@posthog/wizard 2.50.0 → 2.52.0

package/dist/{agent-runner-NzzrOB02.js → agent-runner-B1qCtjid.js}

@@ -1,16 +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-dX5JKcxf.js";
-import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-CArgq7XO.js";
-import { t as getOrAskForProjectData } from "./setup-utils-D5RLhAEe.js";
-import "./wizard-session-tp4Zdbwl.js";
-import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-D_Xcbesm.js";
-import { n as isNonInteractiveEnvironment } from "./environment-D-xSv8Xv.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-C_aRBziP.js";
-import { c as TaskStatus, o as QUEUE_DIR_NAME, s as QueueStore } from "./queue-tools-DW-pUVKK.js";
-import { a as flushScanReport, o as formatScanReport, p as writeScanReport, s as formatYaraAbortMessage } from "./yara-hooks-CGzlyuQA.js";
-import { r as detectNodePackageManagers } from "./package-manager-D3M_H5Pk.js";
-import { a as piBackend, l as isThinkingLevel } from "./pi--w-M8eD-.js";
-import { t as FRAMEWORK_REGISTRY } from "./registry-CYNjgbAh.js";
+import { $ as SONNET_5_MODEL, A as GPT5_6_LUNA_MODEL, D as DEFAULT_AGENT_MODEL, M as GPT5_6_TERRA_MODEL, P as Harness, _ as evaluateWizardReadiness, _t as runtimeEnv, a as getLogFilePath, c as WIZARD_BENCHMARK_FILE, et as Sequence, f as getUI, ft as getSkillsBaseUrl, g as SIGNUP_WIZARD_READINESS_CONFIG, gt as RUN_SURFACE, h as SERVICE_LABELS, i as enableDebugLogs, j as GPT5_6_SOL_MODEL, l as WIZARD_LOG_FILE, nt as WIZARD_CONTACT_EMAIL, o as initLogFile, ot as WIZARD_ORCHESTRATOR_FLAG_KEY, s as logToFile, st as WIZARD_ORCHESTRATOR_OVERRIDE_FLAG_KEY, t as configureLogFile, ut as WIZARD_SELF_DRIVING_USE_PI_HARNESS_FLAG_KEY, v as getBlockingServiceKeys } from "./debug-C3ZR1TVF.js";
+import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-BXz4SIuS.js";
+import { t as getOrAskForProjectData } from "./setup-utils-DDzEnDxJ.js";
+import "./wizard-session-BAZql11N.js";
+import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-AEhoDJ4P.js";
+import { n as isNonInteractiveEnvironment } from "./environment-Bi-zynk6.js";
+import { O as createWizardAskBridge, b as fetchSkillMenu, c as classifySettingsConflicts, i as runAgent$1, n as buildRunTags, o as backupAndFixClaudeSettings, p as AgentSignals, r as initializeAgent, s as checkAllSettingsConflicts, u as restoreClaudeSettings, x as installSkillById } from "./agent-interface-DoFVHRFh.js";
+import { a as flushScanReport, m as writeScanReport, o as formatScanReport, s as formatYaraAbortMessage } from "./yara-hooks-CKsVxWiN.js";
+import { c as TaskStatus, f as requireKnownModel, o as QUEUE_DIR_NAME, s as QueueStore, u as isValidModel } from "./queue-tools-D7ctuzyo.js";
+import { r as detectNodePackageManagers } from "./package-manager-wkSzPJqW.js";
+import { a as piBackend } from "./pi-CGPjm4oE.js";
+import { t as FRAMEWORK_REGISTRY } from "./registry-CsFuKtXn.js";
+import { a as promptModelFor, c as resolveTask, i as loadAgentRegistry, l as taskModelSpec, n as assembleSeedPrompt, r as assembleTaskPrompt, t as agentRunTools } from "./agent-prompt-loader-D-cPVoL5.js";
 import fs, { cpSync, existsSync, mkdirSync, readdirSync, rmSync, writeFileSync } from "fs";
 import * as path$1 from "path";
 import path from "path";
@@ -249,42 +250,29 @@ const anthropicBackend = {
 	}
 };
 //#endregion
-//#region src/lib/agent/runner/switchboard/flags/basic-integration.ts
-/**
-* Basic-integration pi experiment.
