@nick3/copilot-api 1.4.5 → 1.4.9

package/dist/{server-D_7gI9hx.js → server-DqwhClJ-.js}

@@ -1,8 +1,8 @@
-import { PATHS } from "./paths-DoT4SZ8f.js";
-import { listAccountsFromRegistry } from "./accounts-registry-c7rs5Ed9.js";
-import { HTTPError, accountFromState, cacheModels, copilotBaseUrl, copilotHeaders, forwardError, generateRequestIdFromPayload, getCopilotUsage, getRootSessionId, getUUID, isNullish, parseUserIdMetadata, prepareForCompact, prepareInteractionHeaders, sleep, state } from "./utils-BIK3ym34.js";
-import "./get-copilot-token-BySQCue6.js";
-import { PROVIDER_TYPE_ANTHROPIC, accountsManager, getAliasTargetSet, getAnthropicApiKey, getConfig, getExtraPromptForModel, getModelAliases, getModelAliasesInfo, getModelRefreshIntervalMs, getProviderConfig, getReasoningEffortForModel, getSmallModel, isForceAgentEnabled, isFreeModelLoadBalancingEnabled, isMessageStartInputTokensFallbackEnabled, isMessagesApiEnabled, isResponsesApiContextManagementModel, mergeConfigWithDefaults, shouldCompactUseSmallModel } from "./accounts-manager-BsGuQhKM.js";
+import { t as PATHS } from "./paths-Cvzy-eLX.js";
+import { i as listAccountsFromRegistry } from "./accounts-registry-CQYvRe65.js";
+import { C as prepareInteractionHeaders, D as state, E as accountFromState, S as prepareForCompact, T as resolveTraceId, _ as copilotHeaders, a as generateRequestIdFromPayload, c as isNullish, g as copilotBaseUrl, h as forwardError, l as parseUserIdMetadata, m as HTTPError, n as cacheModels, o as getRootSessionId, p as getCopilotUsage, s as getUUID, u as sleep, w as requestContext } from "./utils-DY-jLXwO.js";
+import "./get-copilot-token-BbpphnmV.js";
+import { _ as isMessagesApiEnabled, a as getClaudeTokenMultiplier, b as mergeConfigWithDefaults, c as getModelAliases, d as getProviderConfig, f as getReasoningEffortForModel, g as isMessageStartInputTokensFallbackEnabled, h as isForceAgentEnabled, i as getAnthropicApiKey, l as getModelAliasesInfo, m as isAccountAffinityEnabled, n as PROVIDER_TYPE_ANTHROPIC, o as getConfig, p as getSmallModel, r as getAliasTargetSet, s as getExtraPromptForModel, t as accountsManager, u as getModelRefreshIntervalMs, v as isResponsesApiContextManagementModel, x as shouldCompactUseSmallModel, y as isResponsesApiWebSearchEnabled } from "./accounts-manager-Cjrd_el_.js";
 import consola from "consola";
 import fs, { readFile } from "node:fs/promises";
 import * as path$1 from "node:path";
@@ -12,7 +12,6 @@ import { Hono } from "hono";
 import { cors } from "hono/cors";
 import { logger } from "hono/logger";
 import fs$1, { existsSync } from "node:fs";
-import { AsyncLocalStorage } from "node:async_hooks";
 import { Database } from "bun:sqlite";
 import { fileURLToPath } from "node:url";
 import { streamSSE } from "hono/streaming";
@@ -105,26 +104,6 @@ function createAuthMiddleware(options = {}) {
 	};
 }
 
-//#endregion
-//#region src/lib/request-context.ts
-const TRACE_ID_MAX_LENGTH = 64;
-const TRACE_ID_PATTERN = /^\w[\w.-]*$/;
-const asyncLocalStorage = new AsyncLocalStorage();
-const requestContext = {
-	getStore: () => asyncLocalStorage.getStore(),
-	run: (context, callback) => asyncLocalStorage.run(context, callback)
-};
-function generateTraceId() {
-	const timestamp = Date.now().toString(36);
-	const random = Math.random().toString(36).slice(2, 8);
-	return `${timestamp}-${random}`;
-}
-function resolveTraceId(traceId) {
-	const candidate = traceId?.trim();
-	if (!candidate || candidate.length > TRACE_ID_MAX_LENGTH || !TRACE_ID_PATTERN.test(candidate)) return generateTraceId();
-	return candidate;
-}
-
 //#endregion
 //#region src/lib/trace.ts
 const traceIdMiddleware = async (c, next) => {
@@ -132,7 +111,8 @@ const traceIdMiddleware = async (c, next) => {
 	c.header("x-trace-id", traceId);
 	const context = {
 		traceId,
-		startTime: Date.now()
+		startTime: Date.now(),
+		userAgent: c.req.header("user-agent") || ""
 	};
 	await requestContext.run(context, async () => {
 		await next();
@@ -189,66 +169,69 @@ function getAdminDbUserVersion(db = getAdminDb()) {
 		return 0;
 	}
 }
-function migrateAdminDb(db) {
-	const current = db.query("PRAGMA user_version;").get()?.user_version ?? 0;
-	if (current >= 4) return;
-	if (current < 1) db.run(`
-      CREATE TABLE IF NOT EXISTS request_log (
-        id INTEGER PRIMARY KEY AUTOINCREMENT,
-        request_id TEXT NOT NULL UNIQUE,
+function migrateV1(db) {
+	db.run(`
+    CREATE TABLE IF NOT EXISTS request_log (
+      id INTEGER PRIMARY KEY AUTOINCREMENT,
+      request_id TEXT NOT NULL UNIQUE,
 
