@mupt-ai/dari-gateway 0.2.0 → 0.3.0

package/src/protocol/input/anthropic_messages.ts

@@ -1,11 +1,11 @@
 import { isReasoningEffort, isRecord, type ReasoningEffort } from "@mupt-ai/dari-router/policy-engine";
 
-import { HttpError } from "../../http/errors";
-import { parseDataImageUrl } from "../../interaction/image";
-import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state";
-import { decodeUntrustedPortableReasoningState } from "../../interaction/portable_reasoning_state";
-import { toolsByWireName } from "../../interaction/tool_identity";
-import { explicitModelId, isAnthropicRouterModel, unsupportedModelMessage } from "./router_model";
+import { HttpError } from "../../http/errors.js";
+import { parseDataImageUrl } from "../../interaction/image.js";
+import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import { decodeUntrustedPortableReasoningState } from "../../interaction/portable_reasoning_state.js";
+import { toolsByWireName } from "../../interaction/tool_identity.js";
+import { explicitModelId, isAnthropicRouterModel, unsupportedModelMessage } from "./router_model.js";
 import type {
   ClientTool,
   InteractionContent,
@@ -17,7 +17,7 @@ import type {
   InteractionToolReference,
   ReasoningControls,
   ServerToolKind,
-} from "../../interaction/types";
+} from "../../interaction/types.js";
 
 const SUPPORTED_SERVER_TOOL_TYPES: readonly ServerToolKind[] = [
   "bash_20250124",

package/src/protocol/input/chat_completions.ts

@@ -5,7 +5,7 @@ import {
   type ResponseFormat,
 } from "@mupt-ai/dari-router/policy-engine";
 
-import { HttpError } from "../../http/errors";
+import { HttpError } from "../../http/errors.js";
 import type {
   InteractionContent,
   InteractionItem,
@@ -14,10 +14,10 @@ import type {
   InteractionResponseFormat,
   InteractionTool,
   InteractionToolChoice,
-} from "../../interaction/types";
-import { toolsByWireName } from "../../interaction/tool_identity";
-import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state";
-import { validateChatCompletionsRequest } from "./chat_completions_validation";
+} from "../../interaction/types.js";
+import { toolsByWireName } from "../../interaction/tool_identity.js";
+import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import { validateChatCompletionsRequest } from "./chat_completions_validation.js";
 
 export function chatCompletionsToInteraction(request: ChatCompletionRequest): InteractionRequest {
   validateChatCompletionsRequest(request);
@@ -80,7 +80,9 @@ export function internalChatCompletionsToInteraction(request: ChatCompletionRequ
   const tools = chatTools(request.tools);
   toolsByWireName(tools);
   const responseFormat = toInteractionResponseFormat(request.response_format);
-  const reasoningEffort = request.reasoning_effort;
+  // OpenAI spells the non-reasoning effort "none"; Dari spells it "off".
+  const requestedEffort: string | undefined = request.reasoning_effort;
+  const reasoningEffort = requestedEffort === "none" ? "off" : request.reasoning_effort;
   return {
     sourceProtocol: request.source_protocol ?? "openai_chat",
     model: request.model,

package/src/protocol/input/chat_completions_validation.ts

@@ -1,12 +1,12 @@
 import { isReasoningEffort, isRecord, type ChatCompletionRequest } from "@mupt-ai/dari-router/policy-engine";
 
-import { HttpError } from "../../http/errors";
-import { parseDataImageUrl } from "../../interaction/image";
-import { CANONICAL_ROUTER_MODEL, explicitModelId, unsupportedModelMessage } from "./router_model";
+import { HttpError } from "../../http/errors.js";
+import { parseDataImageUrl } from "../../interaction/image.js";
+import { CANONICAL_ROUTER_MODEL, explicitModelId, unsupportedModelMessage } from "./router_model.js";
 import {
   isProtocolToolCallId,
   TOOL_CALL_ID_MAX_LENGTH,
-} from "../../interaction/tool_identity";
+} from "../../interaction/tool_identity.js";
 
