@nghyane/arcane 0.1.16 → 0.1.18

package/src/modes/interactive-mode.ts

@@ -15,7 +15,7 @@ import {
 	Text,
 	TUI,
 } from "@nghyane/arcane-tui";
-import { hsvToRgb, isEnoent, logger, postmortem } from "@nghyane/arcane-utils";
+import { isEnoent, logger, postmortem } from "@nghyane/arcane-utils";
 import { APP_NAME, getProjectDir } from "@nghyane/arcane-utils/dirs";
 import chalk from "chalk";
 import { KeybindingsManager } from "../config/keybindings";
@@ -27,7 +27,6 @@ import type { AgentSession, AgentSessionEvent } from "../session/agent-session";
 import { HistoryStorage } from "../session/history-storage";
 import type { SessionContext, SessionManager } from "../session/session-manager";
 import { getRecentSessions } from "../session/session-manager";
-import { STTController, type SttState } from "../stt";
 import { setMermaidRenderCallback } from "../theme/mermaid-cache";
 import type { Theme } from "../theme/theme";
 import { getEditorTheme, getMarkdownTheme, onThemeChange, theme } from "../theme/theme";
@@ -145,11 +144,6 @@ export class InteractiveMode implements InteractiveModeContext {
 	readonly #inputController: InputController;
 	readonly #selectorController: SelectorController;
 	readonly #uiHelpers: UiHelpers;
-	#sttController: STTController | undefined;
-	#voiceAnimationInterval: NodeJS.Timeout | undefined;
-	#voiceHue = 0;
-	#voicePreviousShowHardwareCursor: boolean | null = null;
-	#voicePreviousUseTerminalCursor: boolean | null = null;
 	#resizeHandler?: () => void;
 
 	constructor(
@@ -501,11 +495,6 @@ export class InteractiveMode implements InteractiveModeContext {
 			this.loadingAnimation.stop();
 			this.loadingAnimation = undefined;
 		}
-		this.#cleanupMicAnimation();
-		if (this.#sttController) {
-			this.#sttController.dispose();
-			this.#sttController = undefined;
-		}
 		this.#extensionUiController.clearExtensionTerminalInputListeners();
 		this.statusLine.dispose();
 		if (this.#resizeHandler) {
@@ -726,78 +715,6 @@ export class InteractiveMode implements InteractiveModeContext {
 		return this.#commandController.handleMemoryCommand(text);
 	}
 
-	async handleSTTToggle(): Promise<void> {
-		if (!settings.get("stt.enabled")) {
-			this.showWarning("Speech-to-text is disabled. Enable it in settings: stt.enabled");
-			return;
-		}
-		if (!this.#sttController) {
-			this.#sttController = new STTController();
-		}
-		await this.#sttController.toggle(this.editor, {
-			showWarning: (msg: string) => this.showWarning(msg),
-			showStatus: (msg: string) => this.showStatus(msg),
-			onStateChange: (state: SttState) => {
-				if (state === "recording") {
-					this.#voicePreviousShowHardwareCursor = this.ui.getShowHardwareCursor();
-					this.#voicePreviousUseTerminalCursor = this.editor.getUseTerminalCursor();
-					this.ui.setShowHardwareCursor(false);
-					this.editor.setUseTerminalCursor(false);
-					this.#startMicAnimation();
-				} else if (state === "transcribing") {
-					this.#stopMicAnimation();
-					this.editor.cursorOverride = `\x1b[38;2;200;200;200m${theme.icon.mic}\x1b[0m`;
-					this.editor.cursorOverrideWidth = 1;
-				} else {
-					this.#cleanupMicAnimation();
-				}
-				this.updateEditorTopBorder();
-				this.ui.requestRender();
-			},
-		});
-	}
-
-	#updateMicIcon(): void {
-		const { r, g, b } = hsvToRgb({ h: this.#voiceHue, s: 0.9, v: 1.0 });
-		this.editor.cursorOverride = `\x1b[38;2;${r};${g};${b}m${theme.icon.mic}\x1b[0m`;
-		this.editor.cursorOverrideWidth = 1;
-	}
-
-	#startMicAnimation(): void {
-		if (this.#voiceAnimationInterval) return;
-		this.#voiceHue = 0;
-		this.#updateMicIcon();
-		this.#voiceAnimationInterval = setInterval(() => {
-			this.#voiceHue = (this.#voiceHue + 8) % 360;
-			this.#updateMicIcon();
-			this.ui.requestRender();
-		}, 60);
-	}
-
-	#stopMicAnimation(): void {
-		if (this.#voiceAnimationInterval) {
-			clearInterval(this.#voiceAnimationInterval);
-			this.#voiceAnimationInterval = undefined;
-		}
-	}
-
-	#cleanupMicAnimation(): void {
-		if (this.#voiceAnimationInterval) {
-			clearInterval(this.#voiceAnimationInterval);
-			this.#voiceAnimationInterval = undefined;
-		}
-		this.editor.cursorOverride = undefined;
-		this.editor.cursorOverrideWidth = undefined;
-		if (this.#voicePreviousShowHardwareCursor !== null) {
-			this.ui.setShowHardwareCursor(this.#voicePreviousShowHardwareCursor);
-			this.#voicePreviousShowHardwareCursor = null;
-		}
-		if (this.#voicePreviousUseTerminalCursor !== null) {
-			this.editor.setUseTerminalCursor(this.#voicePreviousUseTerminalCursor);
-			this.#voicePreviousUseTerminalCursor = null;
-		}
-	}
-
 	showDebugSelector(): void {
 		this.#selectorController.showDebugSelector();
 	}