-        started_at_ms INTEGER NOT NULL,
-        finished_at_ms INTEGER,
-        duration_ms INTEGER,
-        ttfb_ms INTEGER,
+      started_at_ms INTEGER NOT NULL,
+      finished_at_ms INTEGER,
+      duration_ms INTEGER,
+      ttfb_ms INTEGER,
 
-        method TEXT NOT NULL,
-        path TEXT NOT NULL,
-        upstream_endpoint TEXT,
-        stream INTEGER NOT NULL DEFAULT 0,
+      method TEXT NOT NULL,
+      path TEXT NOT NULL,
+      upstream_endpoint TEXT,
+      stream INTEGER NOT NULL DEFAULT 0,
 
-        account_id TEXT,
-        account_type TEXT,
-        cost_units REAL,
-        client_model TEXT,
-        upstream_model TEXT,
+      account_id TEXT,
+      account_type TEXT,
+      cost_units REAL,
+      client_model TEXT,
+      upstream_model TEXT,
 
-        client_ip TEXT,
-        client_ip_source TEXT,
-        user_agent TEXT,
+      client_ip TEXT,
+      client_ip_source TEXT,
+      user_agent TEXT,
 
-        tokens_input INTEGER,
-        tokens_output INTEGER,
-        tokens_total INTEGER,
-        tokens_cached_input INTEGER,
-        usage_json TEXT,
+      tokens_input INTEGER,
+      tokens_output INTEGER,
+      tokens_total INTEGER,
+      tokens_cached_input INTEGER,
+      usage_json TEXT,
 
-        premium_remaining_before REAL,
-        premium_remaining_after REAL,
-        premium_remaining_diff REAL,
-        premium_unlimited_before INTEGER,
-        premium_unlimited_after INTEGER,
+      premium_remaining_before REAL,
+      premium_remaining_after REAL,
+      premium_remaining_diff REAL,
+      premium_unlimited_before INTEGER,
+      premium_unlimited_after INTEGER,
 
-        http_status INTEGER,
-        error_name TEXT,
-        error_status INTEGER,
-        error_message TEXT,
-        selection_failure_reason TEXT
-      );
+      http_status INTEGER,
+      error_name TEXT,
+      error_status INTEGER,
+      error_message TEXT,
+      selection_failure_reason TEXT
+    );
 
-      CREATE INDEX IF NOT EXISTS idx_request_log_started_at
-        ON request_log(started_at_ms DESC);
-      CREATE INDEX IF NOT EXISTS idx_request_log_account_started_at
-        ON request_log(account_id, started_at_ms DESC);
-      CREATE INDEX IF NOT EXISTS idx_request_log_model_started_at
-        ON request_log(upstream_model, started_at_ms DESC);
-      CREATE INDEX IF NOT EXISTS idx_request_log_endpoint_started_at
-        ON request_log(upstream_endpoint, started_at_ms DESC);
-      CREATE INDEX IF NOT EXISTS idx_request_log_status_started_at
-        ON request_log(http_status, started_at_ms DESC);
+    CREATE INDEX IF NOT EXISTS idx_request_log_started_at
+      ON request_log(started_at_ms DESC);
+    CREATE INDEX IF NOT EXISTS idx_request_log_account_started_at
+      ON request_log(account_id, started_at_ms DESC);
+    CREATE INDEX IF NOT EXISTS idx_request_log_model_started_at
+      ON request_log(upstream_model, started_at_ms DESC);
+    CREATE INDEX IF NOT EXISTS idx_request_log_endpoint_started_at
+      ON request_log(upstream_endpoint, started_at_ms DESC);
+    CREATE INDEX IF NOT EXISTS idx_request_log_status_started_at
+      ON request_log(http_status, started_at_ms DESC);
 
-      PRAGMA user_version = 1;
-    `);
+    PRAGMA user_version = 1;
+  `);
+}
+function migrateAdminDb(db) {
+	const current = db.query("PRAGMA user_version;").get()?.user_version ?? 0;
+	if (current >= 5) return;
+	if (current < 1) migrateV1(db);
 	if (current < 2) db.run(`
       ALTER TABLE request_log ADD COLUMN user_id TEXT;
       ALTER TABLE request_log ADD COLUMN safety_identifier TEXT;
@@ -283,6 +266,12 @@ function migrateAdminDb(db) {
 