 export function validateChatCompletionsRequest(request: ChatCompletionRequest): void {
   if (request.model !== CANONICAL_ROUTER_MODEL && explicitModelId(request.model) === null) {
@@ -214,7 +214,8 @@ function validateReasoning(request: ChatCompletionRequest): void {
       "reasoning"
     );
   }
-  if (request.reasoning_effort !== undefined && !isReasoningEffort(request.reasoning_effort)) {
+  const effort: unknown = request.reasoning_effort;
+  if (effort !== undefined && effort !== "none" && !isReasoningEffort(effort)) {
     throw new HttpError(
       400,
       "reasoning_effort is not supported.",

package/src/protocol/input/openai_responses.ts

@@ -2,15 +2,15 @@ import { createHash } from "node:crypto";
 
 import { isReasoningEffort, isRecord, type ReasoningEffort } from "@mupt-ai/dari-router/policy-engine";
 
-import { HttpError } from "../../http/errors";
-import { parseDataImageUrl } from "../../interaction/image";
-import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state";
-import { CANONICAL_ROUTER_MODEL, explicitModelId, unsupportedModelMessage } from "./router_model";
+import { HttpError } from "../../http/errors.js";
+import { parseDataImageUrl } from "../../interaction/image.js";
+import { decodeUntrustedProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import { CANONICAL_ROUTER_MODEL, explicitModelId, unsupportedModelMessage } from "./router_model.js";
 import {
   isProtocolToolCallId,
   TOOL_CALL_ID_MAX_LENGTH,
   toolsByWireName,
-} from "../../interaction/tool_identity";
+} from "../../interaction/tool_identity.js";
 import type {
   InteractionContent,
   InteractionItem,
@@ -20,7 +20,7 @@ import type {
   InteractionTool,
   InteractionToolChoice,
   ToolKind,
-} from "../../interaction/types";
+} from "../../interaction/types.js";
 
 const UNSUPPORTED_INPUT_ITEM_TYPES = new Set([
   "additional_tools",

package/src/protocol/output/anthropic_messages.ts

@@ -1,13 +1,13 @@
 import type { RoutingDecision } from "@mupt-ai/dari-router/policy-engine";
 
-import type { InteractionOutput, InteractionRequest } from "../../interaction/types";
+import type { InteractionOutput, InteractionRequest } from "../../interaction/types.js";
 import {
   anthropicContent,
   anthropicMessageId,
   anthropicRouting,
   anthropicStopReason,
   anthropicUsage,
-} from "./anthropic_messages_output";
+} from "./anthropic_messages_output.js";
 
 export function anthropicMessagesResponse(args: {
   id: string;
@@ -27,6 +27,6 @@ export function anthropicMessagesResponse(args: {
     stop_sequence: null,
     usage: anthropicUsage(args.output),
     ...(args.output.safeguardResults ? { safeguard_results: args.output.safeguardResults } : {}),
-    dari_routing: anthropicRouting(args.routing, args.conversationId),
+    dari_routing: anthropicRouting(args.routing, args.id, args.conversationId),
   };
 }

package/src/protocol/output/anthropic_messages_output.ts

@@ -1,7 +1,7 @@
-import { parseDataImageUrl } from "../../interaction/image";
-import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state";
-import { encodePortableReasoningState } from "../../interaction/portable_reasoning_state";
-import type { InteractionOutput } from "../../interaction/types";
+import { parseDataImageUrl } from "../../interaction/image.js";
+import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import { encodePortableReasoningState } from "../../interaction/portable_reasoning_state.js";
+import type { InteractionOutput } from "../../interaction/types.js";
 import type { RoutingDecision } from "@mupt-ai/dari-router/policy-engine";
 
 export function anthropicMessageId(requestId: string): string {
@@ -100,6 +100,7 @@ export function anthropicUsage(output: Pick<InteractionOutput, "usage">): Record
 
