@posthog/wizard 2.45.0 → 2.47.0

package/dist/{agent-runner-B0LTVFfw.js → agent-runner-D8pEOs_2.js}

@@ -1,20 +1,20 @@
 import { t as __exportAll } from "./rolldown-runtime-B_-DWIq7.js";
-import { A as GPT5_4_MODEL, 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, St as runtimeEnv, _ as evaluateWizardReadiness, _t as getSkillsBaseUrl, a as getLogFilePath, c as WIZARD_BENCHMARK_FILE, dt as WIZARD_PI_MODEL_FLAG_KEY, f as getUI, g as SIGNUP_WIZARD_READINESS_CONFIG, gt as WIZARD_USE_PI_HARNESS_FLAG_KEY, h as SERVICE_LABELS, i as enableDebugLogs, j as GPT5_5_MODEL, l as WIZARD_LOG_FILE, lt as WIZARD_ORCHESTRATOR_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, xt as RUN_SURFACE } from "./debug-DBmxx59Q.js";
-import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-DQEuL6qS.js";
-import { t as getOrAskForProjectData } from "./setup-utils-Qp-u6BZk.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-BiU-S12q.js";
+import { a as ciExcludedTaskTypes, n as analytics, r as groupsFromUser } from "./analytics-B4dny8FA.js";
+import { t as getOrAskForProjectData } from "./setup-utils-BMy5Zj3Z.js";
 import "./wizard-session-tp4Zdbwl.js";
-import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-BxtdkgI-.js";
-import { n as isNonInteractiveEnvironment } from "./environment-kZtkp7eG.js";
-import { 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 } from "./agent-interface-CF7GsdFf.js";
-import { a as QUEUE_DIR_NAME, o as QueueStore, s as TaskStatus } from "./queue-tools-Bw-AANLM.js";
-import { a as installSkillById, i as fetchSkillMenu, l as fetchWithRetry, u as createWizardAskBridge } from "./wizard-tools-Dg8o2lcH.js";
-import { a as flushScanReport, o as formatScanReport, p as writeScanReport, s as formatYaraAbortMessage } from "./yara-hooks-pw86rZ8y.js";
-import { r as detectNodePackageManagers } from "./package-manager-DKtR3xzM.js";
-import { a as piBackend, l as isThinkingLevel } from "./pi-DJCln9jJ.js";
-import { t as FRAMEWORK_REGISTRY } from "./registry-D2V88NoC.js";
+import { i as wizardAbort, n as registerCleanup, t as WizardError } from "./wizard-abort-D1UxHeiE.js";
+import { n as isNonInteractiveEnvironment } from "./environment-BT0RRheV.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-RZY46R62.js";
+import { a as QUEUE_DIR_NAME, o as QueueStore, s as TaskStatus } from "./queue-tools-rAN2wwDC.js";
+import { a as flushScanReport, o as formatScanReport, p as writeScanReport, s as formatYaraAbortMessage } from "./yara-hooks-D9PSjEqo.js";
+import { r as detectNodePackageManagers } from "./package-manager-L1RLE1ne.js";
+import { a as piBackend, l as isThinkingLevel } from "./pi-BMs3MY7y.js";
+import { t as FRAMEWORK_REGISTRY } from "./registry-Zp3B2L4q.js";
 import fs, { existsSync, rmSync } from "fs";
 import * as path$1 from "path";
 import path from "path";
+import { z } from "zod";
 import { randomUUID } from "crypto";
 //#region src/lib/agent/runner/shared/authenticate.ts
 async function authenticate(session, programId) {
@@ -149,6 +149,7 @@ async function bootstrapProgram(session, config, programConfig) {
 		}
 	}
 	const wizardFlags = await analytics.getAllFlagsForWizard();
+	const wizardFlagPayloads = analytics.getWizardFlagPayloads();
 	const wizardMetadata = buildRunTags({
 		programId: programConfig.id,
 		integration: config.integrationLabel,
@@ -160,6 +161,7 @@ async function bootstrapProgram(session, config, programConfig) {
 		skillsBaseUrl,
 		credentials: session.credentials,
 		wizardFlags,
+		wizardFlagPayloads,
 		wizardMetadata,
 		project
 	};
@@ -247,21 +249,67 @@ const anthropicBackend = {
 	}
 };
 //#endregion
-//#region src/lib/agent/runner/switchboard/harness.ts
+//#region src/lib/agent/runner/switchboard/flags/basic-integration.ts
 /**
-* Harness axis: registry, middleware, resolver. Mirrors `sequence.ts`.
