fix(routing): preserve reasoning overrides across transports and fallbacks (#13556)

Merged. The failure mode was concrete — a matched reasoning rule dropped on native Responses/Anthropic paths, model-suffix/account defaults, or fallback preparation, and `_omnirouteReasoningRule` leaking upstream as `Unsupported parameter` — and the fix is carried in the request-local credential context through dispatch, refreshed credentials and fallbacks, with forced effort winning over defaults and client-forged markers dropped at ingress. The 11-case integration suite exercises the real routing/translation modules.

Validated as a combined board first (this PR merged with the 4 siblings of the JxnLexn wave on the release tip): eslint with the frozen suppressions, typecheck:core, check:open-sse-typecheck, complexity, cognitive-complexity, changelog-integrity, docs-sync, migration-numbering, provider-consistency and a duplicate-identifier audit all green, plus 77 passing / 0 failing focused node:test cases across the test files the wave touches. The wave's i18n fill (new keys carried to all 66 locales), free-tier doc counts and file-size rebaseline land in one follow-up PR right after the wave, as with #13904.

Thank you — and for keeping this a runtime-only change with the editor and service-tier work in their own PRs.

changelog.d/fixes/13556-reasoning-rule-propagation.md

@@ -0,0 +1 @@
+- **fix(routing):** Preserve forced reasoning effort across native requests, account defaults and combo fallbacks while keeping internal routing directives out of upstream payloads. ([#13556](https://github.com/diegosouzapw/OmniRoute/pull/13556)) — thanks @JxnLexn

open-sse/executors/codex.ts

@@ -35,6 +35,7 @@ import {
 import { getAccessToken } from "../services/tokenRefresh.ts";
 import { sanitizeResponsesInputItems } from "../services/responsesInputSanitizer.ts";
 import { applyReasoningInputPolicy } from "../services/reasoningInputPolicy.ts";
+import { getForcedReasoningEffort } from "../utils/reasoningRuleContext.ts";
 import { normalizeCodexVerbosity } from "../services/codexVerbosity.ts";
 import { getThinkingBudgetConfig, ThinkingMode } from "../services/thinkingBudget.ts";
 import { CORS_HEADERS } from "../utils/cors.ts";
@@ -819,6 +820,22 @@ export class CodexExecutor extends BaseExecutor {
       requestInput.body
     );
     const nextInput = { ...requestInput, credentials };
+    const forcedEffort = getForcedReasoningEffort(credentials);
+    if (forcedEffort) {
+      const nextBody =
+        nextInput.body && typeof nextInput.body === "object"
+          ? (nextInput.body as Record<string, unknown>)
+          : {};
+      nextInput.body = {
+        ...nextBody,
+        reasoning: {
+          ...(nextBody.reasoning && typeof nextBody.reasoning === "object"
+            ? nextBody.reasoning
+            : {}),
+          effort: forcedEffort,
+        },
+      };
+    }
 
     if (isCodexAppServerRequired(nextInput.credentials)) {
       if (!this.appServer) {
@@ -1380,8 +1397,13 @@ export class CodexExecutor extends BaseExecutor {
     // Issue #2331: model suffix aliases (for example gpt-5.5-xhigh) represent an
     // explicit model selection, so they must override client-injected defaults such
     // as OpenCode's automatic reasoning.effort=medium for GPT-5-family requests.
+    // A server-selected force rule is stronger than either source.
     const rawEffort =
-      modelEffort || explicitReasoning || requestReasoningEffort || fallbackReasoningEffort;
+      getForcedReasoningEffort(credentials) ||
+      modelEffort ||
+      explicitReasoning ||
+      requestReasoningEffort ||
+      fallbackReasoningEffort;
 
     if (rawEffort) {
       const clampedEffort = clampEffort(cleanModel, rawEffort);