 export function anthropicRouting(
   decision: RoutingDecision,
+  requestId: string,
   conversationId?: string
 ): Record<string, unknown> {
   return {
@@ -107,6 +108,7 @@ export function anthropicRouting(
     selected_model: decision.selectedModel,
     reasoning_effort: decision.reasoningEffort,
     reason: decision.reason,
+    request_id: requestId,
     ...(conversationId ? { conversation_id: conversationId } : {}),
     ...(decision.leaseTurnsRemaining === undefined
       ? {}

package/src/protocol/output/anthropic_messages_stream.ts

@@ -1,20 +1,20 @@
-import { piMessageToInteractionOutput } from "../../providers/pi_output";
-import { providerToolCallIdentity } from "../../interaction/tool_identity";
-import type { InteractionOutput, InteractionRequest, ProviderIdentity } from "../../interaction/types";
+import { piMessageToInteractionOutput } from "../../providers/pi_output.js";
+import { providerToolCallIdentity } from "../../interaction/tool_identity.js";
+import type { InteractionOutput, InteractionRequest, ProviderIdentity } from "../../interaction/types.js";
 import {
   isProviderFailureMessage,
   providerFailureFromEvent,
   providerFailureMessage,
   type ProviderStreamOutcome,
-} from "../../runtime/provider_stream_outcome";
-import { sseStream } from "./sse_stream";
+} from "../../runtime/provider_stream_outcome.js";
+import { sseStream } from "./sse_stream.js";
 import {
   anthropicMessageId,
   anthropicReasoningBlock,
   anthropicRouting,
   anthropicStopReason,
   anthropicUsage,
-} from "./anthropic_messages_output";
+} from "./anthropic_messages_output.js";
 
 export function anthropicMessagesSseStream(args: {
   id: string;
@@ -58,7 +58,7 @@ export function anthropicMessagesSseStream(args: {
           stop_reason: null,
           stop_sequence: null,
           usage: { input_tokens: 0, output_tokens: 0 },
-          dari_routing: anthropicRouting(args.routing, args.conversationId),
+          dari_routing: anthropicRouting(args.routing, args.id, args.conversationId),
         },
       });
 

package/src/protocol/output/chat_completions.ts

@@ -1,17 +1,17 @@
 import { isRecord, type ChatMessage, type RoutingDecision } from "@mupt-ai/dari-router/policy-engine";
 
-import { piMessageToInteractionOutput } from "../../providers/pi_output";
-import { providerToolCallIdentity } from "../../interaction/tool_identity";
-import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state";
-import { chatReasoningDetailId } from "./chat_reasoning_detail";
-import type { InteractionOutput, InteractionRequest, ProviderContinuationState, ProviderIdentity } from "../../interaction/types";
+import { piMessageToInteractionOutput } from "../../providers/pi_output.js";
+import { providerToolCallIdentity } from "../../interaction/tool_identity.js";
+import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import { chatReasoningDetailId } from "./chat_reasoning_detail.js";
+import type { InteractionOutput, InteractionRequest, ProviderContinuationState, ProviderIdentity } from "../../interaction/types.js";
 import {
   isProviderFailureMessage,
   providerFailureFromEvent,
   providerFailureMessage,
   type ProviderStreamOutcome,
-} from "../../runtime/provider_stream_outcome";
-import { sseStream } from "./sse_stream";
+} from "../../runtime/provider_stream_outcome.js";
+import { sseStream } from "./sse_stream.js";
 
 export function chatCompletionResponse(args: {
   id: string;
@@ -34,7 +34,7 @@ export function chatCompletionResponse(args: {
       },
     ],
     usage: usagePayload(args.output, args.modelCost),
-    dari_routing: routingPayload(args.routing, args.conversationId),
+    dari_routing: routingPayload(args.routing, args.id, args.conversationId),
   };
 }
 
@@ -75,7 +75,7 @@ export function chatCompletionsSseStream(args: {
           usage: usagePayload(message, args.modelCost?.(message, true)) });
       };
 
