pi-agent-browser-native 0.2.43 → 0.2.45

package/scripts/platform-smoke/targets.mjs

@@ -1,5 +1,6 @@
 /** Target/suite runner for pi-agent-browser-native platform smoke. */
 
+import { execFileSync } from "node:child_process";
 import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { dirname, resolve } from "node:path";
 
@@ -14,7 +15,8 @@ import {
 	writeManifest,
 	writeSummary,
 } from "./artifacts.mjs";
-import { cleanupStaleTargetState, runOnLease, stopLease, warmupLease } from "./crabbox-runner.mjs";
+import { CAPABILITY_BASELINE } from "../agent-browser-capability-baseline.mjs";
+import { cleanupStaleTargetState, crabboxBin, describeTarget, runOnLease, stopLease, warmupLease } from "./crabbox-runner.mjs";
 
 export function platformFor(targetName) {
 	return targetName === "windows-native" ? "powershell" : "posix";
@@ -38,6 +40,45 @@ function authEnvAllowList(config = {}) {
 	return names.map((name) => String(name).trim()).filter(Boolean);
 }
 
+function packageVersion() {
+	try {
+		return JSON.parse(readFileSync("package.json", "utf8")).version ?? null;
+	} catch {
+		return null;
+	}
+}
+
+function crabboxVersion() {
+	try {
+		return execFileSync(crabboxBin(), ["--version"], { encoding: "utf8", stdio: "pipe", timeout: 10_000 }).trim().split(/\r?\n/)[0] ?? null;
+	} catch {
+		return null;
+	}
+}
+
+function targetEvidence(config, targetName, runId, slug) {
+	const target = describeTarget(targetName, config);
+	return {
+		targetName,
+		platform: platformFor(targetName),
+		runId,
+		slug,
+		packageName: config.packageName,
+		packageVersion: packageVersion(),
+		crabbox: {
+			binary: crabboxBin(),
+			version: crabboxVersion(),
+			provider: target.provider,
+			target: target.crabboxTarget,
+			workRoot: target.workRoot,
+			image: target.image,
+			windowsMode: target.windowsMode,
+			sourceVm: target.sourceVm,
+			snapshot: target.snapshot,
+		},
+	};
+}
+
 function writeRedacted(path, text, secretValues) {
 	writeFileSync(path, redactSecrets(text ?? "", secretValues));
 }
@@ -141,12 +182,11 @@ function finalizeSuite(suiteDir, checks, summary, expectedFiles) {
 	return { assertions: finalAssertions, manifest: writeManifest(suiteDir, [...expectedFiles, "failures.md"]) };
 }
 
-export function createLeaseCleanupResult(config, targetName, leaseId, stopResult, staleCleanupResult = null) {
+export function createLeaseCleanupResult(config, targetName, leaseId, stopResult, staleCleanupResult = null, runId = makeRunId()) {
 	const suiteName = "lease-cleanup";
-	const runId = makeRunId();
 	const suiteDir = createSuiteDir(config.artifactRoot, runId, targetName, suiteName);
 	const secretValues = collectSecretValues(authEnvAllowList(config));
-	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify({ targetName, platform: platformFor(targetName), runId, slug: `${config.packageName}-${targetName}` }, null, 2));
+	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify(targetEvidence(config, targetName, runId, `${config.packageName}-${targetName}`), null, 2));
 	writeFileSync(resolve(suiteDir, "suite.json"), JSON.stringify({ suiteName, leaseId, modelCalls: 0 }, null, 2));
 	writeCommand(suiteDir, `crabbox stop ${targetName} --id ${leaseId}`);
 	writeExitCode(suiteDir, stopResult.code, stopResult.signal);
@@ -180,16 +220,15 @@ export function createLeaseCleanupResult(config, targetName, leaseId, stopResult
 	return { ok: assertions.ok, suiteDir, assertions };
 }
 
-export function createLeaseCleanupFailureResult(config, targetName, leaseId, stopResult) {
-	return createLeaseCleanupResult(config, targetName, leaseId, stopResult);
+export function createLeaseCleanupFailureResult(config, targetName, leaseId, stopResult, runId) {
+	return createLeaseCleanupResult(config, targetName, leaseId, stopResult, null, runId);
 }
 
