@khanglvm/llm-router 2.3.1 → 2.3.2

package/src/runtime/handler/fallback.js

@@ -44,6 +44,18 @@ const POLICY_HINTS = [
   "unsafe",
   "flagged"
 ];
+const MODEL_NOT_FOUND_HINTS = [
+  "model not found",
+  "model does not exist",
+  "model_not_found"
+];
+const VRAM_EXHAUSTION_HINTS = [
+  "insufficient vram",
+  "out of memory",
+  "failed to load model",
+  "insufficient memory"
+];
+const DEFAULT_ORIGIN_MODEL_NOT_FOUND_COOLDOWN_MS = 60 * 60_000;
 const CONTEXT_WINDOW_HINTS = [
   "context window",
   "maximum context length",
@@ -56,6 +68,17 @@ const CONTEXT_WINDOW_HINTS = [
   "too many tokens",
   "ran out of room in the model's context window"
 ];
+const RATE_LIMIT_HINTS = [
+  "tokens per minute",
+  "requests per minute",
+  "rate limit",
+  "rate_limit",
+  "tpm",
+  "rpm",
+  "quota exceeded",
+  "quota_exceeded",
+  "limit exceeded"
+];
 const fallbackCircuitState = new Map();
 
 export function shouldRetryStatus(status) {
@@ -392,6 +415,16 @@ export async function classifyFailureResult(result, retryPolicy) {
   }
 
   if (status === 404 || status === 410) {
+    const hintText404 = await readProviderErrorHint(result);
+    if (hasAnyHint(hintText404, MODEL_NOT_FOUND_HINTS)) {
+      return {
+        category: "model_not_found",
+        retryable: false,
+        retryOrigin: false,
+        allowFallback: true,
+        originCooldownMs: DEFAULT_ORIGIN_MODEL_NOT_FOUND_COOLDOWN_MS
+      };
+    }
     return {
       category: "not_found",
       retryable: false,
@@ -412,9 +445,47 @@ export async function classifyFailureResult(result, retryPolicy) {
         originCooldownMs: 0
       };
     }
+    if (hasAnyHint(hintText, VRAM_EXHAUSTION_HINTS)) {
+      return {
+        category: "vram_exhaustion",
+        retryable: false,
+        retryOrigin: false,
+        allowFallback: true,
+        originCooldownMs: retryPolicy.originFallbackCooldownMs
+      };
+    }
+    if (status === 413 && hasAnyHint(hintText, RATE_LIMIT_HINTS)) {
+      const rateLimitCooldown = retryAfterMs > 0 ? retryAfterMs : retryPolicy.originRateLimitCooldownMs;
+      return {
+        category: "rate_limited",
+        retryable: true,
+        retryOrigin: false,
+        allowFallback: true,
+        originCooldownMs: rateLimitCooldown
+      };
+    }
   }
 
   if (status === 408 || status === 409 || status >= 500) {
+    const hintText5xx = await readProviderErrorHint(result);
+    if (hasAnyHint(hintText5xx, VRAM_EXHAUSTION_HINTS)) {
+      return {
+        category: "vram_exhaustion",
+        retryable: false,
+        retryOrigin: false,
+        allowFallback: true,
+        originCooldownMs: retryPolicy.originFallbackCooldownMs
+      };
+    }
+    if (hasAnyHint(hintText5xx, MODEL_NOT_FOUND_HINTS)) {
+      return {
+        category: "model_not_found",
+        retryable: false,
+        retryOrigin: false,
+        allowFallback: true,
+        originCooldownMs: DEFAULT_ORIGIN_MODEL_NOT_FOUND_COOLDOWN_MS
+      };
+    }
     return {
       category: "temporary_error",
       retryable: true,

