github-router 0.3.311 → 0.3.312

package/dist/{server-setup-OTy-qHfv.js → server-setup-DX27fdcf.js}

@@ -1,4 +1,4 @@
-import { $ as UNKNOWN_EFFORT_ANCHOR, At as shimDefaultsToXhigh, B as resolveAdvisorEffort, Bt as warnOnTokenPriceDrift, Cn as oneMContextDisabled, D as toolbeltEnabled, En as withInstallLock, F as ADVISOR_TOOL_INSTRUCTIONS, Ft as getTokenizerFromModel, G as repairRejectedThinkingHistory, Gt as readResponseBodyCapped, H as formatThinkingRepairDecline, Ht as createResponses, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, It as findLaunchBySecret, J as isControllerClosedError, K as buildAnthropicErrorEvent, Kt as parseJsonOrDiagnose, L as buildAdvisorStream, Mt as createMessages, N as searchWeb, Nt as getTextTokenCount, P as ADVISOR_INTERNAL_TOOL_NAME, Pt as getTokenCount, Q as EFFORT_ORDER, R as injectAdvisorTool, Sn as catalogAdvertises1M, Tn as withOneMSuffixForLead, U as rememberThinkingHistoryRepair, Ut as createChatCompletions, V as resolveAdvisorModel, Vt as resolveMcpToolTimeoutMs, W as repairKnownThinkingHistory, Wt as MAX_RESPONSE_BODY_BYTES, X as readIteratorWithTimeout, Y as logStreamError, Z as relayAnthropicStream, bn as upstreamMaxConnections, cn as BUDGET_SMALL_FAST_SLUG, et as bucketEffort, gn as isBudgetClaudeLead, hn as generateRandomPort, i as assertMcpToolSurfaceConsistent, it as agentToolsEnabled, jt as countTokens, mn as UPSTREAM_INACTIVITY_TIMEOUT_MS, nt as handleMcpDelete, on as toolbeltPathOverride, pn as UPSTREAM_FETCH_TIMEOUT_MS, q as buildOpenAIErrorEvent, qt as normalizeOpenAIUsage, rt as handleMcpPost, tn as provisionTreeSitterAssets, tt as clampEffort, wn as withOneMSuffix, xn as classifyMessagesRoute, yn as upstreamAllowH2, z as isAdvisorRequested, zt as assembleResponsesPayload } from "./peer-mcp-personas-O35uSWsX.js";
+import { $ as UNKNOWN_EFFORT_ANCHOR, An as CHEAP_PROFILE_MODELS, B as resolveAdvisorEffort, Cn as upstreamMaxConnections, D as toolbeltEnabled, Dn as withOneMSuffix, En as oneMContextDisabled, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as withInstallLock, Ft as createMessages, G as repairRejectedThinkingHistory, Gt as createResponses, H as formatThinkingRepairDecline, Ht as assembleResponsesPayload, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, It as getTextTokenCount, J as isControllerClosedError, Jt as readResponseBodyCapped, K as buildAnthropicErrorEvent, Kt as createChatCompletions, L as buildAdvisorStream, Lt as getTokenCount, N as searchWeb, Nt as shimDefaultsToXhigh, On as withOneMSuffixForLead, P as ADVISOR_INTERNAL_TOOL_NAME, Pn as CHEAP_PROFILE_SUBAGENT_CONTEXT_TOKENS, Pt as countTokens, Q as EFFORT_ORDER, R as injectAdvisorTool, Rt as getTokenizerFromModel, Sn as upstreamAllowH2, Tn as catalogAdvertises1M, U as rememberThinkingHistoryRepair, Ut as warnOnTokenPriceDrift, V as resolveAdvisorModel, W as repairKnownThinkingHistory, Wt as resolveMcpToolTimeoutMs, X as readIteratorWithTimeout, Xt as normalizeOpenAIUsage, Y as logStreamError, Yt as parseJsonOrDiagnose, Z as relayAnthropicStream, _n as UPSTREAM_INACTIVITY_TIMEOUT_MS, dn as BUDGET_SMALL_FAST_SLUG, et as bucketEffort, gn as UPSTREAM_FETCH_TIMEOUT_MS, i as assertMcpToolSurfaceConsistent, in as provisionTreeSitterAssets, it as agentToolsEnabled, jn as CHEAP_PROFILE_NATIVE_AGENT_NAMES, kn as CHEAP_PROFILE_ADVISOR_CLIENT_MODEL, ln as toolbeltPathOverride, nt as handleMcpDelete, q as buildOpenAIErrorEvent, qt as MAX_RESPONSE_BODY_BYTES, rt as handleMcpPost, tt as clampEffort, vn as generateRandomPort, wn as classifyMessagesRoute, yn as isBudgetClaudeLead, z as isAdvisorRequested, zt as findLaunchBySecret } from "./peer-mcp-personas-DklYru_1.js";
 import { t as getPackageVersion } from "./version-C8x2hQrZ.js";
 import { i as ensurePaths, t as PATHS } from "./paths-De1Qh9Q9.js";
 import { t as state } from "./state-B8adscCv.js";
@@ -380,6 +380,46 @@ const FAST_PROFILE = Object.freeze({
 	hasCoordinator: false
 });
 /**
+* The `-m cheap1m` roster: the named successor of the original cheap launch.
