github-router 0.3.314 → 0.3.316

package/dist/{server-setup-CP8nGop3.js → server-setup-iLG8LrfL.js}

@@ -1,13 +1,13 @@
-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-0_XU6nm9.js";
+import { $ as EFFORT_ORDER, $t as parseJsonOrDiagnose, An as oneMContextDisabled, B as resolveAdvisorEffort, Bn as CHEAP_PROFILE_MODELS, Bt as getTextTokenCount, Cn as isBudgetClaudeLead, D as toolbeltEnabled, Dn as upstreamMaxConnections, En as upstreamAllowH2, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as CHEAPEST_PROFILE_NATIVE_AGENT_NAMES, G as repairKnownThinkingHistory, Gn as withInstallLock, H as resolveCheapestAdvisorModel, Ht as getTokenizerFromModel, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, J as buildOpenAIErrorEvent, Jt as resolveMcpToolTimeoutMs, K as repairRejectedThinkingHistory, Kt as assembleResponsesPayload, L as buildAdvisorStream, Lt as shimDefaultsToXhigh, Mn as withOneMSuffixForLead, N as searchWeb, Nn as CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL, On as classifyMessagesRoute, P as ADVISOR_INTERNAL_TOOL_NAME, Pn as CHEAPEST_PROFILE_MODELS, Q as relayAnthropicStream, Qt as readResponseBodyCapped, R as injectAdvisorTool, Rn as CHEAPEST_PROFILE_SUBAGENT_CONTEXT_TOKENS, Rt as countTokens, Sn as generateRandomPort, U as formatThinkingRepairDecline, Ut as findLaunchBySecret, V as resolveAdvisorModel, Vn as CHEAP_PROFILE_NATIVE_AGENT_NAMES, Vt as getTokenCount, W as rememberThinkingHistoryRepair, Wn as CHEAP_PROFILE_SUBAGENT_CONTEXT_TOKENS, X as logStreamError, Xt as createChatCompletions, Y as isControllerClosedError, Yt as createResponses, Z as readIteratorWithTimeout, Zt as MAX_RESPONSE_BODY_BYTES, at as agentToolsEnabled, bn as UPSTREAM_FETCH_TIMEOUT_MS, cn as provisionTreeSitterAssets, en as normalizeOpenAIUsage, et as UNKNOWN_EFFORT_ANCHOR, hn as BUDGET_SMALL_FAST_SLUG, i as assertMcpToolSurfaceConsistent, it as handleMcpPost, jn as withOneMSuffix, kn as catalogAdvertises1M, nt as clampEffort, pn as toolbeltPathOverride, q as buildAnthropicErrorEvent, qt as warnOnTokenPriceDrift, rt as handleMcpDelete, tt as bucketEffort, xn as UPSTREAM_INACTIVITY_TIMEOUT_MS, z as isAdvisorRequested, zn as CHEAP_PROFILE_ADVISOR_CLIENT_MODEL, zt as createMessages } from "./peer-mcp-personas-CQ-0Piyr.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";
-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-CCg-Ov5q.js";
-import { t as getCopilotUsage } from "./get-copilot-usage-Cjy8TiFf.js";
-import { a as resolveExecutable, n as killManagedTree, o as runCommandCapture, r as parseBoolEnv, s as runCommandVoid } from "./exec-y8C_MU8A.js";
+import { i as ensurePaths, t as PATHS } from "./paths-CMxlQEfQ.js";
+import { t as state } from "./state-BlMpEFAt.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-C1gthnbV.js";
+import { t as getCopilotUsage } from "./get-copilot-usage-BMD_cJNT.js";
+import { c as runCommandVoid, i as parseBoolEnv, o as resolveExecutable, r as killManagedTree, s as runCommandCapture } from "./exec-COUT4QsS.js";
 import { c as FAST_PROFILE_MODELS, l as FAST_PROFILE_NATIVE_AGENT_NAMES, r as FAST_PROFILE_ADVISOR_MODEL } from "./fast-profile-contract-_DhIWank.js";