+* Basic-integration pi experiment.
+*
+* Flags:  wizard-use-pi-harness (bool) + wizard-pi-model + wizard-pi-effort
+* Routes: ONLY `posthog-integration` — harness, model, and effort axes.
 */
-const HARNESS_OPTIONS = {
-	["anthropic"]: anthropicBackend,
-	["pi"]: piBackend
+const BASIC_INTEGRATION_EXPERIMENT = {
+	program: "posthog-integration",
+	flags: {
+		useFlag: WIZARD_USE_PI_HARNESS_FLAG_KEY,
+		modelFlag: WIZARD_PI_MODEL_FLAG_KEY,
+		effortFlag: WIZARD_PI_EFFORT_FLAG_KEY,
+		fallbackModel: GPT5_4_MODEL
+	}
 };
-function getHarness(name) {
-	const harness = HARNESS_OPTIONS[name];
-	if (!harness) throw new Error(`No harness registered for '${name}'.`);
-	return harness;
+//#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 = {
+	programs: ["posthog-integration"],
+	flag: WIZARD_ORCHESTRATOR_FLAG_KEY,
+	sequence: "orchestrator"
+};
+/** Raw flag read — for telemetry/log lines only, never for routing. */
+function isOrchestratorEnabled(flags = {}) {
+	return flags[ORCHESTRATOR_EXPERIMENT.flag] === "true";
 }
-/** `wizard-pi-model` variant key → gateway id (variant keys can't carry `/` or `.`). */
-const PI_MODEL_FLAG_VARIANTS = {
+//#endregion
+//#region src/lib/agent/runner/switchboard/flags/self-driving.ts
+/**
+* Self-driving pi experiment.
+*
+* Flag:   wizard-self-driving-use-pi-harness (bool; payload `{model, effort?,
+*         harness?, sequence?}`)
+* Routes: ONLY `self-driving` — any axis the payload pins. A missing or
+*         invalid payload keeps the non-flagged binding default.
+*/
+const SELF_DRIVING_EXPERIMENT = {
+	program: "self-driving",
+	flags: { useFlag: WIZARD_SELF_DRIVING_USE_PI_HARNESS_FLAG_KEY }
+};
+//#endregion
+//#region src/lib/agent/runner/switchboard/flags/schemes.ts
+/**
+* Flag config schemes — the shapes a PostHog flag set can take to route a
+* program to the pi harness, plus the shared vocabulary and the per-config
+* 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 `.`). */
+const MODEL_FLAG_VARIANTS = {
 	"gpt-5": GPT5_MODEL,
 	"gpt-5-4": GPT5_4_MODEL,
 	"gpt-5-mini": GPT5_MINI_MODEL,
@@ -272,25 +320,107 @@ const PI_MODEL_FLAG_VARIANTS = {
 	"sonnet-4-6": DEFAULT_AGENT_MODEL,
 	"sonnet-5": SONNET_5_MODEL
 };
-/** Programs the wizard-use-pi-harness flag may switch to pi; off this set the flag is a no-op and the binding default stands. */
-const PI_FLAG_PROGRAMS = new Set(["posthog-integration"]);
+/** Valid effort variants; anything else leaves the model's table default. */
+const EFFORT_FLAG_VARIANTS = [
+	"minimal",
+	"low",
+	"medium",
+	"high",
+	"xhigh"
+];
+/** `{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),
+	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);
+	} catch {
+		return;
+	}
+	const parsed = payloadConfigFlagSchema.safeParse(value);
+	if (!parsed.success) return void 0;
+	return {
+		model: MODEL_FLAG_VARIANTS[parsed.data.model],
+		thinkingLevel: parsed.data.effort,
+		harness: parsed.data.harness,
+		sequence: parsed.data.sequence
+	};
+}
 /**
-* `wizard-use-pi-harness` on → pi, paired with the `wizard-pi-model` variant
-* (unknown/missing variant → gpt-5.4). Off `PI_FLAG_PROGRAMS`, the flag is
-* ignored and the binding default stands.
+* Resolve one config against the flag snapshot, or undefined when it doesn't
+* validly route: use flag off, or a payload that fails validation — the
+* caller then keeps the non-flagged binding default.
+*/
+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
+		};
+	}
+	const route = parseFlagPayload(flagPayloads?.[cfg.useFlag]);
+	if (!route) logToFile(`[switchboard] ${cfg.useFlag} on but payload missing/invalid — keeping the non-flagged default`);
+	return route;
+}
+//#endregion
+//#region src/lib/agent/runner/switchboard/flags/index.ts
+/**
+* Switchboard flags, one module per experiment. Every experiment declares its
+* own scope — which program(s) its flags route — so the flag→program mapping
+* is readable per file, never implied. `schemes.ts` owns the shared config
+* shapes and resolution.