package/src/modes/types.ts

@@ -149,7 +149,6 @@ export interface InteractiveModeContext {
 	handleHandoffCommand(customInstructions?: string): Promise<void>;
 	handleMoveCommand(targetPath: string): Promise<void>;
 	handleMemoryCommand(text: string): Promise<void>;
-	handleSTTToggle(): Promise<void>;
 	executeCompaction(customInstructionsOrOptions?: string | CompactOptions, isAuto?: boolean): Promise<void>;
 	openInBrowser(urlOrPath: string): void;
 	refreshSlashCommandState(cwd?: string): Promise<void>;

package/src/patch/edit-tool.ts

@@ -37,7 +37,6 @@ import {
 	type ReplaceTextEdit,
 } from "./hashline";
 import { detectLineEnding, normalizeToLF, restoreLineEndings, stripBom } from "./normalize";
-import { buildNormativeUpdateInput } from "./normative";
 import {
 	DEFAULT_EDIT_MODE,
 	type EditMode,
@@ -233,9 +232,9 @@ export class EditTool implements AgentTool<TInput, any, Theme> {
 		_toolCallId: string,
 		params: ReplaceParams | PatchParams | HashlineParams,
 		signal?: AbortSignal,
-		_onUpdate?: AgentToolUpdateCallback<EditToolDetails, TInput>,
+		_onUpdate?: AgentToolUpdateCallback<EditToolDetails>,
 		context?: AgentToolContext,
-	): Promise<AgentToolResult<EditToolDetails, TInput>> {
+	): Promise<AgentToolResult<EditToolDetails>> {
 		const batchRequest = getLspBatchRequest(context?.toolCall);
 
 		// ─────────────────────────────────────────────────────────────────
@@ -489,13 +488,6 @@ export class EditTool implements AgentTool<TInput, any, Theme> {
 			}
 			const diffResult = generateDiffString(originalNormalized, result.content);
 
-			const normative = buildNormativeUpdateInput({
-				path,
-				...(rename ? { rename } : {}),
-				oldContent: rawContent,
-				newContent: finalContent,
-			});
-
 			const meta = outputMeta()
 				.diagnostics(diagnostics?.summary ?? "", diagnostics?.messages ?? [])
 				.get();
@@ -516,7 +508,6 @@ export class EditTool implements AgentTool<TInput, any, Theme> {
 					rename,
 					meta,
 				},
-				$normative: normative,
 			};
 		}
 

package/src/patch/hashline.ts

@@ -566,6 +566,37 @@ export function validateLineRef(ref: { line: number; hash: string }, fileLines:
 // Edit Application
 // ═══════════════════════════════════════════════════════════════════════════
 
