@mupt-ai/dari-gateway 0.3.3 → 0.3.5

package/src/providers/key_check.ts

@@ -0,0 +1,76 @@
+import { getBuiltinModels } from "@mupt-ai/pi-ai/providers/all";
+
+const CHECK_TIMEOUT_MS = 10_000;
+
+/**
+ * - `valid`: the provider accepted the key.
+ * - `rejected`: the provider refused it (401 or 403).
+ * - `unchecked`: no answer either way, such as a network error, an outage, or
+ *   a provider without a known check. Callers treat this as not rejected so a
+ *   provider outage never blocks a key.
+ */
+export type ProviderKeyCheckResult =
+  | { result: "valid" }
+  | { result: "rejected"; status: number }
+  | { result: "unchecked" };
+
+// A provider endpoint outside the model catalog, such as a self-run server.
+export type ProviderKeyEndpoint = { api: string; baseUrl: string };
+
+/**
+ * Ask a provider whether an API key works, with the cheapest authenticated
+ * request its wire API offers: list models, or, where listing models needs
+ * no key, an endpoint that does. The self-hosted gateway runs this at startup
+ * and the hosted control plane before it saves a key.
+ */
+export async function checkProviderKey(args: {
+  provider: string;
+  key: string;
+  endpoint?: ProviderKeyEndpoint;
+  fetch: typeof fetch;
+}): Promise<ProviderKeyCheckResult> {
+  const check = keyCheckRequest(args.provider, args.key, args.endpoint);
+  if (check === null) return { result: "unchecked" };
+  try {
+    const response = await args.fetch(check.url, {
+      headers: check.headers,
+      signal: AbortSignal.timeout(CHECK_TIMEOUT_MS),
+    });
+    await response.body?.cancel();
+    if (response.status === 401 || response.status === 403) {
+      return { result: "rejected", status: response.status };
+    }
+    return response.ok ? { result: "valid" } : { result: "unchecked" };
+  } catch {
+    return { result: "unchecked" };
+  }
+}
+
+function keyCheckRequest(
+  provider: string,
+  key: string,
+  endpoint: ProviderKeyEndpoint | undefined,
+): { url: string; headers: Record<string, string> } | null {
+  const bearer = { authorization: `Bearer ${key}` };
+  if (endpoint === undefined) {
+    if (provider === "openrouter") return { url: "https://openrouter.ai/api/v1/key", headers: bearer };
+    if (provider === "vercel-ai-gateway") return { url: "https://ai-gateway.vercel.sh/v1/credits", headers: bearer };
+  }
+  const target = endpoint
+    ?? (getBuiltinModels(provider as never) as unknown as ProviderKeyEndpoint[] | undefined)?.[0];
+  if (target === undefined || !URL.canParse(target.baseUrl)) return null;
+  const base = target.baseUrl.replace(/\/+$/, "");
+  switch (target.api) {
+    case "anthropic-messages":
+      return { url: `${base}/v1/models`, headers: { "x-api-key": key, "anthropic-version": "2023-06-01" } };
+    case "google-generative-ai":
+      return { url: `${base}/models`, headers: { "x-goog-api-key": key } };
+    case "mistral-conversations":
+      return { url: `${base}/v1/models`, headers: bearer };
+    case "openai-completions":
+    case "openai-responses":
+      return { url: `${base}/models`, headers: bearer };
+    default:
+      return null;
+  }
+}

package/src/providers/model_capabilities.ts

@@ -14,6 +14,24 @@ export function modelSupportsImageInput(metadata: ModelCapabilityMetadata): bool
   return Array.isArray(metadata.model.input) && metadata.model.input.includes("image");
 }
 
+/* PDFs are sent natively only where the model owner documents PDF input on
+   that API. The same API can serve other owners (xAI and Meta on Responses,
+   GPT on Bedrock) whose PDF support is unverified, so both must match. */
+const DOCUMENT_INPUT_APIS_BY_OWNER: Readonly<Record<string, ReadonlySet<string>>> = {
+  anthropic: new Set(["anthropic-messages", "bedrock-converse-stream"]),
+  openai: new Set(["openai-responses", "azure-openai-responses", "openai-codex-responses"]),
+};
+
+export function apiSupportsDocumentInput(api: string): boolean {
+  return Object.values(DOCUMENT_INPUT_APIS_BY_OWNER).some((apis) => apis.has(api));
+}
+
+export function modelSupportsDocumentInput(modelId: string, metadata: ModelCapabilityMetadata): boolean {
+  const owner = modelId.split("/", 1)[0] ?? "";
+  const api = metadata.model.api;
+  return typeof api === "string" && DOCUMENT_INPUT_APIS_BY_OWNER[owner]?.has(api) === true;
+}
+
 export function providerApiCanSerializeResponseFormat(metadata: ModelCapabilityMetadata): boolean {
   return (
     metadata.model.api === "openai-responses" ||