       PRAGMA user_version = 4;
     `);
+	if (current < 5) db.run(`
+      ALTER TABLE request_log ADD COLUMN affinity_hit INTEGER;
+      ALTER TABLE request_log ADD COLUMN affinity_cache_key TEXT;
+
+      PRAGMA user_version = 5;
+    `);
 }
 
 //#endregion
@@ -368,14 +357,11 @@ function normalizeMessagesUsage(usage) {
 	const output = usage.output_tokens;
 	const hasInput = typeof input === "number";
 	const hasOutput = typeof output === "number";
-	const tokensInput = hasInput ? Math.max(0, input - cached) : void 0;
-	const tokensOutput = hasOutput ? output : void 0;
-	const tokensTotal = hasInput || hasOutput ? (input ?? 0) + (output ?? 0) : void 0;
 	return {
 		tokensCachedInput: cached,
-		tokensInput,
-		tokensOutput,
-		tokensTotal,
+		tokensInput: hasInput ? Math.max(0, input - cached) : void 0,
+		tokensOutput: hasOutput ? output : void 0,
+		tokensTotal: hasInput || hasOutput ? (input ?? 0) + (output ?? 0) : void 0,
 		usageJson: JSON.stringify(usage)
 	};
 }
@@ -434,13 +420,15 @@ var RequestHistoryStore = class {
         error_name,
         error_status,
         error_message,
-        selection_failure_reason
+        selection_failure_reason,
+        affinity_hit,
+        affinity_cache_key
       ) VALUES (
         ?,?,?,?,?,?,?,?,
         ?,?,?,?,?,?,?,?,
         ?,?,?,?,?,?,?,?,
         ?,?,?,?,?,?,?,?,
-        ?,?,?,?,?
+        ?,?,?,?,?,?,?
       );
     `);
 		this.getByRequestIdStmt = db.query("SELECT * FROM request_log WHERE request_id = ? LIMIT 1;");
@@ -499,7 +487,9 @@ var RequestHistoryStore = class {
 				toDbNull(record.errorName),
 				toDbNull(record.errorStatus),
 				toDbNull(record.errorMessage),
-				toDbNull(record.selectionFailureReason)
+				toDbNull(record.selectionFailureReason),
+				toDbBool(record.affinityHit),
+				toDbNull(record.affinityCacheKey)
 			];
 			this.insertStmt.run(...args);
 		} catch (error) {
@@ -585,10 +575,9 @@ var RequestHistoryStore = class {
 		const rows = this.db.query(sql).all(...values, limit + 1);
 		const items = rows.slice(0, limit);
 		const hasMore = rows.length > limit;
-		const nextCursorId = hasMore ? items.at(-1)?.id : void 0;
 		return {
 			items,
-			nextCursorId,
+			nextCursorId: hasMore ? items.at(-1)?.id : void 0,
 			hasMore
 		};
 	}
@@ -763,7 +752,7 @@ const CONFIG_KEYS = new Set([
 	"auth",
 	"extraPrompts",
 	"smallModel",
-	"freeModelLoadBalancing",
+	"accountAffinity",
 	"apiKey",
 	"anthropicApiKey",
 	"providers",
@@ -776,7 +765,8 @@ const CONFIG_KEYS = new Set([
 	"compactUseSmallModel",
 	"messageStartInputTokensFallback",
 	"modelRefreshIntervalHours",
-	"useMessagesApi"
+	"useMessagesApi",
+	"useResponsesApiWebSearch"
 ]);
 const REASONING_EFFORTS = new Set([
 	"none",
@@ -1124,7 +1114,7 @@ const CONFIG_PATCH_HANDLERS = {
 	auth: applyAuthConfig,
 	extraPrompts: applyExtraPrompts,
 	smallModel: (next, value) => applyOptionalString(next, "smallModel", value),
-	freeModelLoadBalancing: (next, value) => applyOptionalBoolean(next, "freeModelLoadBalancing", value),
+	accountAffinity: (next, value) => applyOptionalBoolean(next, "accountAffinity", value),
 	apiKey: (next, value) => applyOptionalString(next, "apiKey", value),
 	anthropicApiKey: (next, value) => applyOptionalString(next, "anthropicApiKey", value),
 	providers: applyProvidersConfig,
@@ -1137,7 +1127,8 @@ const CONFIG_PATCH_HANDLERS = {
 	compactUseSmallModel: (next, value) => applyOptionalBoolean(next, "compactUseSmallModel", value),
 	messageStartInputTokensFallback: (next, value) => applyOptionalBoolean(next, "messageStartInputTokensFallback", value),
 	modelRefreshIntervalHours: (next, value) => applyOptionalNumber(next, "modelRefreshIntervalHours", value),
-	useMessagesApi: (next, value) => applyOptionalBoolean(next, "useMessagesApi", value)
+	useMessagesApi: (next, value) => applyOptionalBoolean(next, "useMessagesApi", value),
+	useResponsesApiWebSearch: (next, value) => applyOptionalBoolean(next, "useResponsesApiWebSearch", value)
 };
 function applyConfigPatch(base, input) {
 	const next = { ...base };
@@ -1215,7 +1206,7 @@ adminApiRoutes.post("/config", async (c) => {
 	try {
 		await writeConfigFile(result.config);
 		const merged = mergeConfigWithDefaults();
-		accountsManager.setFreeModelLoadBalancingEnabled(isFreeModelLoadBalancingEnabled());
