github-router 0.3.305 → 0.3.308

package/dist/{server-setup-BP_JM0qb.js → server-setup-C7jSESEI.js}

@@ -1,12 +1,12 @@
-import { $ as UNKNOWN_EFFORT_ANCHOR, An as MAX_LUNA_HIGH_ALIAS_ID, At as countTokens, B as resolveAdvisorEffort, Bn as resolveModelAlias, Bt as resolveMcpToolTimeoutMs, Cn as withOneMSuffix, D as toolbeltEnabled, Dn as LUNA_REAL_MODEL_ID, En as LUNA_HAIKU_ALIAS_ID, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as isRetiredFastModelAlias, Ft as findLaunchBySecret, G as repairRejectedThinkingHistory, Gt as parseJsonOrDiagnose, H as formatThinkingRepairDecline, Ht as createChatCompletions, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, In as maxAliasEffort, J as isControllerClosedError, K as buildAnthropicErrorEvent, Kt as normalizeOpenAIUsage, L as buildAdvisorStream, Ln as maxAliasModel, Mn as canonicalizeAliasModel, Mt as getTextTokenCount, N as searchWeb, Nt as getTokenCount, P as ADVISOR_INTERNAL_TOOL_NAME, Pn as isMaxModelAlias, Pt as getTokenizerFromModel, Q as EFFORT_ORDER, R as injectAdvisorTool, Rt as assembleResponsesPayload, Sn as oneMContextDisabled, Tn as LUNA_DRIVER_ALIAS_ID, U as rememberThinkingHistoryRepair, Un as withInstallLock, Ut as MAX_RESPONSE_BODY_BYTES, V as resolveAdvisorModel, Vt as createResponses, W as repairKnownThinkingHistory, Wt as readResponseBodyCapped, X as readIteratorWithTimeout, Y as logStreamError, Z as relayAnthropicStream, an as toolbeltPathOverride, bn as classifyMessagesRoute, en as provisionTreeSitterAssets, et as bucketEffort, fn as UPSTREAM_FETCH_TIMEOUT_MS, hn as isBudgetClaudeLead, i as assertMcpToolSurfaceConsistent, it as agentToolsEnabled, jn as MAX_LUNA_MAX_ALIAS_ID, jt as createMessages, kn as LUNA_SONNET_ALIAS_ID, kt as shimDefaultsToXhigh, mn as generateRandomPort, nt as handleMcpDelete, pn as UPSTREAM_INACTIVITY_TIMEOUT_MS, q as buildOpenAIErrorEvent, rt as handleMcpPost, sn as BUDGET_SMALL_FAST_SLUG, tt as clampEffort, vn as upstreamAllowH2, wn as withOneMSuffixForLead, xn as catalogAdvertises1M, yn as upstreamMaxConnections, z as isAdvisorRequested, zt as warnOnTokenPriceDrift } from "./peer-mcp-personas-C9fDtu69.js";
+import { $ as UNKNOWN_EFFORT_ANCHOR, At as countTokens, B as resolveAdvisorEffort, Bt as resolveMcpToolTimeoutMs, Cn as withOneMSuffix, D as toolbeltEnabled, F as ADVISOR_TOOL_INSTRUCTIONS, Ft as findLaunchBySecret, G as repairRejectedThinkingHistory, Gt as parseJsonOrDiagnose, H as formatThinkingRepairDecline, Ht as createChatCompletions, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, J as isControllerClosedError, K as buildAnthropicErrorEvent, Kt as normalizeOpenAIUsage, L as buildAdvisorStream, Mt as getTextTokenCount, N as searchWeb, Nt as getTokenCount, P as ADVISOR_INTERNAL_TOOL_NAME, Pt as getTokenizerFromModel, Q as EFFORT_ORDER, R as injectAdvisorTool, Rt as assembleResponsesPayload, Sn as oneMContextDisabled, Tn as withInstallLock, U as rememberThinkingHistoryRepair, Ut as MAX_RESPONSE_BODY_BYTES, V as resolveAdvisorModel, Vt as createResponses, W as repairKnownThinkingHistory, Wt as readResponseBodyCapped, X