package/src/providers/pi_context.ts

@@ -17,6 +17,7 @@ import {
   sameProviderIdentity,
 } from "../interaction/provider_continuation_state.js";
 import { isOpenAIResponsesApi } from "./model_capabilities.js";
+import { apiSupportsDocumentInput } from "./model_capabilities.js";
 
 const UNKNOWN_SOURCE: ProviderIdentity = {
   provider: "dari",
@@ -104,7 +105,7 @@ export function buildContext(
     sawConversationItem = true;
     if (item.type === "message" && item.role === "user") {
       flushAssistant();
-      appendUserMessage(messages, userContent(item.content, item.contentShape), timestamp);
+      appendUserMessage(messages, userContent(item.content, item.contentShape, target), timestamp);
       continue;
     }
     if (item.type === "tool_result") {
@@ -297,10 +298,11 @@ function thinkingContent(
 
 function userContent(
   content: InteractionContent[],
-  shape: "string" | "parts" | "null" | "absent"
+  shape: "string" | "parts" | "null" | "absent",
+  target: ProviderIdentity,
 ): string | Array<Record<string, unknown>> {
   if (shape !== "parts") return content.filter((part) => part.type === "text").map((part) => part.text).join("");
-  return content.map(contentToPi);
+  return content.map((part) => contentToPi(part, target));
 }
 
 function appendPositionalNote(
@@ -347,12 +349,19 @@ function toolResultContent(
   const filtered = target.api === "anthropic-messages"
     ? content.filter((part) => part.type !== "tool_reference")
     : content;
-  return filtered.map(contentToPi);
+  return filtered.map((part) => contentToPi(part, target));
 }
 
-function contentToPi(content: InteractionContent): Record<string, unknown> {
+function contentToPi(content: InteractionContent, target: ProviderIdentity): Record<string, unknown> {
   if (content.type === "text") return { type: "text", text: content.text };
   if (content.type === "tool_reference") return { type: "text", text: `Referenced tool: ${content.toolName}` };
+  if (content.type === "document") {
+    // Routing keeps PDFs on document-capable models; this guards direct callers.
+    if (!apiSupportsDocumentInput(target.api)) {
+      throw new HttpError(400, `PDF document input is not supported by the ${target.api} API.`, "unsupported_input_modality", "input.document");
+    }
+    return { type: "document", mimeType: content.mimeType, data: content.data };
+  }
   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");

package/src/runtime/prepare_routed_completion.ts

@@ -1,6 +1,8 @@
 import type { CandidatePruningAudit } from "@mupt-ai/dari-router/policy-engine";
 
 import { interactionConversationPrefix } from "../context/interaction_fingerprint.js";
+import { clampRequestedEffort } from "../context/routing_context.js";
+import { modelMetadataLookup } from "../providers/pi.js";
 import {
   buildRoutingContext,
   derivePrefixLookupHashes,
@@ -134,7 +136,14 @@ export async function prepareRoutedCompletion(args: {
   }
   const access = strictProviderAccess(credentials, args.interaction);
   credentials = access.credentials;
-  args = { ...args, interaction: access.interaction };
+  args = {
+    ...args,
+    interaction: clampRequestedEffort(
+      access.interaction,
+      credentials,
+      modelMetadataLookup(credentials.model_configs),
+    ),
+  };
   credentials = await withProviderCredentials(credentials, args.apiKey, args.clientHarness);
   const billable = true;
   const routing = await buildRoutingContext({
@@ -190,6 +199,8 @@ export async function prepareRoutedCompletion(args: {
     identityHit: routing.identityHit,
     hits: routing.hits,
     costEstimates: routing.reportingCostEstimates,
+    candidates: candidateResolution.candidates,
+    sessionPinKey: sessionPin?.key,
     pruning: reportedPruning(
       candidateResolution.pruning,
       routing.imageCapabilityPrunedModels,

package/src/self_host/config.ts

@@ -12,7 +12,7 @@ import {
  *   default:
  *     models:
  *       - anthropic/claude-sonnet-5-5
- *       - id: openai/gpt-5.5
+ *       - id: openai/gpt-6.1-sol
  *         thinking: [medium, high]
  *     subscriptions: true   # serve with connected Claude Code / ChatGPT plans
  *     sessionPinning: false # route each session and subagent once, then keep its model
@@ -62,12 +62,42 @@ const ROUTER_ID_PATTERN = /^[A-Za-z0-9_-]+$/;
 const PROVIDER_ID_PATTERN = /^[a-z0-9][a-z0-9.-]*$/;
 const MODEL_ID_PATTERN = /^[a-z0-9][a-z0-9.-]*\/\S+$/;
 
