fix(translator): provider thinking compatibility (DeepSeek/Gemini)

open-sse/translator/request/openai-to-gemini.ts

@@ -282,15 +282,23 @@ function openaiToGeminiBase(
   }
 
   // Thinking / Reasoning support (Google Gemini 2.0+ Thinking models)
-  // 1. OpenAI format: reasoning_effort (low/medium/high)
+  // 1. OpenAI format: reasoning_effort (low/medium/high/auto/max/xhigh)
+  // "auto", "max", and "xhigh" are clamped to the high-tier budget because Gemini
+  // does not accept these strings directly. "auto" signals "use max reasonable effort"
+  // which maps to high. "max"/"xhigh" exceed Gemini's accepted range and are clamped.
+  // Port of decolua/9router#2043 by @nguyenxvotanminh3.
   if (body.reasoning_effort) {
+    const highBudget = capThinkingBudget(model, 32768);
     const budgetMap: Record<string, number> = {
       low: 1024,
       medium: getDefaultThinkingBudget(model) || 8192,
-      high: capThinkingBudget(model, 32768),
+      high: highBudget,
+      auto: highBudget,
+      max: highBudget,
+      xhigh: highBudget,
     };
     const budget =
-      budgetMap[body.reasoning_effort as string] || getDefaultThinkingBudget(model) || 8192;
+      budgetMap[body.reasoning_effort as string] ?? getDefaultThinkingBudget(model) ?? 8192;
     result.generationConfig.thinkingConfig = {
       thinkingBudget: budget,
       includeThoughts: true,

