@symerian/symi 3.5.4 → 3.5.6

package/dist/{pi-embedded-BPuUM-gD.js → pi-embedded-C9wLrFjj.js}

@@ -8,17 +8,17 @@ import { _ as DEFAULT_PROVIDER, a as modelKey, d as resolveModelRefFromString, f
 import { i as resolveAckReaction, o as resolveEffectiveMessagesConfig, r as resolveResponsePrefixTemplate, s as resolveHumanDelayConfig, t as createReplyPrefixOptions } from "./reply-prefix-i-FPYcoQ.js";
 import { t as normalizeChatType } from "./chat-type-Acj2OK2p.js";
 import { i as resolveSlackAccount, n as listChannelPlugins, o as resolveSlackAppToken, r as normalizeChannelId$1, s as resolveSlackBotToken, t as getChannelPlugin } from "./plugins-CZ_mYwXq.js";
-import { S as resolveSlackChannelId, _ as resolveSlackWebClientOptions, a as sendMessageSlack, b as buildSlackBlocksFallbackText, c as getDefaultLocalRoots, d as fetchRemoteMedia, f as readResponseWithLimit, g as createSlackWebClient, h as fetchWithTimeout, i as resolveSlackThreadTs, l as loadWebMedia, m as bindAbortRelay, n as deliverReplies, o as renderMarkdownWithMarkers, p as fetchWithSsrFGuard, s as markdownToIRWithMeta, t as createSlackReplyDeliveryPlan, u as MediaFetchError, v as parseSlackBlocksInput, x as parseSlackTarget, y as validateSlackBlocksArray } from "./replies-BYF1gbJf.js";
-import { $ as loadSessionStore, $t as resolveToolProfilePolicy, A as formatRawAssistantErrorForUi, At as resolveMainSessionKey, B as isRateLimitAssistantError, Bt as saveMediaBuffer, C as downgradeOpenAIReasoningBlocks, Ct as resolveChannelResetConfig, D as classifyFailoverReason, Dt as DEFAULT_RESET_TRIGGERS, E as BILLING_ERROR_USER_MESSAGE, Et as resolveThreadFlag, F as isCompactionFailureError, Ft as createBrowserRouteContext, G as parseImageSizeError, Gt as resolveBrowserControlAuth, H as isTimeoutErrorMessage, Ht as resolveExistingPathsWithinRoot, I as isContextOverflowError, It as registerBrowserRoutes, J as resolveSandboxContext, Jt as collectExplicitAllowlist, K as sanitizeUserFacingText, Kt as applyOwnerOnlyToolPolicy, L as isFailoverAssistantError, Lt as resolveBrowserConfig, M as isAuthAssistantError, Mt as resolveGroupSessionKey, N as isBillingAssistantError, Nt as acquireSessionWriteLock, O as formatAssistantErrorText, Ot as resolveFreshSessionTotalTokens, P as isCloudCodeAssistFormatError, Pt as resolveSessionLockMaxHoldFromTimeout, Q as resolveAndPersistSessionFile, Qt as normalizeToolName$1, R as isFailoverErrorMessage, Rt as resolveProfile, S as extractToolResultId, St as evaluateSessionFreshness, T as isGoogleModelApi, Tt as resolveSessionResetType, U as isTransientHttpError, Ut as getBridgeAuthForPort, V as isRawApiErrorPayload, Vt as DEFAULT_UPLOAD_DIR, W as parseImageDimensionError, Wt as ensureBrowserControlAuth, X as extractDeliveryInfo, Xt as expandToolGroups, Y as resolveSandboxRuntimeStatus, Yt as expandPolicyWithPluginGroups, Z as appendAssistantMessageToSessionTranscript, Zt as mergeAlsoAllowPolicy, _ as sanitizeSessionMessagesImages, _t as INPUT_PROVENANCE_KIND_VALUES, a as normalizeChannelTargetInput, an as resolveBootstrapMaxChars, at as updateSessionStoreEntry, b as resolveImageSanitizationLimits, bt as normalizeInputProvenance, c as parseReplyDirectives, cn as getGlobalHookRunner, ct as deliveryContextFromSession, d as parseInlineDirectives$1, dt as normalizeDeliveryContext, en as stripPluginOnlyAllowlist, et as readSessionUpdatedAt, f as validateAnthropicTurns, ft as normalizeSessionDeliveryFields, g as normalizeTextForComparison, gt as extractToolCallNames, h as isMessagingToolDuplicateNormalized, ht as countToolResults, i as buildTargetResolverSignature, in as ensureSessionHeader, it as updateSessionStore, j as getApiErrorPayloadFingerprint, jt as deriveSessionMetaPatch, k as formatBillingErrorMessage, kt as canonicalizeMainSessionAlias, l as MEDIA_TOKEN_RE, ln as initializeGlobalHookRunner, lt as deliveryContextKey, m as pickFallbackThinkingLevel, mt as capArrayByJsonBytes, nn as matchesAnyGlobPattern, o as normalizeTargetForProvider, on as resolveBootstrapTotalMaxChars, ot as isCacheEnabled, p as validateGeminiTurns, pt as archiveSessionTranscripts, q as ensureSandboxWorkspaceForSession, qt