-export function createLeaseWarmupFailureResult(config, targetName, warmupResult) {
+export function createLeaseWarmupFailureResult(config, targetName, warmupResult, runId = makeRunId()) {
 	const suiteName = "lease-warmup";
-	const runId = makeRunId();
 	const suiteDir = createSuiteDir(config.artifactRoot, runId, targetName, suiteName);
 	const secretValues = collectSecretValues(authEnvAllowList(config));
-	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify({ targetName, platform: platformFor(targetName), runId, slug: `${config.packageName}-${targetName}` }, null, 2));
+	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify(targetEvidence(config, targetName, runId, `${config.packageName}-${targetName}`), null, 2));
 	writeFileSync(resolve(suiteDir, "suite.json"), JSON.stringify({ suiteName, modelCalls: 0 }, null, 2));
 	writeCommand(suiteDir, `crabbox warmup ${targetName}`);
 	writeExitCode(suiteDir, warmupResult.code, warmupResult.signal);
@@ -262,7 +301,7 @@ export function buildPlatformBuildCommand(targetName, packageName = "pi-agent-br
 	return lines.join("\n");
 }
 
-export function buildBrowserDogfoodCommand(targetName, agentBrowserVersion = "0.27.1") {
+export function buildBrowserDogfoodCommand(targetName, agentBrowserVersion = CAPABILITY_BASELINE.targetVersion) {
 	if (platformFor(targetName) === "powershell") {
 		return `powershell.exe -NoLogo -NoProfile -ExecutionPolicy Bypass -File .\\scripts\\platform-smoke\\browser-dogfood-windows.ps1 -AgentBrowserVersion ${psSingleQuote(agentBrowserVersion)}`;
 	}
@@ -301,14 +340,13 @@ export function buildBrowserDogfoodCommand(targetName, agentBrowserVersion = "0.
 	return lines.join("\n");
 }
 
-async function runBrowserDogfoodSuite(config, targetName, suiteName, leaseSession) {
-	const runId = makeRunId();
+async function runBrowserDogfoodSuite(config, targetName, suiteName, leaseSession, runId = makeRunId()) {
 	const suiteDir = createSuiteDir(config.artifactRoot, runId, targetName, suiteName);
 	const startedAt = Date.now();
 	const platform = platformFor(targetName);
 	const slug = `${config.packageName}-${targetName}`;
 	const command = buildBrowserDogfoodCommand(targetName, config.agentBrowserVersion);
-	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify({ targetName, platform, runId, slug }, null, 2));
+	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify(targetEvidence(config, targetName, runId, slug), null, 2));
 	writeFileSync(resolve(suiteDir, "suite.json"), JSON.stringify({ suiteName, modelCalls: 0, realBrowser: true }, null, 2));
 	writeCommand(suiteDir, command);
 
@@ -367,15 +405,14 @@ async function runBrowserDogfoodSuite(config, targetName, suiteName, leaseSessio
 	return { ok: assertions.ok, suiteDir, assertions };
 }
 
-async function runPlatformBuildSuite(config, targetName, suiteName, leaseSession) {
-	const runId = makeRunId();
+async function runPlatformBuildSuite(config, targetName, suiteName, leaseSession, runId = makeRunId()) {
 	const suiteDir = createSuiteDir(config.artifactRoot, runId, targetName, suiteName);
 	const startedAt = Date.now();
 	const platform = platformFor(targetName);
 	const slug = `${config.packageName}-${targetName}`;
 	const command = buildPlatformBuildCommand(targetName, config.packageName, config.nodeValidationMajor);
 	mkdirSync(dirname(suiteDir), { recursive: true });
-	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify({ targetName, platform, runId, slug }, null, 2));
+	writeFileSync(resolve(suiteDir, "target.json"), JSON.stringify(targetEvidence(config, targetName, runId, slug), null, 2));
 	writeFileSync(resolve(suiteDir, "suite.json"), JSON.stringify({ suiteName, modelCalls: 0 }, null, 2));
 	writeCommand(suiteDir, command);
 
@@ -436,17 +473,18 @@ async function runPlatformBuildSuite(config, targetName, suiteName, leaseSession
 	return { ok: assertions.ok, suiteDir, assertions };
 }
 
-export async function runTargetSuite(config, targetName, suiteName, leaseSession) {
-	if (suiteName === "platform-build") return await runPlatformBuildSuite(config, targetName, suiteName, leaseSession);
-	if (suiteName === "browser-dogfood-smoke") return await runBrowserDogfoodSuite(config, targetName, suiteName, leaseSession);
+export async function runTargetSuite(config, targetName, suiteName, leaseSession, runId) {
+	if (suiteName === "platform-build") return await runPlatformBuildSuite(config, targetName, suiteName, leaseSession, runId);
+	if (suiteName === "browser-dogfood-smoke") return await runBrowserDogfoodSuite(config, targetName, suiteName, leaseSession, runId);
 	throw new Error(`unknown suite: ${suiteName}`);
 }
 