-*
-* Flags:  wizard-use-pi-harness (bool) + wizard-pi-model + wizard-pi-effort
-* Routes: ONLY `posthog-integration` — harness, model, and effort axes.
-*/
-const BASIC_INTEGRATION_EXPERIMENT = {
-	program: "posthog-integration",
-	flags: {
-		useFlag: WIZARD_USE_PI_HARNESS_FLAG_KEY,
-		modelFlag: WIZARD_PI_MODEL_FLAG_KEY,
-		effortFlag: WIZARD_PI_EFFORT_FLAG_KEY,
-		fallbackModel: GPT5_4_MODEL
-	}
-};
-//#endregion
 //#region src/lib/agent/runner/switchboard/flags/orchestrator.ts
-/**
-* Orchestrator experiment.
-*
-* Flag:   wizard-orchestrator (bool)
-* Routes: ONLY the programs listed below — sequence axis. context-mill
-*         publishes orchestrator agent prompts (seed + per-task) per flow, so
-*         a program joins the experiment by shipping prompts AND being added
-*         here. Other programs can still pin `sequence` via their own flag
-*         route's payload.
-*/
-const ORCHESTRATOR_EXPERIMENT = {
+/** Orchestrator gate: wizard-orchestrator on/off. On routes the declared programs to the orchestrator on pi; every model/effort comes from prompt frontmatter. */
+const ORCHESTRATOR_SEQUENCE_ROUTE = {
 	programs: ["posthog-integration"],
 	flag: WIZARD_ORCHESTRATOR_FLAG_KEY,
 	sequence: "orchestrator"
 };
+/** Same flag, harness axis: the orchestrator drives tasks through pi. */
+const ORCHESTRATOR_HARNESS_ROUTE = {
+	program: "posthog-integration",
+	flags: {
+		useFlag: WIZARD_ORCHESTRATOR_FLAG_KEY,
+		harness: "pi"
+	}
+};
+/** Same programs, per-stage overrides: the variant payload (`{stage: {model?, effort?}}`) outranks prompt frontmatter for the stages it names. */
+const ORCHESTRATOR_STAGE_OVERRIDES = {
+	programs: ORCHESTRATOR_SEQUENCE_ROUTE.programs,
+	flag: WIZARD_ORCHESTRATOR_OVERRIDE_FLAG_KEY
+};
 /** Raw flag read — for telemetry/log lines only, never for routing. */
 function isOrchestratorEnabled(flags = {}) {
-	return flags[ORCHESTRATOR_EXPERIMENT.flag] === "true";
+	return flags[ORCHESTRATOR_SEQUENCE_ROUTE.flag] === "true";
 }
 //#endregion
 //#region src/lib/agent/runner/switchboard/flags/self-driving.ts
@@ -308,12 +296,8 @@ const SELF_DRIVING_EXPERIMENT = {
 * resolution. Experiment modules in this folder declare *which* flags they
 * use; this module owns *how* a config resolves to a route.
