ton/pi_harness
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
packages/ai/src/providers/google-gemini-cli.ts
+17
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) {