pi-ui-extend 1.0.41 → 1.0.45

package/external/pi-tools-suite/src/session-recovery/index.ts

@@ -201,6 +201,14 @@ function toolResultFrom(entry: EntryLike): UnknownRecord | undefined {
 	return message?.role === "toolResult" ? message : undefined;
 }
 
+function recoverableGatewayDetails(message: UnknownRecord): UnknownRecord | undefined {
+	const toolName = typeof message.toolName === "string" ? message.toolName.toLowerCase() : "";
+	if (toolName !== "web_search" && toolName !== "web_fetch") return undefined;
+	if (!isRecord(message.details) || !isRecord(message.details.contextGateway)) return undefined;
+	if (message.details.contextGateway.representation !== "web-recoverable-compact") return undefined;
+	return message.details;
+}
+
 function summaryFrom(entry: EntryLike): string {
 	if ((entry.type === "compaction" || entry.type === "branch_summary") && typeof entry.summary === "string") {
 		return entry.summary;
@@ -218,6 +226,11 @@ function entryText(entry: EntryLike): string {
 		for (const call of toolCallsFrom(entry)) {
 			chunks.push(call.name, serializeJson(call.arguments));
 		}
+		if (message.role === "toolResult") {
+			const toolName = typeof message.toolName === "string" ? message.toolName : "unknown";
+			const callId = typeof message.toolCallId === "string" ? message.toolCallId : "unknown";
+			chunks.push(`tool_result ${toolName}#${callId}`);
+		}
 	} else if (entry.type === "custom_message") {
 		const text = textFromContent(entry.content);
 		if (text) chunks.push(text);
@@ -416,6 +429,10 @@ function renderEntryFull(entry: EntryLike): string {
 			const toolName = typeof message.toolName === "string" ? message.toolName : "unknown";
 			const callId = typeof message.toolCallId === "string" ? message.toolCallId : "unknown";
 			lines.push(`tool_result ${toolName}#${callId}${message.isError === true ? " error" : ""}`);
+			const recoverableDetails = recoverableGatewayDetails(message);
+			if (recoverableDetails) {
+				lines.push("recoverable_raw_details:", serializeJson(recoverableDetails));
+			}
 		}
 	} else if (entry.type === "compaction") {
 		lines.push(summaryFrom(entry));