+* Identical to `cheap` except the Gemini leader keeps its full 1M window
+* (the `-m cheap1m` lead slug is `[1m]`-decorated) and the `astra` peer
+* remains in the persona allowlist next to Oracle. Oracle and Astra run at
+* the 200K default window like every cheap-family role. Hard-denies match
+* fast's: core workers, `orchestrate`, `decide`, `fleet`, and `first-mate`.
+*/
+const CHEAP1M_PROFILE = Object.freeze({
+	id: "cheap1m",
+	nativeRoster: new Set(CHEAP_PROFILE_NATIVE_AGENT_NAMES),
+	personaAllowlist: /* @__PURE__ */ new Set(["oracle", "astra"]),
+	allowedGroups: /* @__PURE__ */ new Set([
+		"peers",
+		"search",
+		"workers",
+		"browser"
+	]),
+	hasCoordinator: false
+});
+/**
+* The `-m cheap` roster: the exact `fast` surface and groups, minus the 1M
+* accounting decoration on the LEAD as well (`-m cheap` selects the BARE
+* `gemini-3.8-flash` slug, so the leader runs at the 200K default window
+* too), plus a cheaper `grok-4.6`/medium Oracle and NO `astra` peer.
+* Hard-denies match fast's: core workers, `orchestrate`, `decide`, `fleet`,
+* and `first-mate`.
+*/
+const CHEAP_PROFILE = Object.freeze({
+	id: "cheap",
+	nativeRoster: new Set(CHEAP_PROFILE_NATIVE_AGENT_NAMES),
+	personaAllowlist: /* @__PURE__ */ new Set(["oracle"]),
+	allowedGroups: /* @__PURE__ */ new Set([
+		"peers",
+		"search",
+		"workers",
+		"browser"
+	]),
+	hasCoordinator: false
+});
+/**
 * The `-m max` profile: Sol/Luna-led, browse-only workers, and explicit
 * cross-lab peer names. The descriptor is a hard projection for bound launch
 * requests; unbound/BYO traffic remains standard because it has no registry
@@ -412,6 +452,8 @@ const MAX_PROFILE = Object.freeze({
 function profileDescriptor(id) {
 	if (id === "fast") return FAST_PROFILE;
 	if (id === "max") return MAX_PROFILE;
+	if (id === "cheap1m") return CHEAP1M_PROFILE;
+	if (id === "cheap") return CHEAP_PROFILE;
 	return STANDARD_PROFILE;
 }
 /**
@@ -429,6 +471,8 @@ function resolveLaunchProfile(modelArg) {
 	const arg = modelArg?.trim().toLowerCase();
 	if (arg === "fast") return "fast";
 	if (arg === "max") return "max";
+	if (arg === "cheap1m") return "cheap1m";
+	if (arg === "cheap") return "cheap";
 	return "standard";
 }
 function validateMaxProfileLaunch(catalog) {
@@ -643,6 +687,104 @@ function validateFastProfilePrerequisites(catalog) {
 function formatFastPrerequisiteFailure(missing) {
 	return "github-router claude -m fast requires the following live-catalog capabilities, which this account's catalog does not fully provide:\n" + missing.map((m) => `  - ${m}`).join("\n") + "\n\nFalling back or silently dropping an agent is not supported for the fast profile's exact roster. Run plain `github-router claude` instead.";
 }
+/** The subagent window floor: every cheap subagent must at least EXCEED the
+*  Claude Code default budget it runs at. Unlike fast, the catalog's real
+*  window is irrelevant to what the client sends (bare slug = 200K either
+*  way); this floor only guarantees the billed model isn't tiny. */
+const CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS = CHEAP_PROFILE_SUBAGENT_CONTEXT_TOKENS;
+/**
+* Shared cheap-family roster prerequisite check, parameterized only by the
+* LEAD's context gate. Both `-m cheap` (200K lead floor) and `-m cheap1m`
+* (1M lead window) validate the EXACT same five-agent roster: the lead gate
+* is the only requirement that differs between the two cheap siblings, and
+* every subagent runs at the 200K default window either way. A non-1M
+* luna/sol/sonnet/grok is acceptable as long as the roster models advertise
+* tool calls, the fixed effort, and a supported endpoint.
