github-router 0.3.293 → 0.3.297

package/dist/{server-setup-DbvbW5Ve.js → server-setup-DhG3YZPj.js}

@@ -1,9 +1,10 @@
-import { $ as UNKNOWN_EFFORT_ANCHOR, $t as provisionTreeSitterAssets, At as shimDefaultsToXhigh, B as resolveAdvisorEffort, Bt as createResponses, Cn as withOneMSuffix, E as toolbeltEnabled, F as FAST_ADVISOR_TOOL_INSTRUCTIONS, G as repairRejectedThinkingHistory, Gt as normalizeOpenAIUsage, H as formatThinkingRepairDecline, Ht as MAX_RESPONSE_BODY_BYTES, I as buildAdvisorStream, J as isControllerClosedError, K as buildAnthropicErrorEvent, L as injectAdvisorTool, Lt as assembleResponsesPayload, M as searchWeb, Mt as createMessages, N as ADVISOR_INTERNAL_TOOL_NAME, Nt as getTokenCount, P as ADVISOR_TOOL_INSTRUCTIONS, Pt as findLaunchBySecret, Q as EFFORT_ORDER, R as isAdvisorRequested, Rt as warnOnTokenPriceDrift, Sn as oneMContextDisabled, Tn as withInstallLock, U as rememberThinkingHistoryRepair, Ut as readResponseBodyCapped, V as resolveAdvisorModel, Vt as createChatCompletions, W as repairKnownThinkingHistory, Wt as parseJsonOrDiagnose, X as readIteratorWithTimeout, Y as logStreamError, Z as relayAnthropicStream, _n as upstreamAllowH2, bn as pickEndpoint, dn as UPSTREAM_FETCH_TIMEOUT_MS, et as bucketEffort, fn as UPSTREAM_INACTIVITY_TIMEOUT_MS, i as assertMcpToolSurfaceConsistent, in as toolbeltPathOverride, jt as countTokens, mn as isBudgetClaudeLead, nt as handleMcpDelete, on as BUDGET_SMALL_FAST_SLUG, pn as generateRandomPort, q as buildOpenAIErrorEvent, rt as handleMcpPost, st as agentToolsEnabled, tt as clampEffort, vn as upstreamMaxConnections, wn as withOneMSuffixForLead, xn as catalogAdvertises1M, yn as classifyMessagesRoute, z as isFastProfileLead, zt as resolveMcpToolTimeoutMs } from "./peer-mcp-personas-DhI7ZPSx.js";
+import { $ as bucketEffort, An as generateRandomPort, B as resolveAdvisorModel, Bn as withOneMSuffixForLead, Cn as BUDGET_SMALL_FAST_SLUG, E as toolbeltEnabled, F as FAST_ADVISOR_TOOL_INSTRUCTIONS, Fn as upstreamMaxConnections, G as buildAnthropicErrorEvent, Gt as shimDefaultsToXhigh, H as rememberThinkingHistoryRepair, Hn as withInstallLock, I as buildAdvisorStream, In as classifyMessagesRoute, It as LUNA_REAL_MODEL_ID, J as logStreamError, Jt as getTextTokenCount, K as buildOpenAIErrorEvent, Kt as countTokens, L as injectAdvisorTool, Ln as catalogAdvertises1M, M as searchWeb, N as ADVISOR_INTERNAL_TOOL_NAME, Nt as LUNA_DRIVER_ALIAS_ID, On as UPSTREAM_FETCH_TIMEOUT_MS, P as ADVISOR_TOOL_INSTRUCTIONS, Pn as upstreamAllowH2, Pt as LUNA_HAIKU_ALIAS_ID, Q as UNKNOWN_EFFORT_ANCHOR, R as isAdvisorRequested, Rn as oneMContextDisabled, Rt as LUNA_SONNET_ALIAS_ID, U as repairKnownThinkingHistory, Ut as resolveModelAlias, V as formatThinkingRepairDecline, Vn as stripTrailingOneMSuffix, W as repairRejectedThinkingHistory, X as relayAnthropicStream, Xt as getTokenizerFromModel, Y as readIteratorWithTimeout, Yt as getTokenCount, Z as EFFORT_ORDER, Zt as findLaunchBySecret, an as MAX_RESPONSE_BODY_BYTES, cn as normalizeOpenAIUsage, ct as agentToolsEnabled, en as assembleResponsesPayload, et as clampEffort, gn as provisionTreeSitterAssets, i as assertMcpToolSurfaceConsistent, in as createChatCompletions, jn as isBudgetClaudeLead, kn as UPSTREAM_INACTIVITY_TIMEOUT_MS, nn as resolveMcpToolTimeoutMs, nt as handleMcpPost, on as readResponseBodyCapped, q as isControllerClosedError, qt as createMessages, rn as createResponses, sn as parseJsonOrDiagnose, tn as warnOnTokenPriceDrift, tt as handleMcpDelete, xn as toolbeltPathOverride, z as resolveAdvisorEffort, zn as withOneMSuffix, zt as canonicalizeAliasModel } from "./peer-mcp-personas-ce6_YMnX.js";
 import { t as getPackageVersion } from "./version-C8x2hQrZ.js";
-import { i as ensurePaths, t as PATHS } from "./paths-BH4J7slC.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-8drORhXg.js";
-import { t as getCopilotUsage } from "./get-copilot-usage-CRrf1ZSC.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 { 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 consola from "consola";
 import * as fs$2 from "node:fs";
 import fs, { existsSync } from "node:fs";
@@ -354,232 +355,6 @@ 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 four native agents (`scout`, `implementer`,
-* `reviewer`, `planner`), the fast-only `oracle` peer tool, no coordinator,
-* and only `peers`/`search` plus the ordinary opt-in `browser` group.