-export async function loadGatewayConfig(path: string): Promise<GatewayConfig> {
+// Served when no config file is set or found. A model is served once its
+// provider has a credential (a connected subscription or the provider's API
+// key); dari/routing without Auto Routing serves the first available one.
+export const DEFAULT_GATEWAY_CONFIG: GatewayConfig = parseGatewayConfig({
+  routers: {
+    default: {
+      models: [
+        "anthropic/claude-sonnet-5-5",
+        "anthropic/claude-opus-5-5",
+        "anthropic/claude-fable-5-2",
+        "anthropic/claude-haiku-5-5",
+        "openai/gpt-6.1-sol",
+        "openai/gpt-6.1-terra",
+      ],
+    },
+  },
+});
+
+const DEFAULT_CONFIG_PATH = "dari-gateway.yaml";
+
+/**
+ * Load the config at `path`, which must exist. Without a path, load
+ * `dari-gateway.yaml` from the working directory when present, and otherwise
+ * serve DEFAULT_GATEWAY_CONFIG.
+ */
+export async function loadGatewayConfig(path: string | undefined): Promise<GatewayConfig> {
+  if (path === undefined) {
+    if (!(await Bun.file(DEFAULT_CONFIG_PATH).exists())) return DEFAULT_GATEWAY_CONFIG;
+    path = DEFAULT_CONFIG_PATH;
+  }
   const file = Bun.file(path);
   if (!(await file.exists())) {
     throw new Error(
       `Gateway config not found at ${path}. Write one from dari-gateway.example.yaml, `
-        + "or set DARI_GATEWAY_CONFIG (in the container image, mount it at /config/dari-gateway.yaml).",
+        + "or unset DARI_GATEWAY_CONFIG to serve the default router "
+        + "(in the container image, mount it at /config/dari-gateway.yaml).",
     );
   }
   const text = await file.text();

package/src/self_host/main.ts

@@ -10,9 +10,12 @@ import { installShutdownHandlers, startServer } from "../http/serve.js";
 import { createAutoRoutingSelectorBackend } from "../strategies/auto_routing_selector.js";
 import { createPrioritySelectorBackend } from "../strategies/priority_selector.js";
 import type { SelectorBackend } from "../strategies/routing.js";
-import { loadGatewayConfig } from "./config.js";
 import { createSelfHostedPlatform } from "./platform.js";
-import { gatewayKeyIndex, providersWithoutCredentials } from "./resolve.js";
+import { providerForModel } from "@mupt-ai/dari-router/policy-engine";
+import { checkProviderKey } from "../providers/key_check.js";
+import { CANONICAL_MODEL_IDS } from "../models/canonical.js";
+import { DEFAULT_GATEWAY_CONFIG, loadGatewayConfig, type GatewayConfig } from "./config.js";
+import { configuredProviderApiKey, gatewayKeyIndex, providersWithoutCredentials } from "./resolve.js";
 import { PrefixStore } from "./prefix_store.js";
 import { SessionPinStore } from "./session_pin_store.js";
 import { handleSubscriptionRequest, SubscriptionStore } from "./subscriptions/index.js";
@@ -23,7 +26,8 @@ const SUBSCRIPTION_ROUTES = "/v1/organizations/current/credentials";
  * Start a self-hosted Dari Gateway from environment variables:
  *
  * - `DARI_GATEWAY_KEYS` (required): comma-separated bearer keys clients use.
- * - `DARI_GATEWAY_CONFIG`: router config file (default `dari-gateway.yaml`).
+ * - `DARI_GATEWAY_CONFIG`: router config file. Unset, `dari-gateway.yaml`
+ *   serves when present, and otherwise the default router.
  * - `DARI_GATEWAY_DB`: SQLite database for subscriptions and prefix-cache
  *   state (default `dari-gateway.db`). It holds subscription tokens.
  * - `DARI_AUTO_ROUTING_API_KEY`: choose among a router's models with Dari
@@ -39,7 +43,7 @@ export async function startSelfHostedGateway(env: Record<string, string> = proce
   if (gatewayKeys.length === 0) {
     throw new Error("Set DARI_GATEWAY_KEYS to the comma-separated keys clients authenticate with.");
   }
-  const config = await loadGatewayConfig(env.DARI_GATEWAY_CONFIG ?? "dari-gateway.yaml");
+  const config = await loadGatewayConfig(env.DARI_GATEWAY_CONFIG?.trim() || undefined);
   const databasePath = env.DARI_GATEWAY_DB ?? "dari-gateway.db";
   const created = databasePath !== ":memory:" && !existsSync(databasePath);
   const db = new Database(databasePath, { create: true });
@@ -76,6 +80,8 @@ export async function startSelfHostedGateway(env: Record<string, string> = proce
       );
     }
   }
+  await checkConfiguredProviderKeys(config, (provider) =>
+    configuredProviderApiKey({ config, providerApiKey, env }, provider));
   const autoRoutingKey = env.DARI_AUTO_ROUTING_API_KEY?.trim();
   const selectorBackend: SelectorBackend = autoRoutingKey
     ? createAutoRoutingSelectorBackend({
@@ -109,7 +115,8 @@ export async function startSelfHostedGateway(env: Record<string, string> = proce
     "dari-gateway listening",
     JSON.stringify({
       hostname,
-      port,
+      // The bound port, so PORT=0 reports the port the OS assigned.
+      port: server.port,
       routers: Object.keys(config.routers),
       auto_routing: autoRoutingKey ? "dari" : "priority",
     }),
@@ -117,7 +124,60 @@ export async function startSelfHostedGateway(env: Record<string, string> = proce
   return { server, close: () => db.close() };
 }
 