-import { A as normalizeTrailingOneMSuffix, a as MAX_PROFILE_NATIVE_AGENT_NAMES, d as maxAdvisorPinIsValid, h as maxOpusModel, i as MAX_PROFILE_MODELS, j as stripTrailingOneMSuffix, k as fastEndpointForModel, l as maxAdvisorEffortForModel, n as MAX_PROFILE_ALLOWED_LEAD_MODEL_IDS, t as MAX_PROFILE_ADVISOR_INSTRUCTIONS, u as maxAdvisorModelFromPin, y as validateMaxProfilePrerequisites } from "./max-profile-contract-BcrMlrmo.js";
-import { d as formatAic, f as nanoAiuToCredits, r as extractAndRecordAic, s as recordAic, u as extractCopilotUsage } from "./aic-ledger-Cx_Ubon3.js";
+import { A as normalizeTrailingOneMSuffix, a as MAX_PROFILE_NATIVE_AGENT_NAMES, d as maxAdvisorPinIsValid, h as maxOpusModel, i as MAX_PROFILE_MODELS, j as stripTrailingOneMSuffix, k as fastEndpointForModel, l as maxAdvisorEffortForModel, n as MAX_PROFILE_ALLOWED_LEAD_MODEL_IDS, t as MAX_PROFILE_ADVISOR_INSTRUCTIONS, u as maxAdvisorModelFromPin, y as validateMaxProfilePrerequisites } from "./max-profile-contract-Bl7I_EOC.js";
+import { c as recordAic, d as extractCopilotUsage, f as formatAic, i as extractAndRecordAic, p as nanoAiuToCredits } from "./aic-ledger-BS47yb_C.js";
 import consola from "consola";
 import fs, { existsSync } from "node:fs";
 import path from "node:path";
@@ -421,6 +421,25 @@ const CHEAP_PROFILE = Object.freeze({
 	hasCoordinator: false
 });
 /**
+* The `-m cheapest` roster: the exact cheap surface and groups, but Luna-led
+* (`gpt-5.6-luna`/max at the 200K default window), a Gemini/high Advisor, a
+* Sol/high Oracle, and a Gemini/high reviewer. Oracle-only peer set, no
+* `astra`. Hard-denies match fast's: core workers, `orchestrate`, `decide`,
+* `fleet`, and `first-mate`.
+*/
+const CHEAPEST_PROFILE = Object.freeze({
+	id: "cheapest",
+	nativeRoster: new Set(CHEAPEST_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
@@ -455,6 +474,7 @@ function profileDescriptor(id) {
 	if (id === "max") return MAX_PROFILE;
 	if (id === "cheap1m") return CHEAP1M_PROFILE;
 	if (id === "cheap") return CHEAP_PROFILE;
+	if (id === "cheapest") return CHEAPEST_PROFILE;
 	return STANDARD_PROFILE;
 }
 /**
@@ -474,6 +494,7 @@ function resolveLaunchProfile(modelArg) {
 	if (arg === "max") return "max";
 	if (arg === "cheap1m") return "cheap1m";
 	if (arg === "cheap") return "cheap";
+	if (arg === "cheapest") return "cheapest";
 	return "standard";
 }
 function validateMaxProfileLaunch(catalog) {
@@ -786,6 +807,53 @@ function formatCheap1mPrerequisiteFailure(missing) {
 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) => `  - ${m}`).join("\n") + "\n\nFalling back or silently dropping an agent is not supported for the cheap profile's exact roster. Run plain `github-router claude` instead.";
 }
+/** The subagent window floor: every cheapest role runs at the 200K default. */
+const CHEAPEST_SUBAGENT_MIN_CONTEXT_TOKENS = CHEAPEST_PROFILE_SUBAGENT_CONTEXT_TOKENS;
+/**
+* Validate the live Copilot catalog for `-m cheapest`: Luna lead at the 200K
+* default window, Luna Explore/Plan/GP/implementer roles, Gemini reviewer and
+* Advisor, and a Sol Oracle — all at the 200K default with their fixed
+* efforts and supported endpoints.
