@mupt-ai/dari-gateway 0.3.3 → 0.3.4

package/dist/providers/key_check.js

@@ -0,0 +1,54 @@
+import { getBuiltinModels } from "@mupt-ai/pi-ai/providers/all";
+const CHECK_TIMEOUT_MS = 10_000;
+/**
+ * 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) {
+    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, key, endpoint) {
+    const bearer = { authorization: `Bearer ${key}` };
+    if (endpoint === undefined) {
+        if (provider === "openrouter")
+            return { url: "https://openrouter.ai/api/v1/key", headers: bearer };
+        if (provider === "vercel-ai-gateway")
+            return { url: "https://ai-gateway.vercel.sh/v1/credits", headers: bearer };
+    }
+    const target = endpoint
+        ?? getBuiltinModels(provider)?.[0];
+    if (target === undefined || !URL.canParse(target.baseUrl))
+        return null;
+    const base = target.baseUrl.replace(/\/+$/, "");
+    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/dist/providers/model_capabilities.d.ts

@@ -4,6 +4,8 @@ type ModelCapabilityMetadata = {
 export declare const BEDROCK_API = "bedrock-converse-stream";
 export declare function isBedrockApi(api: unknown): boolean;
 export declare function modelSupportsImageInput(metadata: ModelCapabilityMetadata): boolean;
+export declare function apiSupportsDocumentInput(api: string): boolean;
+export declare function modelSupportsDocumentInput(modelId: string, metadata: ModelCapabilityMetadata): boolean;
 export declare function providerApiCanSerializeResponseFormat(metadata: ModelCapabilityMetadata): boolean;
 export declare function isOpenAIResponsesApi(api: unknown): boolean;
 export declare function modelSupportsReasoningConfigurationUpdate(metadata: ModelCapabilityMetadata): boolean;

package/dist/providers/model_capabilities.js

@@ -6,6 +6,21 @@ export function isBedrockApi(api) {
 export function modelSupportsImageInput(metadata) {
     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 = {
+    anthropic: new Set(["anthropic-messages", "bedrock-converse-stream"]),
+    openai: new Set(["openai-responses", "azure-openai-responses", "openai-codex-responses"]),
+};
+export function apiSupportsDocumentInput(api) {
+    return Object.values(DOCUMENT_INPUT_APIS_BY_OWNER).some((apis) => apis.has(api));
+}
+export function modelSupportsDocumentInput(modelId, metadata) {
+    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) {
     return (metadata.model.api === "openai-responses" ||
         metadata.model.api === "azure-openai-responses" ||

package/dist/providers/pi_context.js

@@ -5,6 +5,7 @@ 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";
+import { apiSupportsDocumentInput } from "./model_capabilities.js";
 const UNKNOWN_SOURCE = {
     provider: "dari",
     api: "interaction-history",
@@ -72,7 +73,7 @@ export function buildContext(request, target) {
         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") {
@@ -252,10 +253,10 @@ function thinkingContent(item, continuation) {
         redacted: true,
     };
 }
-function userContent(content, shape) {
+function userContent(content, shape, target) {
     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(messages, text, timestamp) {
     const previous = messages.at(-1);
@@ -290,13 +291,20 @@ function toolResultContent(content, target) {
     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) {
+function contentToPi(content, target) {
     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/dist/runtime/prepare_routed_completion.js

@@ -1,4 +1,6 @@
 import { interactionConversationPrefix } from "../context/interaction_fingerprint.js";
+import { clampRequestedEffort } from "../context/routing_context.js";
+import { modelMetadataLookup } from "../providers/pi.js";
 import { buildRoutingContext, derivePrefixLookupHashes, } from "../context/routing_context.js";
 import { isBackgroundRequest } from "../interaction/types.js";
 import { executionProviderForModel } from "../models/execution_provider.js";
@@ -84,7 +86,10 @@ 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({
@@ -138,6 +143,8 @@ 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),
         pricing: routing.pricing,
         chainByModel: routing.prefix.chainByModel,