+/**
+ * Detect suspicious Unicode escape placeholders in edit lines.
+ * LLMs sometimes emit literal `\uDDDD` strings instead of actual Unicode characters.
+ * Returns a warning message if detected, undefined otherwise.
+ */
+function detectUnicodeEscapePlaceholders(lines: string[]): string | undefined {
+	for (const line of lines) {
+		if (/\\u[0-9A-Fa-f]{4}/.test(line)) {
+			return "Warning: edit content contains literal Unicode escape sequences (\\uXXXX). These may be intended as actual Unicode characters.";
+		}
+	}
+	return undefined;
+}
+
+/**
+ * Auto-correct escaped tab indentation in edit lines.
+ * When enabled via ARCANE_HASHLINE_AUTOCORRECT_ESCAPED_TABS=1, replaces
+ * leading `\\t` sequences (literal backslash-t from JSON) with real tab characters.
+ */
+function autocorrectEscapedTabs(lines: string[]): string[] {
+	if (Bun.env.ARCANE_HASHLINE_AUTOCORRECT_ESCAPED_TABS !== "1") {
+		return lines;
+	}
+	return lines.map(line => {
+		const match = line.match(/^((?:\\t)+)/);
+		if (!match) return line;
+		const tabCount = match[1].length / 2; // each \\t is 2 chars
+		return "\t".repeat(tabCount) + line.slice(match[1].length);
+	});
+}
+
 /**
  * Apply an array of hashline edits to file content.
  *
@@ -599,6 +630,16 @@ export function applyHashlineEdits(
 
 	const autocorrect = Bun.env.ARCANE_HL_AUTOCORRECT === "1";
 
+	// Collect warnings and auto-correct edit content
+	const warnings: string[] = [];
+	for (const edit of edits) {
+		const unicodeWarning = detectUnicodeEscapePlaceholders(edit.content);
+		if (unicodeWarning && !warnings.includes(unicodeWarning)) {
+			warnings.push(unicodeWarning);
+		}
+		edit.content = autocorrectEscapedTabs(edit.content);
+	}
+
 	function collectExplicitlyTouchedLines(): Set<number> {
 		const touched = new Set<number>();
 		for (const edit of edits) {
@@ -914,6 +955,7 @@ export function applyHashlineEdits(
 	return {
 		content: finalContent,
 		firstChangedLine,
+		...(warnings.length > 0 ? { warnings } : {}),
 		...(noopEdits.length > 0 ? { noopEdits } : {}),
 	};
 

package/src/prompts/agents/explore.md

@@ -5,7 +5,7 @@ tools: read, grep, find
 model: arcane/fast
 ---
 
-You are a fast, parallel code search agent.
+You are a fast, parallel code search agent running as a subagent inside an AI coding system. Your output goes directly to the main coding agent, not the end user. The main agent invokes you when it needs to locate code by behavior, concept, or multi-step search across the local codebase.
 
 ## Task
 Find files and line ranges relevant to the user's query (provided in the first message).
@@ -30,4 +30,6 @@ Before searching, decompose the query into:
 - Format each file as: `[relativePath#L{start}-L{end}](file://{absolutePath}#L{start}-L{end})`
 - **Use generous line ranges**: Extend ranges to capture complete logical units (full functions, classes, blocks). Add 5-10 lines buffer.
 
-Your final message must contain ONLY the search results — no preamble like "I'll search for...".
+<critical>
+Your final message must contain ONLY the search results — no preamble like "I'll search for...".
+</critical>

package/src/prompts/agents/librarian.md

@@ -6,7 +6,7 @@ model: arcane/fast
 thinking-level: minimal
 ---
 
-<role>Specialized remote repository understanding agent. Explore GitHub repositories, trace code flow across repos, explain architecture, find implementations, and surface relevant history.</role>
+<role>You are the Librarian, a specialized codebase understanding agent that helps answer questions about large, complex codebases across repositories. You are running as a subagent inside an AI coding system — your output goes directly to the main coding agent, not the end user. The main agent invokes you when it needs deep, multi-repository codebase understanding: architecture analysis, cross-repo code tracing, implementation discovery, and history exploration.</role>
 
 <directives>
 - Use the github tool for all repository operations — it handles auth, rate limits, and caching
@@ -16,16 +16,14 @@ thinking-level: minimal
 - Return repository paths (owner/repo + file path) for all referenced files
 </directives>
 
-<github>
+<instruction>
 Use the github tool for all GitHub API operations:
 - `github({ action: "get_file", ... })` for reading remote files
 - `github({ action: "get_tree", ... })` for listing directories
 - `github({ action: "get_issue", ... })` for reading issues with all comments
 - `github({ action: "get_pull", ... })` for PR details and diffs
 - `github({ action: "list_commits", ... })` for commit history
-</github>
 