open-sse/handlers/chatCore.ts

@@ -125,6 +125,8 @@ import { stripStore, usesClaudeBridge } from "./chatCore/agentRouterProtocol.ts"
 import { normalizeClaudeToolsForDispatch } from "./chatCore/claudeToolDefaults.ts";
 import { injectSystemPrompt, injectCustomSystemPrompt } from "../services/systemPrompt.ts";
 import { translateRequest, needsTranslation } from "../translator/index.ts";
+import { applyReasoningRuleDirective } from "@/lib/reasoningRouting/policy";
+import { withReasoningRuleContext } from "../utils/reasoningRuleContext.ts";
 import { FORMATS } from "../translator/formats.ts";
 import { collectCustomToolNamesForSourceFormat } from "../translator/request/openai-responses/additionalTools.ts";
 import { sanitizeKiroTools } from "../utils/kiroSanitizer.ts";
@@ -499,6 +501,9 @@ export async function handleChatCore({
   fallbackAttempts = undefined,
 }) {
   let { provider, model, extendedContext } = modelInfo;
+  // Keep the selected rule across format conversion, retries and refreshed credentials.
+  // Each combo leg gets its own execution context; nothing is written to shared accounts.
+  const reasoningRuleDirective = body?._omnirouteReasoningRule;
   // #12150 P1b: true iff the video-bridge guardrail rendered >=1 transcript
   // cue into a replaced part of this request. Gates both request- and
   // response-derived Memory extraction
@@ -1214,6 +1219,22 @@ export async function handleChatCore({
 
   log?.debug?.("FORMAT", `${sourceFormat} → ${targetFormat} | stream=${stream}`);
 
+  if (reasoningRuleDirective) {
+    // Cache identity must use the effective effort, not the overridden client value.
+    // Retain the directive for the translation step, where general thinking defaults run.
+    body = {
+      ...(applyReasoningRuleDirective(
+        body,
+        sourceFormat === FORMATS.OPENAI_RESPONSES
+          ? "openai-responses"
+          : sourceFormat === FORMATS.CLAUDE
+            ? "claude"
+            : undefined
+      ) as Record<string, unknown>),
+      _omnirouteReasoningRule: reasoningRuleDirective,
+    };
+  }
+
   // Preserve original body for cache signature — the body variable is mutated
   // multiple times below (sanitization, memory/skills injection) before the
   // cache store path runs at Phase 9.1 (non-streaming) / Phase 9.2 (streaming).
@@ -2267,7 +2288,7 @@ export async function handleChatCore({
   try {
     if (nativeResponsesPassthrough) {
       translatedBody = stampNativeResponsesPassthroughBody(
-        body,
+        applyReasoningRuleDirective(body, "openai-responses") as Record<string, unknown>,
         nativeCodexPassthrough
           ? "codex"
           : nativeXaiResponsesPassthrough
@@ -2288,6 +2309,10 @@ export async function handleChatCore({
       // Claude Code-compatible providers expect Anthropic Messages-shaped payloads,
       // but we extract only role/text/max_tokens/effort from an OpenAI-like view first.
       if (sourceFormat === FORMATS.CLAUDE && isClaudeCodeSemanticPassthrough) {
+        normalizedForCc = applyReasoningRuleDirective(
+          normalizedForCc,
+          "claude"
+        ) as typeof normalizedForCc;
         log?.debug?.("FORMAT", "claude-code semantic passthrough enabled for compatible bridge");
       } else if (sourceFormat !== FORMATS.OPENAI) {
         const normalizeToolCallId = getModelNormalizeToolCallId(
@@ -2323,6 +2348,10 @@ export async function handleChatCore({
       const ccRequestDefaults = getClaudeCodeCompatibleRequestDefaults(
         credentials?.providerSpecificData
       );
+      // OpenAI-shaped bridge requests skip translateRequest too.
+      if (sourceFormat === FORMATS.OPENAI) {
+        normalizedForCc = applyReasoningRuleDirective(normalizedForCc) as typeof normalizedForCc;
+      }
       translatedBody = buildClaudeCodeCompatibleRequest({
         sourceBody: body,
         normalizedBody: normalizedForCc,
@@ -2352,7 +2381,7 @@ export async function handleChatCore({
       // payloads at high context (150+ msgs, 100+ tools). Fix: #1359.
       // Claude Code sends well-formed Messages API payloads — trust them
       // regardless of combo strategy or cache_control settings.
-      translatedBody = { ...body };
+      translatedBody = applyReasoningRuleDirective({ ...body }, "claude");
       translatedBody._disableToolPrefix = true;
 