 */
-/** Model variant key → gateway id (variant keys can't carry `/` or `.`). */
+/** Model variant key → gateway id. */
 const MODEL_FLAG_VARIANTS = {
-	"gpt-5": GPT5_MODEL,
-	"gpt-5-4": GPT5_4_MODEL,
-	"gpt-5-mini": GPT5_MINI_MODEL,
-	"gpt-5-5": GPT5_5_MODEL,
 	"gpt-5-6-luna": GPT5_6_LUNA_MODEL,
 	"gpt-5-6-terra": GPT5_6_TERRA_MODEL,
 	"gpt-5-6-sol": GPT5_6_SOL_MODEL,
@@ -328,25 +312,46 @@ const EFFORT_FLAG_VARIANTS = [
 	"high",
 	"xhigh"
 ];
+/** Variant key validated against the shared vocabulary, transformed to its gateway id. */
+const modelVariantSchema = z.string().refine((key) => key in MODEL_FLAG_VARIANTS).transform((key) => MODEL_FLAG_VARIANTS[key]);
 /** `{model, effort?, harness?, sequence?}` payload shape; extra keys tolerated for forward compat. */
 const payloadConfigFlagSchema = z.object({
-	model: z.string().refine((key) => key in MODEL_FLAG_VARIANTS),
+	model: modelVariantSchema,
 	effort: z.enum(EFFORT_FLAG_VARIANTS).optional(),
 	harness: z.nativeEnum(Harness).optional(),
 	sequence: z.nativeEnum(Sequence).optional()
 });
-/** Validate a payload (object or JSON string) into a route; undefined on any unexpected shape. */
-function parseFlagPayload(raw) {
-	let value = raw;
-	if (typeof raw === "string") try {
-		value = JSON.parse(raw);
+/**
+* PostHog serves a variant's payload as a JSON string, e.g. the
+* wizard-orchestrator-override `terra-review` variant arrives as
+* `'{"review":{"model":"gpt-5-6-terra","effort":"medium"}}'`; local CI
+* overrides inject the same payloads as already-parsed objects. Coerce both
+* to the value; anything unparseable is undefined.
+*/
+function coerceJson(raw) {
+	if (typeof raw !== "string") return raw;
+	try {
+		return JSON.parse(raw);
 	} catch {
 		return;
 	}
-	const parsed = payloadConfigFlagSchema.safeParse(value);
+}
+/** Per-stage overrides keyed by stage; a whole-payload parse failure fails closed to frontmatter everywhere. */
+const stageOverridesSchema = z.record(z.object({
+	model: modelVariantSchema.optional(),
+	effort: z.enum(EFFORT_FLAG_VARIANTS).optional()
+}));
+/** Validate a stage-override payload (object or JSON string); undefined on any unexpected shape. */
+function stageOverridesFromPayload(raw) {
+	const parsed = stageOverridesSchema.safeParse(coerceJson(raw));
+	return parsed.success ? parsed.data : void 0;
+}
+/** Validate a payload (object or JSON string) into a route; undefined on any unexpected shape. */
+function parseFlagPayload(raw) {
+	const parsed = payloadConfigFlagSchema.safeParse(coerceJson(raw));
 	if (!parsed.success) return void 0;
 	return {
-		model: MODEL_FLAG_VARIANTS[parsed.data.model],
+		model: parsed.data.model,
 		thinkingLevel: parsed.data.effort,
 		harness: parsed.data.harness,
 		sequence: parsed.data.sequence
@@ -359,13 +364,7 @@ function parseFlagPayload(raw) {
 */
 function routeFromConfigFlag(cfg, flags, flagPayloads) {
 	if (flags[cfg.useFlag] !== "true") return void 0;
-	if (cfg.modelFlag) {
-		const effort = flags[cfg.effortFlag];
-		return {
-			model: MODEL_FLAG_VARIANTS[flags[cfg.modelFlag] ?? ""] ?? cfg.fallbackModel,
-			thinkingLevel: EFFORT_FLAG_VARIANTS.includes(effort) ? effort : void 0
-		};
-	}
+	if (cfg.harness) return { harness: cfg.harness };
 	const route = parseFlagPayload(flagPayloads?.[cfg.useFlag]);
 	if (!route) logToFile(`[switchboard] ${cfg.useFlag} on but payload missing/invalid — keeping the non-flagged default`);
 	return route;
@@ -379,8 +378,8 @@ function routeFromConfigFlag(cfg, flags, flagPayloads) {
 * shapes and resolution.
