gsd-pi 2.57.0-dev.f22a903 → 2.58.0-dev.778d6ac

package/packages/pi-coding-agent/src/core/image-overflow-recovery.test.ts

@@ -0,0 +1,228 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import {
+	isImageDimensionError,
+	MANY_IMAGE_MAX_DIMENSION,
+	downsizeConversationImages,
+} from "./image-overflow-recovery.js";
+import type { Message } from "@gsd/pi-ai";
+
+// ─── isImageDimensionError ────────────────────────────────────────────────────
+
+describe("isImageDimensionError", () => {
+	it("returns true for Anthropic many-image dimension error", () => {
+		const errorMessage =
+			'Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.125.content.38.image.source.base64.data: At least one of the image dimensions exceed max allowed size for many-image requests: 2000 pixels"}}';
+		assert.equal(isImageDimensionError(errorMessage), true);
+	});
+
+	it("returns true for bare dimension exceed message", () => {
+		const errorMessage =
+			"image dimensions exceed max allowed size for many-image requests: 2000 pixels";
+		assert.equal(isImageDimensionError(errorMessage), true);
+	});
+
+	it("returns false for unrelated 400 error", () => {
+		const errorMessage =
+			'Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"max_tokens: 4096 > 2048"}}';
+		assert.equal(isImageDimensionError(errorMessage), false);
+	});
+
+	it("returns false for rate limit error", () => {
+		assert.equal(isImageDimensionError("429 rate limit exceeded"), false);
+	});
+
+	it("returns false for empty string", () => {
+		assert.equal(isImageDimensionError(""), false);
+	});
+
+	it("returns false for undefined", () => {
+		assert.equal(isImageDimensionError(undefined), false);
+	});
+});
+
+// ─── MANY_IMAGE_MAX_DIMENSION ─────────────────────────────────────────────────
+
+describe("MANY_IMAGE_MAX_DIMENSION", () => {
+	it("is less than 2000 (the API-enforced limit)", () => {
+		assert.ok(MANY_IMAGE_MAX_DIMENSION < 2000);
+	});
+
+	it("is a positive integer", () => {
+		assert.ok(MANY_IMAGE_MAX_DIMENSION > 0);
+		assert.equal(MANY_IMAGE_MAX_DIMENSION, Math.floor(MANY_IMAGE_MAX_DIMENSION));
+	});
+});
+
+// ─── helpers ──────────────────────────────────────────────────────────────────
+
+function makeUserMsg(content: Message["content"] & any): Message {
+	return { role: "user", content, timestamp: Date.now() } as Message;
+}
+
+function makeAssistantMsg(text: string): Message {
+	return {
+		role: "assistant",
+		content: [{ type: "text", text }],
+		api: "anthropic-messages",
+		provider: "anthropic",
+		model: "claude-opus-4-6",
+		usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
+		stopReason: "stop",
+		timestamp: Date.now(),
+	} as Message;
+}
+
+function makeToolResultMsg(images: number): Message {
+	const content: any[] = [];
+	for (let i = 0; i < images; i++) {
+		content.push({ type: "image", data: `img${i}`, mimeType: "image/png" });
+	}
+	return {
+		role: "toolResult",
+		toolCallId: `tc${Math.random()}`,
+		toolName: "screenshot",
+		content,
+		isError: false,
+		timestamp: Date.now(),
+	} as Message;
+}
+
+// ─── downsizeConversationImages ───────────────────────────────────────────────
+
+describe("downsizeConversationImages", () => {
+	it("counts images in user and toolResult messages", () => {
+		const messages: Message[] = [
+			makeUserMsg([
+				{ type: "image", data: "img1", mimeType: "image/png" },
+				{ type: "image", data: "img2", mimeType: "image/png" },
+			]),
+			makeAssistantMsg("I see them"),
+			makeToolResultMsg(1),
+		];
+
+		const result = downsizeConversationImages(messages);
+		assert.equal(result.imageCount, 3);
+	});
+
+	it("returns processed=false when no images present", () => {
+		const messages: Message[] = [
+			makeUserMsg("just text"),
+			makeAssistantMsg("reply"),