-      send(chunk({ completion, created, model: args.model, delta: {}, finishReason: null, routing: args.routing, conversationId: args.conversationId }));
+      send(chunk({ completion, created, model: args.model, delta: {}, finishReason: null, routing: routingPayload(args.routing, args.id, args.conversationId) }));
       send(chunk({ completion, created, model: args.model, delta: { role: "assistant" }, finishReason: null }));
 
       const argumentDeltaSeen = new Set<string>();
@@ -208,12 +208,13 @@ export function chatCompletionsSseStream(args: {
   });
 }
 
-function routingPayload(decision: RoutingDecision, conversationId?: string): Record<string, unknown> {
+function routingPayload(decision: RoutingDecision, requestId: string, conversationId?: string): Record<string, unknown> {
   return {
     requested_model: "dari/routing",
     selected_model: decision.selectedModel,
     reasoning_effort: decision.reasoningEffort,
     reason: decision.reason,
+    request_id: requestId,
     ...(conversationId ? { conversation_id: conversationId } : {}),
   };
 }
@@ -224,9 +225,8 @@ function chunk(args: {
   model: string;
   delta: Record<string, unknown>;
   finishReason: string | null;
-  routing?: RoutingDecision;
+  routing?: Record<string, unknown>;
   usage?: Record<string, unknown>;
-  conversationId?: string;
 }): Record<string, unknown> {
   return {
     id: args.completion,
@@ -234,7 +234,7 @@ function chunk(args: {
     created: args.created,
     model: args.model,
     choices: [{ index: 0, delta: args.delta, finish_reason: args.finishReason }],
-    ...(args.routing ? { dari_routing: routingPayload(args.routing, args.conversationId) } : {}),
+    ...(args.routing ? { dari_routing: args.routing } : {}),
     ...(args.usage ? { usage: args.usage } : {}),
   };
 }

package/src/protocol/output/chat_reasoning_detail.ts

@@ -1,4 +1,4 @@
-import type { ProviderContinuationState } from "../../interaction/types";
+import type { ProviderContinuationState } from "../../interaction/types.js";
 
 export function chatReasoningDetailId(
   continuation: ProviderContinuationState,

package/src/protocol/output/openai_responses.ts

@@ -4,8 +4,8 @@ import {
   openAIResponsesRouting,
   openAIResponsesTerminal,
   openAIResponsesUsage,
-} from "./openai_responses_output";
-import type { InteractionOutput, InteractionRequest } from "../../interaction/types";
+} from "./openai_responses_output.js";
+import type { InteractionOutput, InteractionRequest } from "../../interaction/types.js";
 
 export function openAIResponsesResponse(args: {
   id: string;
@@ -24,7 +24,7 @@ export function openAIResponsesResponse(args: {
     output: openAIResponsesOutputItems(args.output),
     usage: openAIResponsesUsage(args.output),
     parallel_tool_calls: args.request.parallelToolCalls ?? true,
-    dari_routing: openAIResponsesRouting(args.routing, args.conversationId),
+    dari_routing: openAIResponsesRouting(args.routing, args.id, args.conversationId),
   };
 }
 import type { RoutingDecision } from "@mupt-ai/dari-router/policy-engine";

package/src/protocol/output/openai_responses_output.ts

@@ -1,5 +1,5 @@
-import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state";
-import type { InteractionOutput } from "../../interaction/types";
+import { encodeProviderContinuationState } from "../../interaction/provider_continuation_state.js";
+import type { InteractionOutput } from "../../interaction/types.js";
 
 export function openAIResponsesResponseId(requestId: string): string {
   return `resp_${requestId}`;
@@ -44,6 +44,7 @@ export function openAIResponsesUsage(output: Pick<InteractionOutput, "usage">):
 
 export function openAIResponsesRouting(
   decision: RoutingDecision,
+  requestId: string,
   conversationId?: string
 ): Record<string, unknown> {
   return {
@@ -51,6 +52,7 @@ export function openAIResponsesRouting(
     selected_model: decision.selectedModel,
     reasoning_effort: decision.reasoningEffort,
     reason: decision.reason,
+    request_id: requestId,
     ...(conversationId ? { conversation_id: conversationId } : {}),
   };
 }