tests/unit/translator-thinking-provider-compat-2043.test.ts

@@ -0,0 +1,187 @@
+/**
+ * Tests for provider thinking compatibility fixes ported from decolua/9router#2043.
+ *
+ * Covers:
+ *   (a) DeepSeek unsigned thinking placeholders — no `signature` field in injected blocks.
+ *   (b) DeepSeek keeps existing thinking blocks as-is (not overwritten by cache/placeholder).
+ *   (c) Gemini `reasoning_effort: "auto"` maps to high budget (not silent fallback to default).
+ *   (d) Gemini `reasoning_effort: "max"` / `"xhigh"` clamp to high budget.
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { prepareClaudeRequest } = await import("../../open-sse/translator/helpers/claudeHelper.ts");
+const reasoningCache = await import("../../open-sse/services/reasoningCache.ts");
+const { openaiToGeminiRequest } =
+  await import("../../open-sse/translator/request/openai-to-gemini.ts");
+const { capThinkingBudget } = await import("../../src/lib/modelCapabilities.ts");
+
+// ──────────────── Fix (a): DeepSeek placeholder has NO signature field ────────────────
+
+test("deepseek: injected thinking placeholder has no signature field", () => {
+  reasoningCache.clearReasoningCacheAll();
+  const body = {
+    model: "deepseek-v4-pro",
+    thinking: { type: "enabled", budget_tokens: 4096 },
+    messages: [
+      { role: "user", content: [{ type: "text", text: "q" }] },
+      {
+        role: "assistant",
+        content: [{ type: "tool_use", id: "toolu_1", name: "Read", input: { file_path: "x" } }],
+      },
+      {
+        role: "user",
+        content: [{ type: "tool_result", tool_use_id: "toolu_1", content: "ok" }],
+      },
+    ],
+  };
+  const out = prepareClaudeRequest(body as any, "deepseek");
+  const assistant = (out as any).messages.find((m: any) => m.role === "assistant");
+  assert.ok(assistant, "assistant message should exist");
+  assert.equal(assistant.content[0].type, "thinking", "injected block should be type=thinking");
+  assert.equal(
+    assistant.content[0].signature,
+    undefined,
+    "DeepSeek placeholder must NOT have a signature field"
+  );
+  assert.equal(assistant.content[1].type, "tool_use", "tool_use should follow the thinking block");
+});
+
+// ──────────────── Fix (b): DeepSeek keeps existing thinking blocks as-is ────────────────
+
+test("deepseek: existing thinking blocks are preserved as-is (text and type unchanged)", () => {
+  reasoningCache.clearReasoningCacheAll();
+  const body = {
+    model: "deepseek-v4-pro",
+    thinking: { type: "enabled", budget_tokens: 4096 },
+    messages: [
+      { role: "user", content: [{ type: "text", text: "q" }] },
+      {
+        role: "assistant",
+        content: [
+          { type: "thinking", thinking: "my actual reasoning", signature: "ds-sig" },
+          { type: "tool_use", id: "toolu_2", name: "Read", input: {} },
+        ],
+      },
+      {
+        role: "user",
+        content: [{ type: "tool_result", tool_use_id: "toolu_2", content: "ok" }],
+      },
+    ],
+  };
+  const out = prepareClaudeRequest(body as any, "deepseek");
+  const assistant = (out as any).messages.find((m: any) => m.role === "assistant");
+  assert.equal(assistant.content[0].type, "thinking");
+  assert.equal(
+    assistant.content[0].thinking,
+    "my actual reasoning",
+    "DeepSeek must keep existing thinking text verbatim"
+  );
+  // Signature stripping is acceptable (DeepSeek doesn't need it), but text must not
+  // be overwritten with placeholder or cache content.
+  assert.ok(
+    assistant.content.length >= 2,
+    "should not drop blocks: thinking + tool_use both present"
+  );
+  assert.equal(assistant.content[1].type, "tool_use");
+});
+
+test("deepseek: injected placeholder thinking text is non-empty dot sentinel", () => {
+  // The upstream uses "." as placeholder text for DeepSeek (not the long NON_ANTHROPIC placeholder)
+  reasoningCache.clearReasoningCacheAll();
+  const body = {
+    model: "deepseek-v4-pro",
+    thinking: { type: "enabled", budget_tokens: 4096 },
+    messages: [
+      { role: "user", content: [{ type: "text", text: "q" }] },
+      {
+        role: "assistant",
+        content: [{ type: "tool_use", id: "toolu_3", name: "Write", input: {} }],
+      },
+      {
+        role: "user",
+        content: [{ type: "tool_result", tool_use_id: "toolu_3", content: "done" }],
+      },
+    ],
+  };
+  const out = prepareClaudeRequest(body as any, "deepseek");
+  const assistant = (out as any).messages.find((m: any) => m.role === "assistant");
+  assert.equal(assistant.content[0].type, "thinking");
+  // Placeholder must be non-empty (DeepSeek rejects empty thinking text).
+  // The value is either "." (upstream canonical) or another non-empty fallback.
+  assert.ok(
+    typeof assistant.content[0].thinking === "string" && assistant.content[0].thinking.length > 0,
+    `thinking placeholder must be non-empty, got: ${JSON.stringify(assistant.content[0].thinking)}`
+  );
+  assert.equal(
+    assistant.content[0].signature,
+    undefined,
+    "no signature on injected DeepSeek block"
+  );
+});
+
+// ──────────────── Fix (c): Gemini reasoning_effort "auto" → high budget ────────────────
+
+test("Gemini: reasoning_effort 'auto' maps to a defined (non-zero) thinking budget", () => {
+  const out = openaiToGeminiRequest(
+    "gemini-3-pro",
+    {
+      messages: [{ role: "user", content: "hello" }],
+      reasoning_effort: "auto",
+    },
+    false
+  ) as any;
+
+  const thinkingBudget = out.generationConfig?.thinkingConfig?.thinkingBudget;
+  assert.ok(
+    typeof thinkingBudget === "number" && thinkingBudget > 0,
+    `reasoning_effort 'auto' should produce a positive thinkingBudget, got: ${thinkingBudget}`
+  );
+  // Specifically should be at least the high-tier budget level (clamped to model max)
+  const highBudget = capThinkingBudget("gemini-3-pro", 32768);
+  assert.equal(
+    thinkingBudget,
+    highBudget,
+    `reasoning_effort 'auto' should map to high budget (${highBudget})`
+  );
+});
+
+// ──────────────── Fix (d): Gemini reasoning_effort "max"/"xhigh" → high budget ────────────────
+
+test("Gemini: reasoning_effort 'max' clamps to high budget (not default fallback)", () => {
+  const out = openaiToGeminiRequest(
+    "gemini-3-pro",
+    {
+      messages: [{ role: "user", content: "hello" }],
+      reasoning_effort: "max",
+    },
+    false
+  ) as any;
+
+  const thinkingBudget = out.generationConfig?.thinkingConfig?.thinkingBudget;
+  const highBudget = capThinkingBudget("gemini-3-pro", 32768);
+  assert.equal(
+    thinkingBudget,
+    highBudget,
+    `reasoning_effort 'max' should clamp to high budget (${highBudget})`
+  );
+});
+
+test("Gemini: reasoning_effort 'xhigh' clamps to high budget (not default fallback)", () => {
+  const out = openaiToGeminiRequest(
+    "gemini-3-pro",
+    {
+      messages: [{ role: "user", content: "hello" }],
+      reasoning_effort: "xhigh",
+    },
+    false
+  ) as any;
+
+  const thinkingBudget = out.generationConfig?.thinkingConfig?.thinkingBudget;
+  const highBudget = capThinkingBudget("gemini-3-pro", 32768);
+  assert.equal(
+    thinkingBudget,
+    highBudget,
+    `reasoning_effort 'xhigh' should clamp to high budget (${highBudget})`
+  );
+});