github-router 0.3.295 → 0.3.302

package/dist/{server-setup-DsGJnJ_O.js → server-setup-Zn5wkosb.js}

@@ -1,9 +1,12 @@
-import { $ as bucketEffort, $t as assembleResponsesPayload, An as isBudgetClaudeLead, B as resolveAdvisorModel, Bn as stripTrailingOneMSuffix, Dn as UPSTREAM_FETCH_TIMEOUT_MS, E as toolbeltEnabled, F as FAST_ADVISOR_TOOL_INSTRUCTIONS, Fn as classifyMessagesRoute, G as buildAnthropicErrorEvent, Gt as countTokens, H as rememberThinkingHistoryRepair, Ht as resolveModelAlias, I as buildAdvisorStream, In as catalogAdvertises1M, J as logStreamError, Jt as getTokenCount, K as buildOpenAIErrorEvent, Kt as createMessages, L as injectAdvisorTool, Ln as oneMContextDisabled, Lt as LUNA_SONNET_ALIAS_ID, M as searchWeb, Mt as LUNA_DRIVER_ALIAS_ID, N as ADVISOR_INTERNAL_TOOL_NAME, Nn as upstreamAllowH2, Nt as LUNA_HAIKU_ALIAS_ID, On as UPSTREAM_INACTIVITY_TIMEOUT_MS, P as ADVISOR_TOOL_INSTRUCTIONS, Pn as upstreamMaxConnections, Q as UNKNOWN_EFFORT_ANCHOR, R as isAdvisorRequested, Rn as withOneMSuffix, Rt as canonicalizeAliasModel, Sn as BUDGET_SMALL_FAST_SLUG, U as repairKnownThinkingHistory, V as formatThinkingRepairDecline, Vn as withInstallLock, W as repairRejectedThinkingHistory, Wt as shimDefaultsToXhigh, X as relayAnthropicStream, Xt as findLaunchBySecret, Y as readIteratorWithTimeout, Yt as getTokenizerFromModel, Z as EFFORT_ORDER, an as readResponseBodyCapped, bn as toolbeltPathOverride, en as warnOnTokenPriceDrift, et as clampEffort, hn as provisionTreeSitterAssets, i as assertMcpToolSurfaceConsistent, in as MAX_RESPONSE_BODY_BYTES, kn as generateRandomPort, nn as createResponses, nt as handleMcpPost, on as parseJsonOrDiagnose, q as isControllerClosedError, qt as getTextTokenCount, rn as createChatCompletions, sn as normalizeOpenAIUsage, st as agentToolsEnabled, tn as resolveMcpToolTimeoutMs, tt as handleMcpDelete, z as resolveAdvisorEffort, zn as withOneMSuffixForLead } from "./peer-mcp-personas-BwNtC8jt.js";
+import { $ as UNKNOWN_EFFORT_ANCHOR, $t as countTokens, An as BUDGET_SMALL_FAST_SLUG, B as resolveAdvisorEffort, Bt as MAX_LUNA_HIGH_ALIAS_ID, Cn as provisionTreeSitterAssets, D as toolbeltEnabled, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as UPSTREAM_FETCH_TIMEOUT_MS, Ft as LUNA_HAIKU_ALIAS_ID, G as repairRejectedThinkingHistory, Gn as oneMContextDisabled, Gt as maxAliasEffort, H as formatThinkingRepairDecline, Hn as upstreamMaxConnections, Ht as canonicalizeAliasModel, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, In as UPSTREAM_INACTIVITY_TIMEOUT_MS, J as isControllerClosedError, Jn as withInstallLock, K as buildAnthropicErrorEvent, Kn as withOneMSuffix, Kt as maxAliasModel, L as buildAdvisorStream, Ln as generateRandomPort, Lt as LUNA_REAL_MODEL_ID, N as searchWeb, On as toolbeltPathOverride, P as ADVISOR_INTERNAL_TOOL_NAME, Pt as LUNA_DRIVER_ALIAS_ID, Q as EFFORT_ORDER, Qt as shimDefaultsToXhigh, R as injectAdvisorTool, Rn as isBudgetClaudeLead, U as rememberThinkingHistoryRepair, Un as classifyMessagesRoute, V as resolveAdvisorModel, Vn as upstreamAllowH2, Vt as MAX_LUNA_MAX_ALIAS_ID, W as repairKnownThinkingHistory, Wn as catalogAdvertises1M, Wt as isMaxModelAlias, X as readIteratorWithTimeout, Y as logStreamError, Yt as resolveModelAlias, Z as relayAnthropicStream, cn as warnOnTokenPriceDrift, dn as createChatCompletions, en as createMessages, et as bucketEffort, fn as MAX_RESPONSE_BODY_BYTES, hn