@gajae-code/coding-agent 0.9.2 → 0.9.3

package/src/session/agent-session.ts

@@ -26,6 +26,7 @@ import {
 	type AgentMessage,
 	type AgentState,
 	type AgentTool,
+	assertImagePlaceholdersHavePayload,
 	resolveTelemetry,
 	type StablePrefixSnapshot,
 	ThinkingLevel,
@@ -151,6 +152,7 @@ import { onAppendOnlyModeChanged } from "../config/settings";
 import { RawSseDebugBuffer } from "../debug/raw-sse-buffer";
 import { loadCapability } from "../discovery";
 import { expandApplyPatchToEntries, normalizeDiff, normalizeToLF, ParseError, previewPatch, stripBom } from "../edit";
+import { MAX_EDIT_FILE_BYTES } from "../edit/read-file";
 import { disposeVmContextsByOwner } from "../eval/js/context-manager";
 import {
 	disposeKernelSessionsByOwner,
@@ -360,6 +362,11 @@ export interface AsyncJobSnapshot {
 // Types
 // ============================================================================
 
+export interface RetainedMemorySample {
+	tuiChatChildren?: number;
+	tuiCachedRenderBytes?: number;
+}
+
 export interface AgentSessionConfig {
 	agent: Agent;
 	sessionManager: SessionManager;
@@ -439,6 +446,8 @@ export interface AgentSessionConfig {
 	 * **MUST NOT** dispose it on their own teardown.
 	 */
 	ownedAsyncJobManager?: AsyncJobManager;
+	/** Cheap TUI retained-memory counters; absent for headless sessions. */
+	retainedMemorySampler?: () => RetainedMemorySample;
 	/**
 	 * MCPManager whose lifecycle this session owns (top-level sessions that
 	 * connected plugin-bundle MCP servers). Only the owned manager is
@@ -1084,6 +1093,68 @@ export function buildContextInjectionSignature(kind: string, parts: readonly str
 	}
 	return hash.digest("base64url");
 }
+
+const STREAMING_EDIT_FILE_CACHE_MAX_ENTRIES = 16;
+const STREAMING_EDIT_FILE_CACHE_MAX_TOTAL_BYTES = 32 * 1024 * 1024;
+
+type StreamingEditFileCacheEntry = {
+	content: string;
+	bytes: number;
+};
+
+export class StreamingEditFileCache {
+	#entries = new Map<string, StreamingEditFileCacheEntry>();
+	#totalBytes = 0;
+
+	get(path: string): string | undefined {
+		const entry = this.#entries.get(path);
+		if (entry === undefined) return undefined;
+		this.#entries.delete(path);
+		this.#entries.set(path, entry);
+		return entry.content;
+	}
+
+	set(path: string, content: string): void {
+		const bytes = Buffer.byteLength(content, "utf8");
+		if (bytes > MAX_EDIT_FILE_BYTES || bytes > STREAMING_EDIT_FILE_CACHE_MAX_TOTAL_BYTES) {
+			this.delete(path);
+			return;
+		}
+
+		this.delete(path);
+		while (
+			this.#entries.size >= STREAMING_EDIT_FILE_CACHE_MAX_ENTRIES ||
+			this.#totalBytes + bytes > STREAMING_EDIT_FILE_CACHE_MAX_TOTAL_BYTES
+		) {
+			const oldestPath = this.#entries.keys().next().value;
+			if (oldestPath === undefined) break;
+			this.delete(oldestPath);
+		}
+
+		this.#entries.set(path, { content, bytes });
+		this.#totalBytes += bytes;
+	}
+
+	delete(path: string): void {
+		const entry = this.#entries.get(path);
+		if (entry === undefined) return;
+		this.#entries.delete(path);
+		this.#totalBytes -= entry.bytes;
+	}
+
+	clear(): void {
+		this.#entries.clear();
+		this.#totalBytes = 0;
+	}
+
+	has(path: string): boolean {
+		return this.#entries.has(path);
+	}
+
+	get totalBytes(): number {
+		return this.#totalBytes;
+	}
+}
 export class AgentSession {
 	readonly agent: Agent;
 	readonly sessionManager: SessionManager;
@@ -1148,6 +1219,7 @@ export class AgentSession {
 	#compactionAbortController: AbortController | undefined = undefined;
 	#autoCompactionAbortController: AbortController | undefined = undefined;
 	#resourceSampler: () => EmergencyCompactionSample = () => this.#defaultResourceSample();
+	#retainedMemorySampler: (() => RetainedMemorySample) | undefined;
 	#prePromptContextCheckPromise: Promise<void> | undefined = undefined;
 