+*/
+function collectCheapPrerequisiteMissing(catalog, leadGateTokens, leadGateMessage) {
+	const missing = [];
+	const gemini = findModel(catalog, CHEAP_PROFILE_MODELS.lead);
+	if (!gemini) missing.push(`${CHEAP_PROFILE_MODELS.lead}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(gemini)) missing.push(`${CHEAP_PROFILE_MODELS.lead}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(gemini, leadGateTokens)) missing.push(leadGateMessage(CHEAP_PROFILE_MODELS.lead));
+		if (!supportsEffort(gemini, "high")) missing.push(`${CHEAP_PROFILE_MODELS.lead}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(gemini, "chat")) missing.push(`${CHEAP_PROFILE_MODELS.lead}: does not advertise a supported chat-completions endpoint`);
+	}
+	const luna = findModel(catalog, CHEAP_PROFILE_MODELS.explore);
+	if (!luna) missing.push(`${CHEAP_PROFILE_MODELS.explore}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(luna)) missing.push(`${CHEAP_PROFILE_MODELS.explore}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(luna, CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAP_PROFILE_MODELS.explore}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(luna, "high") || !supportsEffort(luna, "max")) missing.push(`${CHEAP_PROFILE_MODELS.explore}: does not advertise both "high" and "max" reasoning effort`);
+		if (!supportsEndpoint(luna, "responses")) missing.push(`${CHEAP_PROFILE_MODELS.explore}: does not advertise a supported Responses endpoint`);
+	}
+	const sol = findModel(catalog, CHEAP_PROFILE_MODELS.plan);
+	if (!sol) missing.push(`${CHEAP_PROFILE_MODELS.plan}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(sol)) missing.push(`${CHEAP_PROFILE_MODELS.plan}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(sol, CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAP_PROFILE_MODELS.plan}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(sol, "high")) missing.push(`${CHEAP_PROFILE_MODELS.plan}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(sol, "responses")) missing.push(`${CHEAP_PROFILE_MODELS.plan}: does not advertise a supported Responses endpoint`);
+	}
+	const reviewer = findModel(catalog, CHEAP_PROFILE_MODELS.reviewer);
+	if (!reviewer) missing.push(`${CHEAP_PROFILE_MODELS.reviewer}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(reviewer)) missing.push(`${CHEAP_PROFILE_MODELS.reviewer}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(reviewer, CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAP_PROFILE_MODELS.reviewer}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(reviewer, "max")) missing.push(`${CHEAP_PROFILE_MODELS.reviewer}: does not advertise a "max" reasoning effort`);
+		if (!supportsEndpoint(reviewer, "responses")) missing.push(`${CHEAP_PROFILE_MODELS.reviewer}: does not advertise a supported Responses endpoint`);
+	}
+	const grok = findModel(catalog, CHEAP_PROFILE_MODELS.oracle);
+	if (!grok) missing.push(`${CHEAP_PROFILE_MODELS.oracle}: absent from the live catalog`);
+	else {
+		if (!hasContextAtLeast(grok, CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAP_PROFILE_MODELS.oracle}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(grok, "medium")) missing.push(`${CHEAP_PROFILE_MODELS.oracle}: does not advertise a "medium" reasoning effort`);
+		if (!hasUsablePromptMetadata(grok)) missing.push(`${CHEAP_PROFILE_MODELS.oracle}: no usable max_prompt_tokens metadata`);
+		if (!supportsEndpoint(grok, "responses")) missing.push(`${CHEAP_PROFILE_MODELS.oracle}: does not advertise a supported Responses endpoint`);
+	}
+	return missing;
+}
+/**
+* Validate the live Copilot catalog for `-m cheap1m` (the named successor of
+* the original cheap launch): the Gemini leader must still advertise 1M, so
+* the lead keeps its full window while every subagent runs at the 200K
+* default. Roster and capability checks are otherwise identical to `cheap`.