+		accountsManager.setAccountAffinityEnabled(isAccountAffinityEnabled());
 		accountsManager.setModelsRefreshIntervalMs(getModelRefreshIntervalMs());
 		return c.json({
 			...merged,
@@ -1294,12 +1285,10 @@ function parseAdminModelDetailsItem(raw, aliasesByTarget) {
 	if (!isPlainObject(raw)) return null;
 	const id = parseNonEmptyString(raw.id);
 	if (!id) return null;
-	const name = parseNonEmptyString(raw.name) ?? id;
-	const preview = toBooleanOrUndefined(raw.preview) ?? false;
 	return {
 		id,
-		name,
-		preview,
+		name: parseNonEmptyString(raw.name) ?? id,
+		preview: toBooleanOrUndefined(raw.preview) ?? false,
 		billing: parseBilling(raw.billing),
 		supported_endpoints: parseStringArray(raw.supported_endpoints),
 		capabilities: parseCapabilities(raw.capabilities),
@@ -2192,8 +2181,7 @@ const createHandlerLogger = (name) => {
 		const filePath = path.join(LOG_DIR, `${sanitizedName}-${dateKey}.log`);
 		const message = formatArgs(logObj.args);
 		const traceIdStr = traceId ? ` [${traceId}]` : "";
-		const line = `[${timestamp}] [${logObj.type}] [${logObj.tag || name}]${traceIdStr}${message ? ` ${message}` : ""}`;
-		appendLine(filePath, line);
+		appendLine(filePath, `[${timestamp}] [${logObj.type}] [${logObj.tag || name}]${traceIdStr}${message ? ` ${message}` : ""}`);
 	} });
 	return instance;
 };
@@ -2425,8 +2413,7 @@ const numTokensForTools = (tools, encoder, constants) => {
 * Calculate the token count of messages, supporting multiple GPT encoders
 */
 const getTokenCount = async (payload, model) => {
-	const tokenizer = getTokenizerFromModel(model);
-	const encoder = await getEncodeChatFunction(tokenizer);
+	const encoder = await getEncodeChatFunction(getTokenizerFromModel(model));
 	const simplifiedMessages = payload.messages;
 	const inputMessages = simplifiedMessages.filter((msg) => msg.role !== "assistant");
 	const outputMessages = simplifiedMessages.filter((msg) => msg.role === "assistant");
@@ -2517,10 +2504,11 @@ async function handleCompletion$1(c) {
 		});
 	}
 	logger$6.debug("Request payload:", JSON.stringify(payload).slice(-400));
+	const upstreamRequestId = generateRequestIdFromPayload(payload, normalizedPromptCacheKey);
 	const selection = await accountsManager.selectAccountForRequest([{
 		modelId: clientModel,
 		endpoint: CHAT_COMPLETIONS_ENDPOINT$1
-	}]);
+	}], { requestId: upstreamRequestId });
 	if (!selection.ok) {
 		recordSelectionFailure$2(store, {
 			request,
@@ -2534,6 +2522,8 @@ async function handleCompletion$1(c) {
 		});
 	}
 	const { account, selectedModel } = selection;
+	request.affinityHit = selection.affinityHit;
+	request.affinityCacheKey = selection.affinityCacheKey;
 	const upstreamPayload = {
 		...payload,
 		model: selectedModel.id
@@ -2547,7 +2537,6 @@ async function handleCompletion$1(c) {
 	if (state.manualApprove) await awaitApproval();
 	const payloadWithMaxTokens = applyDefaultMaxTokens(upstreamPayload, selectedModel);
 	const accountCtx = toAccountContext(account);
-	const upstreamRequestId = generateRequestIdFromPayload(payloadWithMaxTokens);
 	const upstreamSessionId = getUUID(upstreamRequestId);
 	request.upstreamRequestId = upstreamRequestId;
 	request.upstreamSessionId = upstreamSessionId;
@@ -2580,7 +2569,6 @@ function buildRequestContext$1(c) {
 	const method = c.req.raw.method;
 	const path$2 = new URL(c.req.url, "http://local").pathname;
 	const { ip: clientIp, source: clientIpSource } = getClientIpInfo(c);
-	const userAgent = c.req.header("user-agent") ?? void 0;
 	return {
 		requestId,
 		startedAtMs,
@@ -2588,7 +2576,7 @@ function buildRequestContext$1(c) {
 		path: path$2,
 		clientIp,
 		clientIpSource,
-		userAgent
+		userAgent: c.req.header("user-agent") ?? void 0
 	};
 }
 function insertRequestLog$2(store, request, record) {
@@ -2605,6 +2593,8 @@ function insertRequestLog$2(store, request, record) {
 		promptCacheKey: request.promptCacheKey,
 		initiator: request.initiator,
 		upstreamRequestId: request.upstreamRequestId,
+		affinityHit: request.affinityHit,
+		affinityCacheKey: request.affinityCacheKey,
 		...record
 	});
 }
@@ -2657,6 +2647,7 @@ async function handleStreamingRequest(params) {
 			upstreamRequestId: request.upstreamRequestId,
 			sessionId: request.upstreamSessionId
 		});
+		selection.confirmAffinity?.();
 	} catch (error) {
 		return handleUpstreamCreateError$1({
 			store,