 	// Branch summarization state
@@ -1321,7 +1393,7 @@ export class AgentSession {
 
 	#streamingEditPrecheckedToolCallIds = new Set<string>();
 
-	#streamingEditFileCache = new Map<string, string>();
+	#streamingEditFileCache = new StreamingEditFileCache();
 	readonly streamingEditDebugCounters = {
 		guardRuns: 0,
 		processedChars: 0,
@@ -1447,6 +1519,7 @@ export class AgentSession {
 		// Power assertions are taken per turn (see #beginInFlight); nothing acquired here.
 		this.#evalKernelOwnerId = config.evalKernelOwnerId ?? `agent-session:${Snowflake.next()}`;
 		this.#ownedAsyncJobManager = config.ownedAsyncJobManager;
+		this.#retainedMemorySampler = config.retainedMemorySampler;
 		this.#ownedMcpManager = config.ownedMcpManager;
 		this.#scopedModels = config.scopedModels ?? [];
 		this.#thinkingLevel = config.thinkingLevel;
@@ -3170,7 +3243,31 @@ export class AgentSession {
 		// which is the optional patch-preview verification gated by
 		// edit.streamingAbort. Skip the read when the setting is off.
 		if (this.settings.get("edit.streamingAbort")) {
-			this.#ensureFileCache(streamingEdit.resolvedPath);
+			// Fire-and-forget async priming: toolcall deltas arrive on the hot
+			// stream path, so never block it on filesystem I/O. The abort-check
+			// path (#maybeAbortStreamingEdit) still falls back to the sync read
+			// if priming hasn't completed by the time removed lines appear.
+			void this.#preCacheFileAsync(streamingEdit.resolvedPath);
+		}
+	}
+
+	#streamingEditPrecachePending = new Set<string>();
+	async #preCacheFileAsync(resolvedPath: string): Promise<void> {
+		if (this.#streamingEditFileCache.has(resolvedPath)) return;
+		if (this.#streamingEditPrecachePending.has(resolvedPath)) return;
+		this.#streamingEditPrecachePending.add(resolvedPath);
+		try {
+			const stat = await fs.promises.stat(resolvedPath);
+			if (stat.size > MAX_EDIT_FILE_BYTES) return;
+
+			const rawText = await fs.promises.readFile(resolvedPath, "utf-8");
+			if (this.#streamingEditFileCache.has(resolvedPath)) return;
+			const { text } = stripBom(rawText);
+			this.#streamingEditFileCache.set(resolvedPath, normalizeToLF(text));
+		} catch {
+			// Don't cache on read errors (including ENOENT) - let the edit tool handle them
+		} finally {
+			this.#streamingEditPrecachePending.delete(resolvedPath);
 		}
 	}
 
@@ -3178,6 +3275,9 @@ export class AgentSession {
 		if (this.#streamingEditFileCache.has(resolvedPath)) return;
 
 		try {
+			const stat = fs.statSync(resolvedPath);
+			if (stat.size > MAX_EDIT_FILE_BYTES) return;
+
 			const rawText = fs.readFileSync(resolvedPath, "utf-8");
 			const { text } = stripBom(rawText);
 			this.#streamingEditFileCache.set(resolvedPath, normalizeToLF(text));
@@ -5298,6 +5398,7 @@ export class AgentSession {
 