       // Sanitize historical thinking-block signatures for Anthropic-native Claude OAuth.
@@ -2996,15 +3025,18 @@ export async function handleChatCore({
   // Get executor for this provider (with optional upstream proxy routing)
   const executor = await resolveExecutorWithProxy(provider);
   const getExecutionCredentials = () =>
-    resolveExecutionCredentialsFor({
-      credentials,
-      nativeCodexPassthrough: nativeResponsesPassthrough,
-      endpointPath,
-      targetFormat,
-      provider,
-      ccSessionId,
-      modelInfo,
-    });
+    withReasoningRuleContext(
+      resolveExecutionCredentialsFor({
+        credentials,
+        nativeCodexPassthrough: nativeResponsesPassthrough,
+        endpointPath,
+        targetFormat,
+        provider,
+        ccSessionId,
+        modelInfo,
+      }),
+      reasoningRuleDirective
+    );
 
   let onPipelineStreamError: streamFailure.PipelineStreamErrorHandler | null = null;
   let onClientDisconnectFinalize:

open-sse/utils/reasoningRuleContext.ts

@@ -0,0 +1,28 @@
+// Request-local execution metadata, not a provider credential or a wire field.
+// Symbol keys survive credential object spreads but cannot be supplied through JSON.
+const FORCED_EFFORT = Symbol.for("omniroute.forcedReasoningEffort");
+const EFFORTS = new Set(["none", "low", "medium", "high", "xhigh", "max", "ultra"]);
+
+export function withReasoningRuleContext<T>(credentials: T, directive: unknown): T {
+  if (
+    !credentials ||
+    typeof credentials !== "object" ||
+    !directive ||
+    typeof directive !== "object"
+  )
+    return credentials;
+  const rule = directive as Record<string, unknown>;
+  if (
+    !rule.id ||
+    rule.effortMode !== "force" ||
+    typeof rule.targetEffort !== "string" ||
+    !EFFORTS.has(rule.targetEffort)
+  )
+    return credentials;
+  return { ...credentials, [FORCED_EFFORT]: rule.targetEffort };
+}
+
+export function getForcedReasoningEffort(credentials: unknown): string | undefined {
+  if (!credentials || typeof credentials !== "object") return undefined;
+  return (credentials as { [FORCED_EFFORT]?: string })[FORCED_EFFORT];
+}

src/app/api/internal/codex-responses-ws/route.ts

@@ -23,6 +23,7 @@ import { logger } from "@omniroute/open-sse/utils/logger.ts";
 import { resolveProxy } from "@omniroute/open-sse/utils/networkProxy.ts";
 import { withCodexFingerprintCredentials } from "@omniroute/open-sse/config/codexIdentity.ts";
 import { proxyConfigToUrl } from "@omniroute/open-sse/utils/proxyDispatcher.ts";
+import { withReasoningRuleContext } from "@omniroute/open-sse/utils/reasoningRuleContext.ts";
 import {
   attachReasoningRuleDirective,
   applyReasoningRuleDirective,
@@ -544,12 +545,17 @@ async function prepare(body: JsonRecord) {
 
   let responseBodyWithMemory = await maybeInjectResponsesWsMemory(responseBody, metadata);
   let reasoningRouting: JsonRecord | null = null;
+  let reasoningRuleDirective: unknown;
   if (reasoningDecision) {
     const withDirective = attachReasoningRuleDirective(responseBodyWithMemory, reasoningDecision);
+    reasoningRuleDirective = withDirective._omnirouteReasoningRule;
     reasoningRouting = isRecord(withDirective._omnirouteReasoningRouteTrace)
       ? withDirective._omnirouteReasoningRouteTrace
       : null;
-    responseBodyWithMemory = applyReasoningRuleDirective(withDirective) as JsonRecord;
+    responseBodyWithMemory = applyReasoningRuleDirective(
+      withDirective,
+      "openai-responses"
+    ) as JsonRecord;
     delete responseBodyWithMemory._omnirouteReasoningRouteTrace;
   }
   // #8052: the WS bridge previously skipped the whole prompt-compression pipeline that the
@@ -561,7 +567,7 @@ async function prepare(body: JsonRecord) {
     requestId: randomUUID(),
   });
   const credentialsWithFingerprint = withCodexFingerprintCredentials(
-    refreshedCredentials,
+    withReasoningRuleContext(refreshedCredentials, reasoningRuleDirective),
     context.clientHeaders,
     responseBodyWithMemory
   );