+*/
+function validateCheapestProfilePrerequisites(catalog) {
+	const missing = [];
+	const luna = findModel(catalog, CHEAPEST_PROFILE_MODELS.lead);
+	if (!luna) missing.push(`${CHEAPEST_PROFILE_MODELS.lead}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(luna)) missing.push(`${CHEAPEST_PROFILE_MODELS.lead}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(luna, CHEAPEST_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAPEST_PROFILE_MODELS.lead}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(luna, "high") || !supportsEffort(luna, "max") || !supportsEffort(luna, "xhigh")) missing.push(`${CHEAPEST_PROFILE_MODELS.lead}: does not advertise "high", "max", and "xhigh" reasoning effort`);
+		if (!supportsEndpoint(luna, "responses")) missing.push(`${CHEAPEST_PROFILE_MODELS.lead}: does not advertise a supported Responses endpoint`);
+	}
+	const sol = findModel(catalog, CHEAPEST_PROFILE_MODELS.plan);
+	if (!sol) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(sol)) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(sol, CHEAPEST_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(sol, "high")) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(sol, "responses")) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: does not advertise a supported Responses endpoint`);
+		if (!hasUsablePromptMetadata(sol)) missing.push(`${CHEAPEST_PROFILE_MODELS.plan}: no usable max_prompt_tokens metadata (Oracle brief)`);
+	}
+	const gemini = findModel(catalog, CHEAPEST_PROFILE_MODELS.reviewer);
+	if (!gemini) missing.push(`${CHEAPEST_PROFILE_MODELS.reviewer}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(gemini)) missing.push(`${CHEAPEST_PROFILE_MODELS.reviewer}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(gemini, CHEAPEST_SUBAGENT_MIN_CONTEXT_TOKENS)) missing.push(`${CHEAPEST_PROFILE_MODELS.reviewer}: advertised context window is below the 200K subagent floor`);
+		if (!supportsEffort(gemini, "high")) missing.push(`${CHEAPEST_PROFILE_MODELS.reviewer}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(gemini, "chat")) missing.push(`${CHEAPEST_PROFILE_MODELS.reviewer}: does not advertise a supported chat-completions endpoint`);
+	}
+	return {
+		ok: missing.length === 0,
+		missing
+	};
+}
+/**
+* Format `validateCheapestProfilePrerequisites`'s failure list into the
+* launch error message.
+*/
+function formatCheapestPrerequisiteFailure(missing) {
+	return "github-router claude -m cheapest 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 cheapest profile's exact roster. Run plain `github-router claude` instead.";
+}
 //#endregion
 //#region src/lib/file-log-reporter.ts
 const MAX_LOG_BYTES = 1048576;
@@ -1150,6 +1218,26 @@ const CHEAP_PICKER_MODELS = Object.freeze([
 		neverOneM: true
 	}
 ]);
+const CHEAPEST_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
+	}
+]);
 /**
 * Return the ordered, profile-specific `/model` rows supported by the live
 * Copilot catalog. Fast intentionally shares Standard's four-row inventory;
@@ -1167,7 +1255,7 @@ 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 : profile === "cheap" || profile === "cheap1m" ? CHEAP_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 : profile === "cheapest" ? CHEAPEST_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
@@ -2108,7 +2196,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.314";
+var version = "0.3.316";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -5159,16 +5247,17 @@ 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";
+const cheapFamily = (profileId) => profileId === "cheap" || profileId === "cheap1m" || profileId === "cheapest";
 /**
 * 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.
+* authenticated fast, cheap, or cheapest launch, so raw/BYO traffic cannot
+* opt itself into private profile semantics. The cheap family (including
+* cheapest) 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 and cheapest reviews on Gemini/high — all 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);
@@ -5539,6 +5628,23 @@ function fastAdvisorMetadataMismatch(rawBody) {
 		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 cheapestAdvisorMetadataMismatch(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 !== CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL) return `the native Advisor tool requested ${JSON.stringify(model)} instead of ${JSON.stringify(CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL)}`;