 		// Expand file-based prompt templates if requested
 		const expandedText = expandPromptTemplates ? expandPromptTemplate(text, [...this.#promptTemplates]) : text;
+		assertImagePlaceholdersHavePayload(expandedText, options?.images);
 		const workflowIntentDiff = options?.synthetic ? null : buildWorkflowIntentDiff(expandedText);
 
 		// If streaming, queue via steer() or followUp() based on option
@@ -5784,6 +5885,7 @@ export class AgentSession {
 		}
 
 		const expandedText = expandPromptTemplate(text, [...this.#promptTemplates]);
+		assertImagePlaceholdersHavePayload(expandedText, images);
 		await this.#queueSteer(expandedText, images);
 	}
 
@@ -5800,6 +5902,7 @@ export class AgentSession {
 		}
 
 		const expandedText = expandPromptTemplate(text, [...this.#promptTemplates]);
+		assertImagePlaceholdersHavePayload(expandedText, images);
 		await this.#queueFollowUp(expandedText, images, {
 			forceOneAtATime: options?.followUpQueuePolicy === "sequential",
 		});
@@ -5809,6 +5912,7 @@ export class AgentSession {
 	 * Internal: Queue a steering message (already expanded, no extension command check).
 	 */
 	async #queueSteer(text: string, images?: ImageContent[]): Promise<void> {
+		assertImagePlaceholdersHavePayload(text, images);
 		const displayText = text || (images && images.length > 0 ? "[Image]" : "");
 		this.#steeringMessages.push(this.#createQueuedDisplayEntry(displayText));
 		const content: (TextContent | ImageContent)[] = [{ type: "text", text }];
@@ -5840,6 +5944,7 @@ export class AgentSession {
 	 * Internal: Queue a follow-up message (already expanded, no extension command check).
 	 */
 	async #queueFollowUp(text: string, images?: ImageContent[], options?: { forceOneAtATime?: boolean }): Promise<void> {
+		assertImagePlaceholdersHavePayload(text, images);
 		const displayText = text || (images && images.length > 0 ? "[Image]" : "");
 		this.#followUpMessages.push(this.#createQueuedDisplayEntry(displayText));
 		const content: (TextContent | ImageContent)[] = [{ type: "text", text }];
@@ -7732,9 +7837,14 @@ export class AgentSession {
 		this.#resourceSampler = sampler;
 	}
 
+	setRetainedMemorySampler(sampler: (() => RetainedMemorySample) | undefined): void {
+		this.#retainedMemorySampler = sampler;
+	}
+
 	#defaultResourceSample(): EmergencyCompactionSample {
 		let providerBytes = 0;
 		let imageBytes = 0;
+		const retainedMemory = this.#retainedMemorySampler?.() ?? {};
 		for (const message of this.state.messages) {
 			const content = (message as { content?: unknown }).content;
 			if (typeof content === "string") {
@@ -7756,6 +7866,10 @@ export class AgentSession {
 			providerBytes,
 			messageCount: this.state.messages.length,
 			imageBytes,
+			sessionResidentImageBytes: this.sessionManager.getResidentImageBytes(),
+			materializedResidentBytes: this.#streamingEditFileCache.totalBytes,
+			tuiChatChildren: retainedMemory.tuiChatChildren ?? 0,
+			tuiCachedRenderBytes: retainedMemory.tuiCachedRenderBytes ?? 0,
 		};
 	}
 
@@ -8071,7 +8185,7 @@ export class AgentSession {
 			attribution: "agent",
 			timestamp: Date.now(),
 		});
-		this.#scheduleAgentContinue({ generation: this.#promptGeneration });
+		this.#scheduleAgentContinue({ skipCompactionCheck: true });
 	}
 
 	/**

package/src/session/session-manager.ts

@@ -45,8 +45,6 @@ import {
 	MemoryBlobStore,
 	parseBlobRef,
 	ResidentBlobMissingError,
-	resolveImageData,
-	resolveImageDataUrl,
 	resolveResidentImageDataSync,
 	resolveResidentImageDataUrlSync,
 	resolveTextBlobSync,
@@ -973,47 +971,46 @@ export async function loadEntriesFromFile(
 }
 