as normalizeOpenAIUsage, i as assertMcpToolSurfaceConsistent, in as findLaunchBySecret, ln as resolveMcpToolTimeoutMs, lt as agentToolsEnabled, mn as parseJsonOrDiagnose, nn as getTokenCount, nt as handleMcpDelete, pn as readResponseBodyCapped, q as buildOpenAIErrorEvent, qn as withOneMSuffixForLead, rn as getTokenizerFromModel, rt as handleMcpPost, sn as assembleResponsesPayload, tn as getTextTokenCount, tt as clampEffort, un as createResponses, z as isAdvisorRequested, zt as LUNA_SONNET_ALIAS_ID } from "./peer-mcp-personas-DwEiYeJS.js";
 import { t as getPackageVersion } from "./version-C8x2hQrZ.js";
-import { i as ensurePaths, t as PATHS } from "./paths-Ci485qSJ.js";
-import { A as state, T as copilotHeaders, _ as forwardError, a as cacheCopilotVersion, c as filterBetaHeader, d as resolveModel, f as sleep, g as HTTPError, i as tryRefreshAndRetry, l as isNullish, n as setupGitHubAgentToken, o as cacheModels, r as setupGitHubToken, s as cacheVSCodeVersion, t as setupCopilotToken, w as copilotBaseUrl, y as fetchWithTransientRetry } from "./token-BtJhjXXu.js";
-import { t as getCopilotUsage } from "./get-copilot-usage-B7Z4XVWj.js";
+import { i as ensurePaths, t as PATHS } from "./paths-IcZW8gZx.js";
+import { t as state } from "./state-B8adscCv.js";
+import { T as copilotHeaders, _ as forwardError, a as cacheCopilotVersion, c as filterBetaHeader, d as resolveModel, f as sleep, g as HTTPError, i as tryRefreshAndRetry, l as isNullish, n as setupGitHubAgentToken, o as cacheModels, r as setupGitHubToken, s as cacheVSCodeVersion, t as setupCopilotToken, w as copilotBaseUrl, y as fetchWithTransientRetry } from "./token-CbuFyx-B.js";
+import { t as getCopilotUsage } from "./get-copilot-usage-ByjQqhdi.js";
 import { a as resolveExecutable, n as killManagedTree, o as runCommandCapture, r as parseBoolEnv, s as runCommandVoid } from "./exec-y8C_MU8A.js";
+import { r as FAST_PROFILE_ADVISOR_MODEL } from "./fast-profile-contract-K3P58nj3.js";
+import { b as stripTrailingOneMSuffix, d as maxAdvisorPinIsValid, i as MAX_PROFILE_MODELS, l as maxAdvisorEffortForModel, m as maxOpusModel, n as MAX_PROFILE_ALLOWED_LEAD_MODEL_IDS, t as MAX_PROFILE_ADVISOR_INSTRUCTIONS, u as maxAdvisorModelFromPin } from "./max-profile-contract-BTXowh6k.js";
 import consola from "consola";
 import * as fs$2 from "node:fs";
 import fs, { existsSync } from "node:fs";
@@ -1089,7 +1092,7 @@ const CLIENT_OUTPUT_RESERVE_TOKENS = 2e4;
 /**
 * The client's additional flat reserve between the output-reduced window and
 * its reactive compaction threshold. Read from the installed Claude Code
-* 2.1.247 bundle: `xWe(e,t)` computes `n = e - 13000` and returns it when no
+* 2.1.251 bundle: `E9(e,t)` computes `r = e - 13000` and returns it when no
 * percentage override is set, where `e` is already
 * `window - min(maxOutput, 20_000)`.
 */
@@ -1098,7 +1101,9 @@ const CLIENT_THRESHOLD_RESERVE_TOKENS = 13e3;
 * Bounds the `CLAUDE_CODE_AUTO_COMPACT_WINDOW` env path applies to whatever it
 * parses: `IWe = 1e5` floor, `ZRt = 1e6` cap. A value below the floor is
 * silently RAISED to it, so deriving something smaller would be a lie; a value
-* above the cap is clamped. Both read from the same 2.1.247 bundle.
+* above the cap is clamped. In the 2.1.251 bundle, `ike = 1e5` and
+* `JNe = 1e6` are passed to the shared env parser and the result is floored
+* again with `Math.max(ike, ...)` before use.