+		];
+
+		const result = downsizeConversationImages(messages);
+		assert.equal(result.imageCount, 0);
+		assert.equal(result.processed, false);
+	});
+
+	it("returns processed=false when image count <= RECENT_IMAGES_TO_KEEP", () => {
+		const messages: Message[] = [
+			makeUserMsg([
+				{ type: "image", data: "img1", mimeType: "image/png" },
+			]),
+			makeAssistantMsg("got it"),
+		];
+
+		const result = downsizeConversationImages(messages);
+		assert.equal(result.imageCount, 1);
+		assert.equal(result.processed, false);
+	});
+
+	it("strips older images when many images present, preserves recent ones", () => {
+		const messages: Message[] = [];
+		for (let i = 0; i < 25; i++) {
+			messages.push(
+				makeUserMsg([
+					{ type: "text", text: `message ${i}` },
+					{ type: "image", data: `img${i}`, mimeType: "image/png" },
+				]),
+			);
+			messages.push(makeAssistantMsg(`reply ${i}`));
+		}
+
+		const result = downsizeConversationImages(messages);
+		assert.ok(result.processed);
+		assert.equal(result.imageCount, 25);
+		assert.equal(result.strippedCount, 20); // 25 - 5 recent
+
+		// Count remaining images
+		let remainingImages = 0;
+		for (const msg of messages) {
+			if (msg.role === "assistant") continue;
+			if (typeof msg.content === "string") continue;
+			const arr = msg.content as any[];
+			for (const block of arr) {
+				if (block.type === "image") remainingImages++;
+			}
+		}
+		assert.equal(remainingImages, 5, "Should keep exactly 5 most recent images");
+
+		// The 5 most recent user messages (indices 40,42,44,46,48) should have images
+		for (let i = 20; i < 25; i++) {
+			const userMsg = messages[i * 2]; // user messages at even indices
+			const arr = userMsg.content as any[];
+			const hasImage = arr.some((c: any) => c.type === "image");
+			assert.ok(hasImage, `Recent message ${i} should retain its image`);
+		}
+	});
+
+	it("adds text placeholder when stripping an image", () => {
+		const messages: Message[] = [];
+		for (let i = 0; i < 10; i++) {
+			messages.push(
+				makeUserMsg([
+					{ type: "image", data: `img${i}`, mimeType: "image/jpeg" },
+				]),
+			);
+			messages.push(makeAssistantMsg(`reply ${i}`));
+		}
+
+		downsizeConversationImages(messages);
+
+		// First message's image should have been replaced with text
+		const firstMsg = messages[0];
+		const arr = firstMsg.content as any[];
+		const placeholder = arr.find(
+			(c: any) => c.type === "text" && c.text.includes("[image removed"),
+		);
+		assert.ok(placeholder, "Stripped image should be replaced with text placeholder");
+		assert.ok(
+			placeholder.text.includes("image/jpeg"),
+			"Placeholder should mention original mime type",
+		);
+	});
+
+	it("handles toolResult messages with images", () => {
+		const messages: Message[] = [];
+		for (let i = 0; i < 10; i++) {
+			messages.push(makeToolResultMsg(1));
+			messages.push(makeAssistantMsg(`reply ${i}`));
+		}
+
+		const result = downsizeConversationImages(messages);
+		assert.equal(result.imageCount, 10);
+		assert.equal(result.strippedCount, 5);
+		assert.ok(result.processed);
+	});
+
+	it("handles mixed user and toolResult images", () => {
+		const messages: Message[] = [];
+		for (let i = 0; i < 8; i++) {
+			messages.push(
+				makeUserMsg([
+					{ type: "text", text: `check ${i}` },
+					{ type: "image", data: `uimg${i}`, mimeType: "image/png" },
+				]),
+			);
+			messages.push(makeAssistantMsg(`processing ${i}`));
+			messages.push(makeToolResultMsg(1));
+			messages.push(makeAssistantMsg(`done ${i}`));
+		}
+
+		const result = downsizeConversationImages(messages);
+		// 8 user images + 8 tool result images = 16 total
+		assert.equal(result.imageCount, 16);
+		assert.equal(result.strippedCount, 11); // 16 - 5 recent
+	});
+});

