@mupt-ai/dari-gateway 0.2.0 → 0.3.0

package/dist/providers/pi.js

@@ -0,0 +1,325 @@
+import { getSupportedThinkingLevels, normalizeContext, } from "@mupt-ai/pi-ai";
+import { anthropicMessagesApi } from "@mupt-ai/pi-ai/api/anthropic-messages.lazy";
+import { azureOpenAIResponsesApi } from "@mupt-ai/pi-ai/api/azure-openai-responses.lazy";
+import { bedrockConverseStreamApi } from "@mupt-ai/pi-ai/api/bedrock-converse-stream.lazy";
+import { googleGenerativeAIApi } from "@mupt-ai/pi-ai/api/google-generative-ai.lazy";
+import { googleVertexApi } from "@mupt-ai/pi-ai/api/google-vertex.lazy";
+import { mistralConversationsApi } from "@mupt-ai/pi-ai/api/mistral-conversations.lazy";
+import { openAICodexResponsesApi } from "@mupt-ai/pi-ai/api/openai-codex-responses.lazy";
+import { openAICompletionsApi } from "@mupt-ai/pi-ai/api/openai-completions.lazy";
+import { openAIResponsesApi } from "@mupt-ai/pi-ai/api/openai-responses.lazy";
+import { getBuiltinModel } from "@mupt-ai/pi-ai/providers/all";
+import { REASONING_EFFORTS, isRecord, nativeModelId, providerForModel, } from "@mupt-ai/dari-router/policy-engine";
+import { HttpError } from "../http/errors.js";
+import { effectiveContextWindow } from "../models/context_limits.js";
+import { defaultBaseUrl } from "../models/provider_defaults.js";
+import { contextWindowOverride, inputModalitiesOverride, providerThinkingLevelOverrides, runtimeCompat, withThinkingLevelOverrides, } from "../models/runtime_overrides.js";
+import { BEDROCK_API } from "./model_capabilities.js";
+import { buildContext } from "./pi_context.js";
+import { providerIdentityFromModel } from "../interaction/provider_continuation_state.js";
+import { buildOptions, supportedThinkingLevels } from "./pi_options.js";
+import { eventsWithSafeguardResults, forwardsAnthropicSafeguards, messageWithSafeguardResults, safeguardResultsTap, } from "./anthropic_safeguards.js";
+import { eventsWithProviderReportedCost, messageWithProviderReportedCost, providerReportedCostTap, reportsProviderCost, } from "./provider_reported_cost.js";
+import { eventsWithSubscriptionCacheWrites, messageWithSubscriptionCacheWrites, modelPricingFromPiModel, } from "./subscription_usage.js";
+// Pi's built-in providers stream through these API implementations unchanged.
+// Dari always passes resolved credentials, so it dispatches on the API rather
+// than through a Models collection with its own auth resolution.
+const PI_APIS = {
+    "anthropic-messages": anthropicMessagesApi(),
+    "azure-openai-responses": azureOpenAIResponsesApi(),
+    "bedrock-converse-stream": bedrockConverseStreamApi(),
+    "google-generative-ai": googleGenerativeAIApi(),
+    "google-vertex": googleVertexApi(),
+    "mistral-conversations": mistralConversationsApi(),
+    "openai-codex-responses": openAICodexResponsesApi(),
+    "openai-completions": openAICompletionsApi(),
+    "openai-responses": openAIResponsesApi(),
+};
+function piStreamSimple(model, context, options) {
+    const api = PI_APIS[String(model.api)];
+    if (!api)
+        throw new Error(`No Pi API implementation for ${String(model.api)}.`);
+    return api.streamSimple(model, normalizeContext(context), options);
+}
+export const pi = {
+    completeSimple: async (model, context, options) => await piStreamSimple(model, context, options).result(),
+    streamSimple: (model, context, options) => piStreamSimple(model, context, options),
+    getModel: (provider, id) => getBuiltinModel(provider, id),
+    getSupportedThinkingLevels: (model) => getSupportedThinkingLevels(model),
+};
+export async function completeInteractionModel(args) {
+    const model = modelAtEndpoint(buildModel(pi, args.model, args.pricing, args.config), args.endpoint);
+    const source = providerIdentityFromModel(model);
+    const context = buildContext(args.request, source);
+    const costTap = reportsProviderCost(model) ? providerReportedCostTap(args.fetch) : undefined;
+    const safeguards = forwardsAnthropicSafeguards(args.request, model)
+        ? safeguardResultsTap(costTap?.fetch ?? args.fetch)
+        : undefined;
+    const options = buildOptions(args.request, args.auth ?? args.apiKey, args.model, model, pi, args.sessionId, args.selectedReasoningEffort, args.timeoutMs, args.signal, safeguards?.fetch ?? costTap?.fetch ?? args.fetch);
+    const usageContext = subscriptionUsageContext(args.model, model, source);
+    const message = options.emitProviderEvents === true
+        ? await completeProviderEventStream(pi.streamSimple(model, context, options))
+        : await pi.completeSimple(model, context, options);
+    return {
+        message: messageWithProviderReportedCost(messageWithSubscriptionCacheWrites(messageWithSafeguardResults(message, safeguards?.results()), usageContext), costTap?.cost()),
+        source,
+    };
+}
+function subscriptionUsageContext(modelId, model, source) {
+    // pi-ai priced this call from model.cost, so the repricing must read it back
+    // rather than the caller's catalog pricing, which fast mode can override.