as buildPluginToolGroups, r as normalizeReplyPayloadsForDelivery, rn as buildBootstrapContextFiles, rt as updateLastRoute, s as throwIfAborted, sn as sanitizeGoogleTurnOrdering, st as resolveCacheTtlMs$1, t as deliverOutboundPayloads, tn as compileGlobPatterns, tt as recordSessionMetaFromInbound, u as splitMediaFromOutput, ut as mergeDeliveryContext, v as sanitizeImageBlocks, vt as applyInputProvenanceToUserMessage, w as isAntigravityClaude, wt as resolveSessionResetPolicy, x as extractToolCallsFromAssistant, xt as resolveSessionKey, y as sanitizeToolResultImages, yt as hasInterSessionUserProvenance, z as isLikelyContextOverflowError, zt as getMediaDir } from "./deliver-Cjyb6h4g.js";
+import { S as resolveSlackChannelId, _ as resolveSlackWebClientOptions, a as sendMessageSlack, b as buildSlackBlocksFallbackText, c as getDefaultLocalRoots, d as fetchRemoteMedia, f as readResponseWithLimit, g as createSlackWebClient, h as fetchWithTimeout, i as resolveSlackThreadTs, l as loadWebMedia, m as bindAbortRelay, n as deliverReplies, o as renderMarkdownWithMarkers, p as fetchWithSsrFGuard, s as markdownToIRWithMeta, t as createSlackReplyDeliveryPlan, u as MediaFetchError, v as parseSlackBlocksInput, x as parseSlackTarget, y as validateSlackBlocksArray } from "./replies-0nzkXt6o.js";
+import { $ as loadSessionStore, $t as resolveToolProfilePolicy, A as formatRawAssistantErrorForUi, At as resolveMainSessionKey, B as isRateLimitAssistantError, Bt as saveMediaBuffer, C as downgradeOpenAIReasoningBlocks, Ct as resolveChannelResetConfig, D as classifyFailoverReason, Dt as DEFAULT_RESET_TRIGGERS, E as BILLING_ERROR_USER_MESSAGE, Et as resolveThreadFlag, F as isCompactionFailureError, Ft as createBrowserRouteContext, G as parseImageSizeError, Gt as resolveBrowserControlAuth, H as isTimeoutErrorMessage, Ht as resolveExistingPathsWithinRoot, I as isContextOverflowError, It as registerBrowserRoutes, J as resolveSandboxContext, Jt as collectExplicitAllowlist, K as sanitizeUserFacingText, Kt as applyOwnerOnlyToolPolicy, L as isFailoverAssistantError, Lt as resolveBrowserConfig, M as isAuthAssistantError, Mt as resolveGroupSessionKey, N as isBillingAssistantError, Nt as acquireSessionWriteLock, O as formatAssistantErrorText, Ot as resolveFreshSessionTotalTokens, P as isCloudCodeAssistFormatError, Pt as resolveSessionLockMaxHoldFromTimeout, Q as resolveAndPersistSessionFile, Qt as normalizeToolName$1, R as isFailoverErrorMessage, Rt as resolveProfile, S as extractToolResultId, St as evaluateSessionFreshness, T as isGoogleModelApi, Tt as resolveSessionResetType, U as isTransientHttpError, Ut as getBridgeAuthForPort, V as isRawApiErrorPayload, Vt as DEFAULT_UPLOAD_DIR, W as parseImageDimensionError, Wt as ensureBrowserControlAuth, X as extractDeliveryInfo, Xt as expandToolGroups, Y as resolveSandboxRuntimeStatus, Yt as expandPolicyWithPluginGroups, Z as appendAssistantMessageToSessionTranscript, Zt as mergeAlsoAllowPolicy, _ as sanitizeSessionMessagesImages, _t as INPUT_PROVENANCE_KIND_VALUES, a as normalizeChannelTargetInput, an as resolveBootstrapMaxChars, at as updateSessionStoreEntry, b as resolveImageSanitizationLimits, bt as normalizeInputProvenance, c as parseReplyDirectives, cn as getGlobalHookRunner, ct as deliveryContextFromSession, d as parseInlineDirectives$1, dt as normalizeDeliveryContext, en as stripPluginOnlyAllowlist, et as readSessionUpdatedAt, f as validateAnthropicTurns, ft as normalizeSessionDeliveryFields, g as normalizeTextForComparison, gt as extractToolCallNames, h as isMessagingToolDuplicateNormalized, ht as countToolResults, i as buildTargetResolverSignature, in as ensureSessionHeader, it as updateSessionStore, j as getApiErrorPayloadFingerprint, jt as deriveSessionMetaPatch, k as formatBillingErrorMessage, kt as canonicalizeMainSessionAlias, l as MEDIA_TOKEN_RE, ln as initializeGlobalHookRunner, lt as deliveryContextKey, m as pickFallbackThinkingLevel, mt as capArrayByJsonBytes, nn as matchesAnyGlobPattern, o as normalizeTargetForProvider, on as resolveBootstrapTotalMaxChars, ot as isCacheEnabled, p as validateGeminiTurns, pt as archiveSessionTranscripts, q