 */
 const ENV_WINDOW_MIN_TOKENS = 1e5;
 const ENV_WINDOW_MAX_TOKENS = 1e6;
@@ -1497,7 +1502,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.295";
+var version = "0.3.302";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -4359,13 +4364,103 @@ function identityPreflightErrorResponse(c, reason, path = "/v1/messages") {
 	}, 403);
 }
 //#endregion
+//#region src/lib/max-request-preprocess.ts
+function maxRequestError(result) {
+	if (result.rejectedAlias) return `Router-owned model alias ${JSON.stringify(result.rejectedAlias)} is valid only for an authenticated -m max launch.`;
+	if (result.rejectedModel) return `Model ${JSON.stringify(result.rejectedModel)} is outside the fixed -m max model set (Sol, Luna, Gemini 3.7 Flash, or Opus 5).`;
+}
+const MAX_MODEL_EFFORTS = Object.freeze({
+	"gpt-5.6-sol": "high",
+	"gpt-5.6-luna": "high",
+	"gemini-3.7-flash": "high",
+	"claude-opus-5": "high"
+});
+function allowedModel(base) {
+	return MAX_PROFILE_ALLOWED_LEAD_MODEL_IDS.includes(base);
+}
+function setEffort(body, effort) {
+	const outputConfig = body.output_config && typeof body.output_config === "object" ? body.output_config : {};
+	if (outputConfig.effort === void 0) body.output_config = {
+		...outputConfig,
+		effort
+	};
+}
+/**
+* Apply the bound max profile's lead-model projection. Max is intentionally
+* separate from the fast preprocessor: fast aliases and fixed fast role rules
+* must not become active for max, while arbitrary old/unknown model selections
+* must fail closed rather than silently escape the max surface.
+*/
+function preprocessMaxRequest(rawBody, launch, subagentRequest = false) {
+	let parsed;
+	try {
+		parsed = JSON.parse(rawBody);
+	} catch {
+		return {
+			body: rawBody,
+			modified: false
+		};
+	}
+	const originalModel = typeof parsed.model === "string" ? parsed.model : void 0;
+	if (!originalModel) return {
+		body: rawBody,
+		modified: false
+	};
+	if (launch?.profileId !== "max") return {
+		body: rawBody,
+		originalModel,
+		modified: false
+	};
+	const { base, hadSuffix } = stripTrailingOneMSuffix(originalModel);
+	const aliasEffort = maxAliasEffort(originalModel);
+	const canonical = maxAliasModel(originalModel);
+	if (canonical) {
+		const effort = aliasEffort ?? "high";
+		parsed.model = hadSuffix ? `${canonical}[1m]` : canonical;
+		setEffort(parsed, effort);
+		if (parsed.thinking && typeof parsed.thinking === "object") {
+			if (parsed.thinking.type === "enabled") parsed.thinking = { type: "adaptive" };
+		}
+		return {
+			body: JSON.stringify(parsed),
+			originalModel,
+			modified: true
+		};
+	}
+	const allowedSubagentModel = subagentRequest && base === "grok-4.6";
+	if (!allowedModel(base) && !allowedSubagentModel) return {
+		body: rawBody,
+		originalModel,
+		modified: false,
+		rejectedModel: originalModel
+	};
+	const effort = allowedSubagentModel ? "medium" : MAX_MODEL_EFFORTS[base];
+	if (!effort) return {
+		body: rawBody,
+		originalModel,
+		modified: false,
+		rejectedModel: originalModel
+	};
+	setEffort(parsed, effort);
+	if (parsed.thinking && typeof parsed.thinking === "object") {
+		if (parsed.thinking.type === "enabled") parsed.thinking = { type: "adaptive" };
+	}
+	return {
+		body: JSON.stringify(parsed),
+		originalModel,
+		modified: true
+	};
+}
+Object.freeze([MAX_LUNA_HIGH_ALIAS_ID, MAX_LUNA_MAX_ALIAS_ID]);
+//#endregion
 //#region src/lib/fast-request-preprocess.ts
 /**
 * Apply authenticated fast-profile model and effort policy before ordinary model
 * resolution. Synthetic aliases are refused outside an authenticated fast
 * launch, so raw/BYO traffic cannot opt itself into private profile semantics.
 */
-function preprocessFastRequest(rawBody, launch) {
+function preprocessFastRequest(rawBody, launch, subagentRequest = false) {
+	if (launch?.profileId === "max") return preprocessMaxRequest(rawBody, launch, subagentRequest);
 	let parsed;
 	try {
 		parsed = JSON.parse(rawBody);
@@ -4381,7 +4476,7 @@ function preprocessFastRequest(rawBody, launch) {
 		modified: false
 	};
 	const alias = resolveModelAlias(originalModel);
-	if (alias && launch?.profileId !== "fast") return {
+	if (alias && (launch?.profileId !== "fast" || isMaxModelAlias(originalModel))) return {
 		body: rawBody,
 		originalModel,
 		modified: false,
@@ -4700,6 +4795,32 @@ function stripWebSearchTool(body) {
 * in probes `shim_advisor_degrade_gpt55` and
 * `shim_advisor_degrade_gemini35flash`.