as readIteratorWithTimeout, Y as logStreamError, Z as relayAnthropicStream, an as toolbeltPathOverride, bn as classifyMessagesRoute, en as provisionTreeSitterAssets, et as bucketEffort, fn as UPSTREAM_FETCH_TIMEOUT_MS, hn as isBudgetClaudeLead, i as assertMcpToolSurfaceConsistent, it as agentToolsEnabled, jt as createMessages, kt as shimDefaultsToXhigh, mn as generateRandomPort, nt as handleMcpDelete, pn as UPSTREAM_INACTIVITY_TIMEOUT_MS, q as buildOpenAIErrorEvent, rt as handleMcpPost, sn as BUDGET_SMALL_FAST_SLUG, tt as clampEffort, vn as upstreamAllowH2, wn as withOneMSuffixForLead, xn as catalogAdvertises1M, yn as upstreamMaxConnections, z as isAdvisorRequested, zt as warnOnTokenPriceDrift } from "./peer-mcp-personas-DAd076hn.js";
 import { t as getPackageVersion } from "./version-C8x2hQrZ.js";
-import { i as ensurePaths, t as PATHS } from "./paths-IcZW8gZx.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-CbuFyx-B.js";
-import { t as getCopilotUsage } from "./get-copilot-usage-ByjQqhdi.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 { r as FAST_PROFILE_ADVISOR_MODEL } from "./fast-profile-contract-OuZQaGif.js";
-import { 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, x as stripTrailingOneMSuffix } from "./max-profile-contract-CiD5wyET.js";
+import { a as FAST_PROFILE_MODELS, o as FAST_PROFILE_NATIVE_AGENT_NAMES, r as FAST_PROFILE_ADVISOR_MODEL } from "./fast-profile-contract-BWCJy6iL.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-v2Q6eN5w.js";
 import consola from "consola";
 import fs, { existsSync } from "node:fs";
 import path from "node:path";
@@ -355,6 +355,295 @@ async function runSelfUpdate(opts) {
 	}
 }
 //#endregion
+//#region src/lib/launch-profile.ts
+const STANDARD_PROFILE = Object.freeze({
+	id: "standard",
+	hasCoordinator: true
+});
+/**
+* The `-m fast` roster: exactly five native agents (`Explore`, `Plan`,
+* `general-purpose`, `implementer`, `reviewer`), the fast-only `oracle` peer
+* tool, no coordinator, and `peers`/`search` plus optional browser and
+* browse-only worker groups. Core workers, `orchestrate`, `decide`, `fleet`,
+* and `first-mate` remain hard denies even when their standard gates pass.
+*/
+const FAST_PROFILE = Object.freeze({
+	id: "fast",
+	nativeRoster: new Set(FAST_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
+* entry carrying this allow-list.
+*/
+const MAX_PROFILE = Object.freeze({
+	id: "max",
+	nativeRoster: new Set(MAX_PROFILE_NATIVE_AGENT_NAMES),
+	personaAllowlist: /* @__PURE__ */ new Set([
+		"sol_critic",
+		"codex_reviewer",
+		"sonnet_reviewer",
+		"opus_critic",
+		"gemini_critic",
+		"gemini_reviewer",
+		"grok_critic",
+		"grok_reviewer"
+	]),
+	allowedGroups: /* @__PURE__ */ new Set([
+		"peers",
+		"search",
+		"workers",
+		"browser",
+		"decide",
+		"fleet",
+		"first-mate"
+	]),
+	hasCoordinator: true
+});
+function profileDescriptor(id) {
+	if (id === "fast") return FAST_PROFILE;
+	if (id === "max") return MAX_PROFILE;
+	return STANDARD_PROFILE;
+}
+/**
+* Resolve the parsed `-m` argument to a launch profile.