+*/
+function validateCheap1mProfilePrerequisites(catalog) {
+	const missing = collectCheapPrerequisiteMissing(catalog, FAST_REQUIRED_CONTEXT_TOKENS, (id) => `${id}: advertised context window is below 1M (leader window)`);
+	return {
+		ok: missing.length === 0,
+		missing
+	};
+}
+/**
+* Validate the live Copilot catalog for `-m cheap`: the Gemini leader runs
+* at the 200K DEFAULT window (bare slug), so the lead only needs to clear
+* the same 200K floor as every subagent — unlike `cheap1m`, which keeps the
+* 1M leader window. Roster and capability checks are otherwise identical.
+*/
+function validateCheapProfilePrerequisites(catalog) {
+	const missing = collectCheapPrerequisiteMissing(catalog, CHEAP_SUBAGENT_MIN_CONTEXT_TOKENS, (id) => `${id}: advertised context window is below the 200K lead floor`);
+	return {
+		ok: missing.length === 0,
+		missing
+	};
+}
+/**
+* Format `validateCheap1mProfilePrerequisites`'s failure list into the launch
+* error message: every missing/invalid model, plus the rollback command.
+*/
+function formatCheap1mPrerequisiteFailure(missing) {
+	return "github-router claude -m cheap1m requires the following live-catalog capabilities, which this account's catalog does not fully provide:\n" + missing.map((m) => `  - ${m}`).join("\n") + "\n\nFalling back or silently dropping an agent is not supported for the cheap1m profile's exact roster. Run plain `github-router claude` instead.";
+}
+/**
+* Format `validateCheapProfilePrerequisites`'s failure list into the launch
+* error message: every missing/invalid model, plus the rollback command.
+*/
+function formatCheapPrerequisiteFailure(missing) {
+	return "github-router claude -m cheap requires the following live-catalog capabilities, which this account's catalog does not fully provide:\n" + missing.map((m) => `  - ${m}`).join("\n") + "\n\nFalling back or silently dropping an agent is not supported for the cheap profile's exact roster. Run plain `github-router claude` instead.";
+}
 //#endregion
 //#region src/lib/file-log-reporter.ts
 const MAX_LOG_BYTES = 1048576;
@@ -981,6 +1123,32 @@ const MAX_PICKER_MODELS = Object.freeze([
 		behavesAs: "claude-opus-5"
 	}
 ]);
+const CHEAP_PICKER_MODELS = Object.freeze([
+	{
+		id: "gpt-5.6-sol",
+		label: "GPT-5.6 Sol",
+		behavesAs: "claude-opus-5",
+		neverOneM: true
+	},
+	{
+		id: "gpt-5.6-luna",
+		label: "GPT-5.6 Luna",
+		behavesAs: "claude-opus-5",
+		neverOneM: true
+	},
+	{
+		id: "gemini-3.8-flash",
+		label: "Gemini 3.8 Flash",
+		behavesAs: "claude-sonnet-5",
+		neverOneM: true
+	},
+	{
+		id: "grok-4.6",
+		label: "Grok 4.6",
+		behavesAs: "claude-sonnet-5",
+		neverOneM: true
+	}
+]);
 /**
 * Return the ordered, profile-specific `/model` rows supported by the live
 * Copilot catalog. Fast intentionally shares Standard's four-row inventory;
@@ -998,7 +1166,7 @@ function selectableModelsInCatalog(profile) {
 	const catalog = state.models?.data;
 	if (!catalog || catalog.length === 0) return [];
 	const present = new Set(catalog.map((entry) => entry.id));
-	return (profile === "max" ? MAX_PICKER_MODELS : STANDARD_PICKER_MODELS).filter((entry) => present.has(entry.id)).map((entry) => ({
+	return (profile === "max" ? MAX_PICKER_MODELS : profile === "cheap" || profile === "cheap1m" ? CHEAP_PICKER_MODELS : STANDARD_PICKER_MODELS).filter((entry) => present.has(entry.id)).map((entry) => ({
 		model: entry.neverOneM ? entry.id : withOneMSuffix(entry.id),
 		label: entry.label,
 		behavesAs: entry.behavesAs
@@ -1938,7 +2106,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.311";
+var version = "0.3.312";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -4910,10 +5078,16 @@ function preprocessMaxRequest(rawBody, launch, subagentRequest = false) {
 Object.freeze([MAX_LUNA_HIGH_ALIAS_ID, MAX_LUNA_MAX_ALIAS_ID]);
 //#endregion
 //#region src/lib/fast-request-preprocess.ts
+const cheapFamily = (profileId) => profileId === "cheap" || profileId === "cheap1m";
 /**
-* 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.