 */
+function hasNonEmptyTools(rawBody) {
+	let body;
+	try {
+		body = JSON.parse(rawBody);
+	} catch {
+		return false;
+	}
+	return Array.isArray(body.tools) && body.tools.length > 0;
+}
+function fastAdvisorMetadataMismatch(rawBody) {
+	let body;
+	try {
+		body = JSON.parse(rawBody);
+	} catch {
+		return;
+	}
+	if (!Array.isArray(body.tools) || body.tools.length === 0) return void 0;
+	for (const tool of body.tools) {
+		if (!tool || typeof tool !== "object") continue;
+		const type = tool.type;
+		if (typeof type !== "string" || !type.startsWith("advisor_")) continue;
+		const model = tool.model;
+		if (typeof model !== "string") return "the native Advisor tool omitted its fixed model";
+		if (stripTrailingOneMSuffix(model).base !== FAST_PROFILE_ADVISOR_MODEL) return `the native Advisor tool requested ${JSON.stringify(model)} instead of ${JSON.stringify(FAST_PROFILE_ADVISOR_MODEL)}`;
+	}
+}
 function stripAdvisorTool(rawBody) {
 	let body;
 	try {
@@ -4833,12 +4954,57 @@ async function handleCompletion(c) {
 	const incomingBeta = c.req.header("anthropic-beta");
 	const advisorRequested = isAdvisorRequested(incomingBeta);
 	const fastProfileRequest = identity.launch?.profileId === "fast";
-	const fastSubagentRequest = fastProfileRequest && Boolean(c.req.header("x-claude-code-agent-id"));
+	const maxProfileRequest = identity.launch?.profileId === "max";
+	const subagentRequest = Boolean(c.req.header("x-claude-code-agent-id"));
+	const fastSubagentRequest = fastProfileRequest && subagentRequest;
+	const maxSubagentRequest = maxProfileRequest && subagentRequest;
 	const fastLeadAdvisor = fastProfileRequest && !fastSubagentRequest;
-	const advisorEnabled = advisorRequested && !fastSubagentRequest;
-	const fastPreprocess = preprocessFastRequest(rawBody, identity.launch);
+	const maxLeadAdvisor = maxProfileRequest && !maxSubagentRequest;
+	const advisorEnabled = advisorRequested && !fastSubagentRequest && !maxSubagentRequest;
+	const fastAdvisorEnabled = fastLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
+	const maxAdvisorEnabled = maxLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
+	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : maxLeadAdvisor ? maxAdvisorEnabled : advisorEnabled;
+	let fastAdvisorChoice;
+	let maxAdvisorChoice;
+	if (maxAdvisorEnabled) {
+		const requestedAdvisor = process.env.GH_ROUTER_ADVISOR_MODEL;
+		if (!maxAdvisorPinIsValid(requestedAdvisor)) return c.json({
+			type: "error",
+			error: {
+				type: "invalid_request_error",
+				message: `Invalid GH_ROUTER_ADVISOR_MODEL ${JSON.stringify(requestedAdvisor)} for max profile; must be gpt-5.6-sol or claude-opus-5.`
+			}
+		}, 400, { "x-should-retry": "false" });
+		const model = maxAdvisorModelFromPin(requestedAdvisor, maxOpusModel());
+		maxAdvisorChoice = {
+			model,
+			effort: maxAdvisorEffortForModel(model)
+		};
+	}
+	if (fastAdvisorEnabled) {
+		const mismatch = fastAdvisorMetadataMismatch(rawBody);
+		if (mismatch) return c.json({
+			type: "error",
+			error: {
+				type: "invalid_request_error",
+				message: `Fast Advisor model mismatch: ${mismatch}. Run \`/advisor ${FAST_PROFILE_ADVISOR_MODEL}[1m]\` to restore the fixed fast profile, or relaunch with \`github-router claude -m fast\`.`
+			}
+		}, 400, { "x-should-retry": "false" });
+		try {
+			fastAdvisorChoice = resolveAdvisorModel(void 0, true);
+		} catch (error) {
+			return c.json({
+				type: "error",
+				error: {
+					type: "api_error",
+					message: error instanceof Error ? error.message : String(error)
+				}
+			}, 503);
+		}
+	}
+	const fastPreprocess = preprocessFastRequest(rawBody, identity.launch, subagentRequest);
 	if (fastPreprocess.rejectedAlias || fastPreprocess.rejectedModel) {
-		const message = fastPreprocess.rejectedAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.rejectedAlias)} is valid only for an authenticated -m fast launch.` : `Model ${JSON.stringify(fastPreprocess.rejectedModel)} is outside the fixed -m fast model set.`;