-<search>
 Use search_code to find code across public GitHub repositories via grep.app:
 - `search_code({ query: "pattern" })` for broad cross-repo search
 - `search_code({ query: "pattern", repo: "owner/repo" })` for searching within a specific repo
@@ -33,7 +31,7 @@ Use search_code to find code across public GitHub repositories via grep.app:
 - Supports regex via `regexp: true`
 - Returns snippets with line numbers and match counts
 - No auth required, better snippets than GitHub Code Search API
-</search>
+</instruction>
 
 <procedure>
 1. Identify target repositories
@@ -59,4 +57,4 @@ Be comprehensive and direct. No filler.
 Only your final message is returned to the caller. It must be self-contained with all findings, paths, and explanations. Do not reference tool names or intermediate steps — present conclusions directly. Your final message must contain ONLY the information found — no preamble.
 
 Use "fluent" linking — embed file/PR/commit references in natural noun phrases, not raw URLs. Example: The [`handleAuth` function](file:///path/to/auth.ts#L42) validates tokens.
-</critical>
+</critical>

package/src/prompts/agents/reviewer.md

@@ -6,7 +6,7 @@ model: arcane/reviewer
 thinking-level: high
 ---
 
-<role>Senior engineer reviewing a proposed change. Identify bugs the author would want fixed before merge.</role>
+<role>You are a senior engineer reviewing a proposed change. You are running as a subagent inside an AI coding system — your output goes directly to the main coding agent, not the end user. The main agent invokes you to identify bugs the author would want fixed before merge.</role>
 
 <procedure>
 1. Run `git diff` (or `gh pr diff <number>`) to view patch

package/src/prompts/agents/task.md

@@ -18,5 +18,9 @@ Do the task end to end. Don’t hand back half-baked work.
 - Prefer edits to existing files over creating new ones. NEVER create documentation files (*.md) unless explicitly requested.
 - When done, write a concise summary of what you did as your final response. This is your output.
 - Use tools to get feedback on your generated code. Run diagnostics and type checks. If build/test commands aren’t known, find them in the environment.
-- Follow the main agent’s instructions and AGENTS.md conventions.
+- Follow the main agent's instructions and AGENTS.md conventions.
+
+<critical>
+Keep going until request is fully fulfilled. This matters.
+</critical>
 </directives>

package/src/prompts/system/system-prompt.md

@@ -1,5 +1,5 @@
 <identity>
-You are a distinguished staff engineer operating inside Arcane, a Pi-based coding harness.
+You are a distinguished staff engineer operating inside Arcane, a Pi-based coding harness. You are the main agent — you interact directly with the user and orchestrate subagents (explore, librarian, oracle, reviewer, task) for complex work.
 
 High-agency. Principled. Decisive.
 Correctness > politeness. Brevity > ceremony.
@@ -23,11 +23,16 @@ Balance initiative with predictability:
 </environment>
 
 ## Tool Usage
-- Use specialized tools instead of Bash for file operations.
+- Use specialized tools instead of Bash for file operations. Use read instead of `cat`/`head`/`tail`, edit instead of `sed`/`awk`, and write instead of echo redirection or heredoc. Reserve Bash for actual system commands.
 - Prefer doing work directly — you retain full context and produce better results.
 - Gather-then-act: collect all needed context first (parallel reads, greps, finds), then make changes. Do not interleave reading and editing one file at a time.
 - When exploring the codebase to gather context, prefer explore over running search commands directly. It reduces context usage and provides better results.
 
+## Editing Files
+- NEVER create files unless absolutely necessary for achieving the goal. ALWAYS prefer editing an existing file to creating a new one.
+- When changing an existing file, use edit. Only use write for files that do not exist yet.
+- Make the smallest reasonable diff. Do not rewrite whole files to change a few lines.
+
 ## Parallel Execution Policy
 Default to **parallel** for all independent work: reads, searches, diagnostics, writes to disjoint files, and subagents. Serialize only when there is a strict dependency (shared file, chained output).
 - Run multiple explore, oracle, or task calls in parallel when concerns are distinct.
@@ -134,6 +139,8 @@ You have three types of subagents (task, oracle, codebase search):
 - Use for: Feature scaffolding, cross-layer refactors, mass migrations, boilerplate generation, changes across many layers after planning.
 - Don't use for: Exploratory work, architectural decisions, debugging analysis, single logical task, reading a single file, editing a single file. Never spawn a single Task call for work you can do yourself.
 - Prompt it with detailed instructions on the goal, enumerate the deliverables, give it step by step procedures and ways to validate the results. Also give it constraints (e.g. coding style) and include relevant context snippets or examples.
+- Include the project's coding conventions relevant to the task — extract from AGENTS.md or surrounding code. Task agents do not internalize project-specific conventions; they rely on what you provide.
+- After a task completes, read its modified files to verify style and correctness. Do not trust task output blindly.
 