as ensureSandboxWorkspaceForSession, qt as buildPluginToolGroups, r as normalizeReplyPayloadsForDelivery, rn as buildBootstrapContextFiles, rt as updateLastRoute, s as throwIfAborted, sn as sanitizeGoogleTurnOrdering, st as resolveCacheTtlMs$1, t as deliverOutboundPayloads, tn as compileGlobPatterns, tt as recordSessionMetaFromInbound, u as splitMediaFromOutput, ut as mergeDeliveryContext, v as sanitizeImageBlocks, vt as applyInputProvenanceToUserMessage, w as isAntigravityClaude, wt as resolveSessionResetPolicy, x as extractToolCallsFromAssistant, xt as resolveSessionKey, y as sanitizeToolResultImages, yt as hasInterSessionUserProvenance, z as isLikelyContextOverflowError, zt as getMediaDir } from "./deliver-f3cIWxXT.js";
 import { a as logMessageProcessed, i as logLaneEnqueue, o as logMessageQueued, r as logLaneDequeue, s as logSessionStateChange, t as diag } from "./diagnostic-BdRnGknC.js";
 import { n as getDiagnosticSessionState } from "./diagnostic-session-state-DpxCUzoM.js";
 import { S as GATEWAY_CLIENT_NAMES, _ as listDeliverableMessageChannels, a as chunkText, b as GATEWAY_CLIENT_IDS, c as resolveChunkMode, d as isSafeFenceBreak, f as parseFenceSpans, g as isMarkdownCapableMessageChannel, h as isInternalMessageChannel, i as chunkMarkdownTextWithMode, l as resolveTextChunkLimit, m as isDeliverableMessageChannel, o as chunkTextWithMode, p as INTERNAL_MESSAGE_CHANNEL, r as chunkMarkdownText, t as chunkByNewline, u as findFenceSpanAt, v as normalizeMessageChannel, x as GATEWAY_CLIENT_MODES, y as resolveGatewayMessageChannel } from "./chunk-CAZujdOi.js";
 import { _ as resolveSymiAgentDir, a as markAuthProfileFailure, c as dedupeProfileIds, d as ensureAuthProfileStore, f as resolveAuthStorePathForDisplay, g as normalizeSecretInput, i as isProfileInCooldown, l as listProfilesForProvider, m as resolveAuthProfileDisplayLabel, n as resolveAuthProfileOrder, o as markAuthProfileUsed, p as withFileLock$1, r as getSoonestCooldownExpiry, s as resolveApiKeyForProfile, u as markAuthProfileGood } from "./auth-profiles-Bt3PyWkt.js";
-import { C as unsetConfigValueAtPath, S as setConfigValueAtPath, T as VERSION, _ as resetConfigOverrides, b as getConfigValueAtPath, c as loadConfig, d as writeConfigFile, f as validateConfigObjectWithPlugins, g as getConfigOverrides, h as validateJsonSchemaValue, i as isBlockedHostnameOrIp, l as readConfigFileSnapshot, o as normalizeHostname, p as parseDurationMs, t as SsrFBlockedError, u as resolveConfigSnapshotHash, v as setConfigOverride, w as DEFAULT_SUBAGENT_MAX_SPAWN_DEPTH, x as parseConfigPath, y as unsetConfigOverride } from "./ssrf-q6hBiAOi.js";
+import { C as unsetConfigValueAtPath, S as setConfigValueAtPath, T as VERSION, _ as resetConfigOverrides, b as getConfigValueAtPath, c as loadConfig, d as writeConfigFile, f as validateConfigObjectWithPlugins, g as getConfigOverrides, h as validateJsonSchemaValue, i as isBlockedHostnameOrIp, l as readConfigFileSnapshot, o as normalizeHostname, p as parseDurationMs, t as SsrFBlockedError, u as resolveConfigSnapshotHash, v as setConfigOverride, w as DEFAULT_SUBAGENT_MAX_SPAWN_DEPTH, x as parseConfigPath, y as unsetConfigOverride } from "./ssrf-BCSnhba8.js";
 import { t as parseBooleanValue } from "./boolean-BW6OTjPi.js";
 import { i as resolveShellEnvFallbackTimeoutMs, n as getShellPathFromLoginShell, s as isTruthyEnvValue } from "./shell-env-DgjeObDZ.js";
 import { c as normalizePluginsConfig, f as isPathInsideWithRealpath, i as safeStatSync, l as resolveEnableState, n as discoverSymiPlugins, p as isDangerousHostEnvVarName, r as isPathInside, s as applyTestPluginDefaults, t as loadPluginManifestRegistry, u as resolveMemorySlotDecision } from "./manifest-registry-CAWGTwb5.js";
-import { C as rawDataToString, O as DEFAULT_AI_SNAPSHOT_MAX_CHARS, T as isSecureWebSocketUrl, x as ensureChromeExtensionRelayServer } from "./chrome-BKUACyeO.js";
+import { C as rawDataToString, O as DEFAULT_AI_SNAPSHOT_MAX_CHARS, T as isSecureWebSocketUrl, x as ensureChromeExtensionRelayServer } from "./chrome-3jl2ulOE.js";
 import { n as resolveCliName, t as formatCliCommand } from "./command-format-DPd9RN2g.js";
