@posthog/wizard 2.46.0 → 2.48.0

package/dist/{agent-runner-CV3qh_2p.js → agent-runner-BbAtU_Ue.js}

@@ -1,17 +1,17 @@
 import { t as __exportAll } from "./rolldown-runtime-B_-DWIq7.js";
-import { A as GPT5_4_MODEL, Ct as runtimeEnv, D as DEFAULT_AGENT_MODEL, F as GPT5_MINI_MODEL, I as GPT5_MODEL, M as GPT5_6_LUNA_MODEL, N as GPT5_6_SOL_MODEL, P as GPT5_6_TERRA_MODEL, R as Harness, St as RUN_SURFACE, _ as evaluateWizardReadiness, _t as WIZARD_USE_PI_HARNESS_FLAG_KEY, a as getLogFilePath, c as WIZARD_BENCHMARK_FILE, dt as WIZARD_PI_MODEL_FLAG_KEY, f as getUI, g as SIGNUP_WIZARD_READINESS_CONFIG, h as SERVICE_LABELS, i as enableDebugLogs, it as Sequence, j as GPT5_5_MODEL, l as WIZARD_LOG_FILE, lt as WIZARD_ORCHESTRATOR_FLAG_KEY, mt as WIZARD_SELF_DRIVING_USE_PI_HARNESS_FLAG_KEY, o as initLogFile, rt as SONNET_5_MODEL, s as logToFile, t as configureLogFile, ut as WIZARD_PI_EFFORT_FLAG_KEY, v as getBlockingServiceKeys, vt as getSkillsBaseUrl } from "./debug-C9HmIrwx.js";
-import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-k76OWsl0.js";
-import { t as getOrAskForProjectData } from "./setup-utils-Cb5uACEO.js";
+import { A as GPT5_4_MODEL, Ct as runtimeEnv, D as DEFAULT_AGENT_MODEL, F as GPT5_MINI_MODEL, I as GPT5_MODEL, M as GPT5_6_LUNA_MODEL, N as GPT5_6_SOL_MODEL, P as GPT5_6_TERRA_MODEL, R as Harness, St as RUN_SURFACE, _ as evaluateWizardReadiness, _t as WIZARD_USE_PI_HARNESS_FLAG_KEY, a as getLogFilePath, c as WIZARD_BENCHMARK_FILE, dt as WIZARD_PI_MODEL_FLAG_KEY, f as getUI, g as SIGNUP_WIZARD_READINESS_CONFIG, h as SERVICE_LABELS, i as enableDebugLogs, it as Sequence, j as GPT5_5_MODEL, l as WIZARD_LOG_FILE, lt as WIZARD_ORCHESTRATOR_FLAG_KEY, mt as WIZARD_SELF_DRIVING_USE_PI_HARNESS_FLAG_KEY, o as initLogFile, rt as SONNET_5_MODEL, s as logToFile, t as configureLogFile, ut as WIZARD_PI_EFFORT_FLAG_KEY, v as getBlockingServiceKeys, vt as getSkillsBaseUrl } from "./debug-CUmAeIdp.js";
+import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-CVmeRvMK.js";
+import { t as getOrAskForProjectData } from "./setup-utils-CvCnesrl.js";
 import "./wizard-session-tp4Zdbwl.js";
-import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-FuyIaI_0.js";
-import { n as isNonInteractiveEnvironment } from "./environment-NJykOp33.js";
-import { O as fetchWithRetry, S as installSkillById, c as classifySettingsConflicts, i as runAgent$1, k as createWizardAskBridge, n as buildRunTags, o as backupAndFixClaudeSettings, p as AgentSignals, r as initializeAgent, s as checkAllSettingsConflicts, u as restoreClaudeSettings, x as fetchSkillMenu } from "./agent-interface-B8roB2UP.js";
-import { a as QUEUE_DIR_NAME, o as QueueStore, s as TaskStatus } from "./queue-tools-CTUgbOU8.js";
-import { a as flushScanReport, o as formatScanReport, p as writeScanReport, s as formatYaraAbortMessage } from "./yara-hooks-DdYRRxfH.js";
-import { r as detectNodePackageManagers } from "./package-manager-CAD9Uklp.js";
-import { a as piBackend, l as isThinkingLevel } from "./pi-C9iHv5uG.js";
-import { t as FRAMEWORK_REGISTRY } from "./registry-DT58jEYr.js";
-import fs, { existsSync, rmSync } from "fs";
+import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-BcG8dYgK.js";
+import { n as isNonInteractiveEnvironment } from "./environment-D-leJpbC.js";
+import { O as fetchWithRetry, S as installSkillById, c as classifySettingsConflicts, i as runAgent$1, k as createWizardAskBridge, n as buildRunTags, o as backupAndFixClaudeSettings, p as AgentSignals, r as initializeAgent, s as checkAllSettingsConflicts, u as restoreClaudeSettings, x as fetchSkillMenu } from "./agent-interface-C9li2RGX.js";
+import { a as QUEUE_DIR_NAME, o as QueueStore, s as TaskStatus } from "./queue-tools-CSTHA8ab.js";
+import { a as flushScanReport, o as formatScanReport, p as writeScanReport, s as formatYaraAbortMessage } from "./yara-hooks-B8GIO827.js";
+import { r as detectNodePackageManagers } from "./package-manager-BhIthBbt.js";