 /**
- * Resolve blob references in loaded entries, restoring both session image blocks and persisted
- * provider image URLs back to the inline data expected by downstream transports. Mutates entries in place.
+ * Convert legacy persisted blob references in loaded entries into resident sentinels.
+ * Images then materialize lazily at provider/display/export chokepoints instead of
+ * pinning every historical base64 string for the lifetime of a resumed session.
  */
 function hasImageUrl(value: unknown): value is { image_url: string | { url?: string } } {
 	return typeof value === "object" && value !== null && "image_url" in value;
 }
 
-async function resolvePersistedBlobRefs(value: unknown, blobStore: BlobStore, key?: string): Promise<void> {
+function residentizePersistedBlobRefs(value: unknown, key?: string): void {
 	if (Array.isArray(value)) {
-		await Promise.all(value.map(item => resolvePersistedBlobRefs(item, blobStore, key)));
+		for (const item of value) residentizePersistedBlobRefs(item, key);
 		return;
 	}
 
 	if (typeof value !== "object" || value === null) return;
 
 	if (isImageBlock(value) && isBlobRef(value.data)) {
-		value.data = await resolveImageData(blobStore, value.data);
+		value.data = residentBlobSentinel("imageData", value.data) as unknown as string;
 	}
 
 	if (hasImageUrl(value)) {
 		if (typeof value.image_url === "string" && isBlobRef(value.image_url)) {
-			value.image_url = await resolveImageDataUrl(blobStore, value.image_url);
+			value.image_url = residentBlobSentinel("imageUrl", value.image_url) as unknown as string;
 		} else if (
 			typeof value.image_url === "object" &&
 			value.image_url !== null &&
 			typeof value.image_url.url === "string" &&
 			isBlobRef(value.image_url.url)
 		) {
-			value.image_url.url = await resolveImageDataUrl(blobStore, value.image_url.url);
+			value.image_url.url = residentBlobSentinel("imageUrl", value.image_url.url) as unknown as string;
 		}
 	}
 
-	await Promise.all(
-		Object.entries(value).map(async ([childKey, item]) => {
-			if (childKey === "data" && typeof item === "string" && isBlobRef(item) && key !== TEXT_CONTENT_KEY) {
-				(value as Record<string, unknown>)[childKey] = await resolveImageDataUrl(blobStore, item);
-				return;
-			}
-			await resolvePersistedBlobRefs(item, blobStore, childKey);
-		}),
-	);
+	for (const [childKey, item] of Object.entries(value)) {
+		if (childKey === "data" && typeof item === "string" && isBlobRef(item) && key !== TEXT_CONTENT_KEY) {
+			(value as Record<string, unknown>)[childKey] = residentBlobSentinel("imageUrl", item);
+			continue;
+		}
+		residentizePersistedBlobRefs(item, childKey);
+	}
 }
 
 /**
@@ -1035,28 +1032,13 @@ async function runWithConcurrency(tasks: Array<() => Promise<void>>, limit: numb
 	await Promise.all(Array.from({ length: workerCount }, () => worker()));
 }
 
-async function resolveBlobRefsInEntries(entries: FileEntry[], blobStore: BlobStore): Promise<void> {
+async function resolveBlobRefsInEntries(entries: FileEntry[], _blobStore: BlobStore): Promise<void> {
 	const tasks: Array<() => Promise<void>> = [];
 
 	for (const entry of entries) {
 		if (entry.type === "session") continue;
-
-		let contentArray: unknown[] | undefined;
-		if (entry.type === "message" && "content" in entry.message && Array.isArray(entry.message.content)) {
-			contentArray = entry.message.content;
-		} else if (entry.type === "custom_message" && Array.isArray(entry.content)) {
-			contentArray = entry.content;
-		}
-
 		tasks.push(async () => {
-			if (contentArray) {
-				for (const block of contentArray) {
-					if (isImageBlock(block) && isBlobRef(block.data)) {
-						block.data = await resolveImageData(blobStore, block.data);
-					}
-				}
-			}
-			await resolvePersistedBlobRefs(entry, blobStore);
+			residentizePersistedBlobRefs(entry);
 		});
 	}
 