 import { c as assertSandboxPath, d as resolveSandboxedMediaSource, f as applySkillEnvOverrides, h as parseFrontmatterBlock, i as resolveSkillsPromptForRun, l as resolveSandboxInputPath, n as buildWorkspaceSkillSnapshot, p as applySkillEnvOverridesFromSnapshot, r as loadWorkspaceSkillEntries, s as assertMediaNotDataUrl } from "./skills-DO7WNsVJ.js";
 import { n as redactToolDetail } from "./redact-DSOAcWMe.js";
@@ -29,12 +29,12 @@ import { n as resolveConversationLabel } from "./conversation-label-DTTqF8gH.js"
 import { i as resolveSessionTranscriptPath, n as resolveSessionFilePath, o as resolveSessionTranscriptsDirForAgent, r as resolveSessionFilePathOptions, s as resolveStorePath, t as resolveDefaultSessionStorePath } from "./paths-BsT3BvfH.js";
 import { t as emitSessionTranscriptUpdate } from "./transcript-events-ChU6IQwp.js";
 import { n as saveJsonFile, t as loadJsonFile } from "./json-file-B7D44OOV.js";
-import { a as parseGeminiAuth, c as resolveMemorySearchConfig, d as requireApiKey, f as resolveApiKeyForProvider, h as resolveModelAuthMode, i as resolveOllamaBaseUrl, l as getApiKeyForModel, m as resolveEnvApiKey, n as retryAsync, o as collectProviderApiKeysForExecution, p as resolveAwsSdkEnvVarName, r as probeOllamaEmbeddingModels, s as executeWithApiKeyRotation, u as getCustomProviderApiKey } from "./manager-rvtFoeFT.js";
+import { a as parseGeminiAuth, c as resolveMemorySearchConfig, d as requireApiKey, f as resolveApiKeyForProvider, h as resolveModelAuthMode, i as resolveOllamaBaseUrl, l as getApiKeyForModel, m as resolveEnvApiKey, n as retryAsync, o as collectProviderApiKeysForExecution, p as resolveAwsSdkEnvVarName, r as probeOllamaEmbeddingModels, s as executeWithApiKeyRotation, u as getCustomProviderApiKey } from "./manager-CnBU0aOR.js";
 import { r as resolveCopilotApiToken, t as DEFAULT_COPILOT_API_BASE_URL } from "./github-copilot-token-dYUr1mDQ.js";
 import { c as normalizeExtraMemoryPaths, f as runTasksWithConcurrency, s as listMemoryFiles } from "./internal-Ce-sg7EN.js";
 import { n as resolveMarkdownTableMode } from "./markdown-tables-jQzXAsf3.js";
-import { a as shouldHandleTextCommands, c as resolveNativeCommandsEnabled, i as normalizeCommandBody, l as resolveNativeSkillsEnabled, n as listChatCommands, o as isCommandFlagEnabled, r as listChatCommandsForConfig, s as isRestartEnabled } from "./commands-registry-DFYDtKr_.js";
-import { c as getSkillsSnapshotVersion, d as createAsyncLock, f as readJsonFile$1, i as resolveSkillCommandInvocation, l as pruneExpiredPending, n as listSkillCommandsForAgents, o as getRemoteSkillEligibility, p as writeJsonAtomic, r as listSkillCommandsForWorkspace, s as ensureSkillsWatcher, t as listReservedChatSlashCommandNames, u as resolvePairingPaths } from "./skill-commands-CKm7eCw6.js";
+import { a as shouldHandleTextCommands, c as resolveNativeCommandsEnabled, i as normalizeCommandBody, l as resolveNativeSkillsEnabled, n as listChatCommands, o as isCommandFlagEnabled, r as listChatCommandsForConfig, s as isRestartEnabled } from "./commands-registry-DG2kGyg5.js";
+import { c as getSkillsSnapshotVersion, d as createAsyncLock, f as readJsonFile$1, i as resolveSkillCommandInvocation, l as pruneExpiredPending, n as listSkillCommandsForAgents, o as getRemoteSkillEligibility, p as writeJsonAtomic, r as listSkillCommandsForWorkspace, s as ensureSkillsWatcher, t as listReservedChatSlashCommandNames, u as resolvePairingPaths } from "./skill-commands-D0xAWG0l.js";
 import { r as normalizeInboundTextNewlines, t as finalizeInboundContext } from "./inbound-context-M0vxsAoa.js";
 import { n as discoverModels, t as discoverAuthStorage } from "./pi-model-discovery-CuZX_Q4t.js";
 import { n as resolveAgentRoute, t as buildAgentSessionKey } from "./resolve-route-0kL4oMLp.js";
@@ -53,7 +53,7 @@ import { fileURLToPath } from "node:url";
 import { CURRENT_SESSION_VERSION, DefaultResourceLoader, SessionManager, SettingsManager, codingTools, createAgentSession, createEditTool, createReadTool, createWriteTool, estimateTokens, generateSummary, readTool } from "@mariozechner/pi-coding-agent";