+	}
+}
 function stripAdvisorTool(rawBody) {
 	let body;
 	try {
@@ -5674,18 +5780,22 @@ async function handleCompletion(c) {
 	const fastProfileRequest = identity.launch?.profileId === "fast";
 	const maxProfileRequest = identity.launch?.profileId === "max";
 	const cheapProfileRequest = identity.launch?.profileId === "cheap" || identity.launch?.profileId === "cheap1m";
+	const cheapestProfileRequest = identity.launch?.profileId === "cheapest";
 	const subagentRequest = Boolean(c.req.header("x-claude-code-agent-id"));
 	const fastSubagentRequest = fastProfileRequest && subagentRequest;
 	const maxSubagentRequest = maxProfileRequest && subagentRequest;
 	const cheapSubagentRequest = cheapProfileRequest && subagentRequest;
+	const cheapestSubagentRequest = cheapestProfileRequest && subagentRequest;
 	const fastLeadAdvisor = fastProfileRequest && !fastSubagentRequest;
 	const maxLeadAdvisor = maxProfileRequest && !maxSubagentRequest;
 	const cheapLeadAdvisor = cheapProfileRequest && !cheapSubagentRequest;
-	const advisorEnabled = advisorRequested && !fastSubagentRequest && !maxSubagentRequest && !cheapSubagentRequest;
+	const cheapestLeadAdvisor = cheapestProfileRequest && !cheapestSubagentRequest;
+	const advisorEnabled = advisorRequested && !fastSubagentRequest && !maxSubagentRequest && !cheapSubagentRequest && !cheapestSubagentRequest;
 	const fastAdvisorEnabled = fastLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
 	const maxAdvisorEnabled = maxLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
 	const cheapAdvisorEnabled = cheapLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
-	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : cheapLeadAdvisor ? cheapAdvisorEnabled : maxLeadAdvisor ? maxAdvisorEnabled : advisorEnabled;
+	const cheapestAdvisorEnabled = cheapestLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
+	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : cheapLeadAdvisor ? cheapAdvisorEnabled : cheapestLeadAdvisor ? cheapestAdvisorEnabled : maxLeadAdvisor ? maxAdvisorEnabled : advisorEnabled;
 	let fastAdvisorChoice;
 	let maxAdvisorChoice;
 	if (maxAdvisorEnabled) {
@@ -5746,10 +5856,31 @@ async function handleCompletion(c) {
 			}, 503);
 		}
 	}
+	if (cheapestAdvisorEnabled) {
+		const mismatch = cheapestAdvisorMetadataMismatch(rawBody);
+		if (mismatch) return c.json({
+			type: "error",
+			error: {
+				type: "invalid_request_error",
+				message: `Cheapest Advisor model mismatch: ${mismatch}. Run \`/advisor ${CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL}\` to restore the fixed cheapest profile, or relaunch with \`github-router claude -m cheapest\`.`
+			}
+		}, 400);
+		try {
+			fastAdvisorChoice = resolveCheapestAdvisorModel();
+		} 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 profileId = identity.launch?.profileId;
-		const profileLabel = profileId === "cheap" || profileId === "cheap1m" ? "cheap" : "fast";
+		const profileLabel = profileId === "cheap" || profileId === "cheap1m" ? "cheap" : profileId === "cheapest" ? "cheapest" : "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",
@@ -5796,7 +5927,7 @@ async function handleCompletion(c) {
 			parsedBase = JSON.parse(promptWindowBody);
 		} catch {}
 		const wantsStream = parsedBase?.stream === true;
-		if ((fastAdvisorEnabled || cheapAdvisorEnabled || maxAdvisorEnabled) && wantsStream) {
+		if ((fastAdvisorEnabled || cheapAdvisorEnabled || cheapestAdvisorEnabled || maxAdvisorEnabled) && wantsStream) {
 			const initialConversation = Array.isArray(parsedBase.messages) ? parsedBase.messages : [];
 			const parsedInitial = parseAnthropicRequest(parsedBase, modelId, selectedModel);
 			const translatedAdvisorAborter = new AbortController();
@@ -5819,7 +5950,7 @@ async function handleCompletion(c) {
 				effort: maxAdvisorChoice.effort,
 				escalated: false,
 				fastProfile: false
-			} : cheapAdvisorEnabled ? {
+			} : cheapAdvisorEnabled || cheapestAdvisorEnabled ? {
 				model: fastAdvisorChoice.model,
 				effort: resolveAdvisorEffort(rawBody, fastAdvisorChoice.model, true),
 				escalated: fastAdvisorChoice.escalated,
@@ -5839,7 +5970,7 @@ async function handleCompletion(c) {