package/src/protocol/output/openai_responses_stream.ts

@@ -5,21 +5,21 @@ import {
   openAIResponsesRouting,
   openAIResponsesTerminal,
   openAIResponsesUsage,
-} from "./openai_responses_output";
-import { piMessageToInteractionOutput } from "../../providers/pi_output";
+} from "./openai_responses_output.js";
+import { piMessageToInteractionOutput } from "../../providers/pi_output.js";
 import type {
   InteractionOutput,
   InteractionOutputItem,
   InteractionRequest,
   ProviderIdentity,
-} from "../../interaction/types";
+} from "../../interaction/types.js";
 import {
   isProviderFailureMessage,
   providerFailureFromEvent,
   providerFailureMessage,
   type ProviderStreamOutcome,
-} from "../../runtime/provider_stream_outcome";
-import { sseStream } from "./sse_stream";
+} from "../../runtime/provider_stream_outcome.js";
+import { sseStream } from "./sse_stream.js";
 
 type StreamArgs = {
   id: string;
@@ -600,7 +600,7 @@ function responseEnvelope(
     output,
     parallel_tool_calls: args.request.parallelToolCalls ?? true,
     ...(interaction ? { usage: openAIResponsesUsage(interaction) } : {}),
-    dari_routing: openAIResponsesRouting(args.routing, args.conversationId),
+    dari_routing: openAIResponsesRouting(args.routing, args.id, args.conversationId),
   };
 }
 

package/src/protocol/output/sse_stream.ts

@@ -1,11 +1,11 @@
-import { ProviderStreamLifecycle } from "../../runtime/provider_stream_lifecycle";
+import { ProviderStreamLifecycle } from "../../runtime/provider_stream_lifecycle.js";
 import {
   providerFailure,
   providerFailureFromCaught,
   providerFailureFromTerminal,
   type ProviderStreamOutcome,
-} from "../../runtime/provider_stream_outcome";
-import { DeliveredStreamSnapshot } from "./delivered_stream_snapshot";
+} from "../../runtime/provider_stream_outcome.js";
+import { DeliveredStreamSnapshot } from "./delivered_stream_snapshot.js";
 
 export type SseStreamLifecycleArgs = {
   onFinal?: (outcome: ProviderStreamOutcome) => void;

package/src/providers/anthropic_safeguards.ts

@@ -1,7 +1,7 @@
 import { isRecord } from "@mupt-ai/dari-router/policy-engine";
 
-import type { AnthropicSafeguards, InteractionRequest } from "../interaction/types";
-import type { ProviderFetch } from "./response_capture";
+import type { AnthropicSafeguards, InteractionRequest } from "../interaction/types.js";
+import type { ProviderFetch } from "./response_capture.js";
 
 // Claude Code's auto mode asks the Anthropic API to run its dangerous-tool-use
 // classifier server-side, at no charge, by sending `safeguards` with a paired

package/src/providers/aws_assume_role.ts

@@ -1,12 +1,12 @@
 import { AssumeRoleCommand, STSClient } from "@aws-sdk/client-sts";
 
-import { HttpError } from "../http/errors";
+import { HttpError } from "../http/errors.js";
 import type {
   AwsAssumeRoleProviderAuth,
   AwsSigV4ProviderAuth,
   ProviderAuth,
   ResolvedProviderAuth,
-} from "../types";
+} from "../types.js";
 
 const ROLE_SESSION_NAME = "dari-routing";
 // AssumeRole always goes through this regional STS endpoint. The customer's

package/src/providers/cancel_estimate.ts

@@ -8,7 +8,7 @@ import type {
   InteractionOutput,
   InteractionUsage,
   ProviderIdentity,
-} from "../interaction/types";
+} from "../interaction/types.js";
 