-* `workers`/`orchestrate`/`decide`/`fleet`/`first-mate` are hard denies even
-* when their independent standard-profile gates pass.
-*/
-const FAST_PROFILE = Object.freeze({
-	id: "fast",
-	nativeRoster: /* @__PURE__ */ new Set([
-		"scout",
-		"implementer",
-		"reviewer",
-		"planner"
-	]),
-	personaAllowlist: /* @__PURE__ */ new Set(["oracle"]),
-	allowedGroups: /* @__PURE__ */ new Set([
-		"peers",
-		"search",
-		"browser"
-	]),
-	hasCoordinator: false
-});
-function profileDescriptor(id) {
-	return id === "fast" ? FAST_PROFILE : 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) {
-	return modelArg?.trim().toLowerCase() === "fast" ? "fast" : "standard";
-}
-/**
-* 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 native-agent alias ids preserve role-specific effort provenance until
-*  the authenticated request boundary. They both canonicalize to Luna, but the
-*  scout is fixed high while the implementer is fixed max. */
-const LUNA_SCOUT_ALIAS_ID = "gh-router-luna-scout-high";
-const LUNA_IMPLEMENTER_ALIAS_ID = "gh-router-luna-implementer-max";
-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";
-/** The real Copilot catalog id every Luna alias (including the driver
-*  itself) canonicalizes to. */
-const LUNA_REAL_MODEL_ID = "gpt-5.6-luna";
-/**
-* 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, so after early canonicalization a
-* table keyed on the real id could no longer tell which absent-effort
-* default applies. Alias provenance — which of the three ids the request
-* actually carried — 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 MODEL_ALIAS_TABLE = /* @__PURE__ */ new Map([
-	[LUNA_DRIVER_ALIAS_ID, {
-		aliasId: LUNA_DRIVER_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_IMPLEMENTER_ALIAS_ID, {
-		aliasId: LUNA_IMPLEMENTER_ALIAS_ID,
-		realModel: LUNA_REAL_MODEL_ID,
-		absentEffortDefault: "max"
-	}],
-	[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 three registered aliases (including the bare
-* `claude-*` ids and every other real Copilot catalog id).
-*/
-function resolveModelAlias(id) {
-	const bare = id.replace(/\[1m\]$/i, "");
-	return MODEL_ALIAS_TABLE.get(bare);
-}
-/**
-* 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 bracket = /\[1m\]$/i.test(id) ? "[1m]" : "";
-	const bare = bracket ? id.slice(0, -bracket.length) : id;
-	const alias = MODEL_ALIAS_TABLE.get(bare);
-	return alias ? `${alias.realModel}${bracket}` : id;
-}
-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, paths) {
-	const endpoints = model?.supported_endpoints;
-	return Array.isArray(endpoints) && endpoints.some((endpoint) => paths.has(endpoint));
-}
-const RESPONSES_ENDPOINTS = /* @__PURE__ */ new Set(["/responses", "/v1/responses"]);
-const MESSAGES_ENDPOINTS = /* @__PURE__ */ new Set(["/messages", "/v1/messages"]);
-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/scout/implementer: tool calls, >=1M, high+max, Responses.
-*   - Sol planner: tool calls, >=1M, high, Responses.
-*   - Grok reviewer: tool calls, medium, Responses, usable prompt metadata.
-*   - Gemini Advisor: >=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(`${LUNA_REAL_MODEL_ID}: does not advertise both "high" and "max" reasoning effort`);
-		if (!supportsEndpoint(luna, RESPONSES_ENDPOINTS)) missing.push(`${LUNA_REAL_MODEL_ID}: does not advertise a supported Responses endpoint`);
-	}
-	const sol = findModel(catalog, "gpt-5.6-sol");
-	if (!sol) missing.push("gpt-5.6-sol: absent from the live catalog");
-	else {
-		if (!hasToolCalls(sol)) missing.push("gpt-5.6-sol: does not advertise tool_calls");
-		if (!hasContextAtLeast(sol, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push("gpt-5.6-sol: advertised context window is below 1M");
-		if (!supportsEffort(sol, "high")) missing.push("gpt-5.6-sol: does not advertise a \"high\" reasoning effort");
-		if (!supportsEndpoint(sol, RESPONSES_ENDPOINTS)) missing.push("gpt-5.6-sol: does not advertise a supported Responses endpoint");
-	}
-	const grok = findModel(catalog, "grok-4.6");
-	if (!grok) missing.push("grok-4.6: absent from the live catalog");
-	else {
-		if (!hasToolCalls(grok)) missing.push("grok-4.6: does not advertise tool_calls");