 #### Oracle
 - Senior engineering advisor with deep reasoning for reviews, architecture, deep debugging, and planning.
@@ -162,11 +169,22 @@ Best practices:
 - Run multiple sub-agents concurrently if tasks are independent with disjoint write targets.
 {{/has}}
 
+### Cross-session Knowledge
+
+Tools: `find_thread`, `read_thread`, `save_memory`
+**Proactive search triggers** — use `find_thread` when:
+- User mentions past work: "we did this before", "last time", "in a previous session"
+- User asks "what did we do about X" or "how did we solve Y"
+- Task seems related to work that may have been done before
+- Handoff context references a parent thread and you need more detail
+**Do NOT search when:**
+- Question is about current session context
+- Generic coding question with no project-specific history
+- User explicitly provides all needed context
+**save_memory**: only when user says "remember this" or states a clear preference. If unsure, ask.
+
 ### Verification
-After completing changes, verify using commands from AGENTS.md or the project's config. Format → typecheck/lint → test (if relevant) → build (if required).
-Report evidence concisely: counts, pass/fail, error summary.
-If unrelated pre-existing failures block you, say so and scope your change.
-Address all errors caused by your changes before yielding.
+Work incrementally. Make a small change, verify it works, then continue. Prefer a sequence of small, validated edits over one large change. Use commands from AGENTS.md or the project's config to verify. Address all errors caused by your changes before yielding.
 
 ### Concurrency Awareness
 You are not alone in the codebase. Others may edit concurrently.
@@ -185,7 +203,7 @@ Never run destructive git commands, bulk overwrites, or delete code you didn't w
 - Resolve blockers before yielding.
 </procedure>
 
-<contract>
+<critical>
 These are inviolable. Violation is system failure.
 1. Never claim unverified correctness. Verify the effect — confirm behavioral changes are observable.
 2. Never yield unless your deliverable is complete. Fix errors you introduced before yielding.
@@ -194,7 +212,10 @@ These are inviolable. Violation is system failure.
 5. Never solve the wished-for problem instead of the actual problem.
 6. Never ask for information obtainable from tools, repo context, or files.
 7. Full cutover within scope — update every call site. No backwards-compat shims.
-</contract>
+
+Keep going until fully resolved. This matters.
+</critical>
+
 
 <project>
 {{#if contextFiles.length}}
@@ -252,16 +273,6 @@ Scan descriptions vs task domain — read skill if ≥50% likely relevant.
 </rules>
 {{/if}}
 
-{{#if memories.length}}
-<memories>
-{{#each memories}}
-<memory path="{{path}}">
-{{content}}
-</memory>
-{{/each}}
-</memories>
-{{/if}}
-
 {{#if preloadedSkills.length}}
 {{#each preloadedSkills}}
 <skill name="{{name}}">

package/src/prompts/thread-extract.md

@@ -0,0 +1,16 @@
+You are helping extract relevant information from a conversation thread based on a goal.
+
+## Task
+
+I am providing a conversation thread rendered as markdown, along with a goal describing what information to extract.
+
+Your job is to:
+1. Analyze the thread content
+2. Identify information that is relevant to the goal
+3. Extract and preserve those relevant parts with full fidelity
+
+## Rules
+- Be concise but complete — include all relevant details
+- Preserve code snippets, file paths, commands, and decisions exactly as they appear
+- Omit pleasantries, failed attempts, and thinking-out-loud unless the goal asks for them
+- If nothing relevant is found, say so briefly

package/src/prompts/tools/render-mermaid.md

@@ -0,0 +1,9 @@
+Convert Mermaid graph source into ASCII diagram output.
+
+Parameters:
+- `mermaid` (required): Mermaid graph text to render.
+- `config` (optional): JSON render configuration (spacing and layout options).
+Behavior:
+- Returns ASCII diagram text.
+- Saves full ASCII output to an artifact URL (`artifact://<id>`) when artifact storage is available.
+- Returns an error when the Mermaid input is invalid or rendering fails.