package/src/runtime/handler/field-filter.js

@@ -0,0 +1,39 @@
+/**
+ * Strips request body fields the target model doesn't support.
+ * Only acts when a capability is explicitly `false` — undefined means "pass through".
+ *
+ * @param {object} providerBody - Request body (already cloned upstream)
+ * @param {object} [capabilities] - Model capabilities from config
+ * @returns {object} The providerBody with unsupported fields deleted
+ */
+export function stripUnsupportedFields(providerBody, capabilities) {
+  if (!capabilities || typeof capabilities !== "object") return providerBody;
+
+  if (capabilities.supportsReasoning === false) {
+    delete providerBody.reasoning_effort;
+    delete providerBody.reasoning;
+  }
+  if (capabilities.supportsThinking === false) {
+    delete providerBody.thinking;
+  }
+  if (capabilities.supportsResponseFormat === false) {
+    delete providerBody.response_format;
+  }
+  if (capabilities.supportsLogprobs === false) {
+    delete providerBody.logprobs;
+    delete providerBody.top_logprobs;
+  }
+  if (capabilities.supportsServiceTier === false) {
+    delete providerBody.service_tier;
+  }
+  if (capabilities.supportsPrediction === false) {
+    delete providerBody.prediction;
+    delete providerBody.predicted_output;
+  }
+  if (capabilities.supportsStreamOptions === false) {
+    delete providerBody.stream_options;
+  }
+
+  return providerBody;
+}
+