+*/
+/** Every experiment on each axis. An experiment routes ONLY by being listed here, and only for the programs it declares. */
+const HARNESS_EXPERIMENTS = [BASIC_INTEGRATION_EXPERIMENT, SELF_DRIVING_EXPERIMENT];
+const SEQUENCE_EXPERIMENTS = [ORCHESTRATOR_EXPERIMENT];
+/** The flag-driven route for a program, or undefined when no experiment covers it or its flags don't validly route. */
+function resolveFlagRoute(program, flags, flagPayloads) {
+	if (RUN_SURFACE === "cloud") return void 0;
+	const experiment = HARNESS_EXPERIMENTS.find((e) => e.program === program);
+	return experiment && routeFromConfigFlag(experiment.flags, flags, flagPayloads);
+}
+/** The flag-driven sequence for a program, or undefined when no sequence experiment covers it with its flag on. */
+function resolveFlagSequence(program, flags) {
+	return SEQUENCE_EXPERIMENTS.find((e) => e.programs.includes(program) && flags[e.flag] === "true")?.sequence;
+}
+//#endregion
+//#region src/lib/agent/runner/switchboard/harness.ts
+/**
+* Harness axis: registry, middleware, resolver. Mirrors `sequence.ts`.
+*/
+const HARNESS_OPTIONS = {
+	["anthropic"]: anthropicBackend,
+	["pi"]: piBackend
+};
+function getHarness(name) {
+	const harness = HARNESS_OPTIONS[name];
+	if (!harness) throw new Error(`No harness registered for '${name}'.`);
+	return harness;
+}
+/**
+* PostHog-flag routing to pi (see `./flags`). No valid route — flag off, no
+* config, or an invalid payload — keeps the non-flagged binding default.
 */
 const flagRunnerOverride = (ctx, next) => {
 	const pick = next();
-	if (ctx.flags["wizard-use-pi-harness"] !== "true") return pick;
-	if (RUN_SURFACE === "cloud") return pick;
-	if (!PI_FLAG_PROGRAMS.has(ctx.program)) return pick;
+	const route = resolveFlagRoute(ctx.program, ctx.flags, ctx.flagPayloads);
+	if (!route) return pick;
 	if (ctx.trace) Object.assign(ctx.trace, {
 		harness: "flag",
 		model: "flag"
 	});
 	return {
-		harness: "pi",
-		model: PI_MODEL_FLAG_VARIANTS[ctx.flags["wizard-pi-model"] ?? ""] ?? "openai/gpt-5.4"
+		harness: route.harness ?? "pi",
+		model: route.model,
+		thinkingLevel: route.thinkingLevel
 	};
 };
 const HARNESS_MIDDLEWARE = [...[], flagRunnerOverride];
@@ -308,6 +438,7 @@ function resolveHarness(ctx, role = "default") {
 		return {
 			harness: binding.harness,
 			model: binding.model,
+			thinkingLevel: binding.thinkingLevel,
 			...binding.contextMillOverride?.[role]
 		};
 	});
@@ -1378,6 +1509,7 @@ async function runLinearProgram(session, config, programConfig, boot, composed =
 	const pick = resolveHarness({
 		program: programConfig.id,
 		flags: wizardFlags,
+		flagPayloads: boot.wizardFlagPayloads,
 		cliHarness: session.harness,
 		cliModel: session.model
 	});
@@ -1391,7 +1523,8 @@ async function runLinearProgram(session, config, programConfig, boot, composed =
 		spinner,
 		askBridge,
 		middleware,
-		model: pick.model
+		model: pick.model,
+		thinkingLevel: pick.thinkingLevel
 	});
 	if (agentResult.error === "WIZARD_ABORT") {
 		const reason = agentResult.message ?? "";
@@ -1910,6 +2043,7 @@ async function runOrchestrator(session, programConfig, boot) {
 	const switchboardCtx = {
 		program: programConfig.id,
 		flags: boot.wizardFlags,
+		flagPayloads: boot.wizardFlagPayloads,
 		cliHarness: session.harness,
 		cliSequence: session.sequence,
 		cliModel: session.model
@@ -2161,15 +2295,29 @@ function getSequence(name) {
 	if (!sequence) throw new Error(`No sequence registered for '${name}'.`);
 	return sequence;
 }
-/** The `wizard-orchestrator` flag is on. */
-function isOrchestratorEnabled(flags = {}) {
-	return flags[WIZARD_ORCHESTRATOR_FLAG_KEY] === "true";
-}
-/** PostHog `wizard-orchestrator` flag → orchestrator. */
-const orchestratorFeatureFlagMw = (ctx, next) => {
-	if (!isOrchestratorEnabled(ctx.flags)) return next();
+/**
+* A composed sub-run (integration inside self-driving) is structurally
+* linear: the orchestrator owns the full run lifecycle (queue, outro) and
+* cannot nest. Sits above every override, including CLI.