-if (import.meta.main) {
+/**
+ * Check each provider API key the routers use before serving: a key its
+ * provider rejects fails startup, and one no check could answer is reported.
+ */
+async function checkConfiguredProviderKeys(
+  config: GatewayConfig,
+  providerApiKey: (provider: string) => Promise<string | undefined>,
+): Promise<void> {
+  const providers = new Set(
+    Object.values(config.routers).flatMap((router) => router.models.map((model) => providerForModel(model.id))),
+  );
+  const checks = await Promise.all([...providers].map(async (provider) => {
+    const endpoint = config.providers[provider];
+    // Endpoints configured without a key variable take no key.
+    if (endpoint !== undefined && endpoint.apiKeyEnv === undefined) return null;
+    const key = await providerApiKey(provider);
+    if (key === undefined) return null;
+    return { provider, ...(await checkProviderKey({ provider, key, ...(endpoint ? { endpoint } : {}), fetch })) };
+  }));
+  const rejected = checks.flatMap((check) => check?.result === "rejected" ? [`${check.provider} (${check.status})`] : []);
+  if (rejected.length > 0) {
+    throw new Error(
+      `Provider API keys were rejected: ${rejected.join(", ")}. Fix or unset each key before starting the gateway.`,
+    );
+  }
+  const unchecked = checks.flatMap((check) => check?.result === "unchecked" ? [check.provider] : []);
+  if (unchecked.length > 0) {
+    console.warn("dari-gateway could not check these provider API keys", JSON.stringify({ providers: unchecked }));
+  }
+}
+
+/** Run the gateway as a process: the `dari-gateway` bin and compiled binary. */
+export async function runSelfHostedGateway(): Promise<void> {
+  // `--print-config` prints the router config the gateway would serve, as
+  // JSON, and exits; `--print-default-config` prints the built-in default.
+  // The Dari CLI edits a local gateway's models from them.
+  if (process.argv.includes("--print-canonical-models")) {
+    console.log(JSON.stringify(CANONICAL_MODEL_IDS));
+    return;
+  }
+  if (process.argv.includes("--print-default-config")) {
+    console.log(JSON.stringify(DEFAULT_GATEWAY_CONFIG));
+    return;
+  }
+  if (process.argv.includes("--print-config")) {
+    try {
+      const config = await loadGatewayConfig(process.env.DARI_GATEWAY_CONFIG?.trim() || undefined);
+      console.log(JSON.stringify(config));
+      return;
+    } catch (error) {
+      console.error(`dari-gateway: ${error instanceof Error ? error.message : String(error)}`);
+      process.exit(1);
+    }
+  }
   // Startup errors are configuration mistakes: print the message, not a stack.
   const started = await startSelfHostedGateway().catch((error: unknown) => {
     console.error(`dari-gateway: ${error instanceof Error ? error.message : String(error)}`);
@@ -125,3 +185,5 @@ if (import.meta.main) {
   });
   installShutdownHandlers(started.server, async () => started.close());
 }
+
+if (import.meta.main) await runSelfHostedGateway();

package/src/self_host/resolve.ts

@@ -86,7 +86,7 @@ export async function resolveSelfHostedRouter(
     if (subscription !== null && deps.subscriptions.list(subscription).length > 0) {
       exhausted.add(subscription);
     }
-    const apiKey = await providerApiKey(deps, provider);
+    const apiKey = await configuredProviderApiKey(deps, provider);
     credentials.set(provider, apiKey ? { key: apiKey, source: "user" } : null);
   }
 