-		if (!supportsEffort(grok, "medium")) missing.push("grok-4.6: does not advertise a \"medium\" reasoning effort");
-		if (!hasUsablePromptMetadata(grok)) missing.push("grok-4.6: no usable max_prompt_tokens metadata");
-		if (!supportsEndpoint(grok, RESPONSES_ENDPOINTS)) missing.push("grok-4.6: does not advertise a supported Responses endpoint");
-	}
-	const gemini = findModel(catalog, "gemini-3.7-flash");
-	if (!gemini) missing.push("gemini-3.7-flash: absent from the live catalog");
-	else {
-		if (!hasContextAtLeast(gemini, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push("gemini-3.7-flash: advertised context window is below 1M");
-		if (!supportsEffort(gemini, "high")) missing.push("gemini-3.7-flash: does not advertise a \"high\" reasoning effort");
-		if (pickEndpoint(gemini) !== "chat") missing.push("gemini-3.7-flash: does not advertise a supported chat-completions endpoint");
-	}
-	const opus = findModel(catalog, "claude-opus-5");
-	if (!opus) missing.push("claude-opus-5: absent from the live catalog");
-	else {
-		if (!hasContextAtLeast(opus, FAST_REQUIRED_CONTEXT_TOKENS)) missing.push("claude-opus-5: advertised context window is below 1M");
-		if (!supportsEffort(opus, "high")) missing.push("claude-opus-5: does not advertise a \"high\" reasoning effort");
-		if (opus.capabilities?.supports?.adaptive_thinking !== true) missing.push("claude-opus-5: does not advertise adaptive_thinking");
-		if (!hasUsablePromptMetadata(opus)) missing.push("claude-opus-5: no usable max_prompt_tokens metadata");
-		if (!supportsEndpoint(opus, MESSAGES_ENDPOINTS)) missing.push("claude-opus-5: 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;
@@ -1306,6 +1081,73 @@ function wireDaemonTeardown(handle, opts = {}) {
 	});
 }
 //#endregion
+//#region src/lib/grok-context.ts
+/** Claude Code's fixed output-token reserve when computing its own
+*  compaction-window assumption. Grok 4.6's 128K max output stays well
+*  above this, so the reserve — not the model's real output ceiling — is
+*  what caps the second term. */
+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
+* percentage override is set, where `e` is already
+* `window - min(maxOutput, 20_000)`.
+*/
+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.
+*/
+const ENV_WINDOW_MIN_TOKENS = 1e5;
+const ENV_WINDOW_MAX_TOKENS = 1e6;
+/**
+* Derive the conservative 85%-of-prompt compaction trigger and the client
+* window that would need to be assumed to make that trigger point natural.
+*
+* `target trigger = floor(max_prompt_tokens * 0.85)`
+* `assumed client window = target trigger + min(max_output_tokens, 20_000)`
+*
+* For Grok 4.6's current catalog entry (max_prompt_tokens=372_000,
+* max_output_tokens=128_000) this yields `{ triggerTokens: 316_200,
+* assumedClientWindowTokens: 336_200 }` — the exact figures the plan
+* document records. Pure and side-effect-free; callers are responsible for
+* sourcing `maxPromptTokens`/`maxOutputTokens` from the live catalog.
+*/
+function computeConservativeCompactionTrigger(maxPromptTokens, maxOutputTokens) {
+	const triggerTokens = Math.floor(maxPromptTokens * .85);
+	return {
+		triggerTokens,
+		assumedClientWindowTokens: triggerTokens + Math.min(maxOutputTokens, CLIENT_OUTPUT_RESERVE_TOKENS)
+	};
+}
+/**
+* The integer `CLAUDE_CODE_AUTO_COMPACT_WINDOW` value that puts the client's
+* REACTIVE compaction trigger at 85% of the provider's real prompt ceiling.
+*
+* The client computes its trigger as
+* `window - min(maxOutput, 20_000) - 13_000`, so we invert that:
+* `window = assumedClientWindowTokens + 13_000`.
+*
+* Worked, against the live catalog at time of writing:
+*   Luna    922_000 prompt / 128_000 out -> trigger 783_700, window 816_700
+*   Opus 5  936_000 prompt /  64_000 out -> trigger 795_600, window 828_600
+*
+* Returns undefined when the catalog metadata is missing or nonsensical, so
+* the caller omits the variable entirely rather than exporting a guess. That
+* degrades to today's behaviour (client budgets against the total window)
+* instead of to a wrong number, which is the safer direction for a value the
+* client silently floors.
