diff --git a/src/adapters/openai-responses/reasoning.ts b/src/adapters/openai-responses/reasoning.ts
index ae585b4c45b..82e20df262d 100644
--- a/src/adapters/openai-responses/reasoning.ts
+++ b/src/adapters/openai-responses/reasoning.ts
@@ -209,8 +209,26 @@ export function mapRoutedResponsesReasoningEffort(
   modelId: string,
 ): unknown {
   if (provider.authMode === "forward") return body;
-  if (configuredReasoningEfforts(provider, modelId) === undefined) return body;
   if (!isPlainObject(body) || !isPlainObject(body.reasoning)) return body;
+  // An unconfigured provider still needs the wire mapper: it folds caller spellings
+  // outside the Codex ladder ("minimal"->"low", "ultra"->"max") and drops unknown
+  // values before they can hit a strict upstream's enum validator. Skipping this on
+  // `configuredReasoningEfforts === undefined` let those spellings leak through
+  // verbatim (observed: DeepSeek-channel 400 "Invalid option" on "minimal"/"ultra").
+  if (configuredReasoningEfforts(provider, modelId) === undefined) {
+    const effort = body.reasoning.effort;
+    // "none" and provider-wire spellings the mapper cannot rank ("enabled"/"disabled"/
+    // "adaptive") pass through verbatim — the mapper's undefined for those means
+    // "not a Codex rung", not "invalid", and downstream adapters translate them.
+    if (typeof effort !== "string" || effort === "none"
+      || effort === "enabled" || effort === "disabled" || effort === "adaptive") return body;
+    const mappedUnconfigured = mapReasoningEffort(provider, modelId, effort);
+    if (mappedUnconfigured === effort) return body;
+    const nextReasoning = { ...body.reasoning };
+    if (mappedUnconfigured === undefined) delete nextReasoning.effort;
+    else nextReasoning.effort = mappedUnconfigured;
+    return { ...body, reasoning: Object.keys(nextReasoning).length > 0 ? nextReasoning : undefined };
+  }
   const declaredEfforts = modelRecordValue(provider.modelReasoningEfforts, modelId) ?? provider.reasoningEfforts;
   // An explicitly empty ladder means no effort control, not no reasoning output.
   // Omit only effort so the upstream default applies; unknown/non-rankable ladders stay untouched.
diff --git a/src/server/chat-native.ts b/src/server/chat-native.ts
index 69361aa9f93..98dcc6fe236 100644
--- a/src/server/chat-native.ts
+++ b/src/server/chat-native.ts
@@ -116,7 +116,22 @@ function normalizePinnedChatEffort(options: HandleNativeChatOptions): void {
   const pinned = !compaction
     ? resolvePinnedEffort(route, selector, config)
     : undefined;
-  let normalizeForWire = false;
+  // Run the wire mapper whenever a reasoning_effort is present on the body, not only
+  // after a pin/cap rewrite: mapReasoningEffort is the identity for valid ladder
+  // spellings and folds "minimal"->"low" / "ultra"->"max" before they can leak to a
+  // strict upstream. "none" is a real caller instruction (thinking disable), not an
+  // invalid spelling — keep it verbatim so downstream adapters can translate it to
+  // thinking:{type:"disabled"} instead of dropping the field (absent ≠ disabled).
+  const callerEffort = typeof chatBody.reasoning_effort === "string" ? chatBody.reasoning_effort : undefined;
+  // Compaction turns keep their exemption: they are maintenance, not agent turns, so
+  // their effort reaches the wire untouched just as pin/cap leave them alone.
+  // "none" and provider-wire spellings the mapper cannot rank ("enabled"/"disabled"/
+  // "adaptive" — consumed by thinkingToggleModels adapters) also pass through: only
+  // spellings the mapper can fold or clamp ("minimal", "ultra", unknown ranks) get
+  // rewritten, while values it returns undefined for would otherwise be deleted.
+  let normalizeForWire = !compaction && callerEffort !== undefined
+    && callerEffort !== "none" && callerEffort !== "enabled"
+    && callerEffort !== "disabled" && callerEffort !== "adaptive";
   if (pinned !== undefined) {
     logCtx.requestedEffort = from ? `${from}->${pinned}` : pinned;
     if (pinned === "none") delete chatBody.reasoning_effort;
diff --git a/tests/codex-integration/model-pinned-effort.test.ts b/tests/codex-integration/model-pinned-effort.test.ts
index 9aa8063f57b..75d7624688d 100644
--- a/tests/codex-integration/model-pinned-effort.test.ts
+++ b/tests/codex-integration/model-pinned-effort.test.ts
@@ -435,10 +435,12 @@ describe("operator pins on the actual request wire", () => {
     expect(wire.reasoning).toEqual({ effort: "max", summary: "auto" });
   });
 