package/external/pi-tools-suite/src/shell-command-policy.ts

@@ -0,0 +1,219 @@
+export type ShellCommandScope = "simple" | "compound" | "unknown";
+export type ShellCommandKind = "inspection" | "test-build" | "mutation" | "unknown";
+
+export interface ShellCommandClassification {
+	scope: ShellCommandScope;
+	kind: ShellCommandKind;
+}
+
+const ENV_ASSIGNMENT_RE = /^[A-Za-z_][A-Za-z0-9_]*=.*$/;
+const SAFE_EPHEMERAL_WRITE_RE = /^\/(?:private\/)?tmp\/[^\0\r\n]+$/;
+const COMMAND_SEPARATORS = new Set([";", "&&", "||", "|"]);
+const REDIRECT_OPERATORS = new Set([">", ">>", "<"]);
+
+interface ShellScan {
+	scope: ShellCommandScope;
+	commands: string[][];
+	unsafeWrite: boolean;
+	unsupported: boolean;
+}
+
+function safeEphemeralWriteTarget(target: string): boolean {
+	return target === "/dev/null" || SAFE_EPHEMERAL_WRITE_RE.test(target);
+}
+
+/**
+ * Minimal conservative shell lexer. It understands quoting only well enough to
+ * keep metacharacters inside quotes literal and to prove a small allowlisted
+ * inspection pipeline safe. Unsupported shell syntax remains unknown.
+ */
+function scanShell(command: string): ShellScan {
+	const tokens: string[] = [];
+	let current = "";
+	let quote: "'" | '"' | undefined;
+	let unsupported = false;
+	const flush = () => {
+		if (current.length > 0) tokens.push(current);
+		current = "";
+	};
+
+	for (let index = 0; index < command.length; index++) {
+		const char = command[index]!;
+		const next = command[index + 1];
+		if (quote) {
+			if (char === quote) {
+				quote = undefined;
+				continue;
+			}
+			if (quote === '"' && char === "\\" && next !== undefined) {
+				current += next;
+				index++;
+				continue;
+			}
+			if (quote === '"' && (char === "`" || (char === "$" && next === "("))) unsupported = true;
+			current += char;
+			continue;
+		}
+
+		if (char === "'" || char === '"') {
+			quote = char;
+			continue;
+		}
+		if (char === "\\" && next !== undefined) {
+			current += next;
+			index++;
+			continue;
+		}
+		if (char === "`" || (char === "$" && next === "(") || ((char === "<" || char === ">") && next === "(")) {
+			unsupported = true;
+		}
+		if (char === "\r" || char === "\n" || char === "&" && next !== "&" || char === "(" || char === ")") {
+			unsupported = true;
+		}
+		if (/\s/.test(char)) {
+			flush();
+			continue;
+		}
+		if (char === ";" || char === "|" || char === ">" || char === "<" || (char === "&" && next === "&")) {
+			flush();
+			let operator = char;
+			if ((char === "&" && next === "&") || (char === "|" && next === "|") || (char === ">" && next === ">") || (char === "<" && next === "<")) {
+				operator += next;
+				index++;
+			}
+			if (operator === "<<") unsupported = true;
+			tokens.push(operator);
+			continue;
+		}
+		current += char;
+	}
+	flush();
+	if (quote) unsupported = true;
+
+	const commands: string[][] = [];
+	let currentCommand: string[] = [];
+	let unsafeWrite = false;
+	let compound = false;
+	for (let index = 0; index < tokens.length; index++) {
+		const token = tokens[index]!;
+		if (COMMAND_SEPARATORS.has(token)) {
+			compound = true;
+			if (currentCommand.length === 0) unsupported = true;
+			else commands.push(currentCommand);
+			currentCommand = [];
+			continue;
+		}
+		if (REDIRECT_OPERATORS.has(token)) {
+			compound = true;
+			const target = tokens[++index];
+			if (!target || COMMAND_SEPARATORS.has(target) || REDIRECT_OPERATORS.has(target)) {
+				unsupported = true;
+				continue;
+			}
+			if ((token === ">" || token === ">>") && !safeEphemeralWriteTarget(target)) unsafeWrite = true;
+			continue;
+		}
+		currentCommand.push(token);
+	}
+	if (currentCommand.length > 0) commands.push(currentCommand);
+	else if (tokens.length > 0 && COMMAND_SEPARATORS.has(tokens.at(-1)!)) unsupported = true;
+
+	for (const words of commands) {
+		while (words.length > 0 && (ENV_ASSIGNMENT_RE.test(words[0]!) || /^\d+$/.test(words[0]!))) words.shift();
+		if (words.length === 0) unsupported = true;
+	}
+	return {
+		scope: unsupported ? "unknown" : compound ? "compound" : "simple",
+		commands,
+		unsafeWrite,
+		unsupported,
+	};
+}
+
+function baseName(value: string): string {
+	return value.replace(/\\/g, "/").split("/").at(-1)?.toLowerCase() ?? "";