 import crypto, { X509Certificate, createHash, createHmac, randomBytes, randomUUID } from "node:crypto";
 import AjvPkg from "ajv";
-import { WebSocket } from "ws";
+import { WebSocket as WebSocket$1 } from "ws";
 import { Buffer as Buffer$1 } from "node:buffer";
 import { complete, createAssistantMessageEventStream, streamSimple } from "@mariozechner/pi-ai";
 import { BedrockClient, ListFoundationModelsCommand } from "@aws-sdk/client-bedrock";
@@ -3390,7 +3390,7 @@ async function getMemorySearchManager(params) {
 				const wrapper = new FallbackMemoryManager({
 					primary,
 					fallbackFactory: async () => {
-						const { MemoryIndexManager } = await import("./manager-rvtFoeFT.js").then((n) => n.t);
+						const { MemoryIndexManager } = await import("./manager-CnBU0aOR.js").then((n) => n.t);
 						return await MemoryIndexManager.get(params);
 					}
 				}, () => QMD_MANAGER_CACHE.delete(cacheKey));
@@ -3403,7 +3403,7 @@ async function getMemorySearchManager(params) {
 		}
 	}
 	try {
-		const { MemoryIndexManager } = await import("./manager-rvtFoeFT.js").then((n) => n.t);
+		const { MemoryIndexManager } = await import("./manager-CnBU0aOR.js").then((n) => n.t);
 		return { manager: await MemoryIndexManager.get(params) };
 	} catch (err) {
 		return {
@@ -5920,7 +5920,7 @@ var GatewayClient = class {
 				if (fingerprint !== expected) return /* @__PURE__ */ new Error("gateway tls fingerprint mismatch");
 			});
 		}
-		this.ws = new WebSocket(url, wsOptions);
+		this.ws = new WebSocket$1(url, wsOptions);
 		this.ws.on("open", () => {
 			if (url.startsWith("wss://") && this.opts.tlsFingerprint) {
 				const tlsError = this.validateTlsFingerprint();
@@ -6139,7 +6139,7 @@ var GatewayClient = class {
 		return null;
 	}
 	async request(method, params, opts) {
-		if (!this.ws || this.ws.readyState !== WebSocket.OPEN) throw new Error("gateway not connected");
+		if (!this.ws || this.ws.readyState !== WebSocket$1.OPEN) throw new Error("gateway not connected");
 		const id = randomUUID();
 		const frame = {
 			type: "req",
@@ -7087,7 +7087,7 @@ async function routeReply(params) {
 	const resolvedReplyToId = replyToId ?? (channelId === "slack" && threadId != null && threadId !== "" ? String(threadId) : void 0);
 	const resolvedThreadId = channelId === "slack" ? null : threadId ?? null;
 	try {
-		const { deliverOutboundPayloads } = await import("./deliver-Cjyb6h4g.js").then((n) => n.n);
+		const { deliverOutboundPayloads } = await import("./deliver-f3cIWxXT.js").then((n) => n.n);
 		return {
 			ok: true,
 			messageId: (await deliverOutboundPayloads({
@@ -16084,13 +16084,21 @@ async function applyMediaUnderstanding(params) {
 * Soft warning appended when context usage crosses the threshold.
 * Fires once per session (tracked via `SessionEntry.contextWarningSent`).
 */
-const SESSION_CONTEXT_WARNING_MESSAGE = "📋 Heads up — this session is approaching the context limit. Start a new session for best results. Past sessions stay in history and can be revisited there.";
+const SESSION_CONTEXT_WARNING_MESSAGE = "📋 Heads up — this session is approaching the context limit. Symi will compact older context automatically when usage reaches 90%.";
 /**
-* Hard limit message returned when the model API rejects the turn for
-* context-too-long or role-ordering reasons, and on every subsequent
-* turn until the user starts a new session.
+* Inline status emitted by Symi before running compaction proactively
+* at the 90% threshold. Followed by the agent's actual response in the
+* same turn.
 */
-const SESSION_OVERFLOW_MESSAGE = "📋 This session has reached the context limit. Please start a new session — your full conversation is saved in history and can be revisited there.";
+const SESSION_COMPACTING_MESSAGE = "📋 Context limit reached for this session. Compacting older context and continuing…";
+/**
+* Returned when compaction can't recover enough budget to continue —
+* either because the proactive compaction at 90% failed, or because a
+* single user message overshot the window in one turn and the reactive
+* fallback compaction also failed. Same message also short-circuits
+* subsequent turns until /reset or /new.