+*
+* Deliberately keyed on the RAW alias string (trimmed, case-insensitive
+* `"fast"`), never on a resolved model id: `resolveLeadSlugArg` maps `fast`
+* to `FAST_LEAD_MODEL` (`./port`) before this is of any use to a caller who
+* only has the resolved id, so callers that already resolved the lead must
+* pass the ORIGINAL `-m` value here, not the resolved one. This is what
+* keeps `-m gpt-5.6-luna` (a direct pin of the same underlying model) a
+* standard-surface launch — only the literal alias narrows the surface.
+*/
+function resolveLaunchProfile(modelArg) {
+	const arg = modelArg?.trim().toLowerCase();
+	if (arg === "fast") return "fast";
+	if (arg === "max") return "max";
+	return "standard";
+}
+function validateMaxProfileLaunch(catalog) {
+	return validateMaxProfilePrerequisites(catalog);
+}
+/**
+* Router-owned alias id for the fast profile's Sonnet-tier row
+* (`ANTHROPIC_DEFAULT_SONNET_MODEL`). Never sent upstream — canonicalized to
+* `LUNA_REAL_MODEL_ID` by `canonicalizeAliasModel` before the request
+* reaches Copilot.
+*/
+const LUNA_DRIVER_ALIAS_ID = "gh-router-luna-driver-max";
+/** Fast Explore alias preserves its high-effort provenance until the
+* authenticated request boundary. Bare Luna remains the max-effort lead and
+* `general-purpose` model. */
+const LUNA_SCOUT_ALIAS_ID = "gh-router-luna-scout-high";
+/** Retired Fast aliases remain recognizable only so old running clients fail
+* with an explicit stale-alias error instead of silently changing semantics. */
+const LUNA_IMPLEMENTER_ALIAS_ID = "gh-router-luna-implementer-max";
+const FAST_CRITIC_ALIAS_ID = "gh-router-fast-critic-medium";
+const LUNA_SONNET_ALIAS_ID = "gh-router-luna-sonnet-xhigh";
+/**
+* Router-owned alias id for the fast profile's Haiku-tier row
+* (`ANTHROPIC_DEFAULT_HAIKU_MODEL` / `ANTHROPIC_SMALL_FAST_MODEL`).
+*/
+const LUNA_HAIKU_ALIAS_ID = "gh-router-luna-haiku-high";
+/** Max-profile aliases are intentionally separate from the fast aliases. They
+* preserve the fixed max role effort until the authenticated max request
+* boundary and are never valid on standard or fast launches. */
+const MAX_LUNA_HIGH_ALIAS_ID = "gh-router-max-luna-high";
+const MAX_LUNA_MAX_ALIAS_ID = "gh-router-max-luna-max";
+/** The real Copilot catalog id every Luna alias (including the driver
+*  itself) canonicalizes to. */
+const LUNA_REAL_MODEL_ID = FAST_PROFILE_MODELS.luna;
+const MAX_ALIAS_IDS = /* @__PURE__ */ new Set([MAX_LUNA_HIGH_ALIAS_ID, MAX_LUNA_MAX_ALIAS_ID]);
+function isMaxModelAlias(id) {
+	return MAX_ALIAS_IDS.has(stripTrailingOneMSuffix(id).base);
+}
+function maxAliasEffort(id) {
+	const base = stripTrailingOneMSuffix(id).base;
+	if (base === "gh-router-max-luna-high") return "high";
+	if (base === "gh-router-max-luna-max") return "max";
+}
+function maxAliasModel(id) {
+	return isMaxModelAlias(id) ? LUNA_REAL_MODEL_ID : void 0;
+}
+/**
+* The full alias table, keyed by `aliasId`. A simpler model-id-only table is
+* rejected by design: the driver, the Sonnet tier, and the Haiku tier all
+* resolve to the SAME Luna catalog id, while the fast critic alias resolves
+* to Gemini. After early canonicalization a table keyed on the real id could
+* no longer tell which absent-effort default applies. Alias provenance is the
+* minimum discriminator that survives from tier selection through to request
+* preprocessing, which is why canonicalization must happen LAST (in the
+* `/v1/messages` identity preflight), after the effort default has already
+* been read off the alias.