package/packages/pi-coding-agent/src/core/image-overflow-recovery.ts

@@ -0,0 +1,118 @@
+/**
+ * Image overflow recovery for many-image sessions.
+ *
+ * When a conversation accumulates many images (screenshots, file reads, etc.),
+ * the Anthropic API enforces a stricter per-image dimension limit (2000px) for
+ * "many-image requests." This module detects the resulting 400 error and
+ * recovers by stripping older images from the conversation history, preserving
+ * the most recent ones to maintain session continuity.
+ *
+ * @see https://github.com/gsd-build/gsd-2/issues/2874
+ */
+
+import type { Message, ImageContent, TextContent } from "@gsd/pi-ai";
+
+/**
+ * Maximum image dimension (px) that the Anthropic API allows in many-image
+ * requests. Images at or above this size in a large conversation will be
+ * rejected with a 400 error. We use 1568 as the safe ceiling (Anthropic's
+ * recommended max for multi-image requests).
+ */
+export const MANY_IMAGE_MAX_DIMENSION = 1568;
+
+/**
+ * Number of recent images to preserve when stripping old images.
+ * Keeps the most recent screenshots/images so the model retains visual context
+ * for the current task.
+ */
+const RECENT_IMAGES_TO_KEEP = 5;
+
+/**
+ * Regex matching the Anthropic API error for oversized images in many-image requests.
+ */
+const IMAGE_DIMENSION_ERROR_RE =
+	/image.dimensions?.exceed.*max.*allowed.*size.*many.image/i;
+
+/**
+ * Detect whether an error message is the Anthropic "image dimensions exceed max
+ * allowed size for many-image requests" 400 error.
+ */
+export function isImageDimensionError(errorMessage: string | undefined | null): boolean {
+	if (!errorMessage) return false;
+	return IMAGE_DIMENSION_ERROR_RE.test(errorMessage);
+}
+
+export interface DownsizeResult {
+	/** Total number of images found in the conversation */
+	imageCount: number;
+	/** Whether any images were stripped */
+	processed: boolean;
+	/** Number of images that were stripped */
+	strippedCount: number;
+}
+
+/**
+ * Strip older images from conversation messages to recover from many-image
+ * dimension errors. Preserves the N most recent images and replaces older ones
+ * with a text placeholder.
+ *
+ * Mutates messages in place (same pattern as replaceMessages/compaction).
+ *
+ * Accepts Message[] (the LLM message union) so it works with both
+ * agent.state.messages and session entries.
+ */
+export function downsizeConversationImages(messages: Message[]): DownsizeResult {
+	// First pass: collect all image locations (message index + content index)
+	const imageLocations: Array<{ msgIdx: number; contentIdx: number }> = [];
+
+	for (let msgIdx = 0; msgIdx < messages.length; msgIdx++) {
+		const msg = messages[msgIdx];
+		if (msg.role === "assistant") continue;
+
+		// UserMessage can have string content; ToolResultMessage always has array
+		if (msg.role === "user" && typeof msg.content === "string") continue;
+
+		const contentArr = msg.content as (TextContent | ImageContent)[];
+		if (!Array.isArray(contentArr)) continue;
+
+		for (let contentIdx = 0; contentIdx < contentArr.length; contentIdx++) {
+			if (contentArr[contentIdx].type === "image") {
+				imageLocations.push({ msgIdx, contentIdx });
+			}
+		}
+	}
+
+	const imageCount = imageLocations.length;
+	if (imageCount === 0) {
+		return { imageCount: 0, processed: false, strippedCount: 0 };
+	}
+
+	// Determine which images to strip (all except the N most recent)
+	const stripCount = Math.max(0, imageCount - RECENT_IMAGES_TO_KEEP);
+	if (stripCount === 0) {
+		return { imageCount, processed: false, strippedCount: 0 };
+	}
+
+	const toStrip = imageLocations.slice(0, stripCount);
+
+	// Second pass: replace stripped images with text placeholder.
+	// Process in reverse order to maintain content indices.
+	for (let i = toStrip.length - 1; i >= 0; i--) {
+		const { msgIdx, contentIdx } = toStrip[i];
+		const msg = messages[msgIdx];
+		if (msg.role === "assistant") continue;
+		if (msg.role === "user" && typeof msg.content === "string") continue;
+
+		const contentArr = msg.content as (TextContent | ImageContent)[];
+		const imageBlock = contentArr[contentIdx] as ImageContent;
+		const mimeType = imageBlock.mimeType || "image/unknown";
+
+		// Replace the image block with a text placeholder
+		(contentArr as any[])[contentIdx] = {
+			type: "text",
+			text: `[image removed to reduce context size — was ${mimeType}]`,
+		} as TextContent;
+	}
+
+	return { imageCount, processed: true, strippedCount: stripCount };
+}