package/dist/self_host/config.d.ts

@@ -7,7 +7,7 @@ import { type ReasoningEffort } from "@mupt-ai/dari-router/policy-engine";
  *   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

package/dist/self_host/main.js

@@ -8,9 +8,11 @@ import { createRequestHandler } from "../http/handler.js";
 import { installShutdownHandlers, startServer } from "../http/serve.js";
 import { createAutoRoutingSelectorBackend } from "../strategies/auto_routing_selector.js";
 import { createPrioritySelectorBackend } from "../strategies/priority_selector.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 { loadGatewayConfig } 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";
@@ -69,6 +71,7 @@ export async function startSelfHostedGateway(env = process.env) {
             console.warn("dari-gateway router has providers without credentials; their models are not served", JSON.stringify({ router: routerId, providers: missing }));
         }
     }
+    await checkConfiguredProviderKeys(config, (provider) => configuredProviderApiKey({ config, providerApiKey, env }, provider));
     const autoRoutingKey = env.DARI_AUTO_ROUTING_API_KEY?.trim();
     const selectorBackend = autoRoutingKey
         ? createAutoRoutingSelectorBackend({
@@ -99,12 +102,38 @@ export async function startSelfHostedGateway(env = process.env) {
     const server = startServer(port, handler, { hostname });
     console.info("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",
     }));
     return { server, close: () => db.close() };
 }
+/**
+ * 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, providerApiKey) {
+    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 }));
+    }
+}
 if (import.meta.main) {
     // Startup errors are configuration mistakes: print the message, not a stack.
     const started = await startSelfHostedGateway().catch((error) => {

package/dist/self_host/resolve.d.ts

@@ -25,6 +25,7 @@ export type SelfHostedResolverDeps = {
  * neither are left out.
  */
 export declare function resolveSelfHostedRouter(deps: SelfHostedResolverDeps, request: ResolveRouterRequest): Promise<ResolvedRouter>;
+export declare function configuredProviderApiKey(deps: Pick<SelfHostedResolverDeps, "config" | "providerApiKey" | "env">, provider: string): Promise<string | undefined>;
 export declare function gatewayKeyIndex(keys: readonly string[], presented: string): number | null;
 /**
  * Providers a router lists that no credential can serve: no API key and no

package/dist/self_host/resolve.js

@@ -44,7 +44,7 @@ export async function resolveSelfHostedRouter(deps, request) {
         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);
     }
     const models = router.models.filter((model) => credentials.get(providerForModel(model.id)) != null);
@@ -112,7 +112,7 @@ export async function resolveSelfHostedRouter(deps, request) {
 }
 // Configured endpoints read their own key variable, or send a placeholder
 // when they take none; known providers use their standard variable.
-async function providerApiKey(deps, provider) {
+export async function configuredProviderApiKey(deps, provider) {
     const endpoint = deps.config.providers[provider];
     if (endpoint === undefined)
         return (await deps.providerApiKey(provider))?.trim() || undefined;
@@ -172,7 +172,7 @@ export async function providersWithoutCredentials(deps, router) {
         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);
     }

package/dist/strategies/selector_projection.js

@@ -145,6 +145,9 @@ function selectorContentField(content, shape, imagePart) {
             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/dist/telemetry/report.d.ts

@@ -1,4 +1,4 @@
-import { type CandidateCostEstimate, type CandidatePruningAudit, type ChainHits, type PrefixHit, type PricingLookup, type RoutingDecision, type RoutingStrategy } from "@mupt-ai/dari-router/policy-engine";
+import { type CandidateCostEstimate, type CandidatePruningAudit, type ChainHits, type PrefixHit, type PricingLookup, type RoutingCandidate, type RoutingDecision, type RoutingStrategy } from "@mupt-ai/dari-router/policy-engine";
 import type { ImageCapabilityPrunedModel } from "../context/routing_context.js";
 import { type InteractionRequest, type ProviderIdentity } from "../interaction/types.js";
 import { type RequestTags } from "../platform/platform.js";
@@ -47,6 +47,8 @@ export declare class RequestReporter {
         identityHit?: PrefixHit;
         hits: ChainHits;
         costEstimates: CandidateCostEstimate[];
+        candidates: readonly RoutingCandidate[];
+        sessionPinKey?: string;
         pruning: ReportedCandidatePruning | null;
         pricing: PricingLookup | null;
         chainByModel: ReadonlyMap<string, string[]>;