+import { a as piBackend, l as isThinkingLevel } from "./pi-CFQC2bp5.js";
+import { t as FRAMEWORK_REGISTRY } from "./registry-qI149NRV.js";
+import fs, { cpSync, existsSync, mkdirSync, readdirSync, rmSync, writeFileSync } from "fs";
 import * as path$1 from "path";
 import path from "path";
 import { z } from "zod";
@@ -1764,7 +1764,7 @@ function commandmentsReference(ctx) {
 	if (!ctx.commandmentsPath) return null;
 	return `Framework rules for this integration are at \`${ctx.commandmentsPath}\`. Read them before you edit and follow them.`;
 }
-const TASK_BASICS = `You are one isolated task in a larger PostHog workflow, run as a fresh agent with no memory of the other tasks beyond the context you are given. Do only your task, then report exactly once by calling complete_task with a structured handoff: what your goal was, what you did, and what the next agent should know. When you are given context from previous steps, trust it — those agents already did their work, so do not re-verify or re-read what their handoffs tell you. Build on it and move fast. Read a file before you edit it, so your own changes do not duplicate what is already there. Work only inside this project's own directory: never read, list, or search (find, ls, grep, glob) outside it — not the OS, not other projects, not global package caches. If your task seems to need something outside this directory, it does not — skip that part and say so in your handoff rather than hunting across the filesystem. If your task does not apply to this project — there is genuinely nothing for it to do — report it with status \`skipped\` and say why, rather than marking it done.`;
+const TASK_BASICS = `You are one isolated task in a larger PostHog workflow, run as a fresh agent with no memory of the other tasks beyond the context you are given. Do only your task, then report exactly once by calling complete_task with a structured handoff: what your goal was, what you did, and what the next agent should know. When you are given context from previous steps, trust it — those agents already did their work, so do not re-verify or re-read what their handoffs tell you. Build on it and move fast. Read a file before you edit it, so your own changes do not duplicate what is already there. Work only inside this project's own directory: never read, list, or search (find, ls, grep, glob) outside it — not the OS, not other projects, not global package caches. If your task seems to need something outside this directory, it does not — skip that part and say so in your handoff rather than hunting across the filesystem. If your task does not apply to this project — there is genuinely nothing for it to do — report it with status \`not needed\` and say why, rather than marking it done.`;
 const SEED_BASICS = `You are the orchestrator. Plan the work and seed the queue with enqueue_task — each call returns an id you can pass as a dependency to a later task. Give each task a short label for the UI — the action in a few words, not file names, class names, or other specifics. You are not a task yourself: do not call complete_task and do not edit the project.`;
 /**
 * Points the agent at its installed task instructions (the HOW). They live under
@@ -1949,6 +1949,8 @@ function renderHandoffContext(task, store) {
 		lines.push(`- did: ${handoff.did}`);
 		lines.push(`- for you: ${handoff.forNextAgent}`);
 		if (handoff.filesTouched?.length) lines.push(`- files: ${handoff.filesTouched.join(", ")}`);
+		if (handoff.evidence) lines.push(`- evidence: ${handoff.evidence}`);
+		if (handoff.assumptions) lines.push(`- assumed: ${handoff.assumptions}`);
 		lines.push("");
 	}
 	if (lines.length === 0) return "";
@@ -1998,6 +2000,36 @@ const EMPTY_PROMPT = {};
 * task resolve to a prompt fetched at startup into the registry. The wizard side
 * stays product-ignorant: it is the queue, the executor, and the loader.