package/packages/pi-coding-agent/src/core/index.ts

@@ -29,6 +29,7 @@ export {
 	type ExecResult,
 	type Extension,
 	type ExtensionAPI,
+	type ExtensionManifest,
 	type ExtensionCommandContext,
 	type ExtensionContext,
 	type ExtensionError,
@@ -53,6 +54,11 @@ export {
 	type SessionSwitchEvent,
 	type SessionTreeEvent,
 	type ToolCallEvent,
+	readManifest,
+	readManifestFromEntryPath,
+	type SortResult,
+	type SortWarning,
+	sortExtensionPaths,
 	type ToolDefinition,
 	type ToolRenderResultOptions,
 	type ToolResultEvent,

package/packages/pi-coding-agent/src/core/lsp/index.ts

@@ -340,6 +340,9 @@ async function runWorkspaceDiagnostics(
 	const proc = spawn(cmd, cmdArgs, {
 		cwd,
 		stdio: ["ignore", "pipe", "pipe"],
+		// On Windows, project-type commands (tsc, cargo, etc.) may be .cmd
+		// wrappers that need shell resolution to avoid ENOENT/EINVAL (#2854).
+		shell: process.platform === "win32",
 	});
 	const abortHandler = () => {
 		proc.kill();

package/packages/pi-coding-agent/src/core/lsp/lspmux.ts

@@ -90,6 +90,9 @@ async function checkServerRunning(binaryPath: string): Promise<boolean> {
 	try {
 		const proc = spawn(binaryPath, ["status"], {
 			stdio: ["ignore", "pipe", "pipe"],
+			// On Windows, the binary may be a .cmd wrapper requiring shell
+			// resolution to avoid ENOENT/EINVAL (#2854).
+			shell: process.platform === "win32",
 		});
 