src/lib/reasoningRouting/policy.ts

@@ -531,7 +531,10 @@ export function attachReasoningRuleDirective(
   return body;
 }
 
-export function applyReasoningRuleDirective(bodyInput: unknown): unknown {
+export function applyReasoningRuleDirective(
+  bodyInput: unknown,
+  targetFormat?: "openai-responses" | "claude"
+): unknown {
   const source = asRecord(bodyInput);
   const directive = asRecord(source._omnirouteReasoningRule);
   if (!directive.id) return bodyInput;
@@ -542,9 +545,11 @@ export function applyReasoningRuleDirective(bodyInput: unknown): unknown {
   if (effortMode === "force" && targetEffort === "none") clearReasoning(body);
   else if ((effortMode === "force" || effortMode === "default") && targetEffort) {
     if (effortMode === "force") clearDiscreteReasoning(body);
-    body.reasoning_effort = targetEffort;
-    body.reasoning = { ...asRecord(body.reasoning), effort: targetEffort };
-    body.output_config = { ...asRecord(body.output_config), effort: targetEffort };
+    if (!targetFormat) body.reasoning_effort = targetEffort;
+    if (targetFormat !== "claude")
+      body.reasoning = { ...asRecord(body.reasoning), effort: targetEffort };
+    if (targetFormat !== "openai-responses")
+      body.output_config = { ...asRecord(body.output_config), effort: targetEffort };
   }
   applyBudget(
     body,

src/sse/handlers/chat.ts

@@ -448,6 +448,14 @@ async function handleChatImplementation(
     return errorResponse(HTTP_STATUS.BAD_REQUEST, "Invalid JSON body");
   }
 
+  // Only the server's policy resolver may attach execution directives or route traces.
+  // Discard lookalike JSON fields supplied by callers before evaluating any rule.
+  if (body && typeof body === "object") {
+    body = { ...body };
+    delete body._omnirouteReasoningRule;
+    delete body._omnirouteReasoningRouteTrace;
+  }
+
   // Feature #6241: fold the canonical `effort` / `thinking` request params onto the
   // per-provider reasoning fields (reasoning_effort / reasoning.effort / thinking) that the
   // existing translators already consume. Done here — right after the body is first