+}
+
+function isKnownInspection(words: readonly string[]): boolean {
+	const exe = baseName(words[0]!);
+	if (["pwd", "ls", "grep", "cat", "head", "tail", "wc", "stat", "du", "which", "sort", "uniq", "cut", "tr", "printf", "echo", "basename", "dirname"].includes(exe)) {
+		return true;
+	}
+	if (exe === "find") {
+		return !words.some((word) => ["-delete", "-exec", "-execdir", "-ok", "-okdir"].includes(word));
+	}
+	if (exe === "fd") {
+		return !words.some((word) => ["-x", "--exec", "-X", "--exec-batch"].includes(word));
+	}
+	if (exe === "rg") {
+		return !words.some((word) => word === "--pre" || word.startsWith("--pre="));
+	}
+	if (exe === "command" && words[1] === "-v") return true;
+	if (exe === "sed") return words.includes("-n") && !words.some((word) => /^-.*i/.test(word));
+	if (exe === "jq") return !words.some((word) => word === "-i" || word === "--in-place");
+	if (exe !== "git") return false;
+	const sub = words[1]?.toLowerCase();
+	if (["status", "diff", "log", "show", "grep", "rev-parse", "ls-files", "ls-tree"].includes(sub ?? "")) return true;
+	if (sub === "branch") return words.includes("--show-current");
+	if (sub === "remote") {
+		const action = words[2]?.toLowerCase();
+		return action === undefined || action === "-v" || action === "--verbose" || action === "get-url";
+	}
+	return false;
+}
+
+function isKnownTestBuild(words: readonly string[]): boolean {
+	const exe = baseName(words[0]!);
+	const first = words[1]?.toLowerCase();
+	if (["pytest", "vitest", "jest", "tsc", "eslint", "stylelint"].includes(exe)) return true;
+	if (exe === "biome") return first === "check" || first === "lint";
+	if (exe === "cargo") return ["test", "check", "build", "clippy", "fmt"].includes(first ?? "");
+	if (exe === "go") return first === "test" || first === "vet";
+	if (exe === "bun") {
+		if (first === "test") return true;
+		if (first === "run") return /^(test|check|lint|typecheck|build)(:|$)/.test(words[2]?.toLowerCase() ?? "");
+	}
+	if (["npm", "pnpm", "yarn"].includes(exe)) {
+		if (first === "test") return true;
+		if (first === "run") return /^(test|check|lint|typecheck|build)(:|$)/.test(words[2]?.toLowerCase() ?? "");
+	}
+	return false;
+}
+
+function isKnownMutation(words: readonly string[]): boolean {
+	const exe = baseName(words[0]!);
+	if (["rm", "mv", "cp", "mkdir", "rmdir", "touch", "chmod", "chown", "ln", "install"].includes(exe)) return true;
+	if (exe === "sed" && words.some((word) => /^-.*i/.test(word))) return true;
+	if (exe === "git") {
+		return [
+			"add", "commit", "reset", "checkout", "switch", "restore", "clean", "merge", "rebase",
+			"cherry-pick", "revert", "tag", "push", "pull", "fetch", "stash", "worktree",
+		].includes(words[1]?.toLowerCase() ?? "");
+	}
+	if (["npm", "pnpm", "yarn", "bun"].includes(exe)) {
+		return ["install", "add", "remove", "uninstall", "update", "upgrade", "link", "publish"].includes(words[1]?.toLowerCase() ?? "");
+	}
+	return false;
+}
+
+export function shellCommandText(input: unknown): string | undefined {
+	if (!input || typeof input !== "object" || Array.isArray(input)) return undefined;
+	const command = (input as Record<string, unknown>).command;
+	return typeof command === "string" && command.trim().length > 0 ? command.trim() : undefined;
+}
+
+/**
+ * Conservative, non-executing shell classifier. A command is considered
+ * inspection/test-build only when both its structure and executable/subcommand
+ * are allowlisted. Everything ambiguous fails closed to unknown.
+ */
+export function classifyShellCommand(input: unknown): ShellCommandClassification {
+	const command = shellCommandText(input);
+	if (!command) return { scope: "unknown", kind: "unknown" };
+	const scan = scanShell(command);
+	if (scan.unsupported || scan.commands.length === 0) return { scope: "unknown", kind: "unknown" };
+	if (scan.unsafeWrite || scan.commands.some(isKnownMutation)) return { scope: scan.scope, kind: "mutation" };
+	if (scan.commands.length === 1 && isKnownTestBuild(scan.commands[0]!)) return { scope: scan.scope, kind: "test-build" };
+	if (scan.commands.every(isKnownInspection)) return { scope: scan.scope, kind: "inspection" };
+	return { scope: scan.scope, kind: "unknown" };
+}