package/dist/telemetry/report.js

@@ -414,6 +414,9 @@ export class RequestReporter {
                 touched_prefix_hash: touchedHash,
             };
             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,
@@ -583,3 +586,23 @@ function usageIngest(usage, pricingKnown) {
         pricing_known: pricingKnown,
     };
 }
+function requestRecordRouter(router) {
+    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,
+    };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mupt-ai/dari-gateway",
-  "version": "0.3.3",
+  "version": "0.3.4",
   "description": "Self-hostable LLM gateway for coding agents and APIs: one endpoint for Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses, with subscription pooling, failover, and prompt-cache-aware routing.",
   "license": "Apache-2.0",
   "repository": {
@@ -50,7 +50,7 @@
   "dependencies": {
     "@aws-sdk/client-sts": "3.1063.0",
     "@mupt-ai/dari-router": "0.2.5",
-    "@mupt-ai/pi-ai": "1.4.0-mupt.2"
+    "@mupt-ai/pi-ai": "1.4.0-mupt.3"
   },
   "devDependencies": {
     "@types/bun": "^1.3.4",

package/src/context/interaction_fingerprint_format.ts

@@ -22,6 +22,7 @@ export type FingerprintContent =
           url: string;
           detail?: "auto" | "low" | "high" | "original";
         }
+      | { type: "document"; mimeType: "application/pdf"; data: string }
       | { type: "tool_reference"; toolName: string }
     >
   | null;
@@ -150,6 +151,9 @@ export function interactionFingerprintContent(
         toolName: part.toolName,
       };
     }
+    if (part.type === "document") {
+      return { type: "document" as const, mimeType: part.mimeType, data: part.data };
+    }
     return {
       type: "image" as const,
       url: part.url,

package/src/context/provider_prompt_accounting.ts

@@ -1,3 +1,5 @@
+import { constants as zlibConstants, inflateSync } from "node:zlib";
+
 import {
   isRecord,
   stableStringify,
@@ -27,6 +29,42 @@ const IMAGE_ESTIMATED_TOKENS = {
   original: 2048,
 } as const;
 const CHARS_PER_ESTIMATED_TOKEN = 4;
+// Anthropic bills each PDF page as its text plus a page image: 1,500-3,000
+// tokens. Use the top of that range so a long PDF cannot look like it fits.
+const PDF_ESTIMATED_TOKENS_PER_PAGE = 3000;
+const PDF_PAGE_PATTERN = /\/Type\s*\/Page(?![a-zA-Z])/g;
+const PDF_OBJECT_STREAM_START = /\/Type\s*\/ObjStm[\s\S]*?stream\r?\n/g;
+// Object streams hold small dictionaries. The cap is total inflated bytes per
+// PDF, so a compression bomb in a request costs at most this much memory.
+const PDF_OBJECT_STREAM_INFLATE_BUDGET = 4 * 1024 * 1024;
+
+function estimatedPdfPages(base64: string): number {
+  const bytes = Buffer.from(base64, "base64");
+  // latin1 maps each byte to one character, so string offsets are byte offsets.
+  const pdf = bytes.toString("latin1");
+  const visible = pdf.match(PDF_PAGE_PATTERN)?.length ?? 0;
+  if (visible > 0) return visible;
+  // PDF 1.5+ may keep every page dictionary in compressed object streams.
+  let hidden = 0;
+  let budget = PDF_OBJECT_STREAM_INFLATE_BUDGET;
+  for (const match of pdf.matchAll(PDF_OBJECT_STREAM_START)) {
+    const start = match.index + match[0].length;
+    const end = pdf.indexOf("endstream", start);
+    if (end === -1 || budget <= 0) break;
+    try {
+      const objects = inflateSync(bytes.subarray(start, end), {
+        finishFlush: zlibConstants.Z_SYNC_FLUSH,
+        maxOutputLength: budget,
+      });
+      budget -= objects.length;
+      hidden += objects.toString("latin1").match(PDF_PAGE_PATTERN)?.length ?? 0;
+    } catch (error) {
+      // Over budget: stop. Another filter or bad data: count what is readable.
+      if (error instanceof RangeError) break;
+    }
+  }
+  return Math.max(1, hidden);
+}
 