+    return { api: source.api, model: modelId, pricing: modelPricingFromPiModel(model) };
+}
+export async function completeProviderEventStream(events) {
+    let providerResponse;
+    const providerOutput = new Map();
+    for await (const event of events) {
+        if (event.type === "provider_event" && isRecord(event.event)) {
+            const raw = event.event;
+            if (raw.type === "response.output_item.done" && isRecord(raw.item)) {
+                const outputIndex = raw.output_index;
+                if (typeof outputIndex !== "number" || !Number.isInteger(outputIndex) || outputIndex < 0) {
+                    throw new Error("Provider output item is missing a valid output index.");
+                }
+                providerOutput.set(outputIndex, raw.item);
+            }
+            if ((raw.type === "response.completed" || raw.type === "response.incomplete") &&
+                isRecord(raw.response)) {
+                providerResponse = raw.response;
+            }
+            continue;
+        }
+        if (event.type === "done" && isRecord(event.message)) {
+            if (providerResponse) {
+                const terminalOutput = providerResponse.output;
+                if ((!Array.isArray(terminalOutput) || terminalOutput.length === 0) && providerOutput.size > 0) {
+                    const entries = [...providerOutput.entries()].sort(([left], [right]) => left - right);
+                    if (entries.some(([index], position) => index !== position)) {
+                        throw new Error("Provider stream ended with non-contiguous output item indexes.");
+                    }
+                    providerResponse = { ...providerResponse, output: entries.map(([, item]) => item) };
+                }
+                return { ...event.message, providerResponse };
+            }
+            return event.message;
+        }
+        if (event.type === "error" && isRecord(event.error))
+            return event.error;
+    }
+    throw new Error("Provider stream ended without a terminal event.");
+}
+export async function streamInteractionModel(args) {
+    const model = modelAtEndpoint(buildModel(pi, args.model, args.pricing, args.config), args.endpoint);
+    const source = providerIdentityFromModel(model);
+    const costTap = reportsProviderCost(model) ? providerReportedCostTap(args.fetch) : undefined;
+    const safeguards = forwardsAnthropicSafeguards(args.request, model)
+        ? safeguardResultsTap(costTap?.fetch ?? args.fetch)
+        : undefined;
+    const events = pi.streamSimple(model, buildContext(args.request, source), buildOptions(args.request, args.auth ?? args.apiKey, args.model, model, pi, args.sessionId, args.selectedReasoningEffort, undefined, args.signal, safeguards?.fetch ?? costTap?.fetch ?? args.fetch));
+    const withCacheWrites = eventsWithSubscriptionCacheWrites(safeguards ? eventsWithSafeguardResults(events, safeguards) : events, subscriptionUsageContext(args.model, model, source));
+    return {
+        events: costTap ? eventsWithProviderReportedCost(withCacheWrites, costTap) : withCacheWrites,
+        source,
+    };
+}
+function modelAtEndpoint(model, endpoint) {
+    if (endpoint === undefined)
+        return model;
+    return {
+        ...model,
+        ...(endpoint.baseUrl ? { baseUrl: endpoint.baseUrl } : {}),
+        ...(endpoint.headers
+            ? { headers: { ...(isRecord(model.headers) ? model.headers : {}), ...endpoint.headers } }
+            : {}),
+    };
+}
+export function modelContextWindow(modelId) {
+    const provider = providerForModel(modelId);
+    const nativeId = nativeModelId(modelId);
+    const registered = resolveModelMetadata(pi, provider, modelId, nativeId);
+    return contextWindowForModel(modelId, registered);
+}
+/**
+ * Pi's catalog price for a canonical model served by `provider`, in USD per
+ * million tokens. Null when Pi does not list the model.
