diff --git a/changelog.d/fixes/12585-codex-strip-temperature.md b/changelog.d/fixes/12585-codex-strip-temperature.md
new file mode 100644
index 0000000000..1e02f64225
--- /dev/null
+++ b/changelog.d/fixes/12585-codex-strip-temperature.md
@@ -0,0 +1 @@
+- **fix(sse):** strip `temperature`/`top_p` on native Codex Responses passthrough so combo `codex-review` traffic no longer 400s with `Unsupported parameter: temperature` ([#12585](https://github.com/diegosouzapw/OmniRoute/pull/12585))
diff --git a/open-sse/executors/codex.ts b/open-sse/executors/codex.ts
index 644a67c08b..82bd7931de 100644
--- a/open-sse/executors/codex.ts
+++ b/open-sse/executors/codex.ts
@@ -19,6 +19,7 @@ import {
 } from "../config/codexInstructions.ts";
 import { FETCH_BODY_TIMEOUT_MS, HTTP_STATUS, PROVIDERS } from "../config/constants.ts";
 import { readCodexPeekChunk, buildCodexTimeoutSafePassthroughBody } from "./codex/bodyTimeout.ts";
+import { stripCodexPassthroughRejectedParams } from "./codex/stripPassthroughRejectedParams.ts";
 import {
   CODEX_CLI_RS_ORIGINATOR,
   getCodexClientVersion,
@@ -1431,16 +1432,7 @@ export class CodexExecutor extends BaseExecutor {
     delete body.truncation;
     delete body.background; // Droid CLI sends this but Codex Responses API rejects it
 
-    // Issue #3317: strip client-only fields the Codex Responses API rejects with
-    // 400 "Unsupported parameter" — for BOTH the native passthrough (early return
-    // below) and the translated path. The chat-completions path already removes
-    // these (base.ts prompt_cache_retention #1884; openai-responses translator
-    // safety_identifier #2770), but the responses->responses passthrough skips
-    // translation. `user` is always rejected by Codex /responses, so it is removed
-    // unconditionally here (unlike base.ts, which only drops it when empty).
-    delete body.prompt_cache_retention;
-    delete body.safety_identifier;
-    delete body.user;
+    stripCodexPassthroughRejectedParams(cleanModel || model, body);
 
     // Inject prompt_cache_key for Codex prompt caching.
     // The official Codex client sets this to conversation_id (a stable UUID per session).
diff --git a/open-sse/executors/codex/stripPassthroughRejectedParams.ts b/open-sse/executors/codex/stripPassthroughRejectedParams.ts
new file mode 100644
index 0000000000..fc998d32bc
--- /dev/null
+++ b/open-sse/executors/codex/stripPassthroughRejectedParams.ts
@@ -0,0 +1,27 @@
+// Strip fields Codex /responses rejects with 400 "Unsupported parameter"
+// BEFORE native passthrough returns (the Responses allowlist never runs on
+// that path). Extracted from CodexExecutor.transformRequest so codex.ts stays
+// under its frozen file-size cap.
+//
+// Client-only (#3317): prompt_cache_retention, safety_identifier, user.
+// The chat-completions path already removes these (base.ts prompt_cache_retention
+// #1884; openai-responses translator safety_identifier #2770), but the
+// responses->responses passthrough skips translation. `user` is always rejected
+// by Codex /responses, so it is removed unconditionally (unlike base.ts, which
+// only drops it when empty).
+//
+// Sampling: temperature, top_p — FastAPI `{"detail":"Unsupported parameter:
+// temperature"}`. Combo codex-review forwarded client temperature onto
+// gpt-5.6-sol-xhigh / gpt-5.6-luna-max.
+
+import { stripUnsupportedParams } from "../../translator/paramSupport.ts";
+
+export function stripCodexPassthroughRejectedParams(
+  model: string,
+  body: Record<string, unknown>
+): void {
+  delete body.prompt_cache_retention;
+  delete body.safety_identifier;
+  delete body.user;
+  stripUnsupportedParams("codex", model, body);
+}
diff --git a/open-sse/services/gpt5SamplingGuard.ts b/open-sse/services/gpt5SamplingGuard.ts
index 729c1ce450..a258dd1b16 100644
--- a/open-sse/services/gpt5SamplingGuard.ts
+++ b/open-sse/services/gpt5SamplingGuard.ts
@@ -12,7 +12,9 @@
  * This guard removes `temperature`/`top_p` only when the resolved effort is active
  * (anything other than `none`). It is scoped to the `openai` provider (raw
  * api.openai.com Chat Completions): the `codex` provider's Responses requests are
- * already covered by the CodexExecutor allowlist (which drops both params), and
+ * already covered by CodexExecutor.transformRequest (STRIP_RULES drops
+ * temperature/top_p before native passthrough returns; the Responses allowlist
+ * also omits them on the translated path), and
  * other providers manage their own sampling rules.
  *
  * Refs: litellm#27351 (GPT-5.1 accepts temperature only when effort=none),
diff --git a/open-sse/translator/paramSupport.ts b/open-sse/translator/paramSupport.ts
index 85dcec35ae..79d2b5219e 100644
--- a/open-sse/translator/paramSupport.ts
+++ b/open-sse/translator/paramSupport.ts
@@ -34,6 +34,13 @@ const STRIP_RULES: StripRule[] = [
   { match: /claude-opus-4/i, drop: ["temperature"] },
   // GitHub Copilot gpt-5.4: temperature unsupported.
   { provider: "github", match: /gpt-5\.4/i, drop: ["temperature"] },
+  // Codex /responses (chatgpt.com backend-api) rejects sampling params with
+  // FastAPI 400 `{"detail":"Unsupported parameter: temperature"}`. Native
+  // Codex passthrough returns before the Responses allowlist, so this rule
+  // must run from CodexExecutor.transformRequest via
+  // stripCodexPassthroughRejectedParams. Live: combo codex-review
+  // gpt-5.6-sol-xhigh / gpt-5.6-luna-max.
+  { provider: "codex", match: /.*/, drop: ["temperature", "top_p"] },
   // GitHub Copilot Claude (except opus/sonnet 4.6): thinking + reasoning_effort rejected. #713
   {
     provider: "github",
diff --git a/tests/unit/codex-responses-passthrough-strip-3317.test.ts b/tests/unit/codex-responses-passthrough-strip-3317.test.ts
index b925eb14d7..4cfcb0b0c2 100644
--- a/tests/unit/codex-responses-passthrough-strip-3317.test.ts
+++ b/tests/unit/codex-responses-passthrough-strip-3317.test.ts
@@ -25,6 +25,8 @@ test("codex native responses passthrough strips client-only params (#3317)", asy
     safety_identifier: "droid-user-123",
     user: "user-abc",
     max_output_tokens: 16,
+    temperature: 0.7,
+    top_p: 0.9,
   };
 