+*/
+function deriveAutoCompactWindowTokens(maxPromptTokens, maxOutputTokens) {
+	if (typeof maxPromptTokens !== "number" || !Number.isFinite(maxPromptTokens) || maxPromptTokens <= 0) return;
+	const { assumedClientWindowTokens } = computeConservativeCompactionTrigger(maxPromptTokens, typeof maxOutputTokens === "number" && Number.isFinite(maxOutputTokens) && maxOutputTokens > 0 ? maxOutputTokens : CLIENT_OUTPUT_RESERVE_TOKENS);
+	const window = assumedClientWindowTokens + CLIENT_THRESHOLD_RESERVE_TOKENS;
+	return Math.min(ENV_WINDOW_MAX_TOKENS, Math.max(ENV_WINDOW_MIN_TOKENS, window));
+}
+//#endregion
 //#region src/lib/proxy.ts
 let socketCounter = 0;
 /**
@@ -1656,7 +1498,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.293";
+var version = "0.3.297";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -1758,7 +1600,7 @@ const NO_LOOP = {
 	tier: null,
 	repeats: 0
 };
-function isRecord(value) {
+function isRecord$1(value) {
 	return typeof value === "object" && value !== null;
 }
 function digest(value) {
@@ -1788,7 +1630,7 @@ function normalizeResultBody(content) {
 }
 function normalizeResultBlock(block) {
 	if (typeof block === "string") return `s:${boundedText(block)}`;
-	if (!isRecord(block)) return `j:${boundedText(safeJson(block))}`;
+	if (!isRecord$1(block)) return `j:${boundedText(safeJson(block))}`;
 	const type = typeof block.type === "string" ? block.type : "unknown";
 	if (type === "text" && typeof block.text === "string") return `t:${boundedText(block.text)}`;
 	return `${type}:${digest(safeJson(block))}`;
@@ -1899,15 +1741,15 @@ function mayContainToolTraffic(rawBody) {
 	return rawBody.includes("tool_result") || rawBody.includes("tool_call_id") || rawBody.includes("function_call_output");
 }
 function blockType(block) {
-	return isRecord(block) && typeof block.type === "string" ? block.type : void 0;
+	return isRecord$1(block) && typeof block.type === "string" ? block.type : void 0;
 }
 /** Anthropic Messages: assistant `tool_use` blocks ↔ user `tool_result` blocks. */
 function extractAnthropicTurns(body) {
-	if (!isRecord(body) || !Array.isArray(body.messages)) return [];
+	if (!isRecord$1(body) || !Array.isArray(body.messages)) return [];
 	const turns = [];
 	for (let i = 0; i < body.messages.length; i++) {
 		const message = body.messages[i];
-		if (!isRecord(message) || message.role !== "assistant") continue;
+		if (!isRecord$1(message) || message.role !== "assistant") continue;
 		if (!Array.isArray(message.content)) continue;
 		const uses = message.content.filter((b) => blockType(b) === "tool_use");
 		if (uses.length === 0) {
@@ -1920,17 +1762,17 @@ function extractAnthropicTurns(body) {
 		const results = /* @__PURE__ */ new Map();
 		for (let j = i + 1; j < body.messages.length; j++) {
 			const next = body.messages[j];
-			if (!isRecord(next) || next.role !== "user") break;
+			if (!isRecord$1(next) || next.role !== "user") break;
 			if (!Array.isArray(next.content)) break;
 			for (const block of next.content) {
-				if (!isRecord(block) || blockType(block) !== "tool_result") continue;
+				if (!isRecord$1(block) || blockType(block) !== "tool_result") continue;
 				const id = block.tool_use_id;
 				if (typeof id === "string") results.set(id, block);
 			}
 		}
 		const calls = [];
 		for (const use of uses) {
-			if (!isRecord(use)) continue;
+			if (!isRecord$1(use)) continue;
 			const name = typeof use.name === "string" ? use.name : "unknown";
 			const id = typeof use.id === "string" ? use.id : void 0;
 			const result = id === void 0 ? void 0 : results.get(id);
@@ -1949,16 +1791,16 @@ function anthropicHasNarration(content) {
 	return content.some((block) => {
 		const type = blockType(block);
 		if (type === "thinking" || type === "redacted_thinking") return true;
-		return type === "text" && isRecord(block) && typeof block.text === "string" && block.text.trim() !== "";
+		return type === "text" && isRecord$1(block) && typeof block.text === "string" && block.text.trim() !== "";
 	});
 }
 /** OpenAI Chat Completions: assistant `tool_calls[]` ↔ `role:"tool"` messages. */
 function extractChatTurns(body) {
-	if (!isRecord(body) || !Array.isArray(body.messages)) return [];
+	if (!isRecord$1(body) || !Array.isArray(body.messages)) return [];
 	const turns = [];
 	for (let i = 0; i < body.messages.length; i++) {
 		const message = body.messages[i];
-		if (!isRecord(message) || message.role !== "assistant") continue;
+		if (!isRecord$1(message) || message.role !== "assistant") continue;
 		if (!Array.isArray(message.tool_calls) || message.tool_calls.length === 0) {
 			turns.push({
 				calls: [],
@@ -1969,14 +1811,14 @@ function extractChatTurns(body) {
 		const results = /* @__PURE__ */ new Map();
 		for (let j = i + 1; j < body.messages.length; j++) {
 			const next = body.messages[j];
-			if (!isRecord(next) || next.role !== "tool") break;
+			if (!isRecord$1(next) || next.role !== "tool") break;
 			const id = next.tool_call_id;
 			if (typeof id === "string") results.set(id, next);
 		}
 		const calls = [];
 		for (const call of message.tool_calls) {
-			if (!isRecord(call)) continue;
-			const fn = isRecord(call.function) ? call.function : void 0;
+			if (!isRecord$1(call)) continue;
+			const fn = isRecord$1(call.function) ? call.function : void 0;
 			const name = typeof fn?.name === "string" ? fn.name : "unknown";
 			const id = typeof call.id === "string" ? call.id : void 0;
 			const result = id === void 0 ? void 0 : results.get(id);
@@ -1999,15 +1841,15 @@ function extractChatTurns(body) {
 function chatHasNarration(content) {
 	if (typeof content === "string") return content.trim() !== "";
 	if (!Array.isArray(content)) return false;
-	return content.some((part) => isRecord(part) && typeof part.text === "string" && part.text.trim() !== "");