 */
+/** Docs page (`django.md`, `nuxt-js-3-6.md`) — steps start with a digit, agent artifacts (`SKILL.md`, `EXAMPLE*`, `COMMANDMENTS.md`) have uppercase. */
+const isDocPage = (name) => name.endsWith(".md") && name === name.toLowerCase() && !/^\d/.test(name);
+/** Copy only the framework docs pages out of the run cache into .claude/skills — the one durable artifact an orchestrator run leaves. Never clobbers an existing install. */
+function promoteReferenceSkill(referenceDir, claudeSkillsDir, referenceSkillId) {
+	const target = path$1.join(claudeSkillsDir, referenceSkillId);
+	const refs = path$1.join(referenceDir, "references");
+	if (!existsSync(refs) || existsSync(target)) return;
+	const docs = readdirSync(refs).filter(isDocPage);
+	if (docs.length === 0) return;
+	mkdirSync(path$1.join(target, "references"), { recursive: true });
+	for (const f of docs) cpSync(path$1.join(refs, f), path$1.join(target, "references", f));
+	writeFileSync(path$1.join(target, ".posthog-wizard"), "");
+}
+/**
+* Remove skills that task agents installed durably mid-run (load_skill):
+* wizard-marked, new this run, and not the framework reference docs.
+* User-authored and pre-existing skills stay.
+*/
+function sweepRunInstalledSkills(claudeSkillsDir, preexistingSkills, referenceSkillId) {
+	if (!existsSync(claudeSkillsDir)) return;
+	for (const id of readdirSync(claudeSkillsDir)) {
+		if (preexistingSkills.has(id) || id === referenceSkillId) continue;
+		if (!existsSync(path$1.join(claudeSkillsDir, id, ".posthog-wizard"))) continue;
+		rmSync(path$1.join(claudeSkillsDir, id), {
+			recursive: true,
+			force: true
+		});
+		logToFile(`[orchestrator] removed run-installed skill ${id}`);
+	}
+}
 function toTodoStatus(status) {
 	switch (status) {
 		case TaskStatus.Running: return "in_progress";
@@ -2103,11 +2135,13 @@ async function runOrchestrator(session, programConfig, boot) {
 	} });
 	let examplePath;
 	let commandmentsPath;
+	let referenceInstallPath;
 	const menuSkillEntries = await fetchSkillMenuEntries(boot.skillsBaseUrl);
 	const referenceSkillId = session.skillId ? resolveReferenceSkillId(menuSkillEntries, session.skillId) : void 0;
 	if (referenceSkillId) {
 		const ref = await installSkillById(referenceSkillId, session.installDir, boot.skillsBaseUrl, path$1.join(QUEUE_DIR_NAME, "reference"));
 		if (ref.kind === "ok") {
+			referenceInstallPath = ref.path;
 			const example = path$1.join(ref.path, "references", "EXAMPLE.md");
 			if (existsSync(path$1.join(session.installDir, example))) examplePath = example;
 			const commandments = path$1.join(ref.path, "references", "COMMANDMENTS.md");
@@ -2189,6 +2223,8 @@ or integrate manually here:
 	});
 	renderQueue();
 	const taskSkillsRoot = path$1.join(QUEUE_DIR_NAME, "skills");
+	const claudeSkillsDir = path$1.join(session.installDir, ".claude", "skills");
+	const preexistingSkills = new Set(existsSync(claudeSkillsDir) ? readdirSync(claudeSkillsDir) : []);
 	let remarkRequested = false;
 	const runTask = async (task) => {
 		renderQueue();
@@ -2232,12 +2268,22 @@ or integrate manually here:
 				}
 			});
 		} finally {
+			try {