package/external/pi-tools-suite/src/todo/index.ts

@@ -1,6 +1,7 @@
 import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { loadPiToolsSuiteConfig, type TodoThinkingLevel as ConfigTodoThinkingLevel } from "../config.js";
 import { isAgentBusyRaceError } from "../context-usage.js";
+import { hasAvailableIndexedProjectRoot } from "../lib/project.js";
 import { autoClearCompletedTodos } from "./state/auto-clear.js";
 import { loadPersistedPlan, syncPersistedPlan } from "./state/persistence.js";
 import { replayFromBranch } from "./state/replay.js";
@@ -32,6 +33,25 @@ const TODO_NUDGE_IDLE_RETRY_DELAY_MS = 100;
 const TODO_NUDGE_MAX_IDLE_ATTEMPTS = 40;
 const ASK_USER_TOOL_NAMES = new Set(["ask_user", "ask_user_question", "question"]);
 const TODO_THINKING_RESTORE_METADATA_KEY = "__piTodoRestoreThinking";
+const REPO_KNOWLEDGE_FINALIZATION_REMINDER =
+	"📚 Before completing the final todo: if behavior/contracts changed, reconcile affected specs and repo knowledge; skip for mechanical changes.";
+
+function completesTodo(info: { action: string; params: TaskMutationParams }): boolean {
+	if (info.action === "update") return info.params.status === "completed";
+	if (info.action !== "batch_update") return false;
+	return (info.params.items ?? []).some((item) => item.status === "completed");
+}
+
+function repoKnowledgeFinalizationReminder(
+	state: ReturnType<typeof getState>,
+	ctx: ExtensionContext,
+	info: { action: string; params: TaskMutationParams },
+): string | undefined {
+	if (!hasAvailableIndexedProjectRoot(ctx.cwd)) return undefined;
+	if (!completesTodo(info)) return undefined;
+	const remainingActive = selectVisibleTasks(state).filter((task) => ACTIVE_STATUSES.has(task.status));
+	return remainingActive.length === 1 ? REPO_KNOWLEDGE_FINALIZATION_REMINDER : undefined;
+}
 
 function isStaleExtensionContextError(error: unknown): boolean {
 	return error instanceof Error && /ctx is stale|stale ctx|stale after session replacement|stale after.*reload/i.test(error.message);
@@ -241,6 +261,7 @@ export default function (pi: ExtensionAPI) {
 		registerTodoTool(pi, {
 			...thinkingPrompt,
 			...(availableThinkingLevels ? { parameters: todoParamsSchemaForThinkingLevels(availableThinkingLevels) } : {}),
+			resultReminder: repoKnowledgeFinalizationReminder,
 			prepareMutation: (state, ctx, info) => {
 				if (!todoThinkingEnabled) return info.params;
 				if (info.action === "create" || info.action === "update") return prepareTodoThinkingMutation(state, ctx, info.params);

package/external/pi-tools-suite/src/todo/todo.ts

@@ -84,6 +84,11 @@ interface TodoToolHooks {
 		ctx: ExtensionContext,
 		info: { action: TaskAction; params: TaskMutationParams },
 	) => TaskMutationParams | Promise<TaskMutationParams>;
+	resultReminder?: (
+		state: ReturnType<typeof getState>,
+		ctx: ExtensionContext,
+		info: { action: TaskAction; params: TaskMutationParams },
+	) => string | undefined | Promise<string | undefined>;
 	afterCommit?: (
 		state: ReturnType<typeof getState>,
 		ctx: ExtensionContext,
@@ -448,7 +453,20 @@ export function registerTodoTool(pi: ExtensionAPI, hooks: TodoToolRegistrationOp
 				params: preparedParams,
 				committedState: autoClear.state,
 			});
-			const toolResult = buildToolResult(params.action, preparedParams, autoClear.state, result.op);
+			let toolResult = buildToolResult(params.action, preparedParams, autoClear.state, result.op);
+			const resultReminder = await hooks.resultReminder?.(result.state, _ctx as ExtensionContext, {
+				action: params.action,
+				params: preparedParams,
+			});
+			if (resultReminder?.trim()) {
+				toolResult = {
+					...toolResult,
+					content: [{
+						type: "text" as const,
+						text: `${toolResult.content[0]?.text ?? ""}\n\n${resultReminder.trim()}`.trim(),
+					}],
+				};
+			}
 			if (!autoClear.cleared) return toolResult;
 			return {
 				...toolResult,

package/external/pi-tools-suite/src/tool-descriptions.ts

@@ -17,6 +17,8 @@ export type RepoDiscoveryToolDescription = ToolDescription & Required<Pick<ToolD
 	targetDescription?: string;
 };
 
