From f19d4b585b927816b37ad4e95ce190c5cd26d135 Mon Sep 17 00:00:00 2001
From: Diego Rodrigues de Sa e Souza
 <8016841+diegosouzapw@users.noreply.github.com>
Date: Thu, 18 Jun 2026 21:25:04 -0300
Subject: [PATCH] fix(sse): map reasoning_effort to DeepSeek V4's native {high,
 max} vocabulary (#4219)

Integrated into release/v3.8.29 (DeepSeek V4 native reasoning_effort {high,max} mapping).
---
 config/quality/file-size-baseline.json        |  2 +-
 open-sse/executors/base.ts                    | 24 +++++
 .../base-executor-sanitize-effort.test.ts     | 99 +++++++++++++++++++
 3 files changed, 124 insertions(+), 1 deletion(-)

diff --git a/config/quality/file-size-baseline.json b/config/quality/file-size-baseline.json
index 552d41029c..b371b7df87 100644
--- a/config/quality/file-size-baseline.json
+++ b/config/quality/file-size-baseline.json
@@ -58,7 +58,7 @@
   "frozen": {
     "open-sse/config/providerRegistry.ts": 4731,
     "open-sse/executors/antigravity.ts": 1664,
-    "open-sse/executors/base.ts": 1334,
+    "open-sse/executors/base.ts": 1358,
     "open-sse/executors/chatgpt-web.ts": 2870,
     "open-sse/executors/claude-web.ts": 1057,
     "open-sse/executors/codex.ts": 1447,
diff --git a/open-sse/executors/base.ts b/open-sse/executors/base.ts
index 46ac640e34..900117470c 100644
--- a/open-sse/executors/base.ts
+++ b/open-sse/executors/base.ts
@@ -288,6 +288,30 @@ export function sanitizeReasoningEffortForProvider(
     return next;
   }
 