+	return content.some((part) => isRecord$1(part) && typeof part.text === "string" && part.text.trim() !== "");
 }
 /** OpenAI Responses: `function_call` items ↔ `function_call_output` items. */
 function extractResponsesTurns(body) {
-	if (!isRecord(body) || !Array.isArray(body.input)) return [];
+	if (!isRecord$1(body) || !Array.isArray(body.input)) return [];
 	const items = body.input;
 	const results = /* @__PURE__ */ new Map();
 	for (const item of items) {
-		if (!isRecord(item) || item.type !== "function_call_output") continue;
+		if (!isRecord$1(item) || item.type !== "function_call_output") continue;
 		const id = item.call_id;
 		if (typeof id === "string") results.set(id, item);
 	}
@@ -2015,14 +1857,14 @@ function extractResponsesTurns(body) {
 	let i = 0;
 	while (i < items.length) {
 		const item = items[i];
-		if (!isRecord(item) || item.type !== "function_call") {
+		if (!isRecord$1(item) || item.type !== "function_call") {
 			i++;
 			continue;
 		}
 		const batch = [];
 		while (i < items.length) {
 			const candidate = items[i];
-			if (!isRecord(candidate) || candidate.type !== "function_call") break;
+			if (!isRecord$1(candidate) || candidate.type !== "function_call") break;
 			batch.push(candidate);
 			i++;
 		}
@@ -2050,7 +1892,7 @@ function extractResponsesTurns(body) {
 function responsesHasNarration(items, batchStart) {
 	for (let i = batchStart - 1; i >= 0; i--) {
 		const item = items[i];
-		if (!isRecord(item)) return false;
+		if (!isRecord$1(item)) return false;
 		if (item.type === "function_call_output") continue;
 		if (item.type === "reasoning") return true;
 		if (item.type === "message" || item.role === "assistant") return responsesItemHasText(item);
@@ -2061,13 +1903,13 @@ function responsesHasNarration(items, batchStart) {
 function responsesItemHasText(item) {
 	if (typeof item.content === "string") return item.content.trim() !== "";
 	if (!Array.isArray(item.content)) return false;
-	return item.content.some((block) => isRecord(block) && typeof block.text === "string" && block.text.trim() !== "");
+	return item.content.some((block) => isRecord$1(block) && typeof block.text === "string" && block.text.trim() !== "");
 }
 function injectAnthropicNudge(body, text) {
-	if (!isRecord(body) || !Array.isArray(body.messages)) return false;
+	if (!isRecord$1(body) || !Array.isArray(body.messages)) return false;
 	const messages = body.messages;
 	const last = messages[messages.length - 1];
-	if (!isRecord(last) || last.role !== "user") return false;
+	if (!isRecord$1(last) || last.role !== "user") return false;
 	if (!Array.isArray(last.content)) return false;
 	messages[messages.length - 1] = {
 		...last,
@@ -2079,7 +1921,7 @@ function injectAnthropicNudge(body, text) {
 	return true;
 }
 function injectChatNudge(body, text) {
-	if (!isRecord(body) || !Array.isArray(body.messages)) return false;
+	if (!isRecord$1(body) || !Array.isArray(body.messages)) return false;
 	body.messages = [...body.messages, {
 		role: "user",
 		content: text
@@ -2087,7 +1929,7 @@ function injectChatNudge(body, text) {
 	return true;
 }
 function injectResponsesNudge(body, text) {
-	if (!isRecord(body) || !Array.isArray(body.input)) return false;
+	if (!isRecord$1(body) || !Array.isArray(body.input)) return false;
 	body.input = [...body.input, {
 		role: "user",
 		content: [{
@@ -4300,10 +4142,10 @@ function isAsyncIterable(x) {
 *     from the catalog (which would also mean re-parsing/re-picking work the
 *     caller already did).
 *   - `makeShimContinueTurn` (below), which `buildAdvisorStream`
-*     (`src/services/advisor/advisor.ts`) injects as its `continueTurn` for
-*     the fast Luna-lead profile, so an advisor continuation on a non-Claude
-*     lead runs through the SAME translation + SSE-synthesis machinery as
-*     the initial turn instead of a parallel, divergent implementation.
+*     (`src/services/advisor/advisor.ts`) injects for any non-Claude model
+*     selected by an authenticated fast primary lead, so its Advisor
+*     continuation runs through the SAME translation + SSE-synthesis machinery
+*     as the initial turn instead of a parallel, divergent implementation.
 */
 async function streamParsedRequestViaShim(parsed, endpoint, opts, signal) {
 	const routePath = opts.routePath ?? "/v1/messages (advisor lead shim)";
@@ -4338,8 +4180,8 @@ async function streamParsedRequestViaShim(parsed, endpoint, opts, signal) {
 * Build an injectable `continueTurn(body, signal)` for `buildAdvisorStream`
 * (`src/services/advisor/advisor.ts`) that routes a continuation turn
 * through THIS module's non-Claude shim instead of Claude passthrough — used
-* for the fast Luna-lead profile's advisor translate-loop. No `onCancel` is
-* threaded through: the advisor loop's own `aborter` (shared with `signal`
+* by any non-Claude model selected in an authenticated fast primary lead.
+* No `onCancel` is threaded through: the advisor loop's own `aborter` (shared with `signal`
 * here) already tears down on consumer cancel via `buildAdvisorStream`'s
 * `cancel()`, so this stream needs no independent teardown hook.
 */
@@ -4551,7 +4393,7 @@ function preprocessFastRequest(rawBody, launch) {
 		originalModel,
 		modified: false
 	};
-	const bare = originalModel.replace(/\[1m\]$/i, "");
+	const { base: bare } = stripTrailingOneMSuffix(originalModel);
 	let effort;
 	if (alias) {
 		effort = alias.absentEffortDefault;
@@ -4581,6 +4423,216 @@ function preprocessFastRequest(rawBody, launch) {
 	};
 }
 //#endregion
+//#region src/lib/prompt-window-salvage.ts
+/**
+* Keep enough room for Copilot's message framing and for small differences
+* between counting the JSON text here and counting the decoded request there.