+export type RepoKnowledgeToolDescription = ToolDescription & Required<Pick<ToolDescription, "promptSnippet" | "promptGuidelines">>;
+
 export type ToolDescriptionSetOptions = {
 	repoDiscovery?: boolean;
 };
@@ -29,10 +31,9 @@ export const COMPRESS_TOOL_DESCRIPTION: ToolDescription = {
 	name: "compress",
 	label: "Compress Context",
 	description: COMPRESS_RANGE_DESCRIPTION,
-	promptSnippet: "Use compress for context-pressure housekeeping: summarize closed, high-yield stale slices when context is meaningfully high or a DCP reminder gives concrete candidates; do not compress just because low-context work produced a small closed slice.",
+	promptSnippet: "Compress only closed, high-yield stale context when pressure/reminders justify it; keep active or still-needed raw context. Omit summary for DCP summarizer.",
 	promptGuidelines: [
-		"Compression candidates: completed implementation, verification, config/doc edits, answered exploration, dead ends, and log inspection when large/stale or context pressure is meaningful; Low context usage by itself does not require compression.",
-		"Summarize large stale shell/read/repo/web outputs, diffs, and passing logs once exact text is no longer useful; keep active or still-needed context raw.",
+		"Prefer completed work and understood large tool/log output; low context alone is not a trigger.",
 	],
 };
 