 */
 /** Every experiment on each axis. An experiment routes ONLY by being listed here, and only for the programs it declares. */
-const HARNESS_EXPERIMENTS = [BASIC_INTEGRATION_EXPERIMENT, SELF_DRIVING_EXPERIMENT];
-const SEQUENCE_EXPERIMENTS = [ORCHESTRATOR_EXPERIMENT];
+const HARNESS_EXPERIMENTS = [ORCHESTRATOR_HARNESS_ROUTE, SELF_DRIVING_EXPERIMENT];
+const SEQUENCE_EXPERIMENTS = [ORCHESTRATOR_SEQUENCE_ROUTE];
 /** The flag-driven route for a program, or undefined when no experiment covers it or its flags don't validly route. */
 function resolveFlagRoute(program, flags, flagPayloads) {
 	if (RUN_SURFACE === "cloud") return void 0;
@@ -391,6 +390,17 @@ function resolveFlagRoute(program, flags, flagPayloads) {
 function resolveFlagSequence(program, flags) {
 	return SEQUENCE_EXPERIMENTS.find((e) => e.programs.includes(program) && flags[e.flag] === "true")?.sequence;
 }
+/** The per-stage overrides for a program's run, or undefined (prompt frontmatter stays). Applied once, where the agent prompts are loaded. */
+function resolveStageOverrides(program, flags, flagPayloads) {
+	if (RUN_SURFACE === "cloud") return void 0;
+	const exp = ORCHESTRATOR_STAGE_OVERRIDES;
+	if (!exp.programs.includes(program)) return void 0;
+	const variant = flags[exp.flag];
+	if (!variant || variant === "false") return void 0;
+	const overrides = stageOverridesFromPayload(flagPayloads?.[exp.flag]);
+	if (!overrides) logToFile(`[switchboard] ${exp.flag}=${variant} but payload missing/invalid — keeping frontmatter`);
+	return overrides;
+}
 //#endregion
 //#region src/lib/agent/runner/switchboard/harness.ts
 /**
@@ -419,8 +429,8 @@ const flagRunnerOverride = (ctx, next) => {
 	});
 	return {
 		harness: route.harness ?? "pi",
-		model: route.model,
-		thinkingLevel: route.thinkingLevel
+		model: route.model ?? pick.model,
+		thinkingLevel: route.thinkingLevel ?? pick.thinkingLevel
 	};
 };
 const HARNESS_MIDDLEWARE = [...[], flagRunnerOverride];
@@ -1745,248 +1755,6 @@ var RunMetrics = class {
 	}
 };
 //#endregion
-//#region src/lib/agent/agent-prompt-loader.ts
-function projectContext(ctx) {
-	return `You have access to the PostHog MCP server and the wizard tools.
-
-Project context:
-- PostHog Project ID: ${ctx.projectId}
-- PostHog public token: ${ctx.projectApiKey}
-- PostHog Host: ${ctx.host.apiHost}`;
-}
-/** Points the agent at the framework's reference integration to learn patterns from. */
-function exampleReference(ctx) {
-	if (!ctx.examplePath) return null;
-	return `A reference PostHog integration for this framework is at \`${ctx.examplePath}\`. It shows the target implementation pattern. Reference its patterns and conventions, adapting them to this codebase.`;
-}
-/** The framework's rules ship with the reference skill; every task follows them. */
-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 \`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
-* the wizard's run dir, not `.claude/skills/`, so the SDK does not auto-load
-* them — the prompt has to name them.