+  // Native DeepSeek (api.deepseek.com) — V4 thinking mode accepts reasoning_effort
+  // ONLY as {high, max} (its own top tier is literally "max"). OmniRoute's internal
+  // scale is low|medium|high|xhigh where xhigh is the top, so map onto DeepSeek's
+  // vocabulary: xhigh → max (top→top), low|medium → high (below the enum floor).
+  // high/max pass through unchanged. Without this, the claude→openai translator's
+  // xhigh (and max-normalized-to-xhigh below) reaches DeepSeek as an unknown value,
+  // silently dropping the client's requested effort. This is the INVERSE of the
+  // OpenRouter-DeepSeek path, whose normalized API expects xhigh, not max (pi#4055).
+  if (provider === "deepseek") {
+    const mapped =
+      effortStr === "xhigh" ? "max" : effortStr === "low" || effortStr === "medium" ? "high" : null;
+    if (mapped && mapped !== effortStr) {
+      log?.info?.(
+        "REASONING_SANITIZE",
+        `deepseek/${modelStr}: normalized reasoning_effort ${effortStr} → ${mapped}`
+      );
+      const next: Record<string, unknown> = { ...b };
+      if (hasTopLevelReasoningEffort) next.reasoning_effort = mapped;
+      if (reasoning) next.reasoning = { ...reasoning, effort: mapped };
+      return next;
+    }
+    return body;
+  }
+
   const supportsXHigh = supportsXHighEffort(provider, modelStr);
   const shouldDowngradeXHigh = effortStr === "xhigh" && !supportsXHigh;
   const supportsXHighForMax = supportsXHigh;
diff --git a/tests/unit/base-executor-sanitize-effort.test.ts b/tests/unit/base-executor-sanitize-effort.test.ts
index 81d6dc51d7..bcdbacff60 100644
--- a/tests/unit/base-executor-sanitize-effort.test.ts
+++ b/tests/unit/base-executor-sanitize-effort.test.ts
@@ -337,3 +337,102 @@ test("sanitizeReasoningEffortForProvider: non-object body returns unchanged", ()
   const arr: unknown[] = [];
   assert.equal(sanitizeReasoningEffortForProvider(arr, "xiaomi-mimo", "x", null), arr);
 });
+
+// ── Native DeepSeek (api.deepseek.com) ───────────────────────────────────────
+// DeepSeek V4 thinking mode accepts reasoning_effort ONLY as {high, max}. The
+// internal OmniRoute scale (low|medium|high|xhigh, xhigh = top) must be mapped
+// onto DeepSeek's native vocabulary so the client's requested effort is honored
+// instead of silently dropped to the default. This is the INVERSE of the
+// OpenRouter-DeepSeek path, whose normalized API expects xhigh, not max.
+
+test("sanitizeReasoningEffortForProvider: native deepseek maps xhigh → max", () => {
+  const log = makeLog();
+  const body = {
+    model: "deepseek-v4-pro",
+    reasoning_effort: "xhigh",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-pro", log);
+  assert.notEqual(result, body, "must return a new object when mutating");
+  assert.equal((result as any).reasoning_effort, "max");
+  assert.equal((result as any).model, "deepseek-v4-pro", "other fields preserved");
+  assert.ok(
+    log.messages.some(([tag, m]) => tag === "REASONING_SANITIZE" && /xhigh → max/.test(m)),
+    "logs the xhigh → max mapping"
+  );
+});
+
+test("sanitizeReasoningEffortForProvider: native deepseek preserves max", () => {
+  const log = makeLog();
+  const body = {
+    model: "deepseek-v4-flash",
+    reasoning_effort: "max",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-flash", log);
+  assert.equal(result, body, "max is DeepSeek's native top tier — passes through unchanged");
+  assert.equal((result as any).reasoning_effort, "max");
+  assert.equal(log.messages.length, 0);
+});
+
+test("sanitizeReasoningEffortForProvider: native deepseek clamps low → high", () => {
+  const body = {
+    model: "deepseek-v4-pro",
+    reasoning_effort: "low",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-pro", null);
+  assert.notEqual(result, body, "must return a new object when mutating");
+  assert.equal((result as any).reasoning_effort, "high", "below the {high, max} floor → high");
+});
+
+test("sanitizeReasoningEffortForProvider: native deepseek clamps medium → high", () => {
+  const body = {
+    model: "deepseek-v4-pro",
+    reasoning_effort: "medium",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-pro", null);
+  assert.equal((result as any).reasoning_effort, "high");
+});
+
+test("sanitizeReasoningEffortForProvider: native deepseek preserves high unchanged", () => {
+  const body = {
+    model: "deepseek-v4-pro",
+    reasoning_effort: "high",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-pro", null);
+  assert.equal(result, body, "high is already valid — passes through unchanged");
+  assert.equal((result as any).reasoning_effort, "high");
+});
+
+test("sanitizeReasoningEffortForProvider: native deepseek maps nested reasoning.effort xhigh → max", () => {
+  const body = {
+    model: "deepseek-v4-pro",
+    reasoning: { effort: "xhigh", summary: "auto" },
+    input: [],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4-pro", null);
+  assert.equal((result as any).reasoning.effort, "max");
+  assert.equal((result as any).reasoning.summary, "auto", "other reasoning fields preserved");
+  assert.equal((result as any).reasoning_effort, undefined);
+});
+
+test("sanitizeReasoningEffortForProvider: OpenRouter DeepSeek still preserves xhigh (not native)", () => {
+  // Regression guard: the native-deepseek mapping must NOT touch openrouter,
+  // whose normalized API expects xhigh (issue earendil-works/pi#4055).
+  const body = {
+    model: "deepseek/deepseek-v4-pro",
+    reasoning_effort: "xhigh",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "openrouter",
+    "deepseek/deepseek-v4-pro",
+    null
+  );
+  assert.equal(result, body);
+  assert.equal((result as any).reasoning_effort, "xhigh");
+});