+*/
+const PROMPT_WINDOW_RESERVE = 2e3;
+const TOOL_RESULT_STUB = "[earlier tool output elided to fit prompt window]";
+const MESSAGE_TEXT_STUB = "[earlier message text elided to fit prompt window]";
+const MARKER_RE = /^\[github-router: elided ~([0-9]+) tokens of older tool output to fit this model's prompt window\]$/;
+function markerText(tokens) {
+	return `[github-router: elided ~${tokens} tokens of older tool output to fit this model's prompt window]`;
+}
+function isRecord(value) {
+	return typeof value === "object" && value !== null;
+}
+function textReplacement(block, replacement) {
+	if (block.type !== "text" || typeof block.text !== "string") return void 0;
+	if (block.text === replacement) return void 0;
+	return {
+		original: block.text,
+		replacement,
+		apply: () => {
+			block.text = replacement;
+		}
+	};
+}
+function collectToolResultReplacements(messages, lastMessageIndex) {
+	const replacements = [];
+	for (let i = 0; i < lastMessageIndex; i += 1) {
+		const message = messages[i];
+		if (!isRecord(message) || message.role !== "user" || !Array.isArray(message.content)) continue;
+		for (const block of message.content) {
+			if (!isRecord(block) || block.type !== "tool_result") continue;
+			if (typeof block.content === "string") {
+				if (block.content === TOOL_RESULT_STUB) continue;
+				const original = block.content;
+				replacements.push({
+					original,
+					replacement: TOOL_RESULT_STUB,
+					apply: () => {
+						block.content = TOOL_RESULT_STUB;
+					}
+				});
+				continue;
+			}
+			if (!Array.isArray(block.content)) continue;
+			for (const nested of block.content) {
+				if (!isRecord(nested)) continue;
+				const replacement = textReplacement(nested, TOOL_RESULT_STUB);
+				if (replacement) replacements.push(replacement);
+			}
+		}
+	}
+	return replacements;
+}
+function collectMessageTextReplacements(messages, lastMessageIndex) {
+	const replacements = [];
+	for (let i = 0; i < lastMessageIndex; i += 1) {
+		const message = messages[i];
+		if (!isRecord(message)) continue;
+		if (message.role !== "assistant" && message.role !== "user") continue;
+		if (typeof message.content === "string") {
+			if (message.content === MESSAGE_TEXT_STUB) continue;
+			const original = message.content;
+			replacements.push({
+				original,
+				replacement: MESSAGE_TEXT_STUB,
+				apply: () => {
+					message.content = MESSAGE_TEXT_STUB;
+				}
+			});
+			continue;
+		}
+		if (!Array.isArray(message.content)) continue;
+		for (const block of message.content) {
+			if (!isRecord(block)) continue;
+			const replacement = textReplacement(block, MESSAGE_TEXT_STUB);
+			if (replacement) replacements.push(replacement);
+		}
+	}
+	return replacements;
+}
+function hasSalvageStub(messages) {
+	for (const message of messages.slice(0, -1)) {
+		if (!isRecord(message)) continue;
+		if (message.content === MESSAGE_TEXT_STUB) return true;
+		if (!Array.isArray(message.content)) continue;
+		for (const block of message.content) {
+			if (!isRecord(block)) continue;
+			if (block.type === "text" && block.text === MESSAGE_TEXT_STUB) return true;
+			if (block.type !== "tool_result") continue;
+			if (block.content === TOOL_RESULT_STUB) return true;
+			if (Array.isArray(block.content) && block.content.some((nested) => isRecord(nested) && nested.type === "text" && nested.text === TOOL_RESULT_STUB)) return true;
+		}
+	}
+	return false;
+}
+function ensureMarker(messages) {
+	const last = messages.at(-1);
+	if (!isRecord(last) || last.role !== "user") return void 0;
+	if (typeof last.content === "string") last.content = [{
+		type: "text",
+		text: last.content
+	}];
+	if (!Array.isArray(last.content)) return void 0;
+	for (const contentBlock of last.content) {
+		if (!isRecord(contentBlock)) continue;
+		if (contentBlock.type !== "text" || typeof contentBlock.text !== "string") continue;
+		const match = MARKER_RE.exec(contentBlock.text);
+		if (!match) continue;
+		if (!hasSalvageStub(messages)) return void 0;
+		const parsed = Number.parseInt(match[1], 10);
+		if (!Number.isSafeInteger(parsed) || parsed < 0) return void 0;
+		return {
+			block: contentBlock,
+			priorElidedTokens: parsed
+		};
+	}
+	const block = {
+		type: "text",
+		text: markerText(0)
+	};
+	last.content.push(block);
+	return {
+		block,
+		priorElidedTokens: 0
+	};
+}
+async function fragmentTokenSavings(replacement, encoding) {
+	const [before, after] = await Promise.all([getTextTokenCount(JSON.stringify(replacement.original), encoding), getTextTokenCount(JSON.stringify(replacement.replacement), encoding)]);
+	return before - after;
+}
+/**
+* Last-resort history salvage for requests that escaped the client's normal
+* compaction path. Protected content is never rewritten, and the original
+* string is returned unless a complete, valid-looking salvage fits the live
+* model budget.