package/src/sdk.ts

@@ -42,13 +42,11 @@ import {
 	ArtifactProtocolHandler,
 	DocsProtocolHandler,
 	InternalUrlRouter,
-	MemoryProtocolHandler,
 	RuleProtocolHandler,
 	SkillProtocolHandler,
 } from "./internal-urls";
 import { disposeAllKernelSessions } from "./ipy/executor";
 import { discoverAndLoadMCPTools, type MCPManager, type MCPToolsLoadResult } from "./mcp";
-import { buildMemoryToolDeveloperInstructions, getMemoryRoot, startMemoryStartupTask } from "./memories";
 import { collectEnvSecrets, loadSecrets, obfuscateMessages, SecretObfuscator } from "./secrets";
 import { AgentSession } from "./session/agent-session";
 import { AuthStorage } from "./session/auth-storage";
@@ -735,7 +733,7 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		settings,
 	};
 
-	// Initialize internal URL router for internal protocols (agent://, artifact://, memory://, skill://, rule://)
+	// Initialize internal URL router for internal protocols (agent://, artifact://, skill://, rule://)
 	const internalRouter = new InternalUrlRouter();
 	const getArtifactsDir = () => {
 		const sessionFile = sessionManager.getSessionFile();
@@ -743,11 +741,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 	};
 	internalRouter.register(new AgentProtocolHandler({ getArtifactsDir }));
 	internalRouter.register(new ArtifactProtocolHandler({ getArtifactsDir }));
-	internalRouter.register(
-		new MemoryProtocolHandler({
-			getMemoryRoot: () => getMemoryRoot(agentDir, settings.getCwd()),
-		}),
-	);
 	internalRouter.register(
 		new SkillProtocolHandler({
 			getSkills: () => skills,
@@ -937,16 +930,10 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		}
 	}
 
-	// Discover custom commands (TypeScript slash commands)
-	const customCommandsResult: CustomCommandsLoadResult = options.disableExtensionDiscovery
-		? { commands: [], errors: [] }
-		: await loadCustomCommandsInternal({ cwd, agentDir });
-	time("discoverCustomCommands");
-	if (!options.disableExtensionDiscovery) {
-		for (const { path, error } of customCommandsResult.errors) {
-			logger.error("Failed to load custom command", { path, error });
-		}
-	}
+	// Start custom commands discovery early (awaited later in parallel)
+	const customCommandsPromise = options.disableExtensionDiscovery
+		? Promise.resolve({ commands: [], errors: [] } as CustomCommandsLoadResult)
+		: loadCustomCommandsInternal({ cwd, agentDir });
 
 	let extensionRunner: ExtensionRunner | undefined;
 	if (extensionsResult.extensions.length > 0) {
@@ -1027,7 +1014,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 
 	const rebuildSystemPrompt = async (toolNames: string[], tools: Map<string, AgentTool>): Promise<string> => {
 		toolContextStore.setToolNames(toolNames);
-		const memoryInstructions = await buildMemoryToolDeveloperInstructions(agentDir, settings);
 		const defaultPrompt = await buildSystemPromptInternal({
 			cwd,
 			skills,
@@ -1037,7 +1023,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			toolNames,
 			rules: rulebookRules,
 			skillsSettings: settings.getGroup("skills") as SkillsSettings,
-			appendSystemPrompt: memoryInstructions,
 		});
 
 		if (options.systemPrompt === undefined) {
@@ -1054,7 +1039,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 				rules: rulebookRules,
 				skillsSettings: settings.getGroup("skills") as SkillsSettings,
 				customPrompt: options.systemPrompt,
-				appendSystemPrompt: memoryInstructions,
 			});
 		}
 		return options.systemPrompt(defaultPrompt);
@@ -1076,16 +1060,15 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		}
 	}
 