+ */
+export function registeredModelPrice(modelId, provider = providerForModel(modelId)) {
+    const registered = resolveModelMetadata(pi, provider, modelId, nativeModelId(modelId));
+    const cost = registered?.cost;
+    if (!isRecord(cost) || typeof cost.input !== "number" || typeof cost.output !== "number") {
+        return null;
+    }
+    return {
+        input: cost.input,
+        output: cost.output,
+        cached_input: typeof cost.cacheRead === "number" ? cost.cacheRead : null,
+        cache_write: typeof cost.cacheWrite === "number" ? cost.cacheWrite : null,
+    };
+}
+export function modelMetadataLookup(configs) {
+    return (modelId, pricing) => {
+        const model = buildModel(pi, modelId, pricing, configs?.[modelId]);
+        return {
+            model,
+            supportedThinkingLevels: supportedThinkingLevels(model, pi),
+        };
+    };
+}
+export function buildModel(pi, modelId, pricing, config) {
+    const provider = config?.provider ?? providerForModel(modelId);
+    const nativeId = config?.native_model_id ?? nativeModelId(modelId);
+    const registered = resolveModelMetadata(pi, provider, modelId, nativeId);
+    const id = registered?.id ?? (provider === "anthropic" ? nativeId.replace(/(?<=\d)\.(?=\d)/g, "-") : nativeId);
+    const price = executionPricing(pricing, config);
+    const configuredThinkingLevels = config?.supported_thinking_levels;
+    // A null catalog value delegates to Pi's registry when it knows the model.
+    // Custom OpenAI-compatible models are not necessarily registered, so keep
+    // the public effort set permissive and let the provider validate the wire
+    // value instead of silently treating an unknown model as off-only. A
+    // built-in provider with its own thinking levels (Meta) serves ids newer
+    // than the registry the same way, with only its key configured.
+    const providerThinking = providerThinkingLevelOverrides(provider);
+    const fallbackThinkingLevels = (config !== undefined || providerThinking !== undefined) &&
+        configuredThinkingLevels == null &&
+        registered === undefined
+        ? REASONING_EFFORTS
+        : undefined;
+    const effectiveThinkingLevels = Array.isArray(configuredThinkingLevels)
+        ? configuredThinkingLevels
+        : fallbackThinkingLevels;
+    const configuredThinking = effectiveThinkingLevels
+        ? thinkingMetadata(effectiveThinkingLevels, registered ?? { thinkingLevelMap: providerThinking })
+        : null;
+    const defaultThinkingLevel = config?.default_thinking_level ?? undefined;
+    const api = config?.api ?? registered?.api ?? defaultApi(provider);
+    const baseUrl = config?.base_url ?? registered?.baseUrl ?? defaultBaseUrl(provider);
+    if (!baseUrl) {
+        throw new HttpError(502, `No base URL is configured for provider ${provider}.`, "provider_configuration_error");
+    }
+    const contextWindow = contextWindowForModel(modelId, registered, config) ?? 128000;
+    const compat = responsesToolContract(api, runtimeCompat(provider, registered, modelId));
+    return {
+        ...(registered ?? {}),
+        ...(compat !== undefined ? { compat } : {}),
+        id,
+        name: registered?.name ?? id,
+        provider,
+        api,
+        baseUrl,
+        reasoning: configuredThinking?.reasoning ?? registered?.reasoning ?? false,
+        input: registered?.input ?? inputModalitiesOverride(modelId) ?? ["text"],
+        cost: price ?? registered?.cost ?? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow,
+        maxTokens: config?.max_output_tokens ?? registered?.maxTokens ?? contextWindow,
+        ...(defaultThinkingLevel !== undefined ? { defaultThinkingLevel } : {}),
+        ...(configuredThinking ? { thinkingLevelMap: configuredThinking.thinkingLevelMap } : {}),
+        ...(effectiveThinkingLevels
+            ? { supportedThinkingLevels: [...effectiveThinkingLevels] }
+            : {}),
+        fastMode: config?.fast_mode ?? false,
+        zeroDataRetention: config?.zdr ?? false,
+    };
+}
+function thinkingMetadata(supportedLevels, registered) {
+    const supported = new Set(supportedLevels);
+    const registeredMap = isRecord(registered?.thinkingLevelMap)
+        ? registered.thinkingLevelMap
+        : {};
+    const thinkingLevelMap = {};
+    for (const level of REASONING_EFFORTS) {
+        if (!supported.has(level)) {
+            thinkingLevelMap[level] = null;
+            continue;
+        }
+        const registeredValue = registeredMap[level];
+        if (typeof registeredValue === "string") {
+            thinkingLevelMap[level] = registeredValue;
+        }
+        else if (level === "xhigh" || level === "max") {
+            // Pi requires extended levels to be explicitly mapped. The catalog
+            // levels use the provider-native OpenAI spelling.