+*/
+const RETIRED_FAST_ALIAS_IDS = /* @__PURE__ */ new Set([LUNA_IMPLEMENTER_ALIAS_ID, FAST_CRITIC_ALIAS_ID]);
+const MODEL_ALIAS_TABLE = /* @__PURE__ */ new Map([
+	[LUNA_DRIVER_ALIAS_ID, {
+		aliasId: LUNA_DRIVER_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "max"
+	}],
+	[MAX_LUNA_HIGH_ALIAS_ID, {
+		aliasId: MAX_LUNA_HIGH_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "high"
+	}],
+	[MAX_LUNA_MAX_ALIAS_ID, {
+		aliasId: MAX_LUNA_MAX_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "max"
+	}],
+	[LUNA_SCOUT_ALIAS_ID, {
+		aliasId: LUNA_SCOUT_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "high"
+	}],
+	[LUNA_SONNET_ALIAS_ID, {
+		aliasId: LUNA_SONNET_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "xhigh"
+	}],
+	[LUNA_HAIKU_ALIAS_ID, {
+		aliasId: LUNA_HAIKU_ALIAS_ID,
+		realModel: LUNA_REAL_MODEL_ID,
+		absentEffortDefault: "high"
+	}]
+]);
+/**
+* Look up the alias descriptor for a wire-facing model id (with or without
+* a trailing `[1m]` bracket — the bracket is stripped before the table
+* lookup and is orthogonal to alias identity). Returns undefined for any
+* id that isn't one of the registered aliases (including bare `claude-*`
+* ids and every other real Copilot catalog id).
+*/
+function resolveModelAlias(id) {
+	const { base } = stripTrailingOneMSuffix(id);
+	return MODEL_ALIAS_TABLE.get(base);
+}
+function isRetiredFastModelAlias(id) {
+	return RETIRED_FAST_ALIAS_IDS.has(stripTrailingOneMSuffix(id).base);
+}
+/**
+* Strip alias provenance and return the real catalog id to send upstream.
+* Idempotent passthrough for any id that isn't a registered alias (a bare
+* `claude-*` slug, an already-real Copilot id, or anything else) — this is
+* safe to call unconditionally on every `body.model` at the outbound
+* boundary. Preserves a trailing `[1m]` bracket: canonicalization only
+* erases ALIAS identity, not the 1M-context accounting decoration.
+*/
+function canonicalizeAliasModel(id) {
+	const { base, hadSuffix } = stripTrailingOneMSuffix(id);
+	const alias = MODEL_ALIAS_TABLE.get(base);
+	if (!alias) return normalizeTrailingOneMSuffix(id);
+	return hadSuffix ? `${alias.realModel}[1m]` : alias.realModel;
+}
+const FAST_REQUIRED_CONTEXT_TOKENS = 1e6;
+function findModel(catalog, id) {
+	return catalog?.data?.find((m) => m.id === id);
+}
+function hasToolCalls(model) {
+	return model?.capabilities?.supports?.tool_calls === true;
+}
+function hasContextAtLeast(model, tokens) {
+	return (model?.capabilities?.limits?.max_context_window_tokens ?? 0) >= tokens;
+}
+function supportsEffort(model, effort) {
+	const list = model?.capabilities?.supports?.reasoning_effort;
+	return Array.isArray(list) && list.includes(effort);
+}
+function supportsEndpoint(model, endpoint) {
+	return model !== void 0 && fastEndpointForModel(model) === endpoint;
+}
+function hasUsablePromptMetadata(model) {
+	const prompt = model?.capabilities?.limits?.max_prompt_tokens;
+	return typeof prompt === "number" && Number.isFinite(prompt) && prompt > 0;
+}
+/**
+* Validate the live Copilot catalog carries every model the fast profile's
+* EXACT roster depends on, with the specific capabilities each assignment
+* needs. These are capability-availability PREREQUISITES for constructing
+* the roster — not an allowlist of models the user may select later in the
+* session — so a partial catalog fails the whole `-m fast` launch rather
+* than silently substituting or dropping an agent.