 /** Prose from current models averages about 4.8 characters per token. */
 export const ESTIMATED_CHARS_PER_TOKEN = 4.8;

package/src/providers/execution.ts

@@ -1,27 +1,27 @@
 import type { RoutingDecision } from "@mupt-ai/dari-router/policy-engine";
-import { HttpError } from "../http/errors";
-import { piMessageToInteractionOutput } from "./pi_output";
-import type { InteractionRequest } from "../interaction/types";
-import { captureOperationalError } from "../observability/errors";
+import { HttpError } from "../http/errors.js";
+import { piMessageToInteractionOutput } from "./pi_output.js";
+import type { InteractionRequest } from "../interaction/types.js";
+import { captureOperationalError } from "../observability/errors.js";
 import {
   CustomerDisconnectedError,
   customerCanceledProviderMessage,
-} from "../runtime/customer_disconnect";
+} from "../runtime/customer_disconnect.js";
 import {
   isProviderFailureMessage,
   providerFailureFromCaught,
   providerFailureFromTerminal,
   providerFailureMessage,
-} from "../runtime/provider_stream_outcome";
+} from "../runtime/provider_stream_outcome.js";
 import {
   reportedExecution,
   type ReportableProviderOutcome,
   type RoutedRuntimeReporter,
-} from "../runtime/runtime_reporter";
-import type { RoutedProtocol } from "../runtime/routed_protocol";
-import { trackBackgroundTelemetry } from "../telemetry/background";
-import { gatewayPlatform, type SubscriptionHold } from "../platform/platform";
-import { reportRoutingUsage } from "../telemetry/usage";
+} from "../runtime/runtime_reporter.js";
+import type { RoutedProtocol } from "../runtime/routed_protocol.js";
+import { trackBackgroundTelemetry } from "../telemetry/background.js";
+import { gatewayPlatform, type SubscriptionHold } from "../platform/platform.js";
+import { reportRoutingUsage } from "../telemetry/usage.js";
 import {
   prepareProviderExecution,
   primeProviderStream,
@@ -32,20 +32,20 @@ import {
   type ProviderExecution,
   type StartedCompletion,
   type StartedStream,
-} from "./provider_call";
+} from "./provider_call.js";
 import {
   runProviderAttemptPlan,
   type ProviderAttemptFailure,
   type ProviderAttemptPlanResult,
   type ProviderAttemptResult,
-} from "./provider_attempt_runner";
+} from "./provider_attempt_runner.js";
 import {
   estimatedCancelUsage,
   NO_DELIVERED_OUTPUT,
   observeDeliveredOutput,
-} from "./cancel_estimate";
-import { providerResponseCaptureFromError, type ProviderResponseCapture } from "./response_capture";
-import { createStreamFinalizer } from "./stream_finalization";
+} from "./cancel_estimate.js";
+import { providerResponseCaptureFromError, type ProviderResponseCapture } from "./response_capture.js";
+import { createStreamFinalizer } from "./stream_finalization.js";
 import {
   detectSubscriptionUsageLimit,
   failureWithSubscriptionUsageLimit,
@@ -53,7 +53,7 @@ import {
   subscriptionUsageLimitFromFailure,
   subscriptionSwitchedMessage,
   subscriptionUsageLimitMessage,
-} from "./subscription_limits";
+} from "./subscription_limits.js";
 
 export type ProviderChoice = {
   decision: RoutingDecision;