@@ -1314,6 +1296,33 @@ function containsResidentSentinel(value: unknown, seen = new WeakSet<object>()):
 	return false;
 }
 
+function containsResidentImageSentinel(value: unknown, seen = new WeakSet<object>()): boolean {
+	if (value === null || value === undefined || typeof value !== "object") return false;
+	if (isResidentBlobSentinel(value)) return value.kind === "imageUrl" || value.kind === "imageData";
+	if (seen.has(value)) return false;
+	seen.add(value);
+	if (Array.isArray(value)) return value.some(item => containsResidentImageSentinel(item, seen));
+	for (const child of Object.values(value)) {
+		if (containsResidentImageSentinel(child, seen)) return true;
+	}
+	return false;
+}
+
+function collectResidentImageRefs(value: unknown, refs: Set<string>, seen = new WeakSet<object>()): void {
+	if (value === null || value === undefined || typeof value !== "object") return;
+	if (isResidentBlobSentinel(value)) {
+		if (value.kind === "imageUrl" || value.kind === "imageData") refs.add(value.ref);
+		return;
+	}
+	if (seen.has(value)) return;
+	seen.add(value);
+	if (Array.isArray(value)) {
+		for (const item of value) collectResidentImageRefs(item, refs, seen);
+		return;
+	}
+	for (const child of Object.values(value)) collectResidentImageRefs(child, refs, seen);
+}
+
 /**
  * Recursively truncate large strings in an object for session persistence.
  * - Truncates any oversized string fields (key-agnostic)
@@ -4148,6 +4157,22 @@ export class SessionManager {
 		return entry ? materializeResidentEntryForReadSync(entry, this.#residentBlobStores(), new Map()) : undefined;
 	}
 
+	getResidentImageBytes(): number {
+		const refs = new Set<string>();
+		for (const entry of this.#fileEntries) collectResidentImageRefs(entry, refs);
+		let bytes = 0;
+		for (const ref of refs) {
+			const hash = parseBlobRef(ref);
+			if (!hash) continue;
+			try {
+				bytes += fs.statSync(path.join(this.#residentImageBlobStore.dir, hash)).size;
+			} catch (err) {
+				if (!isEnoent(err)) throw err;
+			}
+		}
+		return bytes;
+	}
+
 	/**
 	 * Get the most recent model role from the current session path.
 	 * Returns undefined if no model change has been recorded.
@@ -4571,11 +4596,14 @@ export class SessionManager {
 		}
 		this.#materializedEntriesCachePopulateCount++;
 		const resolvedTextBlobCache = new Map<string, string>();
-		const materializedEntries = this.#fileEntries
-			.filter((e): e is SessionEntry => e.type !== "session")
-			.map(entry => materializeResidentEntryForReadSync(entry, this.#residentBlobStores(), resolvedTextBlobCache));
-		this.#materializedEntriesCache = materializedEntries;
-		this.#materializedEntriesRevision = this.#entryRevision;
+		const sourceEntries = this.#fileEntries.filter((e): e is SessionEntry => e.type !== "session");
+		const materializedEntries = sourceEntries.map(entry =>
+			materializeResidentEntryForReadSync(entry, this.#residentBlobStores(), resolvedTextBlobCache),
+		);
+		if (!sourceEntries.some(entry => containsResidentImageSentinel(entry))) {
+			this.#materializedEntriesCache = materializedEntries;
+			this.#materializedEntriesRevision = this.#entryRevision;
+		}
 		return materializedEntries;
 	}
 

package/src/tools/bash-interactive.ts

@@ -12,7 +12,6 @@ import {
 } from "@gajae-code/tui";
 import { sanitizeText } from "@gajae-code/utils";
 import type { Terminal as XtermTerminalType } from "@xterm/headless";
-import xterm from "@xterm/headless";
 import { Settings } from "../config/settings";
 import { NON_INTERACTIVE_ENV } from "../exec/non-interactive-env";
 import type { Theme } from "../modes/theme/theme";
@@ -32,8 +31,6 @@ function normalizeCaptureChunk(chunk: string): string {
 	return sanitizeWithOptionalSixelPassthrough(normalized, sanitizeText);
 }
 