@@ -154,7 +154,7 @@ export async function resolveSelfHostedRouter(
 
 // Configured endpoints read their own key variable, or send a placeholder
 // when they take none; known providers use their standard variable.
-async function providerApiKey(
+export async function configuredProviderApiKey(
   deps: Pick<SelfHostedResolverDeps, "config" | "providerApiKey" | "env">,
   provider: string,
 ): Promise<string | undefined> {
@@ -226,7 +226,7 @@ export async function providersWithoutCredentials(
       : [];
     const connected = plans.some((plan) => plan !== null && deps.subscriptions.list(plan).length > 0);
     if (connected) continue;
-    const apiKey = await providerApiKey(deps, provider);
+    const apiKey = await configuredProviderApiKey(deps, provider);
     if (apiKey === undefined) missing.push(provider);
   }
   return missing;

package/src/strategies/selector_projection.ts

@@ -190,6 +190,8 @@ function selectorContentField(
       if (part.type === "tool_reference") {
         return { type: "text", text: `Referenced tool: ${part.toolName}` };
       }
+      // The selector model is text-only; it sees a marker, not the document bytes.
+      if (part.type === "document") return { type: "text", text: "(PDF document omitted from selector input)" };
       return imagePart(part.url, part.detail);
     }),
   };

package/src/telemetry/report.ts

@@ -11,6 +11,7 @@ import {
   type ChainHits,
   type PrefixHit,
   type PricingLookup,
+  type RoutingCandidate,
   type RoutingDecision,
   type RoutingStrategy,
 } from "@mupt-ai/dari-router/policy-engine";
@@ -25,7 +26,7 @@ import {
   type ProviderIdentity,
 } from "../interaction/types.js";
 import { captureOperationalError } from "../observability/errors.js";
-import { gatewayPlatform, type RequestTags } from "../platform/platform.js";
+import { gatewayPlatform, type RequestRecordRouter, type RequestTags } from "../platform/platform.js";
 import type {
   ProviderAttemptRecord,
   ProviderAttemptReport,
@@ -179,6 +180,10 @@ export class RequestReporter {
       identityHit?: PrefixHit;
       hits: ChainHits;
       costEstimates: CandidateCostEstimate[];
+      // The model and reasoning pairs this request could be routed to.
+      candidates: readonly RoutingCandidate[];
+      // The session pin this request routed under, when the router pins.
+      sessionPinKey?: string;
       // Cache-aware selector-candidate pruning outcome, plus any pre-strategy
       // capability exclusions. Null when custom routing configuration cannot
       // produce candidates.
@@ -603,6 +608,9 @@ export class RequestReporter {
       };
 
       gatewayPlatform().recordRequest({
+        router: requestRecordRouter(a.credentials),
+        candidates: a.candidates.map(({ model, reasoningEffort }) => ({ model, reasoningEffort })),
+        ...(a.sessionPinKey === undefined ? {} : { sessionPinKey: a.sessionPinKey }),
         row: ingestPayload,
         summary: {
           request_id: a.requestId,
@@ -786,3 +794,28 @@ function usageIngest(
     pricing_known: pricingKnown,
   };
 }
+
+function requestRecordRouter(router: ResolvedRouter): RequestRecordRouter {
+  return {
+    enabledModels: [...router.enabled_models],
+    thinkingLevels: Object.fromEntries(
+      router.enabled_models.map((model) => [
+        model,
+        [
+          ...(router.model_thinking_levels?.[model]
+            ?? router.model_configs?.[model]?.supported_thinking_levels
+            ?? []),
+        ],
+      ]),
+    ),
+    defaultThinkingLevels: Object.fromEntries(
+      router.enabled_models.flatMap((model) => {
+        const level = router.model_configs?.[model]?.default_thinking_level;
+        return level ? [[model, level]] : [];
+      }),
+    ),
+    strict: router.strict === true,
+    speculativeRouting: router.speculative_routing === true,
+    sessionPinning: router.session_pinning === true,
+  };
+}