@@ -70,11 +71,11 @@ export function asyncSubagentToolDescriptions(options: ToolDescriptionSetOptions
 			name: "subagents",
 			label: "Subagents",
 			description: [
-				"For every real-browser QA request, immediately spawn subagentType='browser-qa' before checking files, URLs, servers, or other prerequisites; the QA sub-agent owns feasibility checks and blocked reports, so the parent must not attempt browser QA itself.",
-				"If browser-qa reports that credentials are required, it must identify the generated project-local template and explicitly ask the user to fill it; the parent relays that request without reading or editing the credential file.",
-				"After browser testing, browser-qa must return clickable links for every available screenshot, video, and trace; the parent must preserve those links in its user-facing report.",
+				"For every real user-interface QA request (browser, terminal/TUI, or desktop GUI), immediately spawn subagentType='ui-qa' before checking files, URLs, servers, launch commands, or other prerequisites; the QA sub-agent owns feasibility checks and blocked reports, so the parent must not substitute static checks for requested UI QA.",
+				"If ui-qa browser testing reports that credentials are required, it must identify the generated project-local template and explicitly ask the user to fill it; the parent relays that request without reading or editing the credential file.",
+				"After UI testing, ui-qa must return clickable links for every available screenshot, terminal capture, video, trace, or other retained evidence; the parent must preserve those links in its user-facing report.",
 				SUBAGENT_DELEGATION_GUIDANCE,
-				"Presets declare available models; each agent's ordered models selects the first usable model in that pool. AGENTS_PRESET or /subagent-preset session <name> selects the current session pool; /subagent-preset init creates a sample config. Do not override the model merely to choose a role.",
+				"Presets declare available models; each agent's ordered models selects the first usable model in that pool. AGENTS_PRESET or /subagent-preset session <name> selects the current session pool. Bundled pools live with agent files, and projects can override/add pools in .pi/agents/presets.jsonc. Do not override the model merely to choose a role.",
 				SUBAGENT_TYPE_SELECTION_GUIDANCE,
 				repoDiscovery
 					? "Use for broad independent tracks, review axes, or hypotheses even though repo_* tools are available."
@@ -83,18 +84,18 @@ export function asyncSubagentToolDescriptions(options: ToolDescriptionSetOptions
 				"Results are compact with artifact links. Agents run isolated pi processes with extensions disabled to prevent recursive spawning; spawn/task timeoutSeconds can shorten the default 30m watchdog, project concurrency queues excess agents, and retry backoff/fallback models/Antigravity account rotation are config-driven.",
 			].join(" "),
 			promptSnippet:
-				"For every browser-based visual QA, UI bug reproduction, or real-browser fix-verification request, immediately spawn subagentType='browser-qa' even for a single track and before inspecting files or checking prerequisites. The browser-qa sub-agent must discover the target and report missing prerequisites; do not preflight, perform, or substitute browser QA in the parent agent. " +
-				"Give browser-qa a concise acceptance brief: the known target URL/app, user-visible flow, expected observable result, and required artifacts. Do not prescribe repository files, searches, commands, server setup, or mock/synthetic substitutes; unknown setup belongs to the QA sub-agent's discovery. " +
+				"For every real UI QA, UI bug reproduction, or user-facing fix-verification request across browser, terminal/TUI, or desktop GUI, immediately spawn subagentType='ui-qa' even for a single track and before inspecting files or checking prerequisites. The ui-qa sub-agent must discover the actual target/control path and report missing prerequisites; do not preflight, perform, or substitute requested UI QA in the parent agent. " +
+				"Give ui-qa a concise acceptance brief: the known target URL/app/command, user-visible flow, expected observable result, and required artifacts. Do not prescribe repository files, searches, commands, server setup, or mock/synthetic substitutes; unknown setup belongs to the QA sub-agent's discovery. " +
 				"For other work, use subagents action='spawn' for economical execution or context isolation, including one bounded sequential task or explicit delegate/parallelize/split work requests. " +
 				SUBAGENT_TYPE_SELECTION_GUIDANCE + " Avoid trivial reads/edits and do not call status/wait immediately after spawn just for progress. " +
 				(repoDiscovery
 					? "For one semantic code-discovery question, use repo_search; for independent tracks/hypotheses/review axes, delegate even when repo_* tools exist. Read result only after completion when findings are needed."
 					: "For one focused code-discovery question, use direct read/grep. Without repo_* tools, delegate bounded research tracks for broad discovery rather than flooding parent context. Read result only after completion when findings are needed."),
 			promptGuidelines: [
-				"Treat every real-browser QA request as a mandatory delegation trigger and an explicit exception to the large/parallel threshold: immediately spawn with `subagentType: \"browser-qa\"` before checking prerequisites. The QA sub-agent owns target discovery, feasibility checks, browser automation, evidence, and blocked reports; the parent must not inspect the project first or substitute non-browser checks.",
-				"Keep the browser-qa task payload at the user-visible acceptance level: known target URL/app, actions to perform, expected observable outcome, and requested evidence. Do not turn it into a repository investigation plan, name internal files or commands, dictate server setup, or invent a mock/synthetic target. Leave unknown prerequisites to the QA sub-agent.",
-				"When browser-qa reports missing credentials, relay its explicit request and generated `.pi/qa_auth.jsonc` template path; never inspect, populate, or edit that credential file in the parent.",
-				"After browser-qa completes a test, preserve its clickable screenshot, video, and trace links in the final user-facing response whenever those artifacts exist.",
+				"Treat every real UI QA request as a mandatory delegation trigger and an explicit exception to the large/parallel threshold: immediately spawn with `subagentType: \"ui-qa\"` before checking prerequisites. The QA sub-agent owns target discovery, feasibility checks, UI automation, evidence, and blocked reports; the parent must not inspect the project first or substitute non-UI checks.",
+				"Keep the ui-qa task payload at the user-visible acceptance level: known target URL/app/command, actions to perform, expected observable outcome, and requested evidence. Do not turn it into a repository investigation plan, name internal files or commands, dictate setup, or invent a mock/synthetic target. Leave unknown prerequisites to the QA sub-agent.",
+				"When ui-qa browser testing reports missing credentials, relay its explicit request and generated `.pi/qa_auth.jsonc` template path; never inspect, populate, or edit that credential file in the parent.",
+				"After ui-qa completes a test, preserve its clickable screenshot, terminal capture, video, trace, and other evidence links in the final user-facing response whenever those artifacts exist.",
 				SUBAGENT_DELEGATION_GUIDANCE,
 				repoDiscovery
 					? "For one discovery question, use repo_search; spawn for independent tracks/hypotheses/review axes, and do not let repo_* availability suppress delegation."
@@ -222,7 +223,24 @@ export const REPO_DISCOVERY_TOOLS: RepoDiscoveryToolDescription[] = [
 	},
 ];
 
