shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(ai): handle explicit thinking off across providers closes #2490

Browse Source
This commit is contained in:
Mario Zechner
2026-03-22 20:27:44 +01:00
parent 6129971c04
commit d1613e3f53
8 changed files with 246 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
packages/ai/src/providers/azure-openai-responses.ts
View File

@@ -240,18 +240,7 @@ function buildParams(
};
params.include = ["reasoning.encrypted_content"];
} else {
if (model.name.toLowerCase().startsWith("gpt-5")) {
// Jesus Christ, see https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7
messages.push({
role: "developer",
content: [
{
type: "input_text",
text: "# Juice: 0 !important",
},
],
});
}
params.reasoning = { effort: "none" };
}
}

17
packages/ai/src/providers/google-gemini-cli.ts
View File

@@ -889,6 +889,8 @@ export function buildRequest(
} else if (options.thinking.budgetTokens !== undefined) {
generationConfig.thinkingConfig.thinkingBudget = options.thinking.budgetTokens;
}
} else if (model.reasoning && options.thinking && !options.thinking.enabled) {
generationConfig.thinkingConfig = getDisabledThinkingConfig(model.id);
}
const request: CloudCodeAssistRequest["request"] = {
@@ -946,6 +948,21 @@ export function buildRequest(
type ClampedThinkingLevel = Exclude<ThinkingLevel, "xhigh">;
function getDisabledThinkingConfig(modelId: string): ThinkingConfig {
// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite
// do not support full thinking-off either. For Gemini 3 models, use the lowest supported
// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.
if (isGemini3ProModel(modelId)) {
return { thinkingLevel: "LOW" as any };
}
if (isGemini3FlashModel(modelId)) {
return { thinkingLevel: "MINIMAL" as any };
}
// Gemini 2.x supports disabling via thinkingBudget = 0.
return { thinkingBudget: 0 };
}
function getGeminiCliThinkingLevel(effort: ClampedThinkingLevel, modelId: string): GoogleThinkingLevel {
if (isGemini3ProModel(modelId)) {
switch (effort) {

18
packages/ai/src/providers/google-vertex.ts
View File

@@ -436,6 +436,8 @@ function buildParams(
thinkingConfig.thinkingBudget = options.thinking.budgetTokens;
}
config.thinkingConfig = thinkingConfig;
} else if (model.reasoning && options.thinking && !options.thinking.enabled) {
config.thinkingConfig = getDisabledThinkingConfig(model);
}
if (options.signal) {
@@ -464,6 +466,22 @@ function isGemini3FlashModel(model: Model<"google-generative-ai">): boolean {
return /gemini-3(?:\.\d+)?-flash/.test(model.id.toLowerCase());
}
function getDisabledThinkingConfig(model: Model<"google-vertex">): ThinkingConfig {
// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite
// do not support full thinking-off either. For Gemini 3 models, use the lowest supported
// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.
const geminiModel = model as unknown as Model<"google-generative-ai">;
if (isGemini3ProModel(geminiModel)) {
return { thinkingLevel: ThinkingLevel.LOW };
}
if (isGemini3FlashModel(geminiModel)) {
return { thinkingLevel: ThinkingLevel.MINIMAL };
}
// Gemini 2.x supports disabling via thinkingBudget = 0.
return { thinkingBudget: 0 };
}
function getGemini3ThinkingLevel(
effort: ClampedThinkingLevel,
model: Model<"google-generative-ai">,

17
packages/ai/src/providers/google.ts
View File

@@ -371,6 +371,8 @@ function buildParams(
thinkingConfig.thinkingBudget = options.thinking.budgetTokens;
}
config.thinkingConfig = thinkingConfig;
} else if (model.reasoning && options.thinking && !options.thinking.enabled) {
config.thinkingConfig = getDisabledThinkingConfig(model);
}
if (options.signal) {
@@ -399,6 +401,21 @@ function isGemini3FlashModel(model: Model<"google-generative-ai">): boolean {
return /gemini-3(?:\.\d+)?-flash/.test(model.id.toLowerCase());
}
function getDisabledThinkingConfig(model: Model<"google-generative-ai">): ThinkingConfig {
// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite
// do not support full thinking-off either. For Gemini 3 models, use the lowest supported
// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.
if (isGemini3ProModel(model)) {
return { thinkingLevel: "LOW" as any };
}
if (isGemini3FlashModel(model)) {
return { thinkingLevel: "MINIMAL" as any };
}
// Gemini 2.x supports disabling via thinkingBudget = 0.
return { thinkingBudget: 0 };
}
function getGemini3ThinkingLevel(
effort: ClampedThinkingLevel,
model: Model<"google-generative-ai">,

12
packages/ai/src/providers/openai-completions.ts
View File

@@ -410,12 +410,16 @@ function buildParams(model: Model<"openai-completions">, context: Context, optio
(params as any).enable_thinking = !!options?.reasoningEffort;
} else if (compat.thinkingFormat === "qwen-chat-template" && model.reasoning) {
(params as any).chat_template_kwargs = { enable_thinking: !!options?.reasoningEffort };
} else if (compat.thinkingFormat === "openrouter" && options?.reasoningEffort && model.reasoning) {
} else if (compat.thinkingFormat === "openrouter" && model.reasoning) {
// OpenRouter normalizes reasoning across providers via a nested reasoning object.
const openRouterParams = params as typeof params & { reasoning?: { effort?: string } };
openRouterParams.reasoning = {
effort: mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap),
};
if (options?.reasoningEffort) {
openRouterParams.reasoning = {
effort: mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap),
};
} else {
openRouterParams.reasoning = { effort: "none" };
}
} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {
// OpenAI-style reasoning_effort
(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);

13
packages/ai/src/providers/openai-responses.ts
View File

@@ -221,18 +221,7 @@ function buildParams(model: Model<"openai-responses">, context: Context, options
};
params.include = ["reasoning.encrypted_content"];
} else {
if (model.name.startsWith("gpt-5")) {
// Jesus Christ, see https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7
messages.push({
role: "developer",
content: [
{
type: "input_text",
text: "# Juice: 0 !important",
},
],
});
}
params.reasoning = { effort: "none" };
}
}