-*/
-function skillReference(paths) {
-	if (paths.length === 0) return null;
-	return `Your task instructions are at ${paths.map((p) => `\`${p}\``).join(", ")}. Read them before you start and follow them. They are wizard scaffolding, not part of the project.`;
-}
-/** A task agent's full prompt: injected basics, then the authored intent. */
-function assembleTaskPrompt(ctx, body, skillPaths = []) {
-	return [
-		projectContext(ctx),
-		exampleReference(ctx),
-		commandmentsReference(ctx),
-		skillReference(skillPaths),
-		TASK_BASICS,
-		body
-	].filter(Boolean).join("\n\n");
-}
-/** The seed agent's full prompt: injected basics, then the authored intent. */
-function assembleSeedPrompt(ctx, body) {
-	return [
-		projectContext(ctx),
-		SEED_BASICS,
-		body
-	].join("\n\n");
-}
-/** Orchestrator tools are MCP tools under the `posthog-wizard` server. Frontmatter
-*  names them short (e.g. `enqueue_task`); the SDK gates on the full name. */
-const ORCHESTRATOR_TOOL_PREFIX = "mcp__posthog-wizard__";
-const ORCHESTRATOR_TOOLS = new Set([
-	"enqueue_task",
-	"complete_task",
-	"read_handoffs"
-]);
-/** The model + effort an agent runs on for a given harness — `pi` picks the gpt
-* column, anything else the sdk (anthropic) column. */
-function promptModelFor(prompt, harness) {
-	const pi = harness === "pi";
-	return {
-		model: pi ? prompt.modelPi : prompt.modelSdk,
-		effort: pi ? prompt.effortPi : prompt.effortSdk
-	};
-}
-/** The registry for one flow's prompts. Pure; the loader feeds it the fetched set. */
-function buildRegistry(prompts, flow, opts) {
-	const excluded = new Set(opts?.exclude ?? []);
-	const inFlow = prompts.filter((p) => p.flow === flow && !excluded.has(p.type));
-	const byType = new Map(inFlow.map((p) => [p.type, p]));
-	return {
-		types: inFlow.filter((p) => !p.seed).map((p) => p.type),
-		seed: inFlow.find((p) => p.seed),
-		get: (type) => byType.get(type)
-	};
-}
-/** A native tool passes through; an orchestrator tool gets its MCP-qualified name. */
-function expandToolName(name) {
-	return ORCHESTRATOR_TOOLS.has(name) ? `${ORCHESTRATOR_TOOL_PREFIX}${name}` : name;
-}
-/** A prompt's allow/disallow lists with orchestrator tool names MCP-qualified. */
-function agentRunTools(prompt) {
-	return {
-		allowedTools: prompt.allowedTools.map(expandToolName),
-		disallowedTools: prompt.disallowedTools.map(expandToolName)
-	};
-}
-function toStringArray(value) {
-	if (!Array.isArray(value)) return [];
-	return value.filter((v) => typeof v === "string");
-}
-/**
-* Parse the leading `---` frontmatter block and the markdown body. The
-* frontmatter is a small, known schema (scalars and inline `[a, b]` arrays), so
-* a tiny parser covers it without a YAML dependency. Inline `# comments` after a
-* value are stripped. `fallbackType` (the menu id) and `fallbackFlow` (the
-* menu entry's flow) apply when the frontmatter omits `type:`/`flow:`.