+* Apply authenticated fast/cheap-profile model and effort policy before
+* ordinary model resolution. Synthetic aliases are refused outside an
+* authenticated fast or cheap launch, so raw/BYO traffic cannot opt itself
+* into private profile semantics. The cheap family shares fast's effort
+* mapping (same model-to-effort rows, just bare subagent slugs at the
+* wiring layer), so this preprocess is shared. Note the reviewer differs:
+* fast reviews on Sonnet 5/xhigh while cheap reviews on Luna/max — both
+* rows exist here, so each profile's reviewer resolves to its fixed effort.
 */
 function preprocessFastRequest(rawBody, launch, subagentRequest = false) {
 	if (launch?.profileId === "max") return preprocessMaxRequest(rawBody, launch, subagentRequest);
@@ -4938,13 +5112,13 @@ function preprocessFastRequest(rawBody, launch, subagentRequest = false) {
 		retiredAlias: originalModel
 	};
 	const alias = resolveModelAlias(originalModel);
-	if (alias && (launch?.profileId !== "fast" || isMaxModelAlias(originalModel))) return {
+	if (alias && (launch?.profileId !== "fast" && !cheapFamily(launch?.profileId) || isMaxModelAlias(originalModel))) return {
 		body: rawBody,
 		originalModel,
 		modified: false,
 		rejectedAlias: originalModel
 	};
-	if (launch?.profileId !== "fast") return {
+	if (launch?.profileId !== "fast" && !cheapFamily(launch?.profileId)) return {
 		body: rawBody,
 		originalModel,
 		modified: false
@@ -5418,15 +5592,19 @@ async function handleCompletion(c) {
 	const advisorRequested = isAdvisorRequested(incomingBeta);
 	const fastProfileRequest = identity.launch?.profileId === "fast";
 	const maxProfileRequest = identity.launch?.profileId === "max";
+	const cheapProfileRequest = identity.launch?.profileId === "cheap" || identity.launch?.profileId === "cheap1m";
 	const subagentRequest = Boolean(c.req.header("x-claude-code-agent-id"));
 	const fastSubagentRequest = fastProfileRequest && subagentRequest;
 	const maxSubagentRequest = maxProfileRequest && subagentRequest;
+	const cheapSubagentRequest = cheapProfileRequest && subagentRequest;
 	const fastLeadAdvisor = fastProfileRequest && !fastSubagentRequest;
 	const maxLeadAdvisor = maxProfileRequest && !maxSubagentRequest;
-	const advisorEnabled = advisorRequested && !fastSubagentRequest && !maxSubagentRequest;
+	const cheapLeadAdvisor = cheapProfileRequest && !cheapSubagentRequest;
+	const advisorEnabled = advisorRequested && !fastSubagentRequest && !maxSubagentRequest && !cheapSubagentRequest;
 	const fastAdvisorEnabled = fastLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
 	const maxAdvisorEnabled = maxLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
-	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : maxLeadAdvisor ? maxAdvisorEnabled : advisorEnabled;
+	const cheapAdvisorEnabled = cheapLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
+	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : cheapLeadAdvisor ? cheapAdvisorEnabled : maxLeadAdvisor ? maxAdvisorEnabled : advisorEnabled;
 	let fastAdvisorChoice;
 	let maxAdvisorChoice;
 	if (maxAdvisorEnabled) {
@@ -5465,9 +5643,33 @@ async function handleCompletion(c) {
 			}, 503);
 		}
 	}
+	if (cheapAdvisorEnabled) {
+		const mismatch = fastAdvisorMetadataMismatch(rawBody);
+		const relaunchProfile = identity.launch?.profileId === "cheap1m" ? "cheap1m" : "cheap";
+		if (mismatch) return c.json({
+			type: "error",
+			error: {
+				type: "invalid_request_error",
+				message: `Cheap Advisor model mismatch: ${mismatch}. Run \`/advisor ${CHEAP_PROFILE_ADVISOR_CLIENT_MODEL}\` to restore the fixed cheap profile, or relaunch with \`github-router claude -m ${relaunchProfile}\`.`
+			}