-  test("native Chat without pins caps spawned children and preserves unqualified caller spelling", async () => {
+  test("native Chat without pins caps spawned children and maps unqualified caller spelling through the wire map", async () => {
     const c = config({ reasoningEfforts: ["low"], reasoningEffortMap: { max: "enabled" } }, { effortCap: "low", subagentEffortCap: "low" });
     expect((await request(c, "chat", { reasoning_effort: "ultra" }, { "x-openai-subagent": "collab_spawn" })).reasoning_effort).toBe("low");
-    expect((await request(c, "chat", { reasoning_effort: "ultra" })).reasoning_effort).toBe("ultra");
+    // Unpinned main turns still run mapReasoningEffort: ultra folds to max, then the
+    // provider wire alias maps max -> enabled (was: verbatim "ultra" passthrough).
+    expect((await request(c, "chat", { reasoning_effort: "ultra" })).reasoning_effort).toBe("enabled");
     expect(Object.hasOwn(await request(c, "chat", { reasoning_effort: undefined }), "reasoning_effort")).toBe(false);
   });
 
@@ -448,14 +450,18 @@ describe("operator pins on the actual request wire", () => {
     expect((await request(c, "chat", { tools, reasoning_effort: "ultra" })).reasoning_effort).toBe("medium");
     expect((await request(c, "chat", { tools, reasoning_effort: "ultra" }, { "x-openai-subagent": "collab_spawn" })).reasoning_effort).toBe("low");
     c.multiAgentMode = "v1";
-    expect((await request(c, "chat", { tools, reasoning_effort: "ultra" }, { "x-openai-subagent": "collab_spawn" })).reasoning_effort).toBe("ultra");
+    // Cap-exempt v1 turn: effort still passes through the wire mapper, so ultra folds
+    // to max (was: verbatim "ultra" passthrough).
+    expect((await request(c, "chat", { tools, reasoning_effort: "ultra" }, { "x-openai-subagent": "collab_spawn" })).reasoning_effort).toBe("max");
   });
 
   test("native Chat maps newly capped values and preserves lower, non-ladder and absent efforts", async () => {
     const c = config({ reasoningEffortMap: { medium: "enabled", low: "disabled" } }, { subagentEffortCap: "medium" });
     const headers = { "x-codex-turn-metadata": JSON.stringify({ subagent_kind: "thread_spawn" }) };
     expect((await request(c, "chat", { reasoning_effort: "ultra" }, headers)).reasoning_effort).toBe("enabled");
-    expect((await request(c, "chat", { reasoning_effort: "low" }, headers)).reasoning_effort).toBe("low");
+    // A low value the cap leaves alone still resolves through the provider wire map:
+    // low -> disabled (was: verbatim "low" passthrough).
+    expect((await request(c, "chat", { reasoning_effort: "low" }, headers)).reasoning_effort).toBe("disabled");
     expect((await request(c, "chat", { reasoning_effort: "enabled" }, headers)).reasoning_effort).toBe("enabled");
     expect(Object.hasOwn(await request(c, "chat", { reasoning_effort: undefined }, headers), "reasoning_effort")).toBe(false);
     c.providers.fixture!.pinnedReasoningEffort = "medium";