 // Cost and cache-lookback estimates are candidate-specific because compatible
 // continuation state and provider serialization can change what each model
@@ -130,7 +168,7 @@ export function promptMetrics(
       0,
     ),
     serializedChars:
-      stableStringify(prompt).length + accountingInput.imageEstimatedChars,
+      stableStringify(prompt).length + accountingInput.attachmentEstimatedChars,
   };
 }
 
@@ -214,14 +252,22 @@ function providerVisibleTools(
 
 function promptAccountingInput(request: InteractionRequest): {
   request: InteractionRequest;
-  imageEstimatedChars: number;
+  attachmentEstimatedChars: number;
 } {
-  let imageEstimatedChars = 0;
-  const sanitizeContent = (parts: InteractionContent[]) =>
+  let attachmentEstimatedChars = 0;
+  // Attachments are estimated rather than serialized: their base64 is not
+  // what the provider bills, and a PDF must not reach the serializer of a
+  // model that cannot carry it.
+  const sanitizeContent = (parts: InteractionContent[]): InteractionContent[] =>
     parts.map((part) => {
+      if (part.type === "document") {
+        attachmentEstimatedChars +=
+          estimatedPdfPages(part.data) * PDF_ESTIMATED_TOKENS_PER_PAGE * CHARS_PER_ESTIMATED_TOKEN;
+        return { type: "text", text: "" };
+      }
       if (part.type !== "image") return part;
       const detail = part.detail ?? "auto";
-      imageEstimatedChars +=
+      attachmentEstimatedChars +=
         IMAGE_ESTIMATED_TOKENS[detail] * CHARS_PER_ESTIMATED_TOKEN;
       return { ...part, url: ACCOUNTING_IMAGE_URL };
     });
@@ -236,6 +282,6 @@ function promptAccountingInput(request: InteractionRequest): {
   });
   return {
     request: { ...request, items },
-    imageEstimatedChars,
+    attachmentEstimatedChars,
   };
 }

package/src/context/routing_context.ts

@@ -10,6 +10,8 @@ import {
   type HardCapability,
   type PreviousDecision,
   type PricingLookup,
+  isReasoningEffort,
+  type ReasoningEffort,
   type SelectorLeaseHistoryEntry,
   type StrategyCandidateResolution,
 } from "@mupt-ai/dari-router/policy-engine";
@@ -18,12 +20,14 @@ import { randomUUID } from "node:crypto";
 import { HttpError } from "../http/errors.js";
 import { effectiveContextWindow } from "../models/context_limits.js";
 import { keySourceForModel } from "../models/execution_provider.js";
-import { requiredCapabilitiesFor } from "../interaction/capabilities.js";
+import { requestHasDocument, requiredCapabilitiesFor } from "../interaction/capabilities.js";
 import { providerIdentityFromModel } from "../interaction/provider_continuation_state.js";
 import { isBackgroundRequest, type InteractionRequest, type ProviderIdentity } from "../interaction/types.js";
 import { requireCandidateModels } from "../models/ids.js";
+import { explicitModelId } from "../protocol/input/router_model.js";
 import { modelMetadataLookup } from "../providers/pi.js";
 import {
+  modelSupportsDocumentInput,
   modelSupportsImageInput,
   modelSupportsHostedWebSearch,
   modelSupportsReasoningConfigurationUpdate,
@@ -152,6 +156,11 @@ export async function buildRoutingContext(args: {
     enabled: args.credentials.image_capability_filtering !== false,
     supportsImageInput: (model) => capabilityLookup(model).supportsImageInput,
   });
+  const documentCapableModels = documentCapableCandidates(
+    imageFilter.candidateModels,
+    requestHasDocument(args.interaction),
+    (model) => modelSupportsDocumentInput(model, metadataLookup(model)),
+  );
   const prefix = derivePrefixContext({
     routerId: args.routerId,
     interaction: args.interaction,
@@ -171,7 +180,7 @@ export async function buildRoutingContext(args: {
       || hit.conversation_id === args.explicitConversationId,
   );
   const prepared = prepareCandidates({
-    candidateModels: imageFilter.candidateModels,
+    candidateModels: documentCapableModels,
     metadataLookup: capabilityLookup,
     requiredCapabilities: imageFilter.requiredCapabilities,
     requestedReasoningEffort: args.interaction.reasoning?.effort,
@@ -308,6 +317,25 @@ function resolveImageCapableCandidates(args: {
   };
 }
 