@@ -2844,6 +2835,7 @@ async function handleNonStreamingRequest(params) {
 			upstreamRequestId: request.upstreamRequestId,
 			sessionId: request.upstreamSessionId
 		});
+		selection.confirmAffinity?.();
 		finishedAtMs = Date.now();
 		if (!isNonStreaming$1(response)) {
 			logger$6.debug("Unexpected streaming response");
@@ -2930,7 +2922,6 @@ embeddingRoutes.post("/", async (c) => {
 		const method = c.req.raw.method;
 		const path$2 = new URL(c.req.url, "http://local").pathname;
 		const { ip: clientIp, source: clientIpSource } = getClientIpInfo(c);
-		const userAgent = c.req.header("user-agent") ?? void 0;
 		const ctx = {
 			requestId,
 			startedAtMs,
@@ -2938,7 +2929,7 @@ embeddingRoutes.post("/", async (c) => {
 			path: path$2,
 			clientIp,
 			clientIpSource,
-			userAgent
+			userAgent: c.req.header("user-agent") ?? void 0
 		};
 		const payload = await c.req.json();
 		const clientModel = payload.model;
@@ -2962,15 +2953,14 @@ embeddingRoutes.post("/", async (c) => {
 			});
 			return selectionFailureResponse$1(c, clientModel, selection.reason);
 		}
-		const upstreamPayload = {
-			...payload,
-			model: selection.selectedModel.id
-		};
 		return await runEmbeddingsWithAccount({
 			c,
 			store,
 			ctx,
-			payload: upstreamPayload,
+			payload: {
+				...payload,
+				model: selection.selectedModel.id
+			},
 			clientModel,
 			selection
 		});
@@ -3019,8 +3009,7 @@ async function runEmbeddingsWithAccount({ c, store, ctx, payload, clientModel, s
 	let errorMessage;
 	let finishedAtMs;
 	try {
-		const accountCtx = toAccountContext(account);
-		const response = await createEmbeddings(payload, accountCtx);
+		const response = await createEmbeddings(payload, toAccountContext(account));
 		usage = normalizeEmbeddingsUsage(response.usage);
 		finishedAtMs = Date.now();
 		return c.json(response);
@@ -3175,6 +3164,32 @@ const mergeToolResultForClaude = (anthropicPayload) => {
 		msg.content = mergeToolResult(toolResults, textBlocks);
 	}
 };
+const stripUnsupportedCacheControl = (block) => {
+	const cacheControl = block.cache_control;
+	if (!cacheControl || typeof cacheControl !== "object" || Array.isArray(cacheControl)) return;
+	const type = cacheControl.type;
+	if (typeof type === "string") {
+		block.cache_control = { type };
+		return;
+	}
+	delete block.cache_control;
+};
+const stripTextBlockCacheControl = (block) => {
+	if (!block || typeof block !== "object" || Array.isArray(block)) return;
+	const record = block;
+	if (record.type !== "text") return;
+	stripUnsupportedCacheControl(record);
+};
+const stripCacheControl = (anthropicPayload) => {
+	if (Array.isArray(anthropicPayload.system)) for (const block of anthropicPayload.system) stripTextBlockCacheControl(block);
+	for (const msg of anthropicPayload.messages) {
+		if (!Array.isArray(msg.content)) continue;
+		for (const block of msg.content) {
+			stripTextBlockCacheControl(block);
+			if (block.type === "tool_result" && Array.isArray(block.content)) for (const nestedBlock of block.content) stripTextBlockCacheControl(nestedBlock);
+		}
+	}
+};
 const estimateInputTokens = async (payload, selectedModel, logger$7) => {
 	try {
 		return (await getTokenCount(payload, selectedModel)).input;
@@ -3239,6 +3254,21 @@ const maybeBlockOriginalModelName = (context) => {
 		}
 	});
 };
+const compactSystemPromptStart = "You are a helpful AI assistant tasked with summarizing conversations";
+const compactUserMessageStart = "CRITICAL: Respond with TEXT ONLY. Do NOT call any tools.";
+const hasCompactUserMessage = (messages) => {
+	const lastMsg = messages.at(-1);
+	if (!lastMsg || lastMsg.role !== "user") return false;
+	if (typeof lastMsg.content === "string") return lastMsg.content.startsWith(compactUserMessageStart);
+	if (!Array.isArray(lastMsg.content)) return false;
+	return lastMsg.content.some((block) => block.type === "text" && typeof block.text === "string" && block.text.startsWith(compactUserMessageStart));
+};
+const isCompactRequest = (anthropicPayload) => {
+	const system = anthropicPayload.system;
+	if (typeof system === "string" && system.startsWith(compactSystemPromptStart)) return true;
+	if (Array.isArray(system) && system.some((msg) => typeof msg.text === "string" && msg.text.startsWith(compactSystemPromptStart))) return true;
+	return hasCompactUserMessage(anthropicPayload.messages);
+};
 