-export const REPO_DISCOVERY_TOOL_NAMES = REPO_DISCOVERY_TOOLS.map((tool) => tool.name);
+export const REPO_KNOWLEDGE_TOOL_DESCRIPTION: RepoKnowledgeToolDescription = {
+	name: "repo_knowledge",
+	label: "Repo Knowledge",
+	description:
+		"Project behavioral knowledge/spec maintenance backed by idx. Query contracts with context/search/show/status, review task-scoped impact after material behavior changes, discover new/moved docs, and explicitly record/relate/verify metadata only after semantic evidence review. This tool is exposed only when idx is available and the project is indexed.",
+	promptSnippet:
+		"Use repo_knowledge for project behavior/contracts and their maintenance. For a behavior/contract question that also needs implementation/tests/freshness, prefer action=context; use action=search when only the relevant specs/contracts are needed. Before a material behavior change, find the current primary contract; after implementation, run action=impact on this task's changed paths. Update the existing spec, or create a focused new spec with Edit/Write/apply_patch when no suitable contract exists, then record/relate and verify only after checking code/tests. Skip this lifecycle for mechanical/non-behavioral edits.",
+	promptGuidelines: [
+		"Primary specs/contracts are authoritative; catalog/search summaries, embeddings, graph proximity and semantic candidates are routing evidence only. record is classification, not verification; non-fresh status requires review before presenting behavior as current.",
+		"For material externally visible or project-contract behavior changes: query existing knowledge before/while implementing, keep the authoritative spec aligned in the same task, create a focused primary spec when no suitable contract exists, then run task-scoped impact and reconcile uncovered/new/moved docs before finishing.",
+		"Never persist a relation from similarity alone and never add one just to make coverage non-empty. A reviewed no-impact outcome is valid. verify only after reading the primary source and checking relevant implementation/tests; changed code never auto-rewrites spec semantics.",
+	],
+};
+
+export const REPO_DISCOVERY_TOOL_NAMES = [
+	...REPO_DISCOVERY_TOOLS.map((tool) => tool.name),
+	REPO_KNOWLEDGE_TOOL_DESCRIPTION.name,
+];
 
 export const TODO_TOOL_DESCRIPTION: ToolDescription = {
 	name: "todo",
@@ -333,12 +351,16 @@ export function claudeAliasToolDescriptions(options: ToolDescriptionSetOptions |
 		Edit: {
 			name: "Edit",
 			label: "Edit",
-			description: "Replace exact text in an existing file. Use for surgical edits; use Write only for intentional whole-file replacement.",
+			description: repoDiscovery
+				? "Replace exact text in an existing file. Use for surgical edits; use Write only for whole-file replacement. After a material behavior change, do not finish until repo_knowledge impact reviews this task's changed paths and the authoritative spec is updated or created and verified; skip that lifecycle for mechanical/non-behavioral edits."
+				: "Replace exact text in an existing file. Use for surgical edits; use Write only for intentional whole-file replacement.",
 		},
 		Write: {
 			name: "Write",
 			label: "Write",
-			description: "Create or overwrite a file with complete contents. Use only when replacing the whole file is intended.",
+			description: repoDiscovery
+				? "Create or overwrite a file with complete contents. Use only for intentional whole-file writes. When creating/changing material project behavior, keep or create the primary spec in the same task and finish with task-scoped repo_knowledge impact + semantic verification; mechanical/non-behavioral writes do not require it."
+				: "Create or overwrite a file with complete contents. Use only when replacing the whole file is intended.",
 		},
 		Bash: {
 			name: "Bash",
@@ -365,7 +387,9 @@ export function claudeAliasToolDescriptions(options: ToolDescriptionSetOptions |
 export const CLAUDE_ALIAS_TOOL_DESCRIPTIONS = claudeAliasToolDescriptions(false);
 export const CLAUDE_ALIAS_TOOL_DESCRIPTIONS_WITH_REPO = claudeAliasToolDescriptions(true);
 
-export const CODEX_ALIAS_TOOL_DESCRIPTIONS = {
+export function codexAliasToolDescriptions(options: ToolDescriptionSetOptions | boolean = false) {
+	const repoDiscovery = hasRepoDiscovery(options);
+	return {
 	shellCommand: {
 		name: "shell",
 		label: "shell",
@@ -374,7 +398,7 @@ export const CODEX_ALIAS_TOOL_DESCRIPTIONS = {
 	applyPatch: {
 		name: "apply_patch",
 		label: "apply_patch",
-		description: `Apply file edits with a relative-path patch or standard unified diff. Use for creating, updating, moving, or deleting files; keep each patch focused.
+		description: `Apply file edits with a relative-path patch or standard unified diff. Use for creating, updating, moving, or deleting files; keep each patch focused.${repoDiscovery ? " After a material behavior change, keep/create the authoritative primary spec in the same task and finish with task-scoped repo_knowledge impact plus semantic verification; skip this for mechanical/non-behavioral edits." : ""}
 