+*/
+async function salvageOversizedPrompt(rawBody, model) {
+	const unchanged = {
+		body: rawBody,
+		salvaged: false
+	};
+	if (!model) return unchanged;
+	const maxPromptTokens = model.capabilities?.limits?.max_prompt_tokens;
+	if (typeof maxPromptTokens !== "number" || !Number.isFinite(maxPromptTokens) || maxPromptTokens <= 0) return unchanged;
+	const budget = Math.floor(maxPromptTokens) - PROMPT_WINDOW_RESERVE;
+	if (budget <= 0) return unchanged;
+	if (Buffer.byteLength(rawBody, "utf8") <= budget) return unchanged;
+	const encoding = getTokenizerFromModel(model);
+	let initialTokens;
+	try {
+		initialTokens = await getTextTokenCount(rawBody, encoding);
+	} catch (error) {
+		consola.debug("Prompt-window salvage tokenization failed; allowing request:", error);
+		return unchanged;
+	}
+	if (initialTokens <= budget) return unchanged;
+	let body;
+	try {
+		const parsed = JSON.parse(rawBody);
+		if (!isRecord(parsed) || !Array.isArray(parsed.messages) || parsed.messages.length === 0) return unchanged;
+		body = parsed;
+	} catch {
+		return unchanged;
+	}
+	const messages = body.messages;
+	const marker = ensureMarker(messages);
+	if (!marker) return unchanged;
+	const replacements = [...collectToolResultReplacements(messages, messages.length - 1), ...collectMessageTextReplacements(messages, messages.length - 1)];
+	let newlyElidedTokens = 0;
+	let estimatedTokens = initialTokens;
+	for (const replacement of replacements) {
+		let savings;
+		try {
+			savings = await fragmentTokenSavings(replacement, encoding);
+		} catch (error) {
+			consola.debug("Prompt-window salvage tokenization failed; allowing request:", error);
+			return unchanged;
+		}
+		if (savings <= 0) continue;
+		replacement.apply();
+		newlyElidedTokens += savings;
+		estimatedTokens -= savings;
+		const totalElidedTokens = marker.priorElidedTokens + newlyElidedTokens;
+		marker.block.text = markerText(totalElidedTokens);
+		if (estimatedTokens > budget) continue;
+		let serialized;
+		let finalTokens;
+		try {
+			serialized = JSON.stringify(body);
+			finalTokens = await getTextTokenCount(serialized, encoding);
+		} catch (error) {
+			consola.debug("Prompt-window salvage serialization failed; allowing request:", error);
+			return unchanged;
+		}
+		if (finalTokens > budget) {
+			estimatedTokens = finalTokens;
+			continue;
+		}
+		consola.warn(`Prompt-window salvage: model=${model.id} tokens=${initialTokens} budget=${budget} elided=${totalElidedTokens}`);
+		return {
+			body: serialized,
+			salvaged: true,
+			elidedTokens: totalElidedTokens
+		};
+	}
+	return unchanged;
+}
+//#endregion
 //#region src/routes/messages/handler.ts
 const MAX_THINKING_REPAIR_ATTEMPTS = 5;
 const isWebSearchTool$1 = (tool) => !!tool && typeof tool === "object" && (typeof tool.type === "string" && tool.type.startsWith("web_search") || tool.name === "web_search");
@@ -4649,6 +4701,32 @@ 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 {
@@ -4785,6 +4863,30 @@ async function handleCompletion(c) {
 	const fastSubagentRequest = fastProfileRequest && Boolean(c.req.header("x-claude-code-agent-id"));
 	const fastLeadAdvisor = fastProfileRequest && !fastSubagentRequest;
 	const advisorEnabled = advisorRequested && !fastSubagentRequest;
+	const fastAdvisorEnabled = fastLeadAdvisor && advisorRequested && hasNonEmptyTools(rawBody);
+	const advisorBehaviorEnabled = fastLeadAdvisor ? fastAdvisorEnabled : advisorEnabled;
+	let fastAdvisorChoice;
+	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);
 	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.`;
@@ -4808,7 +4910,7 @@ async function handleCompletion(c) {
 		}
 	}, 400, { "x-should-retry": "false" });
 	if (loopGuard.body !== void 0) finalBody = loopGuard.body;
-	if (advisorEnabled) {
+	if (advisorBehaviorEnabled) {
 		finalBody = injectAdvisorTool(finalBody, fastLeadAdvisor ? FAST_ADVISOR_TOOL_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");
 	}
@@ -4824,15 +4926,16 @@ async function handleCompletion(c) {
 	} catch {}
 	const { body: resolvedBody, originalModel, resolvedModel, selectedModel } = resolveModelInBody$1(finalBody);
 	const modelId = resolvedModel ?? originalModel;
-	const messagesRoute = classifyMessagesRoute(modelId, selectedModel, originalModel);
+	const { body: promptWindowBody } = await salvageOversizedPrompt(resolvedBody, selectedModel);
+	const messagesRoute = classifyMessagesRoute(modelId, selectedModel, originalModel, fastProfileRequest);
 	if (messagesRoute !== "claude-passthrough") {
 		const endpoint = messagesRoute === "chat-shim" ? "chat" : "responses";
 		let parsedBase;
 		try {
-			parsedBase = JSON.parse(resolvedBody);
+			parsedBase = JSON.parse(promptWindowBody);
 		} catch {}
 		const wantsStream = parsedBase?.stream === true;
-		if (advisorEnabled && wantsStream && identity.launch?.profileId === "fast" && isFastProfileLead(modelId)) {
+		if (fastAdvisorEnabled && wantsStream) {
 			const initialConversation = Array.isArray(parsedBase.messages) ? parsedBase.messages : [];