 //#endregion
 //#region src/routes/messages/non-stream-translation.ts
@@ -3523,7 +3553,7 @@ async function handleCountTokens(c) {
 			}
 		}
 		let finalTokenCount = tokenCount.input + tokenCount.output;
-		if (anthropicPayload.model.startsWith("claude")) finalTokenCount = Math.round(finalTokenCount * 1.15);
+		if (anthropicPayload.model.startsWith("claude")) finalTokenCount = Math.round(finalTokenCount * getClaudeTokenMultiplier());
 		consola.info("Token count:", finalTokenCount);
 		return c.json({ input_tokens: finalTokenCount });
 	} catch (error) {
@@ -3952,9 +3982,8 @@ const mapResponsesStopReason = (response) => {
 const mapResponsesUsage = (response) => {
 	const inputTokens = response.usage?.input_tokens ?? 0;
 	const outputTokens = response.usage?.output_tokens ?? 0;
-	const inputCachedTokens = response.usage?.input_tokens_details?.cached_tokens;
 	return {
-		input_tokens: inputTokens - (inputCachedTokens ?? 0),
+		input_tokens: inputTokens - (response.usage?.input_tokens_details?.cached_tokens ?? 0),
 		output_tokens: outputTokens,
 		...response.usage?.input_tokens_details?.cached_tokens !== void 0 && { cache_read_input_tokens: response.usage.input_tokens_details.cached_tokens }
 	};
@@ -4374,10 +4403,9 @@ const openFunctionCallBlock = (state$1, params) => {
 	if (!functionCallState) {
 		const blockIndex$1 = state$1.nextContentBlockIndex;
 		state$1.nextContentBlockIndex += 1;
-		const resolvedToolCallId = toolCallId ?? `tool_call_${blockIndex$1}`;
 		functionCallState = {
 			blockIndex: blockIndex$1,
-			toolCallId: resolvedToolCallId,
+			toolCallId: toolCallId ?? `tool_call_${blockIndex$1}`,
 			name: name ?? "function",
 			consecutiveWhitespaceCount: 0
 		};
@@ -4403,26 +4431,20 @@ const openFunctionCallBlock = (state$1, params) => {
 const extractFunctionCallDetails = (rawEvent) => {
 	const item = rawEvent.item;
 	if (item.type !== "function_call") return;
-	const outputIndex = rawEvent.output_index;
-	const toolCallId = item.call_id;
-	const name = item.name;
-	const initialArguments = item.arguments;
 	return {
-		outputIndex,
-		toolCallId,
-		name,
-		initialArguments
+		outputIndex: rawEvent.output_index,
+		toolCallId: item.call_id,
+		name: item.name,
+		initialArguments: item.arguments
 	};
 };
 