 export async function runTargetSuites(config, targetName, suiteNames) {
 	const slug = `${config.packageName}-${targetName}`;
+	const runId = makeRunId();
 	const lease = await warmupLease(targetName, slug, config);
 	if (!lease.ok) {
-		const warmupFailure = createLeaseWarmupFailureResult(config, targetName, lease);
+		const warmupFailure = createLeaseWarmupFailureResult(config, targetName, lease, runId);
 		return { ok: false, results: [warmupFailure] };
 	}
 	const results = [];
@@ -455,7 +493,7 @@ export async function runTargetSuites(config, targetName, suiteNames) {
 	try {
 		let sync = true;
 		for (const suiteName of suiteNames) {
-			const result = await runTargetSuite(config, targetName, suiteName, { ...lease, sync });
+			const result = await runTargetSuite(config, targetName, suiteName, { ...lease, sync }, runId);
 			results.push(result);
 			sync = false;
 			if (!result.ok) break;
@@ -465,7 +503,7 @@ export async function runTargetSuites(config, targetName, suiteNames) {
 		staleCleanupResult = await cleanupStaleTargetState(targetName, config);
 	}
 	if (stopResult) {
-		results.push(createLeaseCleanupResult(config, targetName, lease.leaseId, stopResult, staleCleanupResult));
+		results.push(createLeaseCleanupResult(config, targetName, lease.leaseId, stopResult, staleCleanupResult, runId));
 	}
 	return { ok: results.every((result) => result.ok), results };
 }

package/scripts/platform-smoke.mjs

@@ -61,6 +61,7 @@ Environment:
   PLATFORM_SMOKE_MAC_HOST             macOS SSH host; default localhost
   PLATFORM_SMOKE_MAC_USER             macOS SSH user; default $USER
   PLATFORM_SMOKE_MAC_WORK_ROOT        macOS Crabbox work root
+  PLATFORM_SMOKE_MAC_PORT             macOS SSH port; default 22
   PLATFORM_SMOKE_UBUNTU_IMAGE         Ubuntu local-container image; default pi-agent-browser-native-platform:node24-agent-browser0.27.1
   PLATFORM_SMOKE_WINDOWS_VM           Parallels Windows template VM
   PLATFORM_SMOKE_WINDOWS_SNAPSHOT     Parallels snapshot name

package/extensions/agent-browser/lib/orchestration/browser-run/browser-action-model.ts