+		const message = identity.launch?.profileId === "max" ? maxRequestError(fastPreprocess) ?? "Invalid max request" : fastPreprocess.rejectedAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.rejectedAlias)} is valid only for an authenticated -m fast launch.` : `Model ${JSON.stringify(fastPreprocess.rejectedModel)} is outside the fixed -m fast model set.`;
 		return c.json({
 			type: "error",
 			error: {
@@ -4849,7 +5015,7 @@ async function handleCompletion(c) {
 	}
 	let finalBody = await processWebSearch(fastPreprocess.body);
 	finalBody = sanitizeAnthropicBody(finalBody);
-	if (fastSubagentRequest) finalBody = stripAdvisorTool(finalBody);
+	if (fastSubagentRequest || maxSubagentRequest) finalBody = stripAdvisorTool(finalBody);
 	const loopGuard = guardAnthropicBody(finalBody);
 	if (loopGuard.action === "abort") return c.json({
 		type: "error",
@@ -4859,8 +5025,8 @@ async function handleCompletion(c) {
 		}
 	}, 400, { "x-should-retry": "false" });
 	if (loopGuard.body !== void 0) finalBody = loopGuard.body;
-	if (advisorEnabled) {
-		finalBody = injectAdvisorTool(finalBody, fastLeadAdvisor ? FAST_ADVISOR_TOOL_INSTRUCTIONS : void 0);
+	if (advisorBehaviorEnabled) {
+		finalBody = injectAdvisorTool(finalBody, fastLeadAdvisor ? FAST_ADVISOR_TOOL_INSTRUCTIONS : maxLeadAdvisor ? MAX_PROFILE_ADVISOR_INSTRUCTIONS : void 0);
 		consola.info("ADVISOR enabled for this request — injecting __anthropic_advisor tool; will translate tool_use → server_tool_use{advisor} on the SSE stream");
 	}
 	if (finalBody.includes("\"mcp_servers\"")) try {
@@ -4884,16 +5050,16 @@ async function handleCompletion(c) {
 			parsedBase = JSON.parse(promptWindowBody);
 		} catch {}
 		const wantsStream = parsedBase?.stream === true;
-		if (advisorEnabled && wantsStream && fastLeadAdvisor) {
+		if ((fastAdvisorEnabled || maxAdvisorEnabled) && wantsStream) {
 			const initialConversation = Array.isArray(parsedBase.messages) ? parsedBase.messages : [];
 			const parsedInitial = parseAnthropicRequest(parsedBase, modelId, selectedModel);
-			const fastAdvisorAborter = new AbortController();
+			const translatedAdvisorAborter = new AbortController();
 			const firstResponse = await streamParsedRequestViaShim(parsedInitial, endpoint, {
 				modelId,
 				model: selectedModel,
 				routePath: c.req.path,
-				onCancel: () => fastAdvisorAborter.abort()
-			}, fastAdvisorAborter.signal);
+				onCancel: () => translatedAdvisorAborter.abort()
+			}, translatedAdvisorAborter.signal);
 			logRequest({
 				method: "POST",
 				path: c.req.path,
@@ -4902,7 +5068,17 @@ async function handleCompletion(c) {
 				status: 200,
 				streaming: true
 			}, selectedModel, startTime);
-			const advisorChoice = resolveAdvisorModel(modelId, fastLeadAdvisor);
+			const advisorChoice = maxAdvisorEnabled ? {
+				model: maxAdvisorChoice.model,
+				effort: maxAdvisorChoice.effort,
+				escalated: false,
+				fastProfile: false
+			} : {
+				model: fastAdvisorChoice.model,
+				effort: resolveAdvisorEffort(rawBody, fastAdvisorChoice.model, true),
+				escalated: fastAdvisorChoice.escalated,
+				fastProfile: true
+			};
 			return new Response(buildAdvisorStream({
 				firstResponse,
 				initialConversation,
@@ -4910,9 +5086,10 @@ async function handleCompletion(c) {
 				requestHeaders: {},
 				advisorModel: advisorChoice.model,
 				advisorEscalated: advisorChoice.escalated,
-				advisorFastProfile: fastLeadAdvisor,
-				advisorEffort: resolveAdvisorEffort(rawBody, advisorChoice.model, true),
-				externalAborter: fastAdvisorAborter,
+				advisorFastProfile: advisorChoice.fastProfile,
+				advisorMaxProfile: maxAdvisorEnabled,
+				advisorEffort: advisorChoice.effort,
+				externalAborter: translatedAdvisorAborter,
 				continueTurn: makeShimContinueTurn(endpoint, {
 					modelId,
 					model: selectedModel
@@ -4928,7 +5105,7 @@ async function handleCompletion(c) {
 			});
 		}
 		const shimBody = stripAdvisorTool(promptWindowBody);
-		if (advisorEnabled) consola.info("ADVISOR requested with a non-Claude model — stripping the injected __anthropic_advisor tool and proceeding without advisor (Claude-only feature; gracefully degraded).");
+		if (advisorBehaviorEnabled) consola.info("ADVISOR requested with a non-Claude model — stripping the injected __anthropic_advisor tool and proceeding without advisor (Claude-only feature; gracefully degraded).");