package/src/runtime/handler/large-request-log.js

@@ -0,0 +1,211 @@
+const DEFAULT_TEXT_ENCODER = new TextEncoder();
+
+export const LARGE_REQUEST_LOG_ENABLED_ENV = "LLM_ROUTER_LOG_LARGE_REQUESTS";
+export const LARGE_REQUEST_LOG_THRESHOLD_ENV = "LLM_ROUTER_LARGE_REQUEST_LOG_THRESHOLD_BYTES";
+export const LARGE_REQUEST_LOG_PATH_ENV = "LLM_ROUTER_LARGE_REQUEST_LOG_PATH";
+export const DEFAULT_LARGE_REQUEST_LOG_THRESHOLD_BYTES = 20 * 1024 * 1024;
+const LARGE_STRING_HINT_THRESHOLD_BYTES = 256 * 1024;
+const MAX_LARGE_STRING_HINTS = 8;
+const MAX_SUMMARY_NODES = 50_000;
+
+function toBoolean(value, fallback = false) {
+  if (value === undefined || value === null || value === "") return fallback;
+  if (typeof value === "boolean") return value;
+  const normalized = String(value).trim().toLowerCase();
+  if (["1", "true", "yes", "y", "on"].includes(normalized)) return true;
+  if (["0", "false", "no", "n", "off"].includes(normalized)) return false;
+  return fallback;
+}
+
+function toPositiveInteger(value, fallback) {
+  if (value === undefined || value === null || value === "") return fallback;
+  const parsed = Number.parseInt(String(value), 10);
+  if (!Number.isFinite(parsed) || parsed <= 0) return fallback;
+  return parsed;
+}
+
+function appendToolType(target, value) {
+  const normalized = String(value || "").trim();
+  if (!normalized || target.includes(normalized)) return;
+  target.push(normalized);
+}
+
+function classifyContentType(type) {
+  const normalized = String(type || "").trim().toLowerCase();
+  if (!normalized) return "";
+  if (normalized === "image" || normalized === "image_url" || normalized === "input_image") return "image";
+  if (normalized === "document" || normalized === "input_document") return "document";
+  if (normalized === "audio" || normalized === "input_audio") return "audio";
+  if (normalized === "file" || normalized === "input_file") return "file";
+  if (normalized.includes("attachment")) return "attachment";
+  return "";
+}
+
+function maybeRecordLargeString(summary, value, path, hintType = "string") {
+  if (typeof value !== "string" || value.length === 0) return;
+  const bytes = DEFAULT_TEXT_ENCODER.encode(value).byteLength;
+  if (bytes > summary.largestStringBytes) {
+    summary.largestStringBytes = bytes;
+  }
+  if (bytes < LARGE_STRING_HINT_THRESHOLD_BYTES) return;
+
+  summary.largeStringCount += 1;
+  summary.largeStringHints.push({
+    path,
+    bytes,
+    type: hintType
+  });
+  summary.largeStringHints.sort((left, right) => right.bytes - left.bytes);
+  if (summary.largeStringHints.length > MAX_LARGE_STRING_HINTS) {
+    summary.largeStringHints.length = MAX_LARGE_STRING_HINTS;
+  }
+}
+
+function summarizeProviderBody(body) {
+  const toolTypes = [];
+  for (const tool of Array.isArray(body?.tools) ? body.tools : []) {
+    appendToolType(toolTypes, tool?.type);
+  }
+
+  const summary = {
+    topLevelKeys: body && typeof body === "object" && !Array.isArray(body) ? Object.keys(body).sort() : [],
+    messageCount: Array.isArray(body?.messages) ? body.messages.length : 0,
+    inputCount: Array.isArray(body?.input) ? body.input.length : 0,
+    toolCount: Array.isArray(body?.tools) ? body.tools.length : 0,
+    toolTypes,
+    contentPartCount: 0,
+    attachmentLikeParts: 0,
+    imageParts: 0,
+    documentParts: 0,
+    audioParts: 0,
+    fileParts: 0,
+    dataUrlStrings: 0,
+    base64SourceParts: 0,
+    largeStringCount: 0,
+    largestStringBytes: 0,
+    largeStringHints: [],
+    traversalTruncated: false
+  };
+
+  const stack = [{ value: body, path: "body" }];
+  const seen = new WeakSet();
+  let visited = 0;
+
+  while (stack.length > 0) {
+    const current = stack.pop();
+    visited += 1;
+    if (visited > MAX_SUMMARY_NODES) {
+      summary.traversalTruncated = true;
+      break;
+    }
+
+    const value = current?.value;
+    if (typeof value === "string") {
+      const isDataUrl = value.startsWith("data:");
+      if (isDataUrl) {
+        summary.dataUrlStrings += 1;
+      }
+      maybeRecordLargeString(summary, value, current.path, isDataUrl ? "data-url" : "string");
+      continue;
+    }
+    if (!value || typeof value !== "object") continue;
+    if (seen.has(value)) continue;
+    seen.add(value);
+
+    if (Array.isArray(value)) {
+      for (let index = value.length - 1; index >= 0; index -= 1) {
+        stack.push({
+          value: value[index],
+          path: `${current.path}[${index}]`
+        });
+      }
+      continue;
+    }
+
+    const contentType = classifyContentType(value.type);
+    if (contentType) {
+      summary.attachmentLikeParts += 1;
+      if (contentType === "image") summary.imageParts += 1;
+      if (contentType === "document") summary.documentParts += 1;
+      if (contentType === "audio") summary.audioParts += 1;
+      if (contentType === "file" || contentType === "attachment") summary.fileParts += 1;
+    }
+    if (value?.source && typeof value.source === "object") {
+      const sourceType = String(value.source.type || "").trim().toLowerCase();
+      if (sourceType === "base64") {
+        summary.base64SourceParts += 1;
+        maybeRecordLargeString(summary, value.source.data, `${current.path}.source.data`, "base64");
+      }
+    }
+
+    for (const [key, child] of Object.entries(value)) {
+      const childPath = `${current.path}.${key}`;
+      if (typeof child === "string") {
+        const hintType = key === "data"
+          ? "data"
+          : (key === "text" ? "text" : "string");
+        const isDataUrl = child.startsWith("data:");
+        if (isDataUrl) {
+          summary.dataUrlStrings += 1;
+        }
+        maybeRecordLargeString(summary, child, childPath, isDataUrl ? "data-url" : hintType);
+        continue;
+      }
+      if (key === "content" && Array.isArray(child)) {
+        summary.contentPartCount += child.length;
+      }
+      stack.push({
+        value: child,
+        path: childPath
+      });
+    }
+  }
+
+  return summary;
+}
+
+export function isLargeRequestLoggingEnabled(env = {}) {
+  return toBoolean(env?.[LARGE_REQUEST_LOG_ENABLED_ENV], false);
+}
+
+export function resolveLargeRequestLogThresholdBytes(env = {}) {
+  return toPositiveInteger(
+    env?.[LARGE_REQUEST_LOG_THRESHOLD_ENV],
+    DEFAULT_LARGE_REQUEST_LOG_THRESHOLD_BYTES
+  );
+}
+
+export function measureSerializedRequestBytes(serializedBody = "") {
+  return DEFAULT_TEXT_ENCODER.encode(String(serializedBody || "")).byteLength;
+}
+
+export function buildLargeRequestLogEntry({
+  providerBody,
+  requestBytes,
+  thresholdBytes,
+  providerUrl,
+  candidate,
+  sourceFormat,
+  targetFormat,
+  requestKind,
+  clientType,
+  stream,
+  providerType = "http"
+} = {}) {
+  return {
+    kind: "large-provider-request",
+    providerType: String(providerType || "http").trim() || "http",
+    requestBytes: Number.isFinite(Number(requestBytes)) ? Number(requestBytes) : 0,
+    thresholdBytes: Number.isFinite(Number(thresholdBytes)) ? Number(thresholdBytes) : DEFAULT_LARGE_REQUEST_LOG_THRESHOLD_BYTES,
+    providerUrl: String(providerUrl || "").trim(),
+    clientType: String(clientType || "").trim(),
+    stream: Boolean(stream),
+    sourceFormat: String(sourceFormat || "").trim(),
+    targetFormat: String(targetFormat || "").trim(),
+    requestKind: String(requestKind || "").trim(),
+    requestedModel: String(candidate?.requestModelId || "").trim(),
+    providerId: String(candidate?.providerId || candidate?.provider?.id || "").trim(),
+    backendModel: String(candidate?.backend || candidate?.modelId || providerBody?.model || "").trim(),
+    bodySummary: summarizeProviderBody(providerBody)
+  };
+}