   const result = (await executor.transformRequest("gpt-5.5", body, false, {} as never)) as Record<
@@ -35,6 +37,12 @@ test("codex native responses passthrough strips client-only params (#3317)", asy
   assert.equal(result.prompt_cache_retention, undefined, "prompt_cache_retention must be stripped");
   assert.equal(result.safety_identifier, undefined, "safety_identifier must be stripped");
   assert.equal(result.user, undefined, "user must be stripped");
+  assert.equal(
+    result.temperature,
+    undefined,
+    "temperature must be stripped before native passthrough"
+  );
+  assert.equal(result.top_p, undefined, "top_p must be stripped before native passthrough");
   // The real request payload must survive the strip.
   assert.ok(Array.isArray(result.input), "input array preserved");
 });
diff --git a/tests/unit/executors-strip-unsupported-params.test.ts b/tests/unit/executors-strip-unsupported-params.test.ts
index 9e98a9f9de..914c63be78 100644
--- a/tests/unit/executors-strip-unsupported-params.test.ts
+++ b/tests/unit/executors-strip-unsupported-params.test.ts
@@ -57,6 +57,25 @@ test("stripUnsupportedParams: github + gpt-5 (non-5.4) keeps temperature", () =>
   assert.equal(body.temperature, 1);
 });
 
+test("stripUnsupportedParams: codex strips temperature and top_p (Responses 400)", () => {
+  const body: Record<string, unknown> = {
+    temperature: 0.7,
+    top_p: 0.9,
+    model: "gpt-5.6-luna-max",
+    input: [],
+  };
+  stripUnsupportedParams("codex", "gpt-5.6-sol-xhigh", body);
+  assert.equal(body.temperature, undefined, "Codex /responses rejects temperature");
+  assert.equal(body.top_p, undefined, "Codex /responses rejects top_p");
+  assert.equal(body.model, "gpt-5.6-luna-max", "other params must survive");
+});
+
+test("stripUnsupportedParams: non-codex provider keeps temperature for gpt-5.6-luna-max", () => {
+  const body: Record<string, unknown> = { temperature: 0.7 };
+  stripUnsupportedParams("openai", "gpt-5.6-luna-max", body);
+  assert.equal(body.temperature, 0.7, "codex sampling strip is provider-scoped");
+});
+
 test("stripUnsupportedParams: github + Claude strips thinking + reasoning_effort", () => {
   const body: Record<string, unknown> = {
     thinking: { type: "enabled" },
@@ -193,11 +212,19 @@ test("stripUnsupportedParams: volcengine kimi-k2-5-260127 also clamps max_comple
 test("stripUnsupportedParams: volcengine non-kimi model (glm-4-7-251222) is NOT clamped by the kimi rule", () => {
   const body: Record<string, unknown> = { max_tokens: 65536 };
   stripUnsupportedParams("volcengine", "glm-4-7-251222", body);
-  assert.equal(body.max_tokens, 65536, "kimi-specific cap must not apply to other volcengine models");
+  assert.equal(
+    body.max_tokens,
+    65536,
+    "kimi-specific cap must not apply to other volcengine models"
+  );
 });
 
 test("stripUnsupportedParams: kimi rule is provider-scoped (no-op for non-volcengine providers)", () => {
   const body: Record<string, unknown> = { max_tokens: 65536 };
   stripUnsupportedParams("kimi", "kimi-k2-5-260127", body);
-  assert.equal(body.max_tokens, 65536, "the Ark-specific cap must not leak to other kimi-hosting providers");
+  assert.equal(
+    body.max_tokens,
+    65536,
+    "the Ark-specific cap must not leak to other kimi-hosting providers"
+  );
 });