 //#endregion
 //#region src/routes/responses/utils.ts
 const getResponsesRequestOptions = (payload) => {
-	const vision = hasVisionInput(payload);
-	const initiator = hasAgentInitiator(payload) ? "agent" : "user";
 	return {
-		vision,
-		initiator
+		vision: hasVisionInput(payload),
+		initiator: hasAgentInitiator(payload) ? "agent" : "user"
 	};
 };
 const hasAgentInitiator = (payload) => {
@@ -4482,7 +4504,15 @@ const containsVisionContent = (value) => {
 
 //#endregion
 //#region src/services/copilot/create-messages.ts
+const isAgentMessage = (msg) => {
+	if (msg.role === "assistant") return true;
+	if (Array.isArray(msg.content)) {
+		if (msg.content.every((block) => block.type === "tool_result") && msg.content.length > 0) return true;
+	}
+	return false;
+};
 const getMessagesInitiator = (payload) => {
+	if (isForceAgentEnabled()) return payload.messages.some((msg) => isAgentMessage(msg)) ? "agent" : "user";
 	const lastMessage = payload.messages.at(-1);
 	if (!lastMessage || lastMessage.role !== "user") return "agent";
 	if (!Array.isArray(lastMessage.content)) return "user";
@@ -4839,7 +4869,6 @@ const logger$5 = createHandlerLogger("messages-handler");
 const CHAT_COMPLETIONS_ENDPOINT = "/chat/completions";
 const RESPONSES_ENDPOINT$1 = "/responses";
 const MESSAGES_ENDPOINT = "/v1/messages";
-const compactSystemPromptStart = "You are a helpful AI assistant tasked with summarizing conversations";
 async function handleCompletion(c) {
 	await checkRateLimit(state);
 	const store = getRequestHistoryStore();
@@ -4907,7 +4936,7 @@ async function handleCompletion(c) {
 		modelId: endpointModel?.id ?? openAIPayload.model,
 		endpoint: CHAT_COMPLETIONS_ENDPOINT
 	});
-	const selection = await accountsManager.selectAccountForRequest(candidates);
+	const selection = await accountsManager.selectAccountForRequest(candidates, { requestId: upstreamRequestId });
 	if (!selection.ok) return handleSelectionFailure({
 		c,
 		store,
@@ -4952,7 +4981,10 @@ async function handleCompletion(c) {
 		costUnits,
 		upstreamRequestId,
 		premiumRemainingBefore,
-		premiumUnlimitedBefore
+		premiumUnlimitedBefore,
+		confirmAffinity: selection.confirmAffinity,
+		affinityHit: selection.affinityHit,
+		affinityCacheKey: selection.affinityCacheKey
 	};
 	if (endpoint === MESSAGES_ENDPOINT) return await handleWithMessagesApi({
 		c,
@@ -5002,6 +5034,7 @@ const handleWithChatCompletions = async (params) => {
 			sessionId,
 			isCompact
 		});
+		instr.confirmAffinity?.();
 	} catch (error) {
 		return await handleChatCompletionsCreateError({
 			error,
@@ -5050,6 +5083,7 @@ const handleWithResponsesApi = async (params) => {
 			sessionId,
 			isCompact
 		}, ctx);
+		instr.confirmAffinity?.();
 	} catch (error) {
 		return await handleResponsesCreateError({
 			error,
@@ -5095,6 +5129,8 @@ function insertRequestLog$1(instr, record) {
 		promptCacheKey: instr.promptCacheKey,
 		initiator: instr.initiator,
 		upstreamRequestId: instr.upstreamRequestId,
+		affinityHit: instr.affinityHit,
+		affinityCacheKey: instr.affinityCacheKey,
 		clientModel,
 		upstreamEndpoint,
 		accountId: account.id,
@@ -5109,10 +5145,9 @@ function insertRequestLog$1(instr, record) {
 async function finalizeQuotaAndGetPremiumSnapshot(instr) {
 	await accountsManager.finalizeQuota(instr.account, instr.reservation);
 	const premiumRemainingAfter = instr.account.premiumRemaining;
-	const premiumUnlimitedAfter = instr.account.unlimited;
 	return {
 		premiumRemainingAfter,
-		premiumUnlimitedAfter,
+		premiumUnlimitedAfter: instr.account.unlimited,
 		premiumRemainingDiff: computeDiff(instr.premiumRemainingBefore, premiumRemainingAfter)
 	};
 }
@@ -5500,6 +5535,7 @@ async function streamMessagesAndLog(params) {
 }
 const handleWithMessagesApi = async (params) => {
 	const { c, anthropicPayload, anthropicBetaHeader, initiatorOverride, subagentMarker, sessionId, instr, selectedModel, isCompact } = params;
+	stripCacheControl(anthropicPayload);
 	for (const msg of anthropicPayload.messages) if (msg.role === "assistant" && Array.isArray(msg.content)) msg.content = msg.content.filter((block) => {
 		if (block.type !== "thinking") return true;
 		return block.thinking && block.thinking !== "Thinking..." && block.signature && !block.signature.includes("@");
@@ -5526,6 +5562,7 @@ const handleWithMessagesApi = async (params) => {
 			sessionId,
 			isCompact
 		});
+		instr.confirmAffinity?.();
 	} catch (error) {
 		return await handleMessagesCreateError({
 			error,
@@ -5555,12 +5592,6 @@ const getAnthropicEffortForModel = (model) => {
 	if (reasoningEffort === "none" || reasoningEffort === "minimal") return "low";
 	return reasoningEffort;
 };
-const isCompactRequest = (anthropicPayload) => {
-	const system = anthropicPayload.system;
-	if (typeof system === "string") return system.startsWith(compactSystemPromptStart);
-	if (!Array.isArray(system)) return false;
-	return system.some((msg) => typeof msg.text === "string" && msg.text.startsWith(compactSystemPromptStart));
-};
 
