@ssweens/pi-vertex 1.1.7 → 1.1.9

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 All notable changes to this project will be documented in this file.
 
+## [1.1.9] - 2026-05-19
+### Fixed
+- **Gemini 3/3.5 default thinking preserved** — previously, if no reasoning effort level was specified, we silently overrode Gemini 3/3.5 models to use their lowest thinking level (`MINIMAL`). This severely degraded the intelligence of Gemini 3.5 Flash (defaulting it to no real reasoning/thinking) and Gemini 3 Pro/Flash. We now completely omit `thinkingConfig` when `options.reasoning` is undefined, allowing Gemini 3/3.5 models to use their native GA defaults (`MEDIUM` for 3.5 Flash, `HIGH` for others).
+- **Gemini 2.5 default thinking budget** — Gemini 2.5 has thinking disabled by default on Vertex, so we now apply a healthy thinking budget floor (`2048` tokens for Pro, `1024` for Flash) when `options.reasoning` is undefined.
+
+## [1.1.8] - 2026-05-19
+### Added
+- **Gemini 3.5 Flash** (`gemini-3.5-flash`) — GA release from Google I/O 2026. Most intelligent Flash model; optimized for agentic execution, coding, and long-horizon tasks. 1M context, 65K max output, text/image/video/audio input, reasoning (thinking levels: minimal/low/medium/high), tools. $1.50/$9.00 per 1M tokens (global), $0.15/1M cache read.
+
 ## [1.1.7] - 2026-05-16
 ### Added
 - **Regional pricing for Claude models** — non-global Vertex endpoints (us-east5, europe-west1, asia-southeast1, us/eu multi-region) carry a 10% price premium per GCP's published rates. The streaming layer now automatically selects the correct cost tier based on the resolved endpoint at call time. No config change required — if your `GOOGLE_CLOUD_LOCATION` or config resolves to any non-`global` location, cost tracking reflects the regional rate.

package/models/gemini.ts

@@ -8,6 +8,27 @@
 import type { VertexModelConfig } from "../types.js";
 
 export const GEMINI_MODELS: VertexModelConfig[] = [
+  // --- Gemini 3.5 (GA) ---
+  {
+    id: "gemini-3.5-flash",
+    name: "Gemini 3.5 Flash",
+    apiId: "gemini-3.5-flash",
+    publisher: "google",
+    endpointType: "gemini",
+    contextWindow: 1048576,
+    maxTokens: 65536,
+    input: ["text", "image"],
+    reasoning: true,
+    tools: true,
+    cost: {
+      input: 1.50,
+      output: 9.00,
+      cacheRead: 0.15,
+      cacheWrite: 0,
+    },
+    region: "global",
+  },
+
   // --- Gemini 3.1 (Preview) ---
   {
     id: "gemini-3.1-pro",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ssweens/pi-vertex",
-  "version": "1.1.7",
+  "version": "1.1.9",
   "description": "Google Vertex AI provider for Pi coding agent - supports Gemini, Claude, and all MaaS models",
   "type": "module",
   "main": "index.ts",

package/streaming/gemini.ts

@@ -52,19 +52,6 @@ function getGemini3ThinkingLevel(effort: string, modelId: string): ThinkingLevel
   return THINKING_LEVEL_MAP[effort];
 }
 
-function getLowestThinkingConfig(modelId: string): GeminiThinkingConfig {
-  if (isGemini3ProModel(modelId)) {
-    return { thinkingLevel: ThinkingLevel.LOW };
-  }
-  if (isGemini3FlashModel(modelId)) {
-    return { thinkingLevel: ThinkingLevel.MINIMAL };
-  }
-  if (isGemini25ProModel(modelId)) {
-    return { thinkingBudget: 128 };
-  }
-  return { thinkingBudget: 0 };
-}
-
 function mapGeminiStopReason(reason: string): "stop" | "length" | "toolUse" | "error" {
   switch (reason) {
     case FinishReason.STOP:
@@ -137,8 +124,6 @@ export function streamGemini(
       }
 
       // Add thinking configuration (matches pi-mono's buildParams logic).
-      // For reasoning models: always set a minimum thinking config so the model
-      // doesn't silently suppress thoughts when no effort level is specified.
       if (model.reasoning) {
         if (options?.reasoning) {
           const effort = options.reasoning === "xhigh" ? "high" : options.reasoning;
@@ -161,7 +146,18 @@ export function streamGemini(
 
           config.thinkingConfig = thinkingConfig;
         } else {
-          config.thinkingConfig = getLowestThinkingConfig(model.apiId);
+          // If no reasoning level is specified:
+          // - For Gemini 3.x/3.5 models, omit thinkingConfig entirely so Vertex AI uses
+          //   the model's native default level (e.g. MEDIUM for 3.5, HIGH for others).
+          // - For Gemini 2.5 models, apply a healthy thinking budget floor (thinking is
+          //   disabled by default on 2.5).
+          const isGemini3 = model.apiId.startsWith("gemini-3");
+          if (!isGemini3) {
+            config.thinkingConfig = {
+              includeThoughts: true,
+              thinkingBudget: model.apiId.includes("2.5-pro") ? 2048 : 1024,
+            };
+          }
         }
       }