package/src/providers/pi.ts

@@ -26,35 +26,41 @@ import {
   type ReasoningEffort,
   type RouterModelPrice,
 } from "@mupt-ai/dari-router/policy-engine";
-import { HttpError } from "../http/errors";
-import { effectiveContextWindow } from "../models/context_limits";
-import { defaultBaseUrl } from "../models/provider_defaults";
-import { contextWindowOverride, inputModalitiesOverride, runtimeCompat } from "../models/runtime_overrides";
-import { BEDROCK_API } from "./model_capabilities";
-import { buildContext } from "./pi_context";
-import { providerIdentityFromModel } from "../interaction/provider_continuation_state";
-import { buildOptions, supportedThinkingLevels } from "./pi_options";
-import type { ProviderFetch } from "./response_capture";
+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 type { ProviderFetch } from "./response_capture.js";
 import {
   eventsWithSafeguardResults,
   forwardsAnthropicSafeguards,
   messageWithSafeguardResults,
   safeguardResultsTap,
-} from "./anthropic_safeguards";
+} from "./anthropic_safeguards.js";
 import {
   eventsWithProviderReportedCost,
   messageWithProviderReportedCost,
   providerReportedCostTap,
   reportsProviderCost,
-} from "./provider_reported_cost";
+} from "./provider_reported_cost.js";
 import {
   eventsWithSubscriptionCacheWrites,
   messageWithSubscriptionCacheWrites,
   modelPricingFromPiModel,
   type SubscriptionUsageContext,
-} from "./subscription_usage";
-import type { InteractionRequest, ProviderIdentity } from "../interaction/types";
-import type { ResolvedProviderAuth, ResolvedRouterModelConfig } from "../types";
+} from "./subscription_usage.js";
+import type { InteractionRequest, ProviderIdentity } from "../interaction/types.js";
+import type { ResolvedProviderAuth, ResolvedRouterModelConfig } from "../types.js";
 
 type PiAi = {
   completeSimple: (model: Record<string, unknown>, context: Record<string, unknown>, options: Record<string, unknown>) => Promise<Record<string, unknown>>;
@@ -335,16 +341,21 @@ export function buildModel(
   // 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.
+  // 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 && configuredThinkingLevels == null && registered === undefined
+    (config !== undefined || providerThinking !== undefined) &&
+    configuredThinkingLevels == null &&
+    registered === undefined
       ? REASONING_EFFORTS
       : undefined;
   const effectiveThinkingLevels = Array.isArray(configuredThinkingLevels)
     ? configuredThinkingLevels
     : fallbackThinkingLevels;
   const configuredThinking = effectiveThinkingLevels
-    ? thinkingMetadata(effectiveThinkingLevels, registered)
+    ? thinkingMetadata(effectiveThinkingLevels, registered ?? { thinkingLevelMap: providerThinking })
     : null;
   const defaultThinkingLevel = config?.default_thinking_level ?? undefined;
   const api = config?.api ?? registered?.api ?? defaultApi(provider);
@@ -415,7 +426,10 @@ function resolveModelMetadata(
   nativeId: string
 ): Record<string, unknown> | undefined {
   const registry = registryProvider(provider, modelId);
-  return pi.getModel?.(registry, nativeId) ?? pi.getModel?.(registry, 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
@@ -491,7 +505,8 @@ function executionPricing(
 }
 
 function defaultApi(provider: string): string {
-  if (provider === "openai" || provider === "xai") return "openai-responses";
+  // 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";

package/src/providers/pi_context.ts

@@ -1,7 +1,7 @@
 import { encodeOpenAIReasoningSignature } from "@mupt-ai/dari-router/protocols";
 import { isRecord } from "@mupt-ai/dari-router/policy-engine";
 
-import { HttpError } from "../http/errors";
+import { HttpError } from "../http/errors.js";
 import type {
   InteractionContent,
   InteractionItem,
@@ -9,14 +9,14 @@ import type {
   InteractionTool,
   ProviderContinuationState,
   ProviderIdentity,
-} from "../interaction/types";
-import { parseDataImageUrl } from "../interaction/image";
-import { providerToolCallId, toolWireName } from "../interaction/tool_identity";
+} from "../interaction/types.js";
+import { parseDataImageUrl } from "../interaction/image.js";
+import { providerToolCallId, toolWireName } from "../interaction/tool_identity.js";
 import {
   compatibleProviderContinuation,
   sameProviderIdentity,
-} from "../interaction/provider_continuation_state";
-import { isOpenAIResponsesApi } from "./model_capabilities";
+} from "../interaction/provider_continuation_state.js";
+import { isOpenAIResponsesApi } from "./model_capabilities.js";
 
 const UNKNOWN_SOURCE: ProviderIdentity = {
   provider: "dari",