-const XtermTerminal = xterm.Terminal;
-
 function normalizeInputForPty(data: string, applicationCursorKeysMode: boolean): string {
 	const kitty = parseKittySequence(data);
 	if (kitty?.eventType === 3) {
@@ -113,6 +110,7 @@ class BashInteractiveOverlayComponent implements Component {
 		private readonly command: string,
 		private readonly uiTheme: Theme,
 		private readonly getTerminalRows: () => number,
+		XtermTerminal: typeof XtermTerminalType,
 	) {
 		this.#terminal = new XtermTerminal({
 			cols: 120,
@@ -305,10 +303,17 @@ export async function runInteractiveBashPty(
 		headBytes: resolveOutputSinkHeadBytes(settings),
 		maxColumns: resolveOutputMaxColumns(settings),
 	});
+	const { default: xterm } = await import("@xterm/headless");
+	const XtermTerminal = xterm.Terminal;
 	const result = await ui.custom<BashInteractiveResult>(
 		(tui, uiTheme, _keybindings, done) => {
 			const session = new PtySession();
-			const component = new BashInteractiveOverlayComponent(options.command, uiTheme, () => tui.terminal.rows);
+			const component = new BashInteractiveOverlayComponent(
+				options.command,
+				uiTheme,
+				() => tui.terminal.rows,
+				XtermTerminal,
+			);
 			component.setSession(session);
 			let finished = false;
 			const finalize = (run: PtyRunResult) => {

package/src/tools/browser/readable.ts

@@ -1,5 +1,5 @@
 import { Readability } from "@mozilla/readability";
-import { parseHTML } from "linkedom";
+import { parseHtmlLazy } from "../../utils/linkedom";
 import { htmlToBasicMarkdown } from "../../web/scrapers/types";
 
 export type ReadableFormat = "text" | "markdown";
@@ -31,7 +31,7 @@ export async function extractReadableFromHtml(
 	url: string,
 	format: ReadableFormat,
 ): Promise<ReadableResult | null> {
-	const { document } = parseHTML(html);
+	const { document } = await parseHtmlLazy(html);
 
 	// --- Primary: Readability article extraction ---
 	const article = new Readability(document).parse();

package/src/tools/fetch.ts

@@ -5,7 +5,6 @@ import type { ImageContent, TextContent } from "@gajae-code/ai";
 import { htmlToMarkdown } from "@gajae-code/natives";
 import { type Component, Text } from "@gajae-code/tui";
 import { $which, ptree, truncate } from "@gajae-code/utils";
-import { parseHTML } from "linkedom";
 import type { Settings } from "../config/settings";
 import type { RenderResultOptions } from "../extensibility/custom-tools/types";
 import { type Theme, theme } from "../modes/theme/theme";
@@ -15,6 +14,7 @@ import { DEFAULT_MAX_BYTES, truncateHead } from "../session/streaming-output";
 import { renderStatusLine } from "../tui";
 import { CachedOutputBlock } from "../tui/output-block";
 import { formatDimensionNote, resizeImage } from "../utils/image-resize";
+import { parseHtmlLazy } from "../utils/linkedom";
 import { ensureTool } from "../utils/tools-manager";
 import { INSANE_NOTES, tryInsaneFetch } from "../web/insane/bridge";
 import { validatePublicHttpUrl, validatePublicHttpUrlForInsane } from "../web/insane/url-guard";
@@ -483,9 +483,9 @@ function cleanFeedText(text: string): string {
 /**
  * Parse RSS/Atom feed to markdown
  */
-function parseFeedToMarkdown(content: string, maxItems = 10): string {
+async function parseFeedToMarkdown(content: string, maxItems = 10): Promise<string> {
 	try {
-		const doc = parseHTML(content).document;
+		const doc = (await parseHtmlLazy(content)).document;
 