+*/
+const SESSION_COMPACTION_FAILED_MESSAGE = "📋 Couldn't compact this session further. Please /reset or /new to continue — your full conversation is saved in history.";
 
 //#endregion
 //#region src/sessions/level-overrides.ts
@@ -24404,7 +24412,6 @@ function formatTokenCount$1(value) {
 function formatUsd(value) {
 	if (value === void 0 || !Number.isFinite(value)) return;
 	if (value >= 1) return `$${value.toFixed(2)}`;
-	if (value >= .01) return `$${value.toFixed(2)}`;
 	return `$${value.toFixed(4)}`;
 }
 function resolveModelCostConfig(params) {
@@ -24479,13 +24486,13 @@ function resolveRuntimeLabel(args) {
 		return sessionKey !== resolveMainSessionKey({ session: { scope: args.sessionScope ?? "per-sender" } }).trim();
 	})() ? "docker" : sessionKey ? "direct" : "unknown"}/${sandboxMode}`;
 }
-const formatTokens = (total, contextTokens) => {
+const formatTokens$1 = (total, contextTokens) => {
 	const ctx = contextTokens ?? null;
 	if (total == null) return `?/${ctx ? formatTokenCount(ctx) : "?"}`;
 	const pct = ctx ? Math.min(999, Math.round(total / ctx * 100)) : null;
 	return `${formatTokenCount(total)}/${ctx ? formatTokenCount(ctx) : "?"}${pct !== null ? ` (${pct}%)` : ""}`;
 };
-const formatContextUsageShort = (total, contextTokens) => `Context ${formatTokens(total, contextTokens ?? null)}`;
+const formatContextUsageShort = (total, contextTokens) => `Context ${formatTokens$1(total, contextTokens ?? null)}`;
 const formatQueueDetails = (queue) => {
 	if (!queue) return "";
 	const depth = typeof queue.depth === "number" ? `depth ${queue.depth}` : null;
@@ -24637,7 +24644,7 @@ function buildStatusMessage(args) {
 	const updatedAt = entry?.updatedAt;
 	const sessionLine = [`Session: ${args.sessionKey ?? "unknown"}`, typeof updatedAt === "number" ? `updated ${formatTimeAgo(now - updatedAt)}` : "no activity"].filter(Boolean).join(" • ");
 	const groupActivationValue = entry?.chatType === "group" || entry?.chatType === "channel" || Boolean(args.sessionKey?.includes(":group:")) || Boolean(args.sessionKey?.includes(":channel:")) ? args.groupActivation ?? entry?.groupActivation ?? "mention" : void 0;
-	const contextLine = [`Context: ${formatTokens(totalTokens, contextTokens ?? null)}`, `🧹 Compactions: ${entry?.compactionCount ?? 0}`].filter(Boolean).join(" · ");
+	const contextLine = [`Context: ${formatTokens$1(totalTokens, contextTokens ?? null)}`, `🧹 Compactions: ${entry?.compactionCount ?? 0}`].filter(Boolean).join(" · ");
 	const queueMode = args.queue?.mode ?? "unknown";
 	const queueDetails = formatQueueDetails(args.queue);
 	const verboseLabel = verboseLevel === "full" ? "verbose:full" : verboseLevel === "on" ? "verbose" : null;
@@ -24669,14 +24676,19 @@ function buildStatusMessage(args) {
 		config: args.config
 	}) : void 0;
 	const hasUsage = typeof inputTokens === "number" || typeof outputTokens === "number";
-	const cost = showCost && hasUsage ? estimateUsageCost({
+	const rawCost = showCost && hasUsage ? estimateUsageCost({
 		usage: {
 			input: inputTokens ?? void 0,
 			output: outputTokens ?? void 0
 		},
 		cost: costConfig
 	}) : void 0;
-	const costLabel = showCost && hasUsage ? formatUsd(cost) : void 0;
+	const useOverride = showCost && hasUsage && typeof args.costOverrideUsd === "number" && (rawCost === void 0 || rawCost === 0);
+	const cost = useOverride ? args.costOverrideUsd : rawCost;
+	const baseLabel = showCost && hasUsage ? formatUsd(cost) : void 0;
+	const tag = useOverride ? " (self-hosted)" : "";
+	const suffix = args.costLineSuffix ? ` · ${args.costLineSuffix}` : "";
+	const costLabel = baseLabel ? `${baseLabel}${tag}${suffix}` : void 0;
 	const selectedAuthLabel = selectedAuthLabelValue ? ` · 🔑 ${selectedAuthLabelValue}` : "";
 	const channelModelNote = (() => {
 		if (!args.config || !entry) return;
@@ -24721,6 +24733,8 @@ function buildStatusMessage(args) {
 		fallbackLine,
 		usageCostLine,
 		cacheLine,
+		args.engineLine,
+		args.hardwareLine,
 		`📚 ${contextLine}`,
 		mediaLine,
 		args.usageLine,
@@ -25297,6 +25311,606 @@ const handleDebugCommand = async (params, allowTextCommands) => {
 	return null;
 };
 
+//#endregion
+//#region src/infra/session-cost-usage.ts
+const emptyTotals = () => ({
+	input: 0,
+	output: 0,