 //#endregion
 //#region src/routes/messages/route.ts
@@ -5729,7 +5760,7 @@ async function forwardProviderModels(providerConfig, requestHeaders) {
 //#region src/routes/provider/messages/handler.ts
 const logger$3 = createHandlerLogger("provider-messages-handler");
 async function handleProviderMessages(c) {
-	const provider = c.req.param("provider");
+	const provider = c.req.param("provider") ?? "";
 	const providerConfig = getProviderConfig(provider);
 	if (!providerConfig) return c.json({ error: {
 		message: `Provider '${provider}' not found or disabled`,
@@ -5896,7 +5927,7 @@ const handleResponses = async (c) => {
 	const payload = await c.req.json();
 	const clientModel = payload.model;
 	logger$1.debug("Responses request payload:", JSON.stringify(payload));
-	removeWebSearchTool(payload);
+	if (!isResponsesApiWebSearchEnabled()) removeWebSearchTool(payload);
 	compactInputByLatestCompaction(payload);
 	const streamRequested = Boolean(payload.stream);
 	const { initiator: initialInitiator } = getResponsesRequestOptions(payload);
@@ -5920,10 +5951,11 @@ const handleResponses = async (c) => {
 			message: "This model is only available via an alias. Please use the alias model name."
 		});
 	}
+	const upstreamRequestId = generateRequestIdFromPayload({ messages: payload.input }, normalizedPromptCacheKey);
 	const selection = await accountsManager.selectAccountForRequest([{
 		modelId: clientModel,
 		endpoint: RESPONSES_ENDPOINT
-	}]);
+	}], { requestId: upstreamRequestId });
 	if (!selection.ok) {
 		recordSelectionFailure(store, {
 			request,
@@ -5934,6 +5966,8 @@ const handleResponses = async (c) => {
 		return selectionFailureResponse(c, { reason: selection.reason });
 	}
 	const { account, selectedModel } = selection;
+	request.affinityHit = selection.affinityHit;
+	request.affinityCacheKey = selection.affinityCacheKey;
 	const upstreamPayload = {
 		...payload,
 		model: selectedModel.id
@@ -5947,7 +5981,6 @@ const handleResponses = async (c) => {
 	request.initiator = initiator;
 	if (state.manualApprove) await awaitApproval();
 	const accountCtx = toAccountContext(account);
-	const upstreamRequestId = generateRequestIdFromPayload({ messages: upstreamPayload.input });
 	const upstreamSessionId = getUUID(upstreamRequestId);
 	request.upstreamRequestId = upstreamRequestId;
 	request.upstreamSessionId = upstreamSessionId;
@@ -5984,7 +6017,6 @@ function buildRequestContext(c) {
 	const method = c.req.raw.method;
 	const path$2 = new URL(c.req.url, "http://local").pathname;
 	const { ip: clientIp, source: clientIpSource } = getClientIpInfo(c);
-	const userAgent = c.req.header("user-agent") ?? void 0;
 	return {
 		requestId,
 		startedAtMs,
@@ -5992,7 +6024,7 @@ function buildRequestContext(c) {
 		path: path$2,
 		clientIp,
 		clientIpSource,
-		userAgent
+		userAgent: c.req.header("user-agent") ?? void 0
 	};
 }
 function insertRequestLog(store, request, record) {
@@ -6009,6 +6041,8 @@ function insertRequestLog(store, request, record) {
 		promptCacheKey: request.promptCacheKey,
 		initiator: request.initiator,
 		upstreamRequestId: request.upstreamRequestId,
+		affinityHit: request.affinityHit,
+		affinityCacheKey: request.affinityCacheKey,
 		...record
 	});
 }
@@ -6039,8 +6073,7 @@ function selectionFailureResponse(c, params) {
 function extractUsageFromChunkData(data) {
 	if (!data) return void 0;
 	try {
-		const event = JSON.parse(data);
-		const usage = extractResponsesUsageFromStreamEvent(event);
+		const usage = extractResponsesUsageFromStreamEvent(JSON.parse(data));
 		return usage.usageJson ? usage : void 0;
 	} catch {
 		return;
@@ -6064,6 +6097,7 @@ async function handleStreamingResponses(params) {
 			upstreamRequestId: request.upstreamRequestId,
 			sessionId: request.upstreamSessionId
 		}, accountCtx);
+		selection.confirmAffinity?.();
 	} catch (error) {
 		return handleUpstreamCreateError({
 			store,
@@ -6250,6 +6284,7 @@ async function handleNonStreamingResponses(params) {
 			upstreamRequestId: request.upstreamRequestId,
 			sessionId: request.upstreamSessionId
 		}, accountCtx);
+		selection.confirmAffinity?.();
 		finishedAtMs = Date.now();
 		const streamResponse = handleUnexpectedResponsesStream(c, response);
 		if (streamResponse) return streamResponse;
@@ -6434,4 +6469,4 @@ server.route("/:provider/v1/models", providerModelRoutes);
 
 //#endregion
 export { server };
-//# sourceMappingURL=server-D_7gI9hx.js.map
+//# sourceMappingURL=server-DqwhClJ-.js.map