package/src/runtime/handler/provider-call.js

@@ -21,6 +21,7 @@ import {
 import { maybeRewriteAmpClientResponse } from "./amp-response.js";
 import { applyCachingMapping, mergeCachingHeaders } from "./cache-mapping.js";
 import { applyReasoningEffortMapping } from "./reasoning-effort.js";
+import { stripUnsupportedFields } from "./field-filter.js";
 import { resolveUpstreamTimeoutMs } from "./request.js";
 import { parseJsonSafely } from "./utils.js";
 import { buildTimeoutSignal } from "../../shared/timeout-signal.js";
@@ -35,11 +36,62 @@ import {
   rewriteProviderBodyForAmpWebSearch,
   shouldInterceptAmpWebSearch
 } from "./amp-web-search.js";
+import {
+  buildLargeRequestLogEntry,
+  isLargeRequestLoggingEnabled,
+  measureSerializedRequestBytes,
+  resolveLargeRequestLogThresholdBytes
+} from "./large-request-log.js";
 
 function isSubscriptionProvider(provider) {
   return provider?.type === "subscription";
 }
 
+function queueLargeRequestEvent(onLargeRequestLog, payload) {
+  if (typeof onLargeRequestLog !== "function") return;
+  try {
+    const result = onLargeRequestLog(payload);
+    if (result && typeof result.then === "function") {
+      result.catch(() => {});
+    }
+  } catch {
+  }
+}
+
+function maybeQueueLargeRequestLog({
+  env,
+  onLargeRequestLog,
+  providerBody,
+  serializedBody,
+  providerUrl,
+  candidate,
+  sourceFormat,
+  targetFormat,
+  requestKind,
+  clientType,
+  stream,
+  providerType = "http"
+} = {}) {
+  if (!isLargeRequestLoggingEnabled(env) || typeof onLargeRequestLog !== "function") return;
+  const requestBytes = measureSerializedRequestBytes(serializedBody);
+  const thresholdBytes = resolveLargeRequestLogThresholdBytes(env);
+  if (requestBytes < thresholdBytes) return;
+
+  queueLargeRequestEvent(onLargeRequestLog, buildLargeRequestLogEntry({
+    providerBody,
+    requestBytes,
+    thresholdBytes,
+    providerUrl,
+    candidate,
+    sourceFormat,
+    targetFormat,
+    requestKind,
+    clientType,
+    stream,
+    providerType
+  }));
+}
+
 async function toProviderError(response) {
   const raw = await response.text();
   const parsed = parseJsonSafely(raw);
@@ -97,7 +149,8 @@ async function adaptProviderResponse({
   requestKind,
   requestBody,
   clientType,
-  env
+  env,
+  responsesDowngraded
 }) {
   const buildSuccessResponse = async (resultResponse) => ({
     ok: true,
@@ -111,6 +164,30 @@ async function adaptProviderResponse({
     })
   });
 
+  // Responses API was downgraded to Chat Completions for provider compatibility.
+  // Convert response back: Chat Completions → Claude → Responses API.
+  if (responsesDowngraded) {
+    if (stream) {
+      const claudeStream = handleOpenAIStreamToClaude(response);
+      return buildSuccessResponse(handleClaudeStreamToOpenAIResponses(claudeStream, requestBody, fallbackModel));
+    }
+    const raw = await response.text();
+    const parsed = parseJsonSafely(raw);
+    if (!parsed) {
+      return {
+        ok: false,
+        status: 502,
+        retryable: true,
+        response: jsonResponse({
+          type: "error",
+          error: { type: "api_error", message: "Provider returned invalid JSON." }
+        }, 502)
+      };
+    }
+    const claudeMessage = convertOpenAINonStreamToClaude(parsed, fallbackModel);
+    return buildSuccessResponse(jsonResponse(convertClaudeNonStreamToOpenAIResponses(claudeMessage, requestBody, fallbackModel)));
+  }
+
   if (stream) {
     if (!translate) {
       return buildSuccessResponse(
@@ -489,14 +566,22 @@ function buildProviderRequestPlan({
   requestKind,
   requestHeaders,
   interceptAmpWebSearch,
-  stream
+  stream,
+  forceResponsesDowngrade = false
 }) {
   const normalizedRequestKind = normalizeProviderRequestKind(targetFormat, requestKind);
   const translate = needsTranslation(sourceFormat, targetFormat);
 