+				sweepRunInstalledSkills(claudeSkillsDir, preexistingSkills, referenceSkillId);
+			} catch (err) {
+				logToFile(`[orchestrator] per-task skill sweep failed: ${String(err)}`);
+			}
 			renderQueue();
 		}
 	};
 	try {
 		await drainQueue(store, runTask);
 	} finally {
+		try {
+			if (referenceSkillId && referenceInstallPath) promoteReferenceSkill(path$1.join(session.installDir, referenceInstallPath), claudeSkillsDir, referenceSkillId);
+		} catch (err) {
+			analytics.captureException(err instanceof Error ? err : new Error(String(err)), { step: "orchestrator_reference_promote" });
+		}
 		try {
 			rmSync(path$1.join(session.installDir, QUEUE_DIR_NAME), {
 				recursive: true,
@@ -2246,6 +2292,11 @@ or integrate manually here:
 		} catch (err) {
 			analytics.captureException(err instanceof Error ? err : new Error(String(err)), { step: "orchestrator_cache_cleanup" });
 		}
+		try {
+			sweepRunInstalledSkills(claudeSkillsDir, preexistingSkills, referenceSkillId);
+		} catch (err) {
+			analytics.captureException(err instanceof Error ? err : new Error(String(err)), { step: "orchestrator_skill_sweep" });
+		}
 	}
 	renderQueue();
 	const summary = store.summary();
@@ -2254,7 +2305,7 @@ or integrate manually here:
 		tasks_total: summary.total,
 		tasks_done: summary.done,
 		tasks_failed: summary.failed,
-		tasks_skipped: summary.skipped,
+		tasks_skipped: summary[TaskStatus.Skipped],
 		total_duration_ms: Date.now() - runStartMs,
 		...metrics.summary(),
 		dynamic_enqueue_count: store.list().filter((t) => t.enqueuedBy !== "orchestrator").length,
@@ -2263,7 +2314,8 @@ or integrate manually here:
 	const buildTask = store.list().find((t) => t.type === "build");
 	const conflict = buildTask ? store.readHandoff(buildTask.id)?.conflict : void 0;
 	const reportFile = existsSync(path$1.join(session.installDir, "posthog-setup-report.md")) ? "posthog-setup-report.md" : store.queuePath;
-	const message = conflict ? "PostHog set up, with one conflict to review." : `PostHog set up: ${summary.done}/${summary.total} steps completed.`;
+	const notRequired = summary[TaskStatus.Skipped];
+	const message = conflict ? "PostHog set up, with one conflict to review." : `PostHog set up: ${summary.done}/${summary.total - notRequired} steps completed${notRequired > 0 ? ` (${notRequired} skipped as not required)` : ""}.`;
 	getUI().setOutroData({
 		kind: "success",
 		message,
@@ -2408,6 +2460,11 @@ const PROGRAM_BINDINGS = {
 	"mcp-remove": DEFAULT_BINDING,
 	"mcp-tutorial": DEFAULT_BINDING,
 	"mcp-analytics": DEFAULT_BINDING,
+	"ai-observability": {
+		sequence: "linear",
+		harness: "anthropic",
+		model: SONNET_5_MODEL
+	},
 	slack: DEFAULT_BINDING
 };
 /** Compose both axes. Callers needing only one axis use the per-axis resolver. */
@@ -2519,4 +2576,4 @@ var agent_runner_exports = /* @__PURE__ */ __exportAll({ runAgent: () => runAgen
 //#endregion
 export { formatTokenCount as a, formatCostUsd as i, runAgent as n, authenticate as o, computeTokenCostUsd as r, agent_runner_exports as t };
 
-//# sourceMappingURL=agent-runner-CV3qh_2p.js.map
+//# sourceMappingURL=agent-runner-BbAtU_Ue.js.map