pi-banana 2.5.1 → 2.5.3

package/extensions/index.ts

@@ -30,7 +30,7 @@ import {
 	Text,
 } from "@earendil-works/pi-tui";
 import { GoogleGenAI } from "@google/genai";
-import { existsSync } from "node:fs";
+import { existsSync, readFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { mkdir, readFile, writeFile } from "node:fs/promises";
 import { extname, isAbsolute, resolve } from "node:path";
@@ -53,7 +53,7 @@ const ASPECT_RATIOS = [
 	"21:9",
 ] as const;
 const IMAGE_SIZES = ["1K", "2K", "4K"] as const;
-const IMAGE_QUALITY = ["lite", "fast", "high"] as const;
+const IMAGE_QUALITY = ["lite", "fast", "high", "max"] as const;
 const VISION_QUALITY = ["fast", "high"] as const;
 type ImageQuality = (typeof IMAGE_QUALITY)[number];
 
@@ -67,6 +67,7 @@ const MODEL_FOR_QUALITY: Record<ImageQuality, string> = {
 	lite: "gemini-3.1-flash-lite-image", // Nano Banana 2 Lite (ultra-low latency, cheapest)
 	fast: "gemini-3.1-flash-image", // Nano Banana 2
 	high: "gemini-3-pro-image", // Nano Banana Pro
+	max: "gemini-3-pro-image", // Google fallback; configured provider maps max to GPT Image 2
 };
 
 const DEFAULT_OUTPUT_DIR =
@@ -221,6 +222,19 @@ async function loadReferenceImage(
 	return { data: buf.toString("base64"), mimeType };
 }
 
+// Inline preview bytes, loaded at render time from the file the tool wrote.
+// Nothing image-sized is kept on the tool result itself: content goes to the
+// model and details go into the saved session, and a base64 PNG in either one
+// is counted against the context by request guards for the rest of the session.
+function readPreview(outputPath?: string): string | undefined {
+	if (typeof outputPath !== "string" || !outputPath) return undefined;
+	try {
+		return readFileSync(outputPath).toString("base64");
+	} catch {
+		return undefined; // moved, deleted, or unreadable: show the rest of the card
+	}
+}
+
 async function resolveOutputPath(
 	cwd: string,
 	prompt: string,
@@ -311,7 +325,7 @@ export default function (pi: ExtensionAPI) {
 			quality: Type.Optional(
 				StringEnum(IMAGE_QUALITY, {
 					description:
-						"'lite' = Nano Banana 2 Lite (ultra-low latency, cheapest). 'fast' = Nano Banana 2 (default, ~3s, cheap). 'high' = Nano Banana Pro (slower, top quality).",
+						"'lite' = Nano Banana 2 Lite (ultra-low latency, cheapest). 'fast' = Nano Banana 2 (default, ~3s, cheap). 'high' = Nano Banana Pro (strong editing). 'max' = premium benchmark tier, GPT Image 2 when the configured provider supports it.",
 					default: DEFAULT_QUALITY,
 				}),
 			),
@@ -362,7 +376,7 @@ export default function (pi: ExtensionAPI) {
 			const provider = await resolveProvider(ctx);
 			if (provider.kind === "openai-compat") {
 				const { config } = provider;
-				const oaiModel = config.models[quality] ?? config.models.fast;
+				const oaiModel = config.models[quality] ?? config.models.high ?? config.models.fast;
 				if (!oaiModel) {
 					throw new Error(`banana: no model mapped for quality "${quality}". Run /banana-setup.`);
 				}
@@ -388,7 +402,7 @@ export default function (pi: ExtensionAPI) {
 					details: {
 						prompt: params.prompt, model: oaiModel, quality, aspectRatio, imageSize,
 						mimeType: "image/png", outputPath: outPath, editing,
-						referenceImages: params.referenceImages, imageBase64: buf.toString("base64"),
+						referenceImages: params.referenceImages,
 					},
 				};
 			}
@@ -415,7 +429,7 @@ export default function (pi: ExtensionAPI) {
 				content: [
 					{
 						type: "text",
-						text: `🎨 ${verbLabel} ${aspectRatio} ${imageSize} image with ${quality === "lite" ? "Nano Banana 2 Lite" : quality === "fast" ? "Nano Banana 2" : "Nano Banana Pro"}…`,
+						text: `🎨 ${verbLabel} ${aspectRatio} ${imageSize} image with ${quality === "lite" ? "Nano Banana 2 Lite" : quality === "fast" ? "Nano Banana 2" : quality === "high" ? "Nano Banana Pro" : "GPT Image 2"}…`,
 					},
 				],
 				details: { model, aspectRatio, imageSize, quality, editing },
@@ -509,9 +523,9 @@ export default function (pi: ExtensionAPI) {
 				: `Generated → ${outPath}`;
 
 			return {
-				// Only text in content — the base64 image goes into details
-				// so it does not bloat conversation history sent to the LLM.
-				// renderResult reads the image from details for inline display.
+				// Only text in content, and no image bytes in details either:
+				// details are persisted in the session and counted by request
+				// guards. renderResult loads the preview from outputPath.
 				content: [{ type: "text", text: summary }],
 				details: {
 					prompt: params.prompt,
@@ -523,14 +537,15 @@ export default function (pi: ExtensionAPI) {
 					outputPath: outPath,
 					editing,
 					referenceImages: params.referenceImages,
-					imageBase64: base64,
 				},
 			};
 		},
 
 		renderResult(result, _options, theme) {
-			// Image is in details (not content) so base64 does not bloat
-			// conversation history sent to the LLM on subsequent turns.
+			// No image bytes travel with the result: content stays text so it
+			// does not bloat LLM history, and details stay small so the saved
+			// session does not carry a megabyte per picture. The preview is
+			// read back from the file that was just written.
 			const { details, content } = result;
 			const container = new Container();
 
