From e2b4c2b06e8204936d379571fb7e76b87ae29ed9 Mon Sep 17 00:00:00 2001
From: Anton <39598727+NomenAK@users.noreply.github.com>
Date: Thu, 14 May 2026 05:15:40 +0200
Subject: [PATCH] fix(antigravity): strip generationConfig.thinkingConfig for
 Claude models (#2217)

Integrated into release/v3.8.0
---
 .../translator/request/openai-to-gemini.ts    | 14 ++++++
 .../unit/translator-openai-to-gemini.test.ts  | 45 ++++++++++++++-----
 2 files changed, 49 insertions(+), 10 deletions(-)

diff --git a/open-sse/translator/request/openai-to-gemini.ts b/open-sse/translator/request/openai-to-gemini.ts
index abdd49d133..1d4d016826 100644
--- a/open-sse/translator/request/openai-to-gemini.ts
+++ b/open-sse/translator/request/openai-to-gemini.ts
@@ -547,6 +547,10 @@ export function openaiToAntigravityRequest(model, body, stream, credentials = nu
   // Match real Antigravity client: don't send maxOutputTokens when the user
   // hasn't explicitly specified max_tokens / max_completion_tokens.
   // The Cloud Code server decides the output limit on its own.
+  // Note: read hasThinking BEFORE stripping thinkingConfig below — for Claude
+  // models the Cloud Code envelope still carries a thinkingBudget set upstream
+  // by applyAntigravityGenerationDefaults, which we must consult here so we
+  // do not accidentally drop the maxOutputTokens it bumped for us.
   const clientRequestedMaxTokens = body.max_tokens ?? body.max_completion_tokens;
   const hasThinking = !!envelope.request?.generationConfig?.thinkingConfig?.thinkingBudget;
   if (
@@ -557,6 +561,16 @@ export function openaiToAntigravityRequest(model, body, stream, credentials = nu
     delete envelope.request.generationConfig.maxOutputTokens;
   }
 
+  // Claude models on Antigravity use their own native thinking — Gemini's thinkingConfig
+  // is not understood by the Cloud Code Claude endpoint and must be stripped.
+  // applyAntigravityGenerationDefaults (inside wrapInCloudCodeEnvelope) already bumped
+  // maxOutputTokens to thinkingBudget+1 before we get here, so the budget is preserved.
+  // Must run AFTER the hasThinking-derived maxOutputTokens decision above so the
+  // budget is accounted for before the field is removed.
+  if (isClaude && envelope.request?.generationConfig) {
+    delete envelope.request.generationConfig.thinkingConfig;
+  }
+
   return envelope;
 }
 
diff --git a/tests/unit/translator-openai-to-gemini.test.ts b/tests/unit/translator-openai-to-gemini.test.ts
index 8bf929d858..69f68d3783 100644
--- a/tests/unit/translator-openai-to-gemini.test.ts
+++ b/tests/unit/translator-openai-to-gemini.test.ts
@@ -756,7 +756,11 @@ test("OpenAI -> Antigravity Claude path sanitizes tool names for Gemini schema",
   assert.deepEqual(toolResultBlock.response, { result: { ok: true } });
 });
 
-test("OpenAI -> Antigravity Claude path applies output cap in generationConfig", () => {
+test("OpenAI -> Antigravity Claude path applies output cap and strips thinkingConfig", () => {
+  // For Claude on Antigravity, applyAntigravityGenerationDefaults must bump
+  // maxOutputTokens to thinkingBudget+1 BEFORE the envelope strips thinkingConfig
+  // (because Claude on Cloud Code does not understand Gemini's thinkingConfig
+  // shape but still benefits from the larger output cap derived from it).
   const result = openaiToAntigravityRequest(
     "claude-3-7-sonnet",
     {
@@ -769,15 +773,14 @@ test("OpenAI -> Antigravity Claude path applies output cap in generationConfig",
   );
 
   assert.equal((result as any).request?.generationConfig.maxOutputTokens, 32769);
-  assert.deepEqual((result as any).request?.generationConfig.thinkingConfig, {
-    thinkingBudget: 32768,
-    includeThoughts: true,
-  });
+  // thinkingConfig must be stripped for Claude — Cloud Code Claude endpoint
+  // does not understand the Gemini-shape thinkingConfig field.
+  assert.equal((result as any).request?.generationConfig.thinkingConfig, undefined);
   assert.equal((result as any).request?.max_tokens, undefined);
   assert.equal((result as any).request?.thinking, undefined);
 });
 
-test("OpenAI -> Antigravity Claude path preserves lower requested output", () => {
+test("OpenAI -> Antigravity Claude path preserves lower requested output and strips thinkingConfig", () => {
   const result = openaiToAntigravityRequest(
     "claude-3-7-sonnet",
     {
@@ -790,10 +793,32 @@ test("OpenAI -> Antigravity Claude path preserves lower requested output", () =>
   );
 
   assert.equal((result as any).request?.generationConfig.maxOutputTokens, 32769);
-  assert.deepEqual((result as any).request?.generationConfig.thinkingConfig, {
-    thinkingBudget: 32768,
-    includeThoughts: true,
-  });
+  assert.equal((result as any).request?.generationConfig.thinkingConfig, undefined);
   assert.equal((result as any).request?.max_tokens, undefined);
   assert.equal((result as any).request?.thinking, undefined);
 });
+
+test("OpenAI -> Antigravity Gemini path preserves thinkingConfig (only Claude is stripped)", () => {
+  // Negative-control for the Claude-thinkingConfig-strip behavior. Gemini models
+  // on Antigravity must still receive thinkingConfig — only Claude needs it removed
+  // (Cloud Code Claude endpoint does not understand the Gemini-shape field).
+  const result = openaiToAntigravityRequest(
+    "gemini-2.5-pro",
+    {
+      messages: [{ role: "user", content: "Summarize this" }],
+      max_completion_tokens: 32000,
+      reasoning_effort: "high",
+    },
+    false,
+    { projectId: "proj-gemini-thinking" } as any
+  );
+
+  // For Gemini, thinkingConfig must remain in place because the Cloud Code
+  // Gemini endpoint understands and uses it.
+  assert.ok(
+    (result as any).request?.generationConfig.thinkingConfig,
+    "thinkingConfig must be preserved for Gemini models on Antigravity"
+  );
+  assert.equal((result as any).request?.generationConfig.thinkingConfig.thinkingBudget > 0, true);
+  assert.equal((result as any).request?.generationConfig.thinkingConfig.includeThoughts, true);
+});