 Begin-patch format:
 *** Begin Patch
@@ -388,4 +412,8 @@ Sections: *** Add File (new lines start with +), *** Update File (optionally ***
 
 Unified diff with ---/+++ headers is also supported. Paths must be workspace-relative, never absolute. Provide the complete patch in input.`,
 	},
-} satisfies Record<string, ToolDescription>;
+	} satisfies Record<string, ToolDescription>;
+}
+
+export const CODEX_ALIAS_TOOL_DESCRIPTIONS = codexAliasToolDescriptions(false);
+export const CODEX_ALIAS_TOOL_DESCRIPTIONS_WITH_REPO = codexAliasToolDescriptions(true);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ui-extend",
-  "version": "1.0.41",
+  "version": "1.0.45",
   "description": "Pix: a workspace-first terminal UI for Pi with tabs, readable tool activity, voice input, and bundled agent tools.",
   "private": false,
   "repository": {
@@ -26,7 +26,8 @@
       "import": "./dist/sdk.js"
     },
     "./schemas/pix.json": "./schemas/pix.json",
-    "./schemas/pi-tools-suite.json": "./schemas/pi-tools-suite.json"
+    "./schemas/pi-tools-suite.json": "./schemas/pi-tools-suite.json",
+    "./schemas/tasks.json": "./schemas/tasks.json"
   },
   "files": [
     "assets/screenshots",
@@ -54,6 +55,7 @@
     "prelink:pix": "npm run build:pix --silent",
     "link:pix": "mise exec node@24.16.0 -- npm link --silent",
     "watch:pix": "npm run link:pix --silent && mise exec node@24.16.0 -- node scripts/watch-pix.mjs",
+    "watch:all": "mise exec node@24.16.0 -- node scripts/watch-all.mjs",
     "rebuild:all": "npm run build:desktop",
     "check": "mise exec node@24.16.0 -- npm run check:inner",
     "precheck:inner": "npm run generate-schemas --silent",
@@ -61,6 +63,8 @@
     "test": "mise exec node@24.16.0 -- npm run test:inner",
     "test:inner": "node --import tsx --test \"tests/**/*.test.ts\"",
     "test:tools-suite": "npm --prefix external/pi-tools-suite test",
+    "test:dcp-session-sim": "npm --prefix external/pi-tools-suite run test:dcp-session-sim",
+    "test:dcp-reminder-e2e": "npm --prefix external/pi-tools-suite run test:dcp-reminder-e2e",
     "build:acp": "npm --prefix acp run build",
     "check:acp": "npm --prefix acp run check",
     "dev:desktop": "npm --prefix desktop run dev",
@@ -76,6 +80,7 @@
     "prepack": "npm run generate-schemas --silent",
     "pack:dry-run": "npm pack --dry-run",
     "smoke-test": "mise exec node@24.16.0 -- bash scripts/smoke-test-package.sh",
+    "smoke-test:cli": "mise exec node@24.16.0 -- node scripts/smoke-test-package-cli.mjs",
     "release:check": "npm run check && npm run build:pix && npm run pack:dry-run",
     "publish-npm": "bash scripts/publish.sh",
     "generate-schemas": "tsx scripts/generate-schemas.ts",
@@ -84,10 +89,13 @@
     "prepublishOnly": "npm run check && npm run build:pix && npm run generate-schemas"
   },
   "dependencies": {
+    "@cortexkit/antigravity-auth-core": "2.2.1",
     "@earendil-works/pi-ai": "0.85.1",
     "@earendil-works/pi-coding-agent": "0.85.1",
     "@earendil-works/pi-tui": "0.85.1",
+    "@lydell/node-pty": "1.1.0",
     "@mariozechner/clipboard": "^0.3.9",
+    "@xterm/headless": "^6.0.0",
     "grok-mermaid": "0.2.2",
     "jsonc-parser": "3.3.1",
     "typebox": "1.1.38",
@@ -103,13 +111,6 @@
   "engines": {
     "node": ">=22.19.0 <25"
   },
-  "overrides": {
-    "get-uv-event-loop-napi-h": "1.0.6",
-    "node-addon-api": "8.8.0"
-  },
-  "optionalDependencies": {
-    "vosk": "0.3.39"
-  },
   "keywords": [
     "pi",
     "pix",