+            thinkingLevelMap[level] = level;
+        }
+    }
+    return {
+        reasoning: supportedLevels.some((level) => level !== "off"),
+        thinkingLevelMap,
+    };
+}
+function resolveModelMetadata(pi, provider, modelId, nativeId) {
+    const registry = registryProvider(provider, modelId);
+    return withThinkingLevelOverrides(provider, pi.getModel?.(registry, nativeId) ?? pi.getModel?.(registry, modelId));
+}
+// Dari's "azure" provider has no Pi registry of its own. One Foundry resource
+// serves OpenAI deployments through Pi's Azure OpenAI Responses registry and
+// Claude deployments through the Anthropic registry, whose model metadata
+// (input modalities, adaptive-thinking compat, thinking level maps) is the
+// same on Foundry; the resolved config still owns api, base URL, and cost.
+function registryProvider(provider, modelId) {
+    if (provider !== "azure")
+        return provider;
+    return providerForModel(modelId) === "anthropic" ? "anthropic" : "azure-openai-responses";
+}
+function contextWindowForModel(modelId, registered, config) {
+    const registeredWindow = typeof registered?.contextWindow === "number" ? registered.contextWindow : null;
+    const configWindow = effectiveContextWindow(config, config?.allow_long_context === true);
+    const override = contextWindowOverride(modelId);
+    const configured = configWindow ?? registeredWindow;
+    if (configWindow !== null || override === undefined)
+        return configured;
+    return Math.max(configured ?? 0, override);
+}
+// Pi 1.4 sends Responses tool definitions as conversation items, not the request's
+// `tools` field, when a model supports mid-conversation system messages. The gateway
+// keeps tools at the request level until that native tool-change path is validated.
+const RESPONSES_APIS = new Set(["openai-responses", "azure-openai-responses", "openai-codex-responses"]);
+function responsesToolContract(api, compat) {
+    if (typeof api !== "string" || !RESPONSES_APIS.has(api))
+        return compat;
+    if (!isRecord(compat) || compat.supportsMidConvoSystemMessages !== true)
+        return compat;
+    return { ...compat, supportsMidConvoSystemMessages: false };
+}
+function modelPricingFromConfig(config) {
+    const configuredCost = config?.fast_mode && config.provider === "anthropic" && config.fast_mode_cost
+        ? config.fast_mode_cost
+        : config?.cost;
+    if (!configuredCost || typeof configuredCost.input !== "number" || typeof configuredCost.output !== "number")
+        return null;
+    return {
+        input: configuredCost.input,
+        output: configuredCost.output,
+        cacheRead: configuredCost.cached_input ?? configuredCost.input,
+        cacheWrite: configuredCost.cache_write ?? configuredCost.input,
+        ...(configuredCost.tiers?.length ? { tiers: configuredCost.tiers.map((tier) => ({
+                inputTokensAbove: tier.input_tokens_above,
+                input: tier.input,
+                output: tier.output,
+                cacheRead: tier.cached_input ?? tier.input,
+                cacheWrite: tier.cache_write ?? tier.input,
+            })) } : {}),
+    };
+}
+function executionPricing(pricing, config) {
+    if (config?.fast_mode &&
+        config.provider === "openai" &&
+        config.api === "openai-responses") {
+        // The platform publishes fast prices for routing decisions and the UI,
+        // while Pi multiplies the standard rates using the service tier actually
+        // returned by OpenAI. Feeding Pi premium rates here would charge twice.
+        return modelPricingFromConfig({ ...config, fast_mode: false });
+    }
+    return pricing ?? modelPricingFromConfig(config);
+}
+function defaultApi(provider) {
+    // Meta recommends Responses: its Chat Completions drops reasoning between turns.
+    if (provider === "openai" || provider === "xai" || provider === "meta")
+        return "openai-responses";
+    if (provider === "anthropic")
+        return "anthropic-messages";
+    if (provider === "amazon-bedrock")
+        return BEDROCK_API;
+    if (provider === "azure")
+        return "azure-openai-responses";
+    return "openai-completions";
+}

package/dist/providers/pi_context.d.ts

@@ -0,0 +1,7 @@
+import type { InteractionRequest, ProviderIdentity } from "../interaction/types.js";
+export type PiContext = Record<string, unknown> & {
+    messages: Array<Record<string, unknown>>;
+    systemPrompt?: string;
+    tools?: Array<Record<string, unknown>>;
+};
+export declare function buildContext(request: InteractionRequest, target: ProviderIdentity): PiContext;