   let providerBody = { ...body };
+  let responsesDowngraded = false;
   if (translate) {
     providerBody = translateRequest(sourceFormat, targetFormat, candidate.backend, body, stream);
+  } else if (forceResponsesDowngrade) {
+    // Provider confirmed to not support Responses API — downgrade to Chat Completions
+    // via double-hop: Responses API → Claude → Chat Completions.
+    const intermediateBody = translateRequest(FORMATS.OPENAI, FORMATS.CLAUDE, candidate.backend, body, stream);
+    providerBody = translateRequest(FORMATS.CLAUDE, FORMATS.OPENAI, candidate.backend, intermediateBody, stream);
+    responsesDowngraded = true;
   }
 
   providerBody.model = candidate.backend;
@@ -513,9 +598,19 @@ function buildProviderRequestPlan({
     sourceFormat,
     targetFormat,
     targetModel: candidate.backend,
-    requestHeaders
+    requestHeaders,
+    capabilities: candidate.model?.capabilities
   });
 
+  if (responsesDowngraded) {
+    // Strip Responses-API-only fields that Chat Completions providers reject.
+    delete providerBody.prompt_cache_key;
+    delete providerBody.store;
+    delete providerBody.include;
+    delete providerBody.text;
+    delete providerBody.service_tier;
+  }
+
   const declaredOpenAIHostedWebSearchToolType = getProviderOpenAIHostedWebSearchToolType(candidate.provider, {
     targetFormat,
     requestKind: normalizedRequestKind
@@ -532,11 +627,14 @@ function buildProviderRequestPlan({
     providerBody = rewriteProviderBodyForAmpWebSearch(providerBody, targetFormat, requestKind).providerBody;
   }
 
+  providerBody = stripUnsupportedFields(providerBody, candidate.model?.capabilities);
+
   return {
     targetFormat,
-    requestKind: normalizedRequestKind,
+    requestKind: responsesDowngraded ? undefined : normalizedRequestKind,
     translate,
-    providerBody
+    providerBody,
+    responsesDowngraded
   };
 }
 
@@ -552,7 +650,8 @@ export async function makeProviderCall({
   runtimeConfig,
   stateStore,
   ampContext,
-  runtimeFlags
+  runtimeFlags,
+  onLargeRequestLog
 }) {
   const provider = candidate.provider;
   const targetFormat = candidate.targetFormat;
@@ -576,8 +675,17 @@ export async function makeProviderCall({
     effectiveBody = { ...body, reasoning_effort: ampContext.presets.reasoningEffort };
   }
 