+// A PDF can only be sent natively, so a conversation carrying one must not
+// route (or hold a lease) to a model that cannot read it.
+function documentCapableCandidates(
+  candidateModels: string[],
+  hasDocument: boolean,
+  supportsDocumentInput: (model: string) => boolean,
+): string[] {
+  if (!hasDocument) return candidateModels;
+  const capable = candidateModels.filter(supportsDocumentInput);
+  if (capable.length === 0) {
+    throw new HttpError(
+      400,
+      "This conversation contains a PDF document, but no enabled model on this router accepts PDF input.",
+      "no_document_capable_candidates",
+    );
+  }
+  return capable;
+}
+
 function derivePrefixContext(args: {
   routerId: string;
   interaction: InteractionRequest;
@@ -335,3 +363,29 @@ function derivePrefixContext(args: {
     };
   }
 }
+
+// An explicit `dari/model/<id>` serves the requested thinking level when the
+// model offers it and otherwise its default level, so the request itself drops
+// an unsupported level before routing and provider validation see it. A strict
+// router never substitutes, so it keeps the level and fails on it.
+export function clampRequestedEffort(
+  interaction: InteractionRequest,
+  credentials: { strict?: boolean; model_thinking_levels?: Record<string, ReasoningEffort[]> },
+  metadataLookup: ReturnType<typeof modelMetadataLookup>,
+): InteractionRequest {
+  const requested = interaction.reasoning?.effort;
+  const modelId = explicitModelId(interaction.model);
+  if (requested === undefined || modelId === null || credentials.strict === true) return interaction;
+  const supported = metadataLookup(modelId).supportedThinkingLevels.filter(isReasoningEffort);
+  const configured = credentials.model_thinking_levels?.[modelId] ?? supported;
+  if (configured.includes(requested) && supported.includes(requested)) return interaction;
+  // The effort implied thinking on and its effort-derived fields (baseEffort,
+  // configuration updates) must not leak into the default. Keep only an explicit
+  // thinking-off and any thinking budget.
+  const { enabled, budgetTokens } = interaction.reasoning!;
+  const reasoning = {
+    ...(enabled === false ? { enabled } : {}),
+    ...(budgetTokens === undefined ? {} : { budgetTokens }),
+  };
+  return { ...interaction, reasoning: Object.keys(reasoning).length === 0 ? undefined : reasoning };
+}