 				advisorEscalated: advisorChoice.escalated,
 				advisorFastProfile: advisorChoice.fastProfile,
 				advisorMaxProfile: maxAdvisorEnabled,
-				advisorCheapProfile: cheapAdvisorEnabled,
+				advisorCheapProfile: cheapAdvisorEnabled || cheapestAdvisorEnabled,
 				advisorEffort: advisorChoice.effort,
 				externalAborter: translatedAdvisorAborter,
 				continueTurn: makeShimContinueTurn(endpoint, {
@@ -5962,10 +6093,10 @@ async function handleCompletion(c) {
 				requestHeaders,
 				advisorModel: advisorChoice.model,
 				advisorEscalated: advisorChoice.escalated,
-				advisorFastProfile: fastLeadAdvisor || cheapLeadAdvisor,
+				advisorFastProfile: fastLeadAdvisor || cheapLeadAdvisor || cheapestLeadAdvisor,
 				advisorMaxProfile: maxAdvisorEnabled,
-				advisorCheapProfile: cheapLeadAdvisor,
-				advisorEffort: maxAdvisorChoice ? maxAdvisorChoice.effort : resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor || cheapLeadAdvisor),
+				advisorCheapProfile: cheapLeadAdvisor || cheapestLeadAdvisor,
+				advisorEffort: maxAdvisorChoice ? maxAdvisorChoice.effort : resolveAdvisorEffort(rawBody, advisorChoice.model, fastLeadAdvisor || cheapLeadAdvisor || cheapestLeadAdvisor),
 				externalAborter: advisorAborter
 			}), {
 				status: response.status,
@@ -7113,6 +7244,7 @@ async function setupAndServe(options) {
 	state.showToken = options.showToken;
 	state.extendedBetas = options.extendedBetas;
 	state.browseEnabled = options.browseEnabled || process.env.GH_ROUTER_ENABLE_BROWSE === "1";
+	state.searchEnabled = options.searchEnabled || process.env.GH_ROUTER_ENABLE_SEMANTIC_SEARCH === "1";
 	state.fleetEnabled = options.fleetEnabled || process.env.GH_ROUTER_ENABLE_FLEET === "1";
 	state.agentsEnabled = options.agentsEnabled || process.env.GH_ROUTER_ENABLE_AGENTS === "1";
 	state.powerBrowseEnabled = options.powerBrowseEnabled || process.env.GH_ROUTER_ENABLE_POWER_BROWSE === "1";
@@ -7261,6 +7393,11 @@ const sharedServerArgs = {
 		type: "boolean",
 		default: true,
 		description: "Update github-router itself to the latest npm version on launch (throttled once/hour). Best-effort and non-blocking: the proxy serves immediately and a detached updater applies the new version after this process exits (it takes effect on the NEXT launch; the running process keeps its current build). Disable with --no-self-update or GH_ROUTER_NO_SELF_UPDATE=1. Skipped silently if npm/network unavailable."
+	},
+	search: {
+		type: "boolean",
+		default: false,
+		description: "Enable ColBERT/colgrep semantic code search (meaning-ranked `code` tool results over a per-workspace index, with transparent lexical fallback). Off by default; the `code` tool is lexical-only unless this is set. Can also be enabled with GH_ROUTER_ENABLE_SEMANTIC_SEARCH=1. GH_ROUTER_DISABLE_SEMANTIC_SEARCH=1 hard-disables (wins over --search, for tests)."
 	}
 };
 const allowedAccountTypes = /* @__PURE__ */ new Set([
@@ -7302,7 +7439,8 @@ function parseSharedArgs(args) {
 		fleetEnabled: args.fleet,
 		agentsEnabled: args.agents,
 		powerBrowseEnabled: args["power-browse"],
-		humanlikeEnabled: args.humanlike
+		humanlikeEnabled: args.humanlike,
+		searchEnabled: args.search
 	};
 }
 /**
@@ -7374,7 +7512,7 @@ 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 isCheapProfile = launchProfileId === "cheap" || launchProfileId === "cheap1m";
+	const isCheapProfile = launchProfileId === "cheap" || launchProfileId === "cheap1m" || launchProfileId === "cheapest";
 	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) => {
@@ -7522,6 +7660,6 @@ function getCodexEnvVars(serverUrl) {
 	return vars;
 }
 //#endregion
-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 };
+export { validateMaxProfileLaunch as C, updateClaude as E, validateFastProfilePrerequisites as S, checkClaudeVersion as T, profileDescriptor as _, sharedServerArgs as a, validateCheapProfilePrerequisites as b, stopKeepAwake as c, enableFileLogging as d, LUNA_SCOUT_ALIAS_ID as f, formatFastPrerequisiteFailure as g, formatCheapestPrerequisiteFailure 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, resolveLaunchProfile as v, runSelfUpdate as w, validateCheapestProfilePrerequisites as x, validateCheap1mProfilePrerequisites as y };
 
-//# sourceMappingURL=server-setup-CP8nGop3.js.map
+//# sourceMappingURL=server-setup-iLG8LrfL.js.map