+	cacheRead: 0,
+	cacheWrite: 0,
+	totalTokens: 0,
+	totalCost: 0,
+	inputCost: 0,
+	outputCost: 0,
+	cacheReadCost: 0,
+	cacheWriteCost: 0,
+	missingCostEntries: 0
+});
+const toFiniteNumber = (value) => {
+	if (typeof value !== "number") return;
+	if (!Number.isFinite(value)) return;
+	return value;
+};
+const extractCostBreakdown = (usageRaw) => {
+	if (!usageRaw || typeof usageRaw !== "object") return;
+	const cost = usageRaw.cost;
+	if (!cost) return;
+	const total = toFiniteNumber(cost.total);
+	if (total === void 0 || total < 0) return;
+	return {
+		total,
+		input: toFiniteNumber(cost.input),
+		output: toFiniteNumber(cost.output),
+		cacheRead: toFiniteNumber(cost.cacheRead),
+		cacheWrite: toFiniteNumber(cost.cacheWrite)
+	};
+};
+const parseTimestamp = (entry) => {
+	const raw = entry.timestamp;
+	if (typeof raw === "string") {
+		const parsed = new Date(raw);
+		if (!Number.isNaN(parsed.valueOf())) return parsed;
+	}
+	const message = entry.message;
+	const messageTimestamp = toFiniteNumber(message?.timestamp);
+	if (messageTimestamp !== void 0) {
+		const parsed = new Date(messageTimestamp);
+		if (!Number.isNaN(parsed.valueOf())) return parsed;
+	}
+};
+const parseTranscriptEntry = (entry) => {
+	const message = entry.message;
+	if (!message || typeof message !== "object") return null;
+	const roleRaw = message.role;
+	const role = roleRaw === "user" || roleRaw === "assistant" ? roleRaw : void 0;
+	if (!role) return null;
+	const usageRaw = message.usage ?? entry.usage;
+	const usage = usageRaw ? normalizeUsage(usageRaw) ?? void 0 : void 0;
+	const provider = (typeof message.provider === "string" ? message.provider : void 0) ?? (typeof entry.provider === "string" ? entry.provider : void 0);
+	const model = (typeof message.model === "string" ? message.model : void 0) ?? (typeof entry.model === "string" ? entry.model : void 0);
+	const costBreakdown = extractCostBreakdown(usageRaw);
+	const stopReason = typeof message.stopReason === "string" ? message.stopReason : void 0;
+	const durationMs = toFiniteNumber(message.durationMs ?? entry.durationMs);
+	return {
+		message,
+		role,
+		timestamp: parseTimestamp(entry),
+		durationMs,
+		usage,
+		costTotal: costBreakdown?.total,
+		costBreakdown,
+		provider,
+		model,
+		stopReason,
+		toolNames: extractToolCallNames(message),
+		toolResultCounts: countToolResults(message)
+	};
+};
+const formatDayKey = (date) => date.toLocaleDateString("en-CA", { timeZone: Intl.DateTimeFormat().resolvedOptions().timeZone });
+const computeLatencyStats = (values) => {
+	if (!values.length) return;
+	const sorted = values.toSorted((a, b) => a - b);
+	const total = sorted.reduce((sum, v) => sum + v, 0);
+	const count = sorted.length;
+	const p95Index = Math.max(0, Math.ceil(count * .95) - 1);
+	return {
+		count,
+		avgMs: total / count,
+		p95Ms: sorted[p95Index] ?? sorted[count - 1],
+		minMs: sorted[0],
+		maxMs: sorted[count - 1]
+	};
+};
+const applyUsageTotals = (totals, usage) => {
+	totals.input += usage.input ?? 0;
+	totals.output += usage.output ?? 0;
+	totals.cacheRead += usage.cacheRead ?? 0;
+	totals.cacheWrite += usage.cacheWrite ?? 0;
+	const totalTokens = usage.total ?? (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0);
+	totals.totalTokens += totalTokens;
+};
+const applyCostBreakdown = (totals, costBreakdown) => {
+	if (costBreakdown === void 0 || costBreakdown.total === void 0) return;
+	totals.totalCost += costBreakdown.total;
+	totals.inputCost += costBreakdown.input ?? 0;
+	totals.outputCost += costBreakdown.output ?? 0;
+	totals.cacheReadCost += costBreakdown.cacheRead ?? 0;
+	totals.cacheWriteCost += costBreakdown.cacheWrite ?? 0;
+};
+const applyCostTotal = (totals, costTotal) => {
+	if (costTotal === void 0) {
+		totals.missingCostEntries += 1;
+		return;
+	}
+	totals.totalCost += costTotal;
+};
+async function* readJsonlRecords(filePath) {
+	const fileStream = fs.createReadStream(filePath, { encoding: "utf-8" });
+	const rl = readline.createInterface({
+		input: fileStream,
+		crlfDelay: Infinity
+	});
+	try {
+		for await (const line of rl) {
+			const trimmed = line.trim();
+			if (!trimmed) continue;
+			try {
+				const parsed = JSON.parse(trimmed);
+				if (!parsed || typeof parsed !== "object") continue;