@@ -541,11 +556,12 @@ export default function (pi: ExtensionAPI) {
 				container.addChild(new Text(theme.fg("success", "✔ " + summaryText), 1, 0));
 			}
 
-			// 2. Image from details
-			if (details && (details as any).imageBase64) {
+			// 2. Image, read from the saved file
+			const preview = readPreview((details as any)?.outputPath);
+			if (preview) {
 				container.addChild(new Spacer(1));
 				container.addChild(
-					new Image((details as any).imageBase64, (details as any).mimeType ?? "image/png", {
+					new Image(preview, (details as any).mimeType ?? "image/png", {
 						fallbackColor: (s: string) => theme.fg("muted", s),
 					}, {
 						maxWidthCells: 80,

package/extensions/providers.ts

@@ -18,7 +18,7 @@ import { resolve } from "path";
 export interface OpenAiCompatConfig {
 	baseUrl: string;
 	apiKey: string;
-	models: { lite?: string; fast?: string; high?: string };
+	models: { lite?: string; fast?: string; high?: string; max?: string };
 	visionModel?: string;
 	/** Optional provider id whose key lives in pi's auth.json (e.g. "mantice"). */
 	providerName?: string;
@@ -69,7 +69,11 @@ export async function resolveProvider(ctx: any): Promise<Provider> {
 
 // ─── OpenAI-compat calls ─────────────────────────────────────────────
 
-/** Map aspect ratio + size tier to a generic WxH string. */
+/** Round to the nearest multiple of 16 (never below 16): the OpenAI-compatible
+ * image upstream rejects sizes whose width or height is not divisible by 16. */
+const round16 = (n: number): number => Math.max(16, Math.round(n / 16) * 16);
+
+/** Map aspect ratio + size tier to a generic WxH string (multiples of 16). */
 export function sizeFor(aspectRatio: string, imageSize: string): string {
 	const base = imageSize === "4K" ? 4096 : imageSize === "2K" ? 2048 : 1024;
 	const ratio = (() => {
@@ -77,8 +81,8 @@ export function sizeFor(aspectRatio: string, imageSize: string): string {
 		return w && h ? w / h : 1;
 	})();
 	if (Math.abs(ratio - 1) < 0.01) return `${base}x${base}`;
-	if (ratio > 1) return `${base}x${Math.round(base / ratio)}`;
-	return `${Math.round(base * ratio)}x${base}`;
+	if (ratio > 1) return `${round16(base)}x${round16(Math.round(base / ratio))}`;
+	return `${round16(Math.round(base * ratio))}x${round16(base)}`;
 }
 
 async function asBuffer(data: any): Promise<Buffer> {
@@ -121,10 +125,15 @@ export async function openaiEditImage(
 	form.append("model", opts.model);
 	form.append("prompt", opts.prompt);
 	form.append("size", opts.size);
+	// A single reference image is the plain "image" file part. Multiple
+	// references must be appended as repeated "image[]" parts: the
+	// /images/edits upstream rejects duplicate "image" fields with
+	// 400 "Duplicate parameter image".
+	const field = opts.images.length > 1 ? "image[]" : "image";
 	for (const img of opts.images) {
 		const bytes = Buffer.from(img.data, "base64");
 		const ext = img.mimeType.split("/")[1] ?? "png";
-		form.append("image", new Blob([bytes], { type: img.mimeType }), `ref.${ext}`);
+		form.append(field, new Blob([bytes], { type: img.mimeType }), `ref.${ext}`);
 	}
 	const res = await fetch(`${config.baseUrl}/images/edits`, {
 		method: "POST",

package/extensions/wizard.ts

@@ -52,7 +52,8 @@ export async function configureBananaProvider(ctx: any, opts: ConfigureOptions):
 	const models = {
 		lite: pick(/lite|fast|turbo|mini/i),
 		fast: pick(/^(?!.*(lite|high|max|pro|ultra)).*$/i),
-		high: pick(/max|high|pro|ultra/i),
+		high: pick(/high|pro/i),
+		max: pick(/max|ultra/i),
 	};
 	const visionModel = opts.visionModel ?? probe.visionModels[0];
 	persistBananaConfig({ baseUrl, apiKey, models, ...(visionModel ? { visionModel } : {}) });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-banana",
-  "version": "2.5.1",
+  "version": "2.5.3",
   "description": "Generate, edit, and analyze images in pi using Google Nano Banana (image gen) and Gemini Vision (analysis). Inline terminal preview, reference-image editing, auto-save.",
   "author": "Francesco Frapporti <effedue@gmail.com>",
   "license": "MIT",