 		const exited = await Promise.race([

package/packages/pi-coding-agent/src/core/messages.test.ts

@@ -0,0 +1,114 @@
+/**
+ * messages.test.ts — Tests for convertToLlm custom message handling.
+ *
+ * Reproduction test for #3026: background job completion notifications
+ * delivered as custom messages must be clearly distinguishable from
+ * user-typed input when converted to LLM messages.
+ */
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { convertToLlm, type CustomMessage } from "./messages.js";
+
+/** Extract the first content block from a message, asserting array content. */
+function firstTextBlock(msg: ReturnType<typeof convertToLlm>[number]) {
+	const { content } = msg;
+	assert.ok(Array.isArray(content), "Expected content to be an array");
+	const block = content[0];
+	assert.ok(typeof block === "object" && block !== null, "Expected first block to be an object");
+	return block;
+}
+
+test("convertToLlm wraps custom messages with system notification prefix", () => {
+	const customMsg: CustomMessage = {
+		role: "custom",
+		customType: "async_job_result",
+		content: "**Background job done: bg_abc123** (sleep 2, 2.1s)\n\ndone",
+		display: true,
+		timestamp: Date.now(),
+	};
+
+	const result = convertToLlm([customMsg]);
+	assert.equal(result.length, 1);
+	assert.equal(result[0].role, "user");
+
+	// The content must include a system notification wrapper so the LLM
+	// does not confuse it with user input (#3026).
+	const text = firstTextBlock(result[0]);
+	assert.equal(text.type, "text");
+	assert.ok(
+		"text" in text && text.text.includes("[system notification"),
+		"Custom message should be wrapped with system notification marker",
+	);
+});
+
+test("convertToLlm wraps custom messages with array content", () => {
+	const customMsg: CustomMessage = {
+		role: "custom",
+		customType: "bg-shell-status",
+		content: [{ type: "text", text: "Background processes:\n  ✓ bg1 dev-server :3000" }],
+		display: false,
+		timestamp: Date.now(),
+	};
+
+	const result = convertToLlm([customMsg]);
+	assert.equal(result.length, 1);
+	assert.equal(result[0].role, "user");
+
+	const text = firstTextBlock(result[0]);
+	assert.equal(text.type, "text");
+	assert.ok(
+		"text" in text && text.text.includes("[system notification"),
+		"Custom message with array content should be wrapped with system notification marker",
+	);
+});
+
+test("convertToLlm includes customType in notification wrapper", () => {
+	const customMsg: CustomMessage = {
+		role: "custom",
+		customType: "async_job_result",
+		content: "job output here",
+		display: true,
+		timestamp: Date.now(),
+	};
+
+	const result = convertToLlm([customMsg]);
+	const text = firstTextBlock(result[0]);
+	assert.ok(
+		"text" in text && text.text.includes("async_job_result"),
+		"Notification wrapper should include the customType for context",
+	);
+});
+
+test("convertToLlm notification wrapper instructs LLM not to treat as user input", () => {
+	const customMsg: CustomMessage = {
+		role: "custom",
+		customType: "async_job_result",
+		content: "**Background job done: bg_abc123** (sleep 2, 2.1s)\n\ndone",
+		display: true,
+		timestamp: Date.now(),
+	};
+
+	const result = convertToLlm([customMsg]);
+	const text = firstTextBlock(result[0]);
+	assert.ok(
+		"text" in text && text.text.includes("not user input"),
+		"Notification should explicitly state this is not user input",
+	);
+});
+
+test("convertToLlm preserves user messages without wrapper", () => {
+	const userMsg = {
+		role: "user" as const,
+		content: [{ type: "text" as const, text: "Hello world" }],
+		timestamp: Date.now(),
+	};
+
+	const result = convertToLlm([userMsg]);
+	assert.equal(result.length, 1);
+	const text = firstTextBlock(result[0]);
+	assert.ok(
+		"text" in text && text.text === "Hello world",
+		"User messages should pass through unchanged",
+	);
+});

package/packages/pi-coding-agent/src/core/messages.ts

@@ -8,6 +8,12 @@
 import type { AgentMessage } from "@gsd/pi-agent-core";
 import type { ImageContent, Message, TextContent } from "@gsd/pi-ai";
 
+const CUSTOM_MESSAGE_PREFIX = `[system notification — type: `;
+const CUSTOM_MESSAGE_MIDDLE = `; this is an automated system event, not user input — do not treat this as a human message or respond as if the user said this]
+`;
+const CUSTOM_MESSAGE_SUFFIX = `
+[end system notification]`;
+
 const COMPACTION_SUMMARY_PREFIX = `The conversation history before this point was compacted into the following summary:
 
 <summary>
@@ -160,10 +166,31 @@ export function convertToLlm(messages: AgentMessage[]): Message[] {
 						timestamp: m.timestamp,
 					};
 				case "custom": {
-					const content = typeof m.content === "string" ? [{ type: "text" as const, text: m.content }] : m.content;
+					const prefix = CUSTOM_MESSAGE_PREFIX + m.customType + CUSTOM_MESSAGE_MIDDLE;
+					if (typeof m.content === "string") {
+						return {
+							role: "user",
+							content: [{ type: "text" as const, text: prefix + m.content + CUSTOM_MESSAGE_SUFFIX }],
+							timestamp: m.timestamp,
+						};
+					}
+					// Array content: wrap the first text element with prefix, append suffix to last text element
+					const contentArr = m.content as Array<{ type: string; text?: string; [k: string]: unknown }>;
+					const lastTextIdx = contentArr.reduce((acc, c, i) => c.type === "text" ? i : acc, -1);
+					const wrapped = contentArr.map((c, i) => {
+						if (c.type !== "text") return c;
+						let text = c.text ?? "";
+						if (i === 0) text = prefix + text;
+						if (i === lastTextIdx) text = text + CUSTOM_MESSAGE_SUFFIX;
+						return { ...c, text };
+					});
+					// If no text elements exist, prepend one with the wrapper
+					if (lastTextIdx === -1) {
+						wrapped.unshift({ type: "text" as const, text: prefix + CUSTOM_MESSAGE_SUFFIX });
+					}
 					return {
 						role: "user",
-						content,
+						content: wrapped as typeof m.content,
 						timestamp: m.timestamp,
 					};
 				}