+*
+* Checks, per the fast-launch-profile design:
+*   - Luna lead/Explore/general-purpose: tool calls, >=1M, high+max, Responses.
+*   - Sol Plan: tool calls, >=1M, high, Responses.
+*   - Grok reviewer: tool calls, medium, Responses, usable prompt metadata.
+*   - Gemini implementer/Advisor: tool calls, >=1M, high, chat-completions.
+*   - Opus Oracle: exact Opus 5, >=1M, adaptive/high, Messages, prompt metadata.
+*
+* Pure over the passed-in catalog snapshot so it's unit-testable without
+* `state` — callers pass `state.models` at call time.
+*/
+function validateFastProfilePrerequisites(catalog) {
+	const missing = [];
+	const luna = findModel(catalog, LUNA_REAL_MODEL_ID);
+	if (!luna) missing.push(`${LUNA_REAL_MODEL_ID}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(luna)) missing.push(`${LUNA_REAL_MODEL_ID}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(luna, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push(`${LUNA_REAL_MODEL_ID}: advertised context window is below 1M`);
+		if (!supportsEffort(luna, "high") || !supportsEffort(luna, "max")) missing.push(`${LUNA_REAL_MODEL_ID}: does not advertise both "high" and "max" reasoning effort`);
+		if (!supportsEndpoint(luna, "responses")) missing.push(`${LUNA_REAL_MODEL_ID}: does not advertise a supported Responses endpoint`);
+	}
+	const sol = findModel(catalog, FAST_PROFILE_MODELS.plan);
+	if (!sol) missing.push(`${FAST_PROFILE_MODELS.plan}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(sol)) missing.push(`${FAST_PROFILE_MODELS.plan}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(sol, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push(`${FAST_PROFILE_MODELS.plan}: advertised context window is below 1M`);
+		if (!supportsEffort(sol, "high")) missing.push(`${FAST_PROFILE_MODELS.plan}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(sol, "responses")) missing.push(`${FAST_PROFILE_MODELS.plan}: does not advertise a supported Responses endpoint`);
+	}
+	const sonnet = findModel(catalog, FAST_PROFILE_MODELS.reviewer);
+	if (!sonnet) missing.push(`${FAST_PROFILE_MODELS.reviewer}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(sonnet)) missing.push(`${FAST_PROFILE_MODELS.reviewer}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(sonnet, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push(`${FAST_PROFILE_MODELS.reviewer}: advertised context window is below 1M`);
+		if (!supportsEffort(sonnet, "xhigh")) missing.push(`${FAST_PROFILE_MODELS.reviewer}: does not advertise an "xhigh" reasoning effort`);
+		if (sonnet.capabilities?.supports?.adaptive_thinking !== true) missing.push(`${FAST_PROFILE_MODELS.reviewer}: does not advertise adaptive_thinking`);
+		if (!hasUsablePromptMetadata(sonnet)) missing.push(`${FAST_PROFILE_MODELS.reviewer}: no usable max_prompt_tokens metadata`);
+		if (!supportsEndpoint(sonnet, "messages")) missing.push(`${FAST_PROFILE_MODELS.reviewer}: does not advertise a supported Messages endpoint`);
+	}
+	const gemini = findModel(catalog, FAST_PROFILE_MODELS.implementer);
+	if (!gemini) missing.push(`${FAST_PROFILE_MODELS.implementer}: absent from the live catalog`);
+	else {
+		if (!hasToolCalls(gemini)) missing.push(`${FAST_PROFILE_MODELS.implementer}: does not advertise tool_calls`);
+		if (!hasContextAtLeast(gemini, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push(`${FAST_PROFILE_MODELS.implementer}: advertised context window is below 1M`);