@@ -1,154 +0,0 @@
-/**
- * Purpose: Normalize planned browser argv into a small action model for prompt-derived guards.
- * Responsibilities: Map command tokens and batch stdin steps to click-like and keyboard-submit actions with target labels.
- * Scope: Best-effort finalizing-action detection only; does not model eval, generic fill/type, or non-Enter keyboard flows.
- */
-
-import type { SessionRefSnapshot } from "../../session-page-state.js";
-import { parseValidBatchStepEntries } from "../batch-stdin.js";
-
-const FINAL_ACTION_PATTERN = /\b(?:finish|place\s+(?:the\s+)?order|submit\s+(?:the\s+)?order|complete\s+(?:the\s+)?order|confirm\s+(?:the\s+)?order|purchase|buy\s+now|pay\s+now|finali[sz]e|submit\s+payment|checkout\s+complete)\b/i;
-
-const CLICK_LIKE_COMMANDS = new Set(["click", "dblclick", "tap"]);
-const FIND_CLICK_ACTIONS = new Set(["click", "dblclick", "tap"]);
-const KEYBOARD_SUBMIT_KEYS = new Set(["enter", "return"]);
-
-export type BrowserFinalizingActionKind = "click-like" | "keyboard-submit";
-
-export interface BrowserFinalizingAction {
-	command: string[];
-	kind: BrowserFinalizingActionKind;
-	stepIndex?: number;
-	targetLabel?: string;
-}
-
-export const STOP_BOUNDARY_GUARD_SCOPE = {
-	covered: [
-		"standalone click, dblclick, and tap",
-		"find … click|dblclick|tap",
-		"batch steps with the click-like shapes above",
-		"press <key> and key <key> when key is Enter or Return",
-	],
-	excluded: [
-		"eval --stdin and other scripted activation",
-		"fill, type, select, drag, and upload without an explicit click-like command",
-		"keyboard type/inserttext and keyboard shortcuts other than Enter/Return",
-		"semanticAction and job/qa compiled plans unless their batch stdin contains a covered step",
-	],
-} as const;
-
-function normalizeTargetText(value: string): string {
-	return value
-		.replace(/[_-]+/g, " ")
-		.replace(/[\[\]{}()#.'\"=:/]+/g, " ")
-		.replace(/\s+/g, " ")
-		.trim();
-}
-
-export function matchesFinalActionLabel(value: string | undefined): boolean {
-	return value !== undefined && FINAL_ACTION_PATTERN.test(normalizeTargetText(value));
-}
-
-function parseRefId(value: string | undefined): string | undefined {
-	if (!value) return undefined;
-	const trimmed = value.trim();
-	const candidate = trimmed.startsWith("@") ? trimmed.slice(1) : trimmed.startsWith("ref=") ? trimmed.slice(4) : trimmed;
-	return /^e\d+$/.test(candidate) ? candidate : undefined;
-}
-
-function getRefTargetLabel(refSnapshot: SessionRefSnapshot | undefined, refId: string | undefined): string | undefined {
-	if (!refId) return undefined;
-	const ref = refSnapshot?.refs?.[refId];
-	return ref ? [ref.role, ref.name].filter(Boolean).join(" ") : undefined;
-}
-
-function getFlagValue(tokens: string[], flag: string): string | undefined {
-	for (const [index, token] of tokens.entries()) {
-		if (token === flag) return tokens[index + 1];
-		if (token.startsWith(`${flag}=`)) return token.slice(flag.length + 1);
-	}
-	return undefined;
-}
-
-function getClickLikeTargetLabel(command: string[], refSnapshot: SessionRefSnapshot | undefined): string | undefined {
-	const target = command[1];
-	return getRefTargetLabel(refSnapshot, parseRefId(target)) ?? target;
-}
-
-function getFindClickTargetLabel(command: string[]): string | undefined {
-	if (command[0] !== "find") return undefined;
-	const actionIndex = command.findIndex((token, index) => index >= 3 && FIND_CLICK_ACTIONS.has(token));
-	if (actionIndex === -1) return undefined;
-	return getFlagValue(command, "--name") ?? command[2];
-}
-
-function getKeyboardSubmitKey(command: string[]): string | undefined {
-	const commandName = command[0];
-	if (commandName === "press" || commandName === "key") return command[1];
-	return undefined;
-}
-
-function collectActionsFromCommand(command: string[], refSnapshot: SessionRefSnapshot | undefined, stepIndex?: number): BrowserFinalizingAction[] {
-	const actions: BrowserFinalizingAction[] = [];
-	if (CLICK_LIKE_COMMANDS.has(command[0] ?? "")) {
-		actions.push({
-			command,
-			kind: "click-like",
-			stepIndex,
-			targetLabel: getClickLikeTargetLabel(command, refSnapshot),
-		});
-		return actions;
-	}
-	if (command[0] === "find") {
-		const actionIndex = command.findIndex((token, index) => index >= 3 && FIND_CLICK_ACTIONS.has(token));
-		if (actionIndex !== -1) {
-			actions.push({
-				command,
-				kind: "click-like",
-				stepIndex,
-				targetLabel: getFindClickTargetLabel(command),
-			});
-		}
-		return actions;
-	}
-	const submitKey = getKeyboardSubmitKey(command)?.trim().toLowerCase();
-	if (submitKey && KEYBOARD_SUBMIT_KEYS.has(submitKey)) {
-		actions.push({
-			command,
-			kind: "keyboard-submit",
-			stepIndex,
-			targetLabel: submitKey,
-		});
-	}
-	return actions;
-}
-
-export function collectBrowserFinalizingActions(options: {
-	commandTokens: string[];
-	refSnapshot?: SessionRefSnapshot;
-	stdin?: string;
-}): BrowserFinalizingAction[] {
-	const actions = collectActionsFromCommand(options.commandTokens, options.refSnapshot);
-	if (options.commandTokens[0] !== "batch") return actions;
-	for (const { index, step } of parseValidBatchStepEntries(options.stdin)) {
-		actions.push(...collectActionsFromCommand(step, options.refSnapshot, index));
-	}
-	return actions;
-}
-
-export function shouldBlockFinalizingAction(action: BrowserFinalizingAction): boolean {
-	if (action.kind === "keyboard-submit") return true;
-	return matchesFinalActionLabel(action.targetLabel);
-}
-
-export function findBlockedFinalizingAction(options: {
-	commandTokens: string[];
-	refSnapshot?: SessionRefSnapshot;
-	stdin?: string;
-}): BrowserFinalizingAction | undefined {
-	for (const action of collectBrowserFinalizingActions(options)) {
-		if (!shouldBlockFinalizingAction(action)) continue;
-		return action;
-	}
-	return undefined;
-}