package/packages/pi-coding-agent/src/core/model-resolver.ts

@@ -37,6 +37,7 @@ const defaultModelPerProvider: Record<KnownProvider, string> = {
 	"opencode-go": "kimi-k2.5",
 	"kimi-coding": "kimi-k2-thinking",
 	"alibaba-coding-plan": "qwen3.5-plus",
+	ollama: "llama3.1:8b",
 	"ollama-cloud": "qwen3:32b",
 };
 

package/packages/pi-coding-agent/src/core/resource-loader.ts

@@ -129,6 +129,12 @@ export interface DefaultResourceLoaderOptions {
 	appendSystemPrompt?: string;
 	/** Names of bundled extensions (used to identify built-in extensions in conflict detection). */
 	bundledExtensionNames?: Set<string>;
+	/**
+	 * Transform extension paths before loading. Receives the merged list of all
+	 * discovered extension paths and returns a (possibly reordered/filtered) list.
+	 * Use this to apply dependency sorting or registry-based filtering.
+	 */
+	extensionPathsTransform?: (paths: string[]) => { paths: string[]; diagnostics?: string[] };
 	extensionsOverride?: (base: LoadExtensionsResult) => LoadExtensionsResult;
 	skillsOverride?: (base: { skills: Skill[]; diagnostics: ResourceDiagnostic[] }) => {
 		skills: Skill[];
@@ -167,6 +173,7 @@ export class DefaultResourceLoader implements ResourceLoader {
 	private systemPromptSource?: string;
 	private appendSystemPromptSource?: string;
 	private bundledExtensionNames: Set<string>;
+	private extensionPathsTransform?: (paths: string[]) => { paths: string[]; diagnostics?: string[] };
 	private extensionsOverride?: (base: LoadExtensionsResult) => LoadExtensionsResult;
 	private skillsOverride?: (base: { skills: Skill[]; diagnostics: ResourceDiagnostic[] }) => {
 		skills: Skill[];
@@ -223,6 +230,7 @@ export class DefaultResourceLoader implements ResourceLoader {
 		this.systemPromptSource = options.systemPrompt;
 		this.appendSystemPromptSource = options.appendSystemPrompt;
 		this.bundledExtensionNames = options.bundledExtensionNames ?? new Set();
+		this.extensionPathsTransform = options.extensionPathsTransform;
 		this.extensionsOverride = options.extensionsOverride;
 		this.skillsOverride = options.skillsOverride;
 		this.promptsOverride = options.promptsOverride;
@@ -378,10 +386,21 @@ export class DefaultResourceLoader implements ResourceLoader {
 		const cliEnabledPrompts = getEnabledPaths(cliExtensionPaths.prompts);
 		const cliEnabledThemes = getEnabledPaths(cliExtensionPaths.themes);
 
-		const extensionPaths = this.noExtensions
+		let extensionPaths = this.noExtensions
 			? cliEnabledExtensions
 			: this.mergePaths(cliEnabledExtensions, enabledExtensions);
 
+		// Apply path transform (dependency sorting, registry filtering) if provided
+		if (this.extensionPathsTransform) {
+			const transformed = this.extensionPathsTransform(extensionPaths);
+			extensionPaths = transformed.paths;
+			if (transformed.diagnostics?.length) {
+				for (const msg of transformed.diagnostics) {
+					process.stderr.write(`[extensions] ${msg}\n`);
+				}
+			}
+		}
+
 		const extensionsResult = await loadExtensions(extensionPaths, this.cwd, this.eventBus);
 		const inlineExtensions = await this.loadExtensionFactories(extensionsResult.runtime);
 		extensionsResult.extensions.push(...inlineExtensions.extensions);