tests/integration/reasoning-routing-reliability.test.ts

@@ -0,0 +1,389 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { createChatPipelineHarness } from "./_chatPipelineHarness.ts";
+
+const h = await createChatPipelineHarness("reasoning-reliability");
+const providers = await import("../../src/lib/db/providers.ts");
+const { flushProxyLogsSync } = await import("../../src/lib/proxyLogger.ts");
+type RecordedBody = Record<string, unknown> & { reasoning?: { effort?: string } };
+let calls: RecordedBody[] = [];
+let firstConnectionId: string;
+
+test.beforeEach(async () => {
+  await h.resetStorage();
+  h.BaseExecutor.RETRY_CONFIG.delayMs = 0;
+  calls = [];
+  const connection = await providers.createProviderConnection({
+    provider: "codex",
+    authType: "oauth",
+    name: "Reasoning fixture",
+    accessToken: "fixture-access-token",
+    refreshToken: "fixture-refresh-token",
+    expiresAt: new Date(Date.now() + 3600000).toISOString(),
+    isActive: true,
+    testStatus: "active",
+    providerSpecificData: { requestDefaults: { reasoningEffort: "high" } },
+  });
+  firstConnectionId = String(connection.id);
+  globalThis.fetch = async (url, init: RequestInit = {}) => {
+    assert.match(String(url), /chatgpt\.com\/backend-api\/codex\/responses/);
+    calls.push(JSON.parse(String(init.body)) as RecordedBody);
+    return new Response(
+      "data: " +
+        JSON.stringify({
+          type: "response.completed",
+          response: {
+            id: "resp_fixture",
+            object: "response",
+            status: "completed",
+            model: "gpt-5.6-luna",
+            output: [
+              {
+                id: "msg_fixture",
+                type: "message",
+                role: "assistant",
+                content: [{ type: "output_text", text: "ok", annotations: [] }],
+              },
+            ],
+            usage: { input_tokens: 1, output_tokens: 1, total_tokens: 2 },
+          },
+        }) +
+        "\n\ndata: [DONE]\n\n",
+      { headers: { "Content-Type": "text/event-stream" } }
+    );
+  };
+});
+test.afterEach(async () => {
+  flushProxyLogsSync();
+  h.BaseExecutor.RETRY_CONFIG.delayMs = h.originalRetryDelayMs;
+  await h.resetStorage();
+});
+test.after(() => {
+  flushProxyLogsSync();
+  return h.cleanup();
+});
+
+async function forceLow(modelPattern: string, apiKeyId: string) {
+  return h.reasoningRulesDb.createReasoningRoutingRule({
+    name: "Force low",
+    description: "",
+    scope: "apiKey",
+    apiKeyId,
+    comboId: null,
+    connectionId: null,
+    modelPattern,
+    sourceEffort: "any",
+    requestTags: [],
+    tagMatchMode: "any",
+    effortMode: "force",
+    targetEffort: "low",
+    targetKind: "keep",
+    targetModel: null,
+    targetComboId: null,
+    budgetAction: "preserve",
+    budgetTokens: null,
+    priority: 0,
+    enabled: true,
+  });
+}
+
+test("reliable reasoning: forced effort wins over explicit effort and adaptive header opt-in", async () => {
+  const key = await h.seedApiKey();
+  await forceLow("codex/gpt-5.6-luna", key.id);
+  const response = await h.handleChat(
+    h.buildRequest({
+      authKey: key.key,
+      headers: { "x-omniroute-effort": "auto" },
+      body: {
+        model: "codex/gpt-5.6-luna",
+        stream: false,
+        reasoning_effort: "high",
+        messages: [
+          { role: "user", content: "Prove the theorem step by step and verify every edge case." },
+        ],
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls.at(-1)?.reasoning?.effort, "low");
+});
+
+test("reliable reasoning: explicit client effort survives adaptive header opt-in", async () => {
+  const key = await h.seedApiKey();
+  await providers.updateProviderConnection(firstConnectionId, { providerSpecificData: {} });
+  const response = await h.handleChat(
+    h.buildRequest({
+      authKey: key.key,
+      headers: { "x-omniroute-effort": "auto" },
+      body: {
+        model: "codex/gpt-5.6-luna",
+        stream: false,
+        reasoning_effort: "medium",
+        messages: [{ role: "user", content: "Reply ok." }],
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls.at(-1)?.reasoning?.effort, "medium");
+});
+
+test("reliable reasoning: native Responses honors key force without leaking cross-format fields", async () => {