+		}, 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.retiredAlias || fastPreprocess.rejectedAlias || fastPreprocess.rejectedModel) {
-		const message = fastPreprocess.retiredAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.retiredAlias)} belongs to a retired Fast role and is no longer valid. Relaunch or select a current Fast role.` : 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.`;
+		const profileId = identity.launch?.profileId;
+		const profileLabel = profileId === "cheap" || profileId === "cheap1m" ? "cheap" : "fast";
+		const message = fastPreprocess.retiredAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.retiredAlias)} belongs to a retired Fast role and is no longer valid. Relaunch or select a current Fast role.` : profileId === "max" ? maxRequestError(fastPreprocess) ?? "Invalid max request" : fastPreprocess.rejectedAlias ? `Router-owned model alias ${JSON.stringify(fastPreprocess.rejectedAlias)} is valid only for an authenticated -m ${profileLabel} launch.` : `Model ${JSON.stringify(fastPreprocess.rejectedModel)} is outside the fixed -m ${profileLabel} model set.`;
 		return c.json({
 			type: "error",
 			error: {
@@ -5513,7 +5715,7 @@ async function handleCompletion(c) {
 			parsedBase = JSON.parse(promptWindowBody);
 		} catch {}
 		const wantsStream = parsedBase?.stream === true;
-		if ((fastAdvisorEnabled || maxAdvisorEnabled) && wantsStream) {
+		if ((fastAdvisorEnabled || cheapAdvisorEnabled || maxAdvisorEnabled) && wantsStream) {
 			const initialConversation = Array.isArray(parsedBase.messages) ? parsedBase.messages : [];
 			const parsedInitial = parseAnthropicRequest(parsedBase, modelId, selectedModel);
 			const translatedAdvisorAborter = new AbortController();
@@ -5536,6 +5738,11 @@ async function handleCompletion(c) {
 				effort: maxAdvisorChoice.effort,
 				escalated: false,
 				fastProfile: false
+			} : cheapAdvisorEnabled ? {
+				model: fastAdvisorChoice.model,
+				effort: resolveAdvisorEffort(rawBody, fastAdvisorChoice.model, true),
+				escalated: fastAdvisorChoice.escalated,
+				fastProfile: true
 			} : {
 				model: fastAdvisorChoice.model,
 				effort: resolveAdvisorEffort(rawBody, fastAdvisorChoice.model, true),
@@ -5551,6 +5758,7 @@ async function handleCompletion(c) {
 				advisorEscalated: advisorChoice.escalated,
 				advisorFastProfile: advisorChoice.fastProfile,
 				advisorMaxProfile: maxAdvisorEnabled,
+				advisorCheapProfile: cheapAdvisorEnabled,
 				advisorEffort: advisorChoice.effort,
 				externalAborter: translatedAdvisorAborter,
 				continueTurn: makeShimContinueTurn(endpoint, {
@@ -5673,9 +5881,10 @@ async function handleCompletion(c) {
 				requestHeaders,
 				advisorModel: advisorChoice.model,
 				advisorEscalated: advisorChoice.escalated,
-				advisorFastProfile: fastLeadAdvisor,
+				advisorFastProfile: fastLeadAdvisor || cheapLeadAdvisor,
 				advisorMaxProfile: maxAdvisorEnabled,
-				advisorEffort: maxAdvisorChoice ? maxAdvisorChoice.effort : resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor),
+				advisorCheapProfile: cheapLeadAdvisor,
+				advisorEffort: maxAdvisorChoice ? maxAdvisorChoice.effort : resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor || cheapLeadAdvisor),
 				externalAborter: advisorAborter
 			}), {
 				status: response.status,
@@ -7066,8 +7275,9 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard", pi
 	if (process.env.MCP_TOOL_TIMEOUT === void 0) vars.MCP_TOOL_TIMEOUT = mcpToolTimeoutMs;
 	const isFastProfile = launchProfileId === "fast";
 	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 isCheapProfile = launchProfileId === "cheap" || launchProfileId === "cheap1m";
