reasoning" || next.type === "tool_call")
+            continue;
+        if (next.type === "message" && next.role === "assistant") {
+            return next.content.some((part) => part.type === "text" && part.text.length > 0)
+                ? `${readable}\n\n`
+                : readable;
+        }
+        break;
+    }
+    return readable;
+}
+function readableContent(content) {
+    return content.map((part) => (part.type === "text" ? part.text : "(image omitted)")).join("\n");
+}
+function toolToPi(tool, targetApi) {
+    if (tool.kind === "client_tool" && targetApi === "anthropic-messages") {
+        return {
+            name: tool.name,
+            type: tool.type,
+            serverTool: true,
+        };
+    }
+    const result = {
+        name: (tool.kind === "server_tool" || tool.kind === "client_tool"
+            ? tool.name
+            : toolWireName(tool.namespace, tool.name)),
+        // Anthropic rejects description on server tool definitions as an extra input.
+        ...(tool.kind === "server_tool" ? {} : { description: tool.description ?? "" }),
+    };
+    if (tool.kind === "function") {
+        result.parameters = tool.parameters ?? { type: "object", properties: {} };
+        if (tool.deferLoading !== undefined)
+            result.deferLoading = tool.deferLoading;
+    }
+    else if (tool.kind === "server_tool") {
+        // Server tools are Anthropic-specific; Pi doesn't handle them. We pass
+        // the type through so the Pi adapter can preserve it for Anthropic.
+        result.type = tool.type;
+        // Server tools don't have input_schema; but Pi's convertTools would add one
+        // if we set parameters. Mark it to skip the schema.
+        result.serverTool = true;
+    }
+    else if (tool.kind === "client_tool") {
+        result.parameters = { type: "object", properties: {}, additionalProperties: true };
+    }
+    else {
+        result.parameters = {
+            type: "object",
+            properties: { input: { type: "string" } },
+            required: ["input"],
+            additionalProperties: false,
+        };
+    }
+    return result;
+}
+function toolArguments(kind, input) {
+    if (kind === "custom")
+        return { input: typeof input === "string" ? input : "" };
+    if (input && typeof input === "object")
+        return input;
+    if (typeof input !== "string" || input.trim() === "")
+        return {};
+    let parsed;
+    try {
+        parsed = JSON.parse(input);
+    }
+    catch {
+        throw new HttpError(400, "Function tool arguments must contain a JSON object.", "invalid_request_error", "input.arguments");
+    }
+    if (!isRecord(parsed)) {
+        throw new HttpError(400, "Function tool arguments must contain a JSON object.", "invalid_request_error", "input.arguments");
+    }
+    return parsed;
+}
+function textSignature(id, phase) {
+    return JSON.stringify({ v: 1, id, ...(phase ? { phase } : {}) });
+}
+function identityOf(value) {
+    return { provider: value.provider, api: value.api, model: value.model };
+}
+function zeroUsage() {
+    return {
+        input: 0,
+        output: 0,
+        cacheRead: 0,
+        cacheWrite: 0,
+        totalTokens: 0,
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+    };
+}

package/dist/providers/pi_options.d.ts

@@ -0,0 +1,11 @@
+import { type ReasoningEffort } from "@mupt-ai/dari-router/policy-engine";
+import type { InteractionRequest } from "../interaction/types.js";
+import type { ResolvedProviderAuth } from "../types.js";
+import type { ProviderFetch } from "./response_capture.js";
+type PiOptionsSupport = {
+    getSupportedThinkingLevels?: (model: Record<string, unknown>) => string[];
+};
+type Model = Record<string, unknown>;
+export declare function buildOptions(request: InteractionRequest, auth: ResolvedProviderAuth | string, selectedModel: string, model: Model, pi: PiOptionsSupport, sessionId?: string, selectedReasoningEffort?: ReasoningEffort, timeoutMs?: number, signal?: AbortSignal, fetch?: ProviderFetch): Record<string, unknown>;
+export declare function supportedThinkingLevels(model: Model, pi: PiOptionsSupport): string[];
+export {};