+*/
+const composedClampMw = (ctx, next) => {
+	if (!ctx.composed) return next();
+	if (ctx.trace) ctx.trace.sequence = "composed";
+	return "linear";
+};
+/** A program's own flag route may pin the sequence; wins over the global orchestrator flag. Traced as 'payload' to stay distinguishable from sequence experiments ('flag'). */
+const flagRouteSequenceMw = (ctx, next) => {
+	const route = resolveFlagRoute(ctx.program, ctx.flags, ctx.flagPayloads);
+	if (!route?.sequence) return next();
+	if (ctx.trace) ctx.trace.sequence = "payload";
+	return route.sequence;
+};
+/** Sequence experiments (e.g. wizard-orchestrator), each inert outside its declared programs. */
+const sequenceExperimentMw = (ctx, next) => {
+	const sequence = resolveFlagSequence(ctx.program, ctx.flags);
+	if (!sequence) return next();
 	if (ctx.trace) ctx.trace.sequence = "flag";
-	return "orchestrator";
+	return sequence;
 };
 /**
 * The orchestrator drives harnesses through `runTask`; a harness that has not
@@ -2186,9 +2334,11 @@ const runTaskCapabilityClampMw = (ctx, next) => {
 	return "linear";
 };
 const SEQUENCE_MIDDLEWARE = [
+	composedClampMw,
 	...[],
 	runTaskCapabilityClampMw,
-	orchestratorFeatureFlagMw
+	flagRouteSequenceMw,
+	sequenceExperimentMw
 ];
 /** CLI wins over `wizard-orchestrator` flag wins over binding default. */
 function resolveSequence(ctx) {
@@ -2258,17 +2408,23 @@ 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. */
 function resolveBinding(ctx, role = "default") {
 	ctx.trace ??= {};
 	const sequence = resolveSequence(ctx);
-	const { harness, model } = resolveHarness(ctx, role);
+	const { harness, model, thinkingLevel } = resolveHarness(ctx, role);
 	return {
 		sequence,
 		harness,
-		model
+		model,
+		thinkingLevel
 	};
 }
 //#endregion
@@ -2292,7 +2448,7 @@ async function runProgram(session, config, programConfig, options = {}) {
 	const boot = await bootstrapProgram(session, config, programConfig);
 	registerCleanup(() => flushScanReport(session));
 	try {
-		const binding = resolveProgramRunner(session, programConfig, boot);
+		const binding = resolveProgramRunner(session, programConfig, boot, options.composed ?? false);
 		if (binding.sequence === "orchestrator") getUI().log.info("Task-queue orchestrator enabled.");
 		return await getSequence(binding.sequence).run(session, config, programConfig, boot, options.composed ?? false);
 	} finally {
@@ -2308,10 +2464,12 @@ async function runProgram(session, config, programConfig, options = {}) {
 * concern (bootstrap, cleanup, dispatch, per-task per-role harness picks) is
 * either upstream or downstream of this call.
 */
-function resolveProgramRunner(session, programConfig, boot) {
+function resolveProgramRunner(session, programConfig, boot, composed) {
 	const ctx = {
 		program: programConfig.id,
+		composed,
 		flags: boot.wizardFlags,
+		flagPayloads: boot.wizardFlagPayloads,
 		cliHarness: session.harness,
 		cliSequence: session.sequence,
 		cliModel: session.model
@@ -2332,6 +2490,8 @@ function captureSwitchboardDecision(ctx, binding) {
 		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],
 		cli_harness: ctx.cliHarness,
 		cli_sequence: ctx.cliSequence,
@@ -2341,6 +2501,7 @@ function captureSwitchboardDecision(ctx, binding) {
 		sequence_source: trace.sequence,
 		harness: binding.harness,
 		model: binding.model,
+		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 ?? "?"})`);
@@ -2363,4 +2524,4 @@ 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-B0LTVFfw.js.map
+//# sourceMappingURL=agent-runner-D8pEOs_2.js.map