-	const systemPrompt = await rebuildSystemPrompt(initialToolNames, toolRegistry);
-	time("buildSystemPrompt");
-
-	const promptTemplates = options.promptTemplates ?? (await discoverPromptTemplates(cwd, agentDir));
-	time("discoverPromptTemplates");
+	const [systemPrompt, promptTemplates, slashCommands, customCommandsResult] = await Promise.all([
+		rebuildSystemPrompt(initialToolNames, toolRegistry),
+		options.promptTemplates ?? discoverPromptTemplates(cwd, agentDir),
+		options.slashCommands ?? discoverSlashCommands(cwd),
+		customCommandsPromise,
+	]);
+	time("buildSystemPrompt+discoverPromptTemplates+discoverSlashCommands");
 	toolSession.promptTemplates = promptTemplates;
 
-	const slashCommands = options.slashCommands ?? (await discoverSlashCommands(cwd));
-	time("discoverSlashCommands");
-
 	// Create convertToLlm wrapper that filters images if blockImages is enabled (defense-in-depth)
 	const convertToLlmWithBlockImages = (messages: AgentMessage[]): Message[] => {
 		const converted = convertToLlm(messages);
@@ -1258,14 +1241,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		}
 	}
 
-	startMemoryStartupTask({
-		session,
-		settings,
-		modelRegistry,
-		agentDir,
-		isSubagent,
-	});
-
 	return {
 		session,
 		extensionsResult,

package/src/session/agent-session.ts

@@ -115,7 +115,6 @@ import {
 	maybeAbortStreamingEdit,
 	preCacheStreamingEditFile,
 	resetStreamingEditState,
-	rewriteToolCallArgs,
 } from "./streaming-edit";
 import {
 	addPendingTtsrInjections,
@@ -523,17 +522,12 @@ export class AgentSession {
 			}
 
 			if (event.message.role === "toolResult") {
-				const { toolName, $normative, toolCallId, details, isError, content } = event.message as {
+				const { toolName, details, isError, content } = event.message as {
 					toolName?: string;
-					toolCallId?: string;
 					details?: { path?: string };
-					$normative?: Record<string, unknown>;
 					isError?: boolean;
 					content?: Array<TextContent | ImageContent>;
 				};
-				if ($normative && toolCallId && this.settings.get("normativeRewrite")) {
-					await rewriteToolCallArgs(this.agent, this.sessionManager, toolCallId, $normative);
-				}
 				// Invalidate streaming edit cache when edit tool completes to prevent stale data
 				if (toolName === "edit" && details?.path) {
 					invalidateFileCacheForPath(this.#streamingEdit, details.path, this.sessionManager.getCwd());
@@ -2067,9 +2061,10 @@ Be thorough - include exact file paths, function names, error messages, and tech
 				return undefined;
 			}
 
-			// Start a new session
+			// Start a new session with parent reference
+			const parentThreadId = this.sessionManager.getSessionId();
 			await this.sessionManager.flush();
-			await this.sessionManager.newSession();
+			await this.sessionManager.newSession({ parentSession: parentThreadId });
 			this.agent.reset();
 			this.agent.sessionId = this.sessionManager.getSessionId();
 			this.#steeringMessages = [];
@@ -2078,7 +2073,7 @@ Be thorough - include exact file paths, function names, error messages, and tech
 			this.#todoReminderCount = 0;
 
 			// Inject the handoff document as a custom message
-			const handoffContent = `<handoff-context>\n${handoffText}\n</handoff-context>\n\nThe above is a handoff document from a previous session. Use this context to continue the work seamlessly.`;
+			const handoffContent = `<handoff-context thread="${parentThreadId}">\n${handoffText}\n</handoff-context>\n\nThe above is a handoff document from thread \`${parentThreadId}\`. Use this context to continue the work seamlessly. If you need additional details not covered above, use \`read_thread("${parentThreadId}", "your specific question")\` to query the original session.`;
 			this.sessionManager.appendCustomMessageEntry("handoff", handoffContent, true);
 
 			// Rebuild agent messages from session

package/src/session/retry-utils.ts

@@ -17,7 +17,7 @@ export function isRetryableErrorMessage(errorMessage: string): boolean {
  * Check if an error message indicates a usage/billing limit (non-transient).
  */
 export function isUsageLimitErrorMessage(errorMessage: string): boolean {
-	return /usage.?limit|usage_limit_reached|limit_reached/i.test(errorMessage);
+	return /usage.?limit|usage_limit_reached|limit_reached|quota.?exhaust/i.test(errorMessage);
 }
 
 /**