+		if (!supportsEffort(gemini, "high")) missing.push(`${FAST_PROFILE_MODELS.implementer}: does not advertise a "high" reasoning effort`);
+		if (!supportsEndpoint(gemini, "chat")) missing.push(`${FAST_PROFILE_MODELS.implementer}: does not advertise a supported chat-completions endpoint`);
+	}
+	const opus = findModel(catalog, FAST_PROFILE_MODELS.oracle);
+	if (!opus) missing.push(`${FAST_PROFILE_MODELS.oracle}: absent from the live catalog`);
+	else {
+		if (!hasContextAtLeast(opus, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push(`${FAST_PROFILE_MODELS.oracle}: advertised context window is below 1M`);
+		if (!supportsEffort(opus, "high")) missing.push(`${FAST_PROFILE_MODELS.oracle}: does not advertise a "high" reasoning effort`);
+		if (opus.capabilities?.supports?.adaptive_thinking !== true) missing.push(`${FAST_PROFILE_MODELS.oracle}: does not advertise adaptive_thinking`);
+		if (!hasUsablePromptMetadata(opus)) missing.push(`${FAST_PROFILE_MODELS.oracle}: no usable max_prompt_tokens metadata`);
+		if (!supportsEndpoint(opus, "messages")) missing.push(`${FAST_PROFILE_MODELS.oracle}: does not advertise a supported Messages endpoint`);
+	}
+	return {
+		ok: missing.length === 0,
+		missing
+	};
+}
+/**
+* Format `validateFastProfilePrerequisites`'s failure list into the launch
+* error message: every missing/invalid model, plus the rollback command.
+*/
+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.";
+}
+//#endregion
 //#region src/lib/file-log-reporter.ts
 const MAX_LOG_BYTES = 1048576;
 const DEDUP_MAX = 1e3;
@@ -1649,7 +1938,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.305";
+var version = "0.3.308";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -4574,14 +4863,14 @@ function preprocessMaxRequest(rawBody, launch, subagentRequest = false) {
 			modified: true
 		};
 	}
-	const allowedSubagentModel = subagentRequest && (base === "grok-4.6" || base === "gpt-5.6-luna");
+	const allowedSubagentModel = subagentRequest && (base === "grok-4.6" || base === "gpt-5.6-luna" || base === "claude-sonnet-5");
 	if (!allowedModel(base) && !allowedSubagentModel) return {
 		body: rawBody,
 		originalModel,
 		modified: false,
 		rejectedModel: originalModel
 	};
-	const effort = allowedSubagentModel ? base === "gpt-5.6-luna" ? "max" : "high" : MAX_MODEL_EFFORTS[base];
+	const effort = allowedSubagentModel ? base === "claude-sonnet-5" ? "xhigh" : base === "gpt-5.6-luna" ? "max" : "high" : MAX_MODEL_EFFORTS[base];
 	if (!effort) return {
 		body: rawBody,
 		originalModel,
@@ -4649,6 +4938,7 @@ function preprocessFastRequest(rawBody, launch, subagentRequest = false) {
 	else if (bare === "gpt-5.6-sol") effort = "high";
 	else if (bare === "grok-4.6") effort = "medium";
 	else if (bare === "gemini-3.8-flash") effort = "high";
+	else if (bare === "claude-sonnet-5") effort = "xhigh";
 	else if (bare === "claude-opus-5") effort = "high";
 	if (!effort && !alias) return {
 		body: rawBody,
@@ -6887,6 +7177,6 @@ function getCodexEnvVars(serverUrl) {
 	return vars;
 }
 //#endregion
-export { sharedServerArgs as a, stopKeepAwake as c, enableFileLogging as d, runSelfUpdate as f, setupAndServe as i, injectModelPickerSettingsFile as l, updateClaude as m, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, checkClaudeVersion as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, listModelsForEndpoint as u };
+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 };
 
-//# sourceMappingURL=server-setup-BP_JM0qb.js.map
+//# sourceMappingURL=server-setup-C7jSESEI.js.map