 		const shimOpts = {
 			rawBody: shimBody,
 			modelId,
@@ -4940,7 +5117,7 @@ async function handleCompletion(c) {
 	}
 	if (modelId) logEndpointMismatch(modelId, "/v1/messages");
 	const effectiveBetas = applyDefaultBetas(betaHeaders, resolvedModel ?? originalModel);
-	const advisorAborter = advisorEnabled ? new AbortController() : void 0;
+	const advisorAborter = advisorBehaviorEnabled ? new AbortController() : void 0;
 	const requestHeaders = {
 		...selectedModel?.requestHeaders,
 		...effectiveBetas
@@ -5014,13 +5191,18 @@ async function handleCompletion(c) {
 		if (requestId) streamHeaders["x-request-id"] = requestId;
 		const reqId = response.headers.get("request-id");
 		if (reqId) streamHeaders["request-id"] = reqId;
-		if (advisorEnabled && response.body) {
+		if (advisorBehaviorEnabled && response.body) {
 			let parsedBase = {};
 			try {
 				parsedBase = JSON.parse(nativeBody);
 			} catch {}
 			const initialConversation = Array.isArray(parsedBase.messages) ? parsedBase.messages : [];
-			const advisorChoice = resolveAdvisorModel(originalModel, fastLeadAdvisor);
+			const advisorChoice = fastAdvisorChoice ?? (maxAdvisorChoice ? {
+				model: maxAdvisorChoice.model,
+				effort: maxAdvisorChoice.effort,
+				escalated: false,
+				fastProfile: false
+			} : resolveAdvisorModel(originalModel, false));
 			return new Response(buildAdvisorStream({
 				firstResponse: response,
 				initialConversation,
@@ -5029,7 +5211,8 @@ async function handleCompletion(c) {
 				advisorModel: advisorChoice.model,
 				advisorEscalated: advisorChoice.escalated,
 				advisorFastProfile: fastLeadAdvisor,
-				advisorEffort: resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor),
+				advisorMaxProfile: maxAdvisorEnabled,
+				advisorEffort: maxAdvisorChoice ? maxAdvisorChoice.effort : resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor),
 				externalAborter: advisorAborter
 			}), {
 				status: response.status,
@@ -5365,9 +5548,11 @@ async function handleCountTokens(c) {
 	const startTime = Date.now();
 	const identity = runMessagesIdentityPreflight(c);
 	if (!identity.ok) return identityPreflightErrorResponse(c, identity.reason, c.req.path);
-	const fastPreprocess = preprocessFastRequest(await c.req.text(), identity.launch);
+	const rawBody = await c.req.text();
+	const subagentRequest = Boolean(c.req.header("x-claude-code-agent-id"));
+	const fastPreprocess = preprocessFastRequest(rawBody, identity.launch, subagentRequest);
 	if (fastPreprocess.rejectedAlias || fastPreprocess.rejectedModel) {
-		const message = fastPreprocess.rejectedAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.rejectedAlias)} is valid only for an authenticated -m fast launch.` : `Model ${JSON.stringify(fastPreprocess.rejectedModel)} is outside the fixed -m fast model set.`;
+		const message = identity.launch?.profileId === "max" ? maxRequestError(fastPreprocess) ?? "Invalid max request" : fastPreprocess.rejectedAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.rejectedAlias)} is valid only for an authenticated -m fast launch.` : `Model ${JSON.stringify(fastPreprocess.rejectedModel)} is outside the fixed -m fast model set.`;
 		return c.json({
 			type: "error",
 			error: {
@@ -6425,6 +6610,30 @@ function nativeSelectableModelsInCatalog() {
 		display_name: m.displayName
 	}));
 }
+const MAX_NATIVE_MODELS = [
+	{
+		id: MAX_PROFILE_MODELS.sol,
+		displayName: "GPT-5.6 Sol"
+	},
+	{
+		id: MAX_PROFILE_MODELS.luna,
+		displayName: "GPT-5.6 Luna"
+	},
+	{
+		id: MAX_PROFILE_MODELS.gemini,
+		displayName: "Gemini 3.7 Flash"
+	},
+	{
+		id: MAX_PROFILE_MODELS.opus,
+		displayName: "Claude Opus 5"
+	}
+];
+function maxSelectableModelsInCatalog() {
+	return MAX_NATIVE_MODELS.filter((model) => state.models?.data.some((entry) => entry.id === model.id)).map((model) => ({
+		id: withOneMSuffix(model.id),
+		display_name: model.displayName
+	}));
+}
 /**
 * Pre-seed Claude Code's gateway-model discovery cache so the non-Claude
 * models appear as selectable picker rows WITHOUT the network fetch.