package/src/ingress/anthropic_messages.ts

@@ -5,6 +5,7 @@ import { anthropicToken } from "../http/auth.js";
 import { errorResponseHeaders, normalizeError } from "../http/errors.js";
 import { captureRequestError } from "../observability/errors.js";
 import { anthropicMessagesToInteraction } from "../protocol/input/anthropic_messages.js";
+import { explicitModelId } from "../protocol/input/router_model.js";
 import { handleRoutedCompletion } from "../runtime/routed_completion.js";
 import type { SelectorBackend } from "../strategies/routing.js";
 import { claudeCodeRequestContext } from "./claude_code/request_context.js";
@@ -12,7 +13,7 @@ import {
   claudeCodeConversationIdentity,
 } from "../runtime/harness_identity/index.js";
 import { parseJsonRequest } from "./json.js";
-import type { AnthropicSafeguards } from "../interaction/types.js";
+import type { AnthropicSafeguards, ReasoningControls } from "../interaction/types.js";
 
 type AnthropicErrorType =
   | "authentication_error"
@@ -42,6 +43,7 @@ export async function handleAnthropicMessages(
     const payload = isClaudeCode
       ? {
           ...forwarded,
+          reasoning: routedReasoning(forwarded),
           agentContext: claudeCodeRequestContext(request.headers),
           // Native tool results may be followed by harness-injected text.
           // Preserve their structural continuation semantics, not prompt inference.
@@ -79,6 +81,18 @@ export async function handleAnthropicMessages(
   }
 }
 
+// Claude Code sends its session effort on every turn. That preference
+// describes a direct model, not a heterogeneous router, so let Dari select the
+// effort when the router routes; a `dari/model/<id>` keeps the chosen effort.
+function routedReasoning(
+  interaction: ReturnType<typeof anthropicMessagesToInteraction>,
+): ReasoningControls | undefined {
+  const reasoning = interaction.reasoning;
+  if (reasoning === undefined || explicitModelId(interaction.model) !== null) return reasoning;
+  // Keep only an explicit "thinking off"; effort alone must not force thinking on.
+  return reasoning.enabled === false ? { enabled: false } : undefined;
+}
+
 function latestMessageHasToolResult(messages: unknown): boolean {
   if (!Array.isArray(messages)) return false;
   const message = messages.at(-1);

package/src/ingress/codex.ts

@@ -35,12 +35,15 @@ export async function handleResponses(
       ...(museClientId === undefined ? {} : { museCodeClientId: museClientId }),
     };
     const codexIdentity = codexConversationIdentity(request.headers, rawPayload);
-    // Codex applies its global model_reasoning_effort setting to custom model
+    // Codex and Muse Code apply their session effort to custom model
     // providers. That preference describes a direct model, not a heterogeneous
-    // router, so let Dari select the reasoning effort for verified Codex turns.
-    // Parsing above still validates malformed Responses reasoning controls.
-    if (codexIdentity !== undefined && explicitModelId(payload.model) === null) {
-      payload.reasoning = undefined;
+    // router, so let Dari select the reasoning effort for their routed turns;
+    // a `dari/model/<id>` keeps the chosen effort. Parsing above still
+    // validates malformed Responses reasoning controls.
+    const harness = codexIdentity !== undefined || museClientId !== undefined;
+    if (harness && explicitModelId(payload.model) === null) {
+      // Keep only an explicit thinking-off; a session effort must not force thinking on.
+      payload.reasoning = payload.reasoning?.enabled === false ? { enabled: false } : undefined;
     }
     const harnessConversationIdentity = codexIdentity
       ?? (museClientId === undefined ? undefined : museCodeConversationIdentity(request.headers));

package/src/interaction/capabilities.ts

@@ -24,6 +24,13 @@ function requiresWebSearch(item: InteractionItem): boolean {
   return item.type === "hosted_tool_call" && item.tool === "web_search";
 }
 
+export function requestHasDocument(request: Pick<InteractionRequest, "items">): boolean {
+  return request.items.some((item) => {
+    const parts = item.type === "message" ? item.content : item.type === "tool_result" ? item.output : null;
+    return Array.isArray(parts) && parts.some((part) => part.type === "document");
+  });
+}
+
 function itemHasImage(item: InteractionItem): boolean {
   if (item.type === "message") return item.content.some((part) => part.type === "image");
   return item.type === "tool_result" && Array.isArray(item.output) && item.output.some((part) => part.type === "image");