+	const smallFastModel = isMaxProfile ? MAX_LUNA_HIGH_ALIAS_ID : isFastProfile ? LUNA_HAIKU_ALIAS_ID : isCheapProfile ? 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) && !isCheapProfile ? oneMSuffixForAlias(smallFastModel) : smallFastModel;
 	const seedTierRow = (modelKey, nameKey, bareSlug) => {
 		if (process.env[modelKey] !== void 0) return;
 		vars[modelKey] = withOneMSuffixForLead(bareSlug);
@@ -7078,6 +7288,11 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard", pi
 		vars[modelKey] = oneMSuffixForAlias(aliasId);
 		if (process.env[nameKey] === void 0) vars[nameKey] = displayName;
 	};
+	const seedCheapAliasTierRow = (modelKey, nameKey, aliasId, displayName) => {
+		if (process.env[modelKey] !== void 0) return;
+		vars[modelKey] = aliasId;
+		if (process.env[nameKey] === void 0) vars[nameKey] = displayName;
+	};
 	if (isMaxProfile) {
 		const seedMaxRow = (modelKey, nameKey, id) => {
 			if (process.env[modelKey] !== void 0) return;
@@ -7091,6 +7306,17 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard", pi
 			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 (isCheapProfile) {
+		seedCheapAliasTierRow("ANTHROPIC_DEFAULT_SONNET_MODEL", "ANTHROPIC_DEFAULT_SONNET_MODEL_NAME", LUNA_SONNET_ALIAS_ID, "GPT-5.6 Luna (xhigh)");
+		seedCheapAliasTierRow("ANTHROPIC_DEFAULT_HAIKU_MODEL", "ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME", LUNA_HAIKU_ALIAS_ID, "GPT-5.6 Luna (high)");
+		const cheapAliasCapabilities = "effort,xhigh_effort,max_effort,thinking,adaptive_thinking,interleaved_thinking";
+		if (process.env.ANTHROPIC_DEFAULT_SONNET_MODEL_SUPPORTED_CAPABILITIES === void 0) vars.ANTHROPIC_DEFAULT_SONNET_MODEL_SUPPORTED_CAPABILITIES = cheapAliasCapabilities;
+		if (process.env.ANTHROPIC_DEFAULT_HAIKU_MODEL_SUPPORTED_CAPABILITIES === void 0) vars.ANTHROPIC_DEFAULT_HAIKU_MODEL_SUPPORTED_CAPABILITIES = cheapAliasCapabilities;
+		if (process.env.ANTHROPIC_CUSTOM_MODEL_OPTION === void 0) {
+			vars.ANTHROPIC_CUSTOM_MODEL_OPTION = LUNA_DRIVER_ALIAS_ID;
+			if (process.env.ANTHROPIC_CUSTOM_MODEL_OPTION_NAME === void 0) vars.ANTHROPIC_CUSTOM_MODEL_OPTION_NAME = "GPT-5.6 Luna (max)";
+			if (process.env.ANTHROPIC_CUSTOM_MODEL_OPTION_SUPPORTED_CAPABILITIES === void 0) vars.ANTHROPIC_CUSTOM_MODEL_OPTION_SUPPORTED_CAPABILITIES = cheapAliasCapabilities;
+		}
 	} 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)");
@@ -7197,6 +7423,6 @@ function getCodexEnvVars(serverUrl) {
 	return vars;
 }
 //#endregion
-export { validateMaxProfileLaunch as _, sharedServerArgs as a, updateClaude as b, stopKeepAwake as c, enableFileLogging as d, LUNA_SCOUT_ALIAS_ID as f, validateFastProfilePrerequisites as g, resolveLaunchProfile as h, setupAndServe as i, injectModelPickerSettingsFile as l, profileDescriptor as m, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, formatFastPrerequisiteFailure as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, listModelsForEndpoint as u, runSelfUpdate as v, checkClaudeVersion as y };
+export { checkClaudeVersion as C, runSelfUpdate as S, resolveLaunchProfile as _, sharedServerArgs as a, validateFastProfilePrerequisites as b, stopKeepAwake as c, enableFileLogging as d, LUNA_SCOUT_ALIAS_ID as f, profileDescriptor as g, formatFastPrerequisiteFailure as h, setupAndServe as i, injectModelPickerSettingsFile as l, formatCheapPrerequisiteFailure as m, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, formatCheap1mPrerequisiteFailure as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, listModelsForEndpoint as u, validateCheap1mProfilePrerequisites as v, updateClaude as w, validateMaxProfileLaunch as x, validateCheapProfilePrerequisites as y };
 
-//# sourceMappingURL=server-setup-OTy-qHfv.js.map
+//# sourceMappingURL=server-setup-DX27fdcf.js.map