@@ -6552,7 +6761,10 @@ function clearGatewayModelCache(configDir = PATHS.CLAUDE_CONFIG_DIR) {
 */
 function oneMSuffixForAlias(aliasId) {
 	if (oneMContextDisabled()) return aliasId;
-	return catalogAdvertises1M("gpt-5.6-luna") ? `${aliasId}[1m]` : aliasId;
+	return catalogAdvertises1M(LUNA_REAL_MODEL_ID) ? `${aliasId}[1m]` : aliasId;
+}
+function oneMSuffixForMaxModel(id) {
+	return withOneMSuffix(id);
 }
 function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 	const vars = {
@@ -6567,8 +6779,9 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 	if (process.env.MCP_TIMEOUT === void 0) vars.MCP_TIMEOUT = mcpToolTimeoutMs;
 	if (process.env.MCP_TOOL_TIMEOUT === void 0) vars.MCP_TOOL_TIMEOUT = mcpToolTimeoutMs;
 	const isFastProfile = launchProfileId === "fast";
-	const smallFastModel = isFastProfile ? LUNA_HAIKU_ALIAS_ID : isBudgetClaudeLead(model) && (state.models?.data?.some((m) => m.id === "claude-haiku-4.5") ?? false) ? BUDGET_SMALL_FAST_SLUG : "claude-sonnet-5";
-	if (process.env.ANTHROPIC_SMALL_FAST_MODEL === void 0) vars.ANTHROPIC_SMALL_FAST_MODEL = isFastProfile ? oneMSuffixForAlias(smallFastModel) : smallFastModel;
+	const isMaxProfile = launchProfileId === "max";
+	const smallFastModel = isMaxProfile ? MAX_LUNA_HIGH_ALIAS_ID : isFastProfile ? LUNA_HAIKU_ALIAS_ID : isBudgetClaudeLead(model) && (state.models?.data?.some((m) => m.id === "claude-haiku-4.5") ?? false) ? BUDGET_SMALL_FAST_SLUG : "claude-sonnet-5";
+	if (process.env.ANTHROPIC_SMALL_FAST_MODEL === void 0) vars.ANTHROPIC_SMALL_FAST_MODEL = isFastProfile || isMaxProfile ? oneMSuffixForAlias(smallFastModel) : smallFastModel;
 	const seedTierRow = (modelKey, nameKey, bareSlug) => {
 		if (process.env[modelKey] !== void 0) return;
 		vars[modelKey] = withOneMSuffixForLead(bareSlug);
@@ -6579,7 +6792,20 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 		vars[modelKey] = oneMSuffixForAlias(aliasId);
 		if (process.env[nameKey] === void 0) vars[nameKey] = displayName;
 	};
-	if (isFastProfile) {
+	if (isMaxProfile) {
+		const seedMaxRow = (modelKey, nameKey, id) => {
+			if (process.env[modelKey] !== void 0) return;
+			vars[modelKey] = oneMSuffixForMaxModel(id);
+			if (process.env[nameKey] === void 0) vars[nameKey] = id;
+		};
+		seedMaxRow("ANTHROPIC_DEFAULT_SONNET_MODEL", "ANTHROPIC_DEFAULT_SONNET_MODEL_NAME", MAX_PROFILE_MODELS.luna);
+		seedMaxRow("ANTHROPIC_DEFAULT_HAIKU_MODEL", "ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME", MAX_PROFILE_MODELS.luna);
+		if (maxOpusModel()) seedMaxRow("ANTHROPIC_DEFAULT_OPUS_MODEL", "ANTHROPIC_DEFAULT_OPUS_MODEL_NAME", MAX_PROFILE_MODELS.opus);
+		if (process.env.ANTHROPIC_CUSTOM_MODEL_OPTION === void 0) {
+			vars.ANTHROPIC_CUSTOM_MODEL_OPTION = oneMSuffixForMaxModel(MAX_PROFILE_MODELS.sol);
+			if (process.env.ANTHROPIC_CUSTOM_MODEL_OPTION_NAME === void 0) vars.ANTHROPIC_CUSTOM_MODEL_OPTION_NAME = "GPT-5.6 Sol";
+		}
+	} else if (isFastProfile) {
 		seedFastAliasTierRow("ANTHROPIC_DEFAULT_SONNET_MODEL", "ANTHROPIC_DEFAULT_SONNET_MODEL_NAME", LUNA_SONNET_ALIAS_ID, "GPT-5.6 Luna (xhigh)");
 		seedFastAliasTierRow("ANTHROPIC_DEFAULT_HAIKU_MODEL", "ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME", LUNA_HAIKU_ALIAS_ID, "GPT-5.6 Luna (high)");