 		// Try RSS
 		const channel = doc.querySelector("channel");
@@ -1026,7 +1026,7 @@ async function renderUrl(
 	}
 
 	if (isFeed || (isXml && (rawContent.includes("<rss") || rawContent.includes("<feed")))) {
-		const parsed = parseFeedToMarkdown(rawContent);
+		const parsed = await parseFeedToMarkdown(rawContent);
 		const output = finalizeOutput(parsed);
 		return {
 			url,
@@ -1119,7 +1119,7 @@ async function renderUrl(
 			const altResult = await loadPage(resolved, { timeout, signal });
 			if (altResult.ok && altResult.content.trim().length > 200) {
 				notes.push(`Used feed alternate: ${resolved}`);
-				const parsed = parseFeedToMarkdown(altResult.content);
+				const parsed = await parseFeedToMarkdown(altResult.content);
 				const output = finalizeOutput(parsed);
 				return {
 					url,

package/src/tools/skill.ts

@@ -4,15 +4,16 @@
  * Lets the agent hand off to another available skill in the current turn. The
  * callee's SKILL.md is dispatched through the same custom-message path used by
  * `/skill:<name>` typing, as a user-attribution message delivered same-turn
- * (without `deliverAs: "nextTurn"`). Before dispatch, the tool calls
- * `gjc state <caller> handoff --to <callee>` in-process via the state-runtime
- * function so caller and callee mode-states plus `skill-active-state.json`
- * transition atomically.
+ * (without `deliverAs: "nextTurn"`). When the caller is one of the canonical
+ * workflow skills, the tool first runs the native state handoff so workflow
+ * mode-state and `skill-active-state.json` transition atomically. Runtime
+ * project/user skills have no native workflow mode-state, so they are dispatched
+ * directly and tracked by the prompt observer instead.
  *
- * Chaining is refused unless the caller's `current_phase` is in
- * `{complete, completed, handoff, failed, cancelled, canceled, inactive}`. The
- * agent declares readiness either by writing `current_phase: "handoff"` to its
- * mode-state or by running the handoff verb directly.
+ * Canonical workflow chaining is refused unless the caller's `current_phase` is
+ * in `{complete, completed, handoff, failed, cancelled, canceled, inactive}`.
+ * The agent declares readiness either by writing `current_phase: "handoff"` to
+ * its mode-state or by running the handoff verb directly.
  */
 
 import type { AgentTool, AgentToolResult } from "@gajae-code/agent-core";
@@ -24,6 +25,7 @@ import { buildSkillPromptMessage } from "../extensibility/skills";
 import { runNativeStateCommand } from "../gjc-runtime/state-runtime";
 import skillDescription from "../prompts/tools/skill.md" with { type: "text" };
 import { SKILL_PROMPT_MESSAGE_TYPE } from "../session/messages";
+import { isCanonicalGjcWorkflowSkill } from "../skill-state/active-state";
 import type { ToolSession } from ".";
 import { ToolError } from "./tool-errors";
 
@@ -130,9 +132,10 @@ export class SkillTool implements AgentTool<typeof skillSchema, SkillToolDetails
 				throw new ToolError(`skill tool: unknown skill "${requestedName}".${hint}`);
 			}
 
-			// Phase guard + atomic handoff. Only runs when transitioning between
-			// distinct skills (same-skill recursion was already refused above).
-			if (activeSkill) {
+			// Phase guard + atomic native handoff only apply to canonical workflow
+			// skills. Runtime project/user skills do not have a native mode-state,
+			// so there is no `gjc state <skill>` command to run for them.
+			if (activeSkill && isCanonicalGjcWorkflowSkill(activeSkill)) {
 				const phase = (this.#session.getActiveSkillPhase?.() ?? "running").trim().toLowerCase();
 				if (!TERMINAL_PHASES.has(phase)) {
 					throw new ToolError(