-*/
-function parseAgentPrompt(text, fallbackType, fallbackFlow) {
-	const match = text.match(/^---\r?\n([\s\S]*?)\r?\n---\r?\n?([\s\S]*)$/);
-	const frontmatter = match ? match[1] : "";
-	const body = (match ? match[2] : text).trim();
-	const fields = {};
-	for (const rawLine of frontmatter.split(/\r?\n/)) {
-		const line = rawLine.replace(/\s+#.*$/, "").trim();
-		if (!line || line.startsWith("#")) continue;
-		const kv = line.match(/^([\w-]+):\s*(.*)$/);
-		if (!kv) continue;
-		const [, key, raw] = kv;
-		if (raw.startsWith("[") && raw.endsWith("]")) fields[key] = raw.slice(1, -1).split(",").map((s) => s.trim().replace(/^['"]|['"]$/g, "")).filter(Boolean);
-		else fields[key] = raw.replace(/^['"]|['"]$/g, "");
-	}
-	const str = (v) => typeof v === "string" ? v : void 0;
-	const effort = (v, key) => {
-		if (v === void 0) return void 0;
-		if (isThinkingLevel(v)) return v;
-		logToFile(`[agent-prompt] ${fallbackType}: ignoring invalid ${key} "${String(v)}"`);
-		analytics.wizardCapture("agent prompt invalid effort", {
-			task_type: fallbackType,
-			key,
-			value: String(v)
-		});
-	};
-	return {
-		type: typeof fields.type === "string" ? fields.type : fallbackType,
-		label: typeof fields.label === "string" ? fields.label : void 0,
-		flow: typeof fields.flow === "string" ? fields.flow : fallbackFlow,
-		seed: fields.seed === "true",
-		modelPi: str(fields.model_pi),
-		effortPi: effort(fields.effort_pi, "effort_pi"),
-		modelSdk: str(fields.model_sdk),
-		effortSdk: effort(fields.effort_sdk, "effort_sdk"),
-		skills: toStringArray(fields.skills),
-		allowedTools: toStringArray(fields.allowedTools),
-		disallowedTools: toStringArray(fields.disallowedTools),
-		dependsOn: toStringArray(fields.dependsOn),
-		body
-	};
-}
-async function fetchText(url) {
-	return (await fetchWithRetry(url)).text();
-}
-/**
-* Fetch the agent menu and every agent prompt it lists, parse them, and build
-* the registry for one flow. Throws if the menu cannot be fetched — the
-* orchestrator cannot run without its prompts.
-*/
-async function loadAgentRegistry(skillsBaseUrl, flow, opts) {
-	const menuRaw = await fetchText(`${skillsBaseUrl}/agent-menu.json`);
-	const entries = (JSON.parse(menuRaw).agents ?? []).filter((entry) => !entry.flow || entry.flow === flow);
-	return buildRegistry(await Promise.all(entries.map(async (entry) => {
-		return parseAgentPrompt(await fetchText(entry.downloadUrl), entry.id, entry.flow);
-	})), flow, opts);
-}
-/**
-* Render a task's own inputs into a section, so a fanned-out task (e.g. one
-* `capture` per event) sees the specific thing it owns. Empty when there are none.
-*/
-function renderInputs(task) {
-	const entries = Object.entries(task.inputs ?? {});
-	if (entries.length === 0) return "";
-	return `## Your task input\n\n${entries.map(([k, v]) => `- ${k}: ${formatInputValue(v)}`).join("\n")}`;
-}
-function formatInputValue(value) {
-	if (typeof value === "string") return value;
-	return JSON.stringify(value);
-}
-/**
-* The ids of every task `task` transitively depends on — the full upstream
-* chain, not just direct dependencies — ordered roots-first, each once. A `seen`
-* set dedupes diamonds and guards against cycles.
-*/
-function ancestorIds(task, store) {
-	const seen = /* @__PURE__ */ new Set();
-	const ordered = [];
-	const visit = (id) => {
-		if (seen.has(id)) return;
-		seen.add(id);
-		const t = store.get(id);
-		if (!t) return;
-		for (const dep of t.dependsOn) visit(dep);
-		ordered.push(id);
-	};
-	for (const dep of task.dependsOn) visit(dep);
-	return ordered;
-}
-/**
-* Render the handoffs of every step `task` transitively depends on into a context
-* section, so a fresh agent sees the whole upstream chain — not just its direct
-* dependencies. Reliability over token economy: a step must never have to
-* re-discover what any ancestor already established just because an intermediate
-* handoff happened to omit it. Empty when there are no completed ancestors.
-*/
-function renderHandoffContext(task, store) {
-	const lines = [];
-	for (const id of ancestorIds(task, store)) {
-		const dep = store.get(id);
-		const handoff = store.readHandoff(id);
-		if (!dep || !handoff) continue;
-		lines.push(`### ${dep.type}`);
-		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: ${handoff.assumptions}`);
-		lines.push("");
-	}
-	if (lines.length === 0) return "";
-	return `## Context from previous steps\n\n${lines.join("\n")}`.trim();
-}
-/**
-* Resolve a queued task to its run config: the prompt body (with upstream
-* handoffs appended), the model, and the tool lists with orchestrator tool names
-* MCP-qualified. The model precedence is enqueue override, then prompt, then
-* default. Throws if no prompt is registered for the task's type.