 		const fastAliasCapabilities = "effort,xhigh_effort,max_effort,thinking,adaptive_thinking,interleaved_thinking";
@@ -6594,7 +6820,7 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 		seedTierRow("ANTHROPIC_DEFAULT_SONNET_MODEL", "ANTHROPIC_DEFAULT_SONNET_MODEL_NAME", "claude-sonnet-5");
 		seedTierRow("ANTHROPIC_DEFAULT_HAIKU_MODEL", "ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME", smallFastModel);
 	}
-	seedTierRow("ANTHROPIC_DEFAULT_OPUS_MODEL", "ANTHROPIC_DEFAULT_OPUS_MODEL_NAME", "claude-opus-5");
+	if (!isMaxProfile) seedTierRow("ANTHROPIC_DEFAULT_OPUS_MODEL", "ANTHROPIC_DEFAULT_OPUS_MODEL_NAME", "claude-opus-5");
 	if (process.env.CLAUDE_CODE_PLAN_V2_AGENT_COUNT === void 0) vars.CLAUDE_CODE_PLAN_V2_AGENT_COUNT = "7";
 	for (const key of [
 		"CLAUDE_CODE_ENABLE_EXPERIMENTAL_ADVISOR_TOOL",
@@ -6603,11 +6829,11 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 		"CLAUDE_CODE_ENABLE_FINE_GRAINED_TOOL_STREAMING",
 		"CLAUDE_CODE_ENABLE_TASKS"
 	]) if (process.env[key] === void 0) vars[key] = "1";
-	const nativeModels = nativeSelectableModelsInCatalog();
+	const nativeModels = isMaxProfile ? maxSelectableModelsInCatalog() : nativeSelectableModelsInCatalog();
 	if (nativeModels.length > 0) {
 		if (seedGatewayModelCache(serverUrl, nativeModels) && process.env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY === void 0 && vars.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY === void 0) vars.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1";
 	} else clearGatewayModelCache();
-	applyAutoCompactWindow(vars);
+	applyAutoCompactWindow(vars, launchProfileId);
 	if (toolbeltEnabled()) Object.assign(vars, toolbeltPathOverride(process.env, PATHS.TOOLBELT_BIN_DIR));
 	return vars;
 }
@@ -6652,7 +6878,7 @@ function catalogEntryForSeededModel(value) {
 *
 * Presence-guarded on the parent env, symmetric with every other guard here.
 */
-function applyAutoCompactWindow(vars) {
+function applyAutoCompactWindow(vars, launchProfileId) {
 	if (process.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW !== void 0) return;
 	let tightestWindow;
 	const consider = (value) => {
@@ -6662,7 +6888,8 @@ function applyAutoCompactWindow(vars) {
 		if (candidateWindow !== void 0 && (tightestWindow === void 0 || candidateWindow < tightestWindow)) tightestWindow = candidateWindow;
 	};
 	for (const key of LEAD_CAPABLE_MODEL_ENV_KEYS) consider(vars[key] ?? process.env[key]);
-	for (const model of nativeSelectableModelsInCatalog()) consider(model.id);
+	const gatewayModels = launchProfileId === "max" ? maxSelectableModelsInCatalog() : nativeSelectableModelsInCatalog();
+	for (const model of gatewayModels) consider(model.id);
 	if (tightestWindow === void 0) return;
 	vars.CLAUDE_CODE_AUTO_COMPACT_WINDOW = String(tightestWindow);
 }
@@ -6690,4 +6917,4 @@ function getCodexEnvVars(serverUrl) {
 //#endregion
 export { sharedServerArgs as a, stopKeepAwake as c, runSelfUpdate as d, checkClaudeVersion as f, setupAndServe as i, listModelsForEndpoint as l, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, updateClaude as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, enableFileLogging as u };
 
-//# sourceMappingURL=server-setup-DsGJnJ_O.js.map
+//# sourceMappingURL=server-setup-Zn5wkosb.js.map