+  const key = await h.seedApiKey();
+  await forceLow("codex/gpt-5.6-luna", key.id);
+  const input = [
+    { type: "message", role: "user", content: [{ type: "input_text", text: "Reply ok" }] },
+  ];
+  const response = await h.handleChat(
+    h.buildRequest({
+      url: "http://localhost/v1/responses",
+      authKey: key.key,
+      body: { model: "codex/gpt-5.6-luna", input, stream: false, reasoning: { effort: "high" } },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].reasoning?.effort, "low");
+  assert.deepEqual(calls[0].input, input);
+  for (const name of [
+    "_omnirouteReasoningRule",
+    "_omnirouteReasoningRouteTrace",
+    "output_config",
+    "reasoning_effort",
+  ]) {
+    assert.equal(calls[0][name], undefined, name);
+  }
+});
+
+test("reliable reasoning: combo member effort suffix cannot override the selected key rule", async () => {
+  const key = await h.seedApiKey();
+  await h.combosDb.createCombo({
+    name: "luna-combo",
+    models: ["codex/gpt-5.6-luna-high"],
+    strategy: "priority",
+  });
+  await forceLow("luna-combo", key.id);
+  const response = await h.handleChat(
+    h.buildRequest({
+      authKey: key.key,
+      body: {
+        model: "luna-combo",
+        stream: false,
+        reasoning_effort: "high",
+        messages: [{ role: "user", content: "Reply ok" }],
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].model, "gpt-5.6-luna");
+  assert.equal(calls[0].reasoning?.effort, "low");
+});
+
+test("reliable reasoning: client metadata cannot impersonate an administrator rule", async () => {
+  const response = await h.handleChat(
+    h.buildRequest({
+      body: {
+        model: "codex/gpt-5.6-luna",
+        stream: false,
+        reasoning_effort: "high",
+        messages: [{ role: "user", content: "Reply ok" }],
+        _omnirouteReasoningRule: { id: "forged", effortMode: "force", targetEffort: "low" },
+        _omnirouteReasoningRouteTrace: { ruleId: "forged" },
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls[0].reasoning?.effort, "high");
+  assert.equal(calls[0]._omnirouteReasoningRule, undefined);
+});
+
+test("reliable reasoning: fallback attempts retain the forced effort", async () => {
+  const key = await h.seedApiKey();
+  const second = await providers.createProviderConnection({
+    provider: "codex",
+    authType: "oauth",
+    name: "Fallback fixture",
+    accessToken: "fixture-fallback-token",
+    isActive: true,
+    testStatus: "active",
+    expiresAt: new Date(Date.now() + 3600000).toISOString(),
+    providerSpecificData: {},
+  });
+  await h.combosDb.createCombo({
+    name: "fallback-combo",
+    models: [
+      { model: "codex/gpt-5.6-luna-high", connectionId: firstConnectionId },
+      { model: "codex/gpt-5.5-xhigh", connectionId: String(second.id) },
+    ],
+    strategy: "priority",
+  });
+  await forceLow("fallback-combo", key.id);
+  const successFetch = globalThis.fetch;
+  globalThis.fetch = async (url, init: RequestInit = {}) => {
+    const body = JSON.parse(String(init.body)) as RecordedBody;
+    if (body.model === "gpt-5.6-luna") {
+      calls.push(body);
+      return new Response(JSON.stringify({ error: { message: "Fixture unavailable" } }), {
+        status: 503,
+        headers: { "Content-Type": "application/json" },
+      });
+    }
+    return successFetch(url, init);
+  };
+  const response = await h.handleChat(
+    h.buildRequest({
+      authKey: key.key,
+      body: {
+        model: "fallback-combo",
+        stream: false,
+        reasoning_effort: "high",
+        messages: [{ role: "user", content: "Reply ok" }],
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.ok(calls.length >= 2);
+  assert.equal(calls.at(-1)?.model, "gpt-5.5");
+  assert.ok(calls.every((call) => call.reasoning?.effort === "low"));
+});
+
+test("reliable reasoning: force is request-local and does not affect another API key", async () => {
+  const restricted = await h.seedApiKey({ name: "Forced key" });
+  const ordinary = await h.seedApiKey({ name: "Ordinary key" });
+  await forceLow("codex/gpt-5.6-luna", restricted.id);
+  for (const key of [restricted, ordinary]) {
+    const response = await h.handleChat(
+      h.buildRequest({
+        authKey: key.key,
+        body: {