+  // For Responses API requests to OpenAI-format providers, try the native endpoint first.
+  // If the provider doesn't support /v1/responses (returns 404/400), fall back to a
+  // downgraded Chat Completions plan with double-hop translation.
+  const needsResponsesDowngradeFallback = !isSubscriptionProvider(provider)
+    && sourceFormat === FORMATS.OPENAI
+    && targetFormat === FORMATS.OPENAI
+    && requestKind === "responses";
+
   let activePlan;
   let fallbackPlan = null;
+  let responsesDowngradedPlan = null;
   try {
     activePlan = buildProviderRequestPlan({
       body: effectiveBody,
@@ -601,6 +709,19 @@ export async function makeProviderCall({
         stream
       });
     }
+    if (needsResponsesDowngradeFallback) {
+      responsesDowngradedPlan = buildProviderRequestPlan({
+        body: effectiveBody,
+        sourceFormat,
+        targetFormat,
+        candidate,
+        requestKind,
+        requestHeaders,
+        interceptAmpWebSearch,
+        stream,
+        forceResponsesDowngrade: true
+      });
+    }
   } catch (error) {
     return {
       ok: false,
@@ -651,13 +772,33 @@ export async function makeProviderCall({
         prompt_cache_key: activePlan.providerBody.prompt_cache_key || ampContext.threadId
       };
     }
-    const executeSubscriptionRequest = async (requestBody) => makeSubscriptionProviderCall({
-      provider,
-      body: requestBody,
-      // ChatGPT Codex backend expects stream=true; non-stream responses are reconstructed from SSE.
-      stream: subscriptionType === "chatgpt-codex" ? true : Boolean(stream),
-      env
-    });
+    const executeSubscriptionRequest = async (requestBody) => {
+      const requestStream = subscriptionType === "chatgpt-codex" ? true : Boolean(stream);
+      const providerUrl = subscriptionType === "chatgpt-codex"
+        ? "https://chatgpt.com/backend-api/codex/responses"
+        : "https://console.anthropic.com/v1/messages?beta=true";
+      maybeQueueLargeRequestLog({
+        env,
+        onLargeRequestLog,
+        providerBody: requestBody,
+        serializedBody: JSON.stringify(requestBody),
+        providerUrl,
+        candidate,
+        sourceFormat,
+        targetFormat: activePlan.targetFormat,
+        requestKind: activePlan.requestKind,
+        clientType,
+        stream: requestStream,
+        providerType: subscriptionType
+      });
+      return makeSubscriptionProviderCall({
+        provider,
+        body: requestBody,
+        // ChatGPT Codex backend expects stream=true; non-stream responses are reconstructed from SSE.
+        stream: requestStream,
+        env
+      });
+    };
     const subscriptionResult = await executeSubscriptionRequest(activePlan.providerBody);
 
     if (!subscriptionResult?.ok) {
@@ -854,11 +995,26 @@ export async function makeProviderCall({
     const timeoutMs = resolveUpstreamTimeoutMs(env);
     const timeoutControl = buildTimeoutSignal(timeoutMs);
     try {
+      const serializedBody = JSON.stringify(plan.providerBody);
       const init = {
         method: "POST",
         headers,
-        body: JSON.stringify(plan.providerBody)
+        body: serializedBody
       };
+      maybeQueueLargeRequestLog({
+        env,
+        onLargeRequestLog,
+        providerBody: plan.providerBody,
+        serializedBody,
+        providerUrl,
+        candidate,
+        sourceFormat,
+        targetFormat: plan.targetFormat,
+        requestKind: plan.requestKind,
+        clientType,
+        stream,
+        providerType: "http"
+      });
       if (timeoutControl.signal) {
         init.signal = timeoutControl.signal;
       }
@@ -934,6 +1090,19 @@ export async function makeProviderCall({
     };
   }
 
+  // Provider doesn't support native /v1/responses — retry with Chat Completions downgrade.
+  if ((!response || !response.ok) && responsesDowngradedPlan) {
+    try {
+      const downgradedResponse = await executeHttpProviderRequest(responsesDowngradedPlan);
+      if (downgradedResponse instanceof Response && downgradedResponse.ok) {
+        response = downgradedResponse;
+        activePlan = responsesDowngradedPlan;
+      }
+    } catch {
+      // Keep the original failure if the downgraded request also fails.
+    }
+  }
+
   if (!response.ok) {
     const hostedWebSearchErrorKind = await resolveHostedWebSearchErrorKind(response, activePlan.providerBody, {
       targetFormat: activePlan.targetFormat,
@@ -983,6 +1152,7 @@ export async function makeProviderCall({
     requestKind: activePlan.requestKind,
     requestBody: body,
     clientType,
-    env
+    env,
+    responsesDowngraded: activePlan.responsesDowngraded
   });
 }

package/src/runtime/handler/reasoning-effort.js

@@ -295,8 +295,18 @@ export function applyReasoningEffortMapping({
   sourceFormat,
   targetFormat,
   targetModel,
-  requestHeaders
+  requestHeaders,
+  capabilities
 }) {
+  if (capabilities) {
+    if (targetFormat === FORMATS.OPENAI && capabilities.supportsReasoning === false) {
+      return providerBody;
+    }
+    if (targetFormat === FORMATS.CLAUDE && capabilities.supportsThinking === false) {
+      return providerBody;
+    }
+  }
+
   const effort = resolveRequestedEffort(originalBody, requestHeaders);
   if (!effort) return providerBody;