 			const parsedInitial = parseAnthropicRequest(parsedBase, modelId, selectedModel);
 			const fastAdvisorAborter = new AbortController();
@@ -4850,7 +4953,7 @@ async function handleCompletion(c) {
 				status: 200,
 				streaming: true
 			}, selectedModel, startTime);
-			const advisorChoice = resolveAdvisorModel(modelId, true);
+			const advisorChoice = fastAdvisorChoice;
 			return new Response(buildAdvisorStream({
 				firstResponse,
 				initialConversation,
@@ -4875,8 +4978,8 @@ async function handleCompletion(c) {
 				}
 			});
 		}
-		const shimBody = stripAdvisorTool(resolvedBody);
-		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).");
+		const shimBody = stripAdvisorTool(promptWindowBody);
+		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,
@@ -4888,12 +4991,12 @@ 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
 	};
-	let nativeBody = resolvedBody;
+	let nativeBody = promptWindowBody;
 	const knownThinkingRepair = repairKnownThinkingHistory(nativeBody);
 	if (knownThinkingRepair) {
 		nativeBody = knownThinkingRepair.body;
@@ -4962,13 +5065,13 @@ 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 ?? resolveAdvisorModel(originalModel, false);
 			return new Response(buildAdvisorStream({
 				firstResponse: response,
 				initialConversation,
@@ -6500,7 +6603,7 @@ 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 getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 	const vars = {
@@ -6555,10 +6658,66 @@ function getClaudeCodeEnvVars(serverUrl, model, launchProfileId = "standard") {
 	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);
 	if (toolbeltEnabled()) Object.assign(vars, toolbeltPathOverride(process.env, PATHS.TOOLBELT_BIN_DIR));
 	return vars;
 }
 /**
+* Every env var whose value can become the ACTIVE main-loop model id, and so
+* whose prompt ceiling the compaction window has to respect. `ANTHROPIC_MODEL`
+* is the lead; the tier rows and the custom option each become the active id
+* verbatim when picked from `/model`.
+*
+* `ANTHROPIC_SMALL_FAST_MODEL` is deliberately absent: background ops run
+* there, but compaction itself runs on the main-loop model (verified in the
+* 2.1.247 bundle — the summarizer starts from `r.options.mainLoopModel`).
+*/
+const LEAD_CAPABLE_MODEL_ENV_KEYS = [
+	"ANTHROPIC_MODEL",
+	"ANTHROPIC_DEFAULT_OPUS_MODEL",
+	"ANTHROPIC_DEFAULT_SONNET_MODEL",
+	"ANTHROPIC_DEFAULT_HAIKU_MODEL",
+	"ANTHROPIC_CUSTOM_MODEL_OPTION"
+];
+/**
+* Resolve one seeded env value to its live catalog entry: strip the `[1m]`
+* accounting bracket, erase router-owned alias provenance, then translate the
+* Anthropic-dashed slug onto Copilot's catalog id.
+*/
+function catalogEntryForSeededModel(value) {
+	const { base } = stripTrailingOneMSuffix(canonicalizeAliasModel(value));
+	const id = resolveModel(base);
+	return state.models?.data?.find((m) => m.id === id);
+}
+/**
+* Set `CLAUDE_CODE_AUTO_COMPACT_WINDOW` to the smallest complete derived
+* window across every 1M-accounted model this launch can reach. Each candidate
+* puts the client's reactive trigger at 85% of that model's prompt ceiling;
+* the minimum is therefore safe after any `/model` switch.
+*
+* Only `[1m]`-decorated candidates participate. An undecorated row already
+* budgets at the client's conservative 200K default, which is below every
+* prompt ceiling in the lineup, so including it would drag the window down for
+* no benefit. When nothing is decorated, or no candidate carries usable
+* catalog limits, the variable is omitted entirely rather than guessed.
+*
+* Presence-guarded on the parent env, symmetric with every other guard here.
+*/
+function applyAutoCompactWindow(vars) {
+	if (process.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW !== void 0) return;
+	let tightestWindow;
+	const consider = (value) => {
+		if (!value || !/\[1m\]/i.test(value)) return;
+		const limits = catalogEntryForSeededModel(value)?.capabilities?.limits;
+		const candidateWindow = deriveAutoCompactWindowTokens(limits?.max_prompt_tokens, limits?.max_output_tokens);
+		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);
+	if (tightestWindow === void 0) return;
+	vars.CLAUDE_CODE_AUTO_COMPACT_WINDOW = String(tightestWindow);
+}
+/**
 * Build environment variables for Codex CLI.
 *
 * Like `getClaudeCodeEnvVars`, the parent env is sanitized of
@@ -6580,6 +6739,6 @@ function getCodexEnvVars(serverUrl) {
 	return vars;
 }
 //#endregion
-export { validateFastProfilePrerequisites as _, sharedServerArgs as a, updateClaude as b, stopKeepAwake as c, LUNA_DRIVER_ALIAS_ID as d, LUNA_IMPLEMENTER_ALIAS_ID as f, resolveLaunchProfile as g, profileDescriptor as h, setupAndServe as i, listModelsForEndpoint as l, formatFastPrerequisiteFailure as m, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, LUNA_SCOUT_ALIAS_ID as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, enableFileLogging as u, runSelfUpdate as v, checkClaudeVersion as y };
+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-DbvbW5Ve.js.map
+//# sourceMappingURL=server-setup-DhG3YZPj.js.map