+          model: "codex/gpt-5.6-luna",
+          stream: false,
+          reasoning_effort: "high",
+          messages: [{ role: "user", content: "Reply " + key.id }],
+        },
+      })
+    );
+    await response.text();
+    assert.equal(response.status, 200);
+  }
+  assert.deepEqual(
+    calls.map((call) => call.reasoning?.effort),
+    ["low", "high"]
+  );
+});
+
+test("reliable reasoning: force none survives Codex account defaults", async () => {
+  const key = await h.seedApiKey();
+  const rule = await forceLow("codex/gpt-5.6-luna", key.id);
+  await h.reasoningRulesDb.updateReasoningRoutingRule(rule.id, { targetEffort: "none" });
+  const response = await h.handleChat(
+    h.buildRequest({
+      authKey: key.key,
+      url: "http://localhost/v1/responses",
+      body: {
+        model: "codex/gpt-5.6-luna",
+        input: "Reply ok",
+        stream: false,
+        reasoning: { effort: "high" },
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal(calls[0].reasoning?.effort, "none");
+});
+
+test("reliable reasoning: native default applies only when client effort is missing", async () => {
+  const key = await h.seedApiKey();
+  const rule = await forceLow("codex/gpt-5.6-luna", key.id);
+  await h.reasoningRulesDb.updateReasoningRoutingRule(rule.id, { effortMode: "default" });
+  for (const effort of [undefined, "high"]) {
+    const response = await h.handleChat(
+      h.buildRequest({
+        url: "http://localhost/v1/responses",
+        authKey: key.key,
+        body: {
+          model: "codex/gpt-5.6-luna",
+          input: "Default " + effort,
+          stream: false,
+          ...(effort ? { reasoning: { effort } } : {}),
+        },
+      })
+    );
+    await response.text();
+    assert.equal(response.status, 200);
+  }
+  assert.deepEqual(
+    calls.map((call) => call.reasoning?.effort),
+    ["low", "high"]
+  );
+  assert.ok(calls.every((call) => call.output_config === undefined));
+});
+
+test("reliable reasoning: WebSocket prepare preserves force in the final Codex payload", async () => {
+  process.env.OMNIROUTE_WS_BRIDGE_SECRET = "fixture-bridge-secret";
+  const { POST } = await import("../../src/app/api/internal/codex-responses-ws/route.ts");
+  const key = await h.seedApiKey();
+  await forceLow("codex/gpt-5.6-luna", key.id);
+  const response = await POST(
+    new Request("http://localhost/api/internal/codex-responses-ws", {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        "x-omniroute-ws-bridge-secret": "fixture-bridge-secret",
+      },
+      body: JSON.stringify({
+        action: "prepare",
+        requestUrl: "http://localhost/v1/responses",
+        headers: { authorization: "Bearer " + key.key },
+        response: {
+          model: "codex/gpt-5.6-luna-high",
+          input: "Reply ok",
+          reasoning: { effort: "high" },
+        },
+      }),
+    })
+  );
+  const data = await response.json();
+  assert.equal(response.status, 200);
+  assert.equal(data.response.reasoning.effort, "low");
+  assert.equal(data.response.output_config, undefined);
+  assert.equal(data.response._omnirouteReasoningRule, undefined);
+  assert.equal(calls.length, 0);
+});
+
+test("reliable reasoning: native Anthropic applies effort without OpenAI wire fields", async () => {
+  await h.seedConnection("anthropic");
+  const key = await h.seedApiKey();
+  await forceLow("anthropic/claude-sonnet-4-6", key.id);
+  globalThis.fetch = async (_url, init: RequestInit = {}) => {
+    calls.push(JSON.parse(String(init.body)));
+    return h.buildClaudeResponse("ok");
+  };
+  const response = await h.handleChat(
+    h.buildRequest({
+      url: "http://localhost/v1/messages",
+      authKey: key.key,
+      body: {
+        model: "anthropic/claude-sonnet-4-6",
+        max_tokens: 128,
+        stream: false,
+        thinking: { type: "adaptive" },
+        output_config: { effort: "high" },
+        messages: [{ role: "user", content: "Reply ok" }],
+      },
+    })
+  );
+  await response.text();
+  assert.equal(response.status, 200);
+  assert.equal((calls[0].output_config as { effort?: string })?.effort, "low");
+  assert.equal(calls[0].reasoning, undefined);
+  assert.equal(calls[0].reasoning_effort, undefined);
+  assert.equal(calls[0]._omnirouteReasoningRule, undefined);
+});