package/src/utils/linkedom.ts

@@ -0,0 +1,14 @@
+type LinkedomModule = typeof import("linkedom");
+
+type ParseHtml = LinkedomModule["parseHTML"];
+
+let parseHtml: ParseHtml | undefined;
+let linkedomImport: Promise<LinkedomModule> | undefined;
+
+export async function parseHtmlLazy(markup: string): Promise<ReturnType<ParseHtml>> {
+	if (!parseHtml) {
+		linkedomImport ??= import("linkedom");
+		parseHtml = (await linkedomImport).parseHTML;
+	}
+	return parseHtml(markup);
+}

package/src/utils/pasted-image-path.ts

@@ -1,12 +1,12 @@
 /**
  * Resolve pasted editor text to an image file path.
  *
- * Terminals insert a shell-escaped filesystem path when a file is drag-dropped
- * onto them (e.g. macOS `Screenshot\ 2026-07-07\ at\ 11.06.38 PM.png`, where
- * the visible "space" before AM/PM is U+202F and stays unescaped). When the
- * entire paste is a single path to an existing image file, the interactive
- * editor attaches the image and inserts an `[image N]` placeholder instead of
- * leaving the raw path in the prompt.
+ * Some terminal/clipboard integrations paste a temporary filesystem path after a
+ * copied image is pasted into the editor (for example `/tmp/clipboard-...png`).
+ * Only those recognized clipboard temp files are auto-attached and replaced with
+ * an `[image N] source="/path"` reference so the model receives both the image
+ * payload and the raw temp file path. Ordinary image paths remain literal prompt
+ * text; users attach saved files explicitly with `@path/to/image.png`.
  */
 import * as fs from "node:fs";
 import * as os from "node:os";
@@ -14,6 +14,7 @@ import * as path from "node:path";
 import { fileURLToPath } from "node:url";
 
 const IMAGE_FILE_EXTENSION_PATTERN = /\.(?:png|jpe?g|gif|webp)$/i;
+const CLIPBOARD_TEMP_BASENAME_PATTERN = /^clipboard-\d{4}-\d{2}-\d{2}-\d{6}-[A-Za-z0-9-]+\.(?:png|jpe?g|gif|webp)$/i;
 
 export interface DecodePastedPathOptions {
 	/**
@@ -150,16 +151,25 @@ function hasSupportedImageMagic(filePath: string): boolean {
 	}
 }
 
+function isRecognizedClipboardTempPath(filePath: string): boolean {
+	const resolved = path.resolve(filePath);
+	const tempRoot = path.resolve(os.tmpdir());
+	if (resolved !== tempRoot && !resolved.startsWith(`${tempRoot}${path.sep}`)) return false;
+	return CLIPBOARD_TEMP_BASENAME_PATTERN.test(path.basename(resolved));
+}
+
 /**
- * Returns the resolved path when the whole pasted text is a single path to an
- * existing image file (verified by extension AND content signature),
- * otherwise `undefined` (the paste is inserted as text).
+ * Returns the resolved path when the whole pasted text is a recognized clipboard
+ * temp path to an existing image file, otherwise `undefined` (the paste is
+ * inserted as text). Arbitrary saved image paths are not auto-attached here;
+ * users attach them explicitly with `@path/to/image.png`.
  */
 export function resolvePastedImagePath(text: string, options?: ResolvePastedImagePathOptions): string | undefined {
 	const candidate = decodePastedPathCandidate(text, options);
 	if (!candidate || !IMAGE_FILE_EXTENSION_PATTERN.test(candidate)) return undefined;
 
 	const resolved = path.resolve(options?.cwd ?? process.cwd(), candidate);
+	if (!isRecognizedClipboardTempPath(resolved)) return undefined;
 	try {
 		if (!fs.statSync(resolved).isFile()) return undefined;
 	} catch {
@@ -168,3 +178,7 @@ export function resolvePastedImagePath(text: string, options?: ResolvePastedImag
 	if (!hasSupportedImageMagic(resolved)) return undefined;
 	return resolved;
 }
+
+export function formatPastedImageReference(placeholder: string, imagePath: string): string {
+	return `${placeholder} source=${JSON.stringify(imagePath)}`;
+}