-*/
-function resolveTask(registry, task, store) {
-	const prompt = registry.get(task.type);
-	if (!prompt) throw new Error(`No agent prompt registered for task type "${task.type}"`);
-	const body = [
-		renderInputs(task),
-		prompt.body,
-		renderHandoffContext(task, store)
-	].filter(Boolean).join("\n\n");
-	return {
-		...agentRunTools(prompt),
-		prompt: body,
-		skills: prompt.skills
-	};
-}
-/** The model + effort a task runs on for a harness: enqueue override, then the
-* prompt's per-profile frontmatter; the caller's switchboard pick is the fallback. */
-function taskModelSpec(registry, task, harness) {
-	const picked = promptModelFor(registry.get(task.type) ?? EMPTY_PROMPT, harness);
-	return {
-		model: task.model ?? picked.model,
-		effort: picked.effort
-	};
-}
-const EMPTY_PROMPT = {};
-//#endregion
 //#region src/lib/agent/runner/sequence/orchestrator/orchestrator-runner.ts
 /**
 * Experimental task-queue orchestrator runner.
@@ -2081,7 +1849,10 @@ async function runOrchestrator(session, programConfig, boot) {
 		cliModel: session.model
 	};
 	const flow = programConfig.agentFlow ?? programConfig.id;
-	const registry = await loadAgentRegistry(boot.skillsBaseUrl, flow, { exclude: ciExcludedTaskTypes() });
+	const registry = await loadAgentRegistry(boot.skillsBaseUrl, flow, {
+		exclude: ciExcludedTaskTypes(),
+		overrides: resolveStageOverrides(programConfig.id, boot.wizardFlags, boot.wizardFlagPayloads)
+	});
 	const seedPrompt = registry.seed;
 	if (!seedPrompt) throw new Error(`No seed agent prompt (frontmatter \`seed: true\`) for flow "${flow}" is available from ${boot.skillsBaseUrl}.`);
 	const runStartMs = Date.now();
@@ -2089,9 +1860,10 @@ async function runOrchestrator(session, programConfig, boot) {
 	const durationMs = (t) => t.startedAt && t.finishedAt ? Date.parse(t.finishedAt) - Date.parse(t.startedAt) : void 0;
 	const store = new QueueStore(session.installDir, runId, { onTransition: (event, task) => {
 		const pick = resolveHarness(switchboardCtx, task.type);
+		const specModel = taskModelSpec(registry, task, pick.harness).model;
 		const base = {
 			type: task.type,
-			model: taskModelSpec(registry, task, pick.harness).model ?? pick.model,
+			model: isValidModel(specModel) ? specModel : pick.model,
 			attempts: task.attempts
 		};
 		switch (event) {
@@ -2203,7 +1975,7 @@ or integrate manually here:
 		boot,
 		prompt: assembleSeedPrompt(promptContext, seedPrompt.body),
 		spinner,
-		model: seedModel.model ?? seedPick.model,
+		model: requireKnownModel(seedModel.model, seedPick.model),
 		effort: seedModel.effort,
 		...agentRunTools(seedPrompt),
 		orchestrator: orchestratorCtx(),
@@ -2248,7 +2020,7 @@ or integrate manually here:
 				boot,
 				prompt: assembleTaskPrompt(promptContext, resolved.prompt, skillPaths),
 				spinner,
-				model: taskModel.model ?? taskPick.model,
+				model: requireKnownModel(taskModel.model, taskPick.model),
 				effort: taskModel.effort,
 				allowedTools: resolved.allowedTools,
 				disallowedTools: resolved.disallowedTools,
@@ -2306,10 +2078,23 @@ or integrate manually here:
 		dynamic_enqueue_count: store.list().filter((t) => t.enqueuedBy !== "orchestrator").length,
 		retried_task_count: store.list().filter((t) => t.attempts > 1).length
 	});
-	const buildTask = store.list().find((t) => t.type === "build");
-	const conflict = buildTask ? store.readHandoff(buildTask.id)?.conflict : void 0;
+	const reviewTask = store.list().find((t) => t.type === "review");
+	const conflict = reviewTask ? store.readHandoff(reviewTask.id)?.conflict : void 0;
 	const reportFile = existsSync(path$1.join(session.installDir, "posthog-setup-report.md")) ? "posthog-setup-report.md" : store.queuePath;
 	const notRequired = summary[TaskStatus.Skipped];
+	const blocked = summary[TaskStatus.Pending];
+	if (summary.failed > 0 || blocked > 0) {
+		const failedTypes = store.list().filter((t) => t.status === TaskStatus.Failed).map((t) => t.type).join(", ");
+		await wizardAbort({
+			message: `The wizard was unable to set up PostHog: ${failedTypes ? `the ${failedTypes} step failed` : `${blocked} steps never ran`}.\n\nPlease report this to: ${WIZARD_CONTACT_EMAIL}`,
+			error: new WizardError("orchestrator drain ended with failed tasks", {
+				tasks_failed: summary.failed,
+				tasks_blocked: blocked,
+				failed_types: failedTypes,
+				queue_state: JSON.stringify(store.list())
+			})
+		});
+	}
 	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",
@@ -2534,9 +2319,6 @@ function captureSwitchboardDecision(ctx, binding) {
 	const trace = ctx.trace ?? {};
 	analytics.wizardCapture("switchboard resolved", {
 		program: ctx.program,
-		flag_use_pi_harness: ctx.flags[WIZARD_USE_PI_HARNESS_FLAG_KEY],
-		flag_pi_model: ctx.flags[WIZARD_PI_MODEL_FLAG_KEY],
-		flag_pi_effort: ctx.flags[WIZARD_PI_EFFORT_FLAG_KEY],
 		flag_self_driving_use_pi_harness: ctx.flags[WIZARD_SELF_DRIVING_USE_PI_HARNESS_FLAG_KEY],
 		flag_self_driving_pi_payload: JSON.stringify(ctx.flagPayloads?.["wizard-self-driving-use-pi-harness"] ?? null),
 		flag_orchestrator: ctx.flags[WIZARD_ORCHESTRATOR_FLAG_KEY],
@@ -2551,7 +2333,7 @@ function captureSwitchboardDecision(ctx, binding) {
 		thinking_level: binding.thinkingLevel,
 		sequence: binding.sequence
 	});
-	logToFile(`[switchboard] decision: program=${ctx.program} in(use-pi-harness=${ctx.flags["wizard-use-pi-harness"] ?? "-"}, orchestrator=${ctx.flags["wizard-orchestrator"] ?? "-"}, cli=${ctx.cliHarness ?? "-"}/${ctx.cliSequence ?? "-"}/${ctx.cliModel ?? "-"}) → harness=${binding.harness} (${trace.harness ?? "?"}) model=${binding.model} (${trace.model ?? "?"}) sequence=${binding.sequence} (${trace.sequence ?? "?"})`);
+	logToFile(`[switchboard] decision: program=${ctx.program} in(orchestrator=${ctx.flags["wizard-orchestrator"] ?? "-"}, cli=${ctx.cliHarness ?? "-"}/${ctx.cliSequence ?? "-"}/${ctx.cliModel ?? "-"}) → harness=${binding.harness} (${trace.harness ?? "?"}) model=${binding.model} (${trace.model ?? "?"}) sequence=${binding.sequence} (${trace.sequence ?? "?"})`);
 }
 /**
 * Tag the run with its two routing axes. Sequence is stable for the whole
@@ -2571,4 +2353,4 @@ 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-NzzrOB02.js.map
+//# sourceMappingURL=agent-runner-B1qCtjid.js.map