+				yield parsed;
+			} catch {}
+		}
+	} finally {
+		rl.close();
+		fileStream.destroy();
+	}
+}
+async function scanTranscriptFile(params) {
+	for await (const parsed of readJsonlRecords(params.filePath)) {
+		const entry = parseTranscriptEntry(parsed);
+		if (!entry) continue;
+		if (entry.usage && entry.costTotal === void 0) {
+			const cost = resolveModelCostConfig({
+				provider: entry.provider,
+				model: entry.model,
+				config: params.config
+			});
+			entry.costTotal = estimateUsageCost({
+				usage: entry.usage,
+				cost
+			});
+		}
+		params.onEntry(entry);
+	}
+}
+async function scanUsageFile(params) {
+	await scanTranscriptFile({
+		filePath: params.filePath,
+		config: params.config,
+		onEntry: (entry) => {
+			if (!entry.usage) return;
+			params.onEntry({
+				usage: entry.usage,
+				costTotal: entry.costTotal,
+				costBreakdown: entry.costBreakdown,
+				provider: entry.provider,
+				model: entry.model,
+				timestamp: entry.timestamp
+			});
+		}
+	});
+}
+async function loadCostUsageSummary(params) {
+	const now = /* @__PURE__ */ new Date();
+	let sinceTime;
+	let untilTime;
+	if (params?.startMs !== void 0 && params?.endMs !== void 0) {
+		sinceTime = params.startMs;
+		untilTime = params.endMs;
+	} else {
+		const days = Math.max(1, Math.floor(params?.days ?? 30));
+		const since = new Date(now);
+		since.setDate(since.getDate() - (days - 1));
+		sinceTime = since.getTime();
+		untilTime = now.getTime();
+	}
+	const dailyMap = /* @__PURE__ */ new Map();
+	const totals = emptyTotals();
+	const sessionsDir = resolveSessionTranscriptsDirForAgent(params?.agentId);
+	const entries = await fs.promises.readdir(sessionsDir, { withFileTypes: true }).catch(() => []);
+	const files = (await Promise.all(entries.filter((entry) => entry.isFile() && entry.name.endsWith(".jsonl")).map(async (entry) => {
+		const filePath = path.join(sessionsDir, entry.name);
+		const stats = await fs.promises.stat(filePath).catch(() => null);
+		if (!stats) return null;
+		if (stats.mtimeMs < sinceTime) return null;
+		return filePath;
+	}))).filter((filePath) => Boolean(filePath));
+	for (const filePath of files) await scanUsageFile({
+		filePath,
+		config: params?.config,
+		onEntry: (entry) => {
+			const ts = entry.timestamp?.getTime();
+			if (!ts || ts < sinceTime || ts > untilTime) return;
+			const dayKey = formatDayKey(entry.timestamp ?? now);
+			const bucket = dailyMap.get(dayKey) ?? emptyTotals();
+			applyUsageTotals(bucket, entry.usage);
+			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(bucket, entry.costBreakdown);
+			else applyCostTotal(bucket, entry.costTotal);
+			dailyMap.set(dayKey, bucket);
+			applyUsageTotals(totals, entry.usage);
+			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(totals, entry.costBreakdown);
+			else applyCostTotal(totals, entry.costTotal);
+		}
+	});
+	const daily = Array.from(dailyMap.entries()).map(([date, bucket]) => Object.assign({ date }, bucket)).toSorted((a, b) => a.date.localeCompare(b.date));
+	const days = Math.ceil((untilTime - sinceTime) / (1440 * 60 * 1e3)) + 1;
+	return {
+		updatedAt: Date.now(),
+		days,
+		daily,
+		totals
+	};
+}
+async function loadSessionCostSummary(params) {
+	const sessionFile = params.sessionFile ?? (params.sessionId ? resolveSessionFilePath(params.sessionId, params.sessionEntry, { agentId: params.agentId }) : void 0);
+	if (!sessionFile || !fs.existsSync(sessionFile)) return null;
+	const totals = emptyTotals();
+	let firstActivity;
+	let lastActivity;
+	const activityDatesSet = /* @__PURE__ */ new Set();
+	const dailyMap = /* @__PURE__ */ new Map();
+	const dailyMessageMap = /* @__PURE__ */ new Map();
+	const dailyLatencyMap = /* @__PURE__ */ new Map();
+	const dailyModelUsageMap = /* @__PURE__ */ new Map();
+	const messageCounts = {
+		total: 0,
+		user: 0,
+		assistant: 0,
+		toolCalls: 0,
+		toolResults: 0,
+		errors: 0
+	};
+	const toolUsageMap = /* @__PURE__ */ new Map();
+	const modelUsageMap = /* @__PURE__ */ new Map();
+	const errorStopReasons = new Set([
+		"error",
+		"aborted",
+		"timeout"
+	]);
+	const latencyValues = [];
+	let lastUserTimestamp;
+	const MAX_LATENCY_MS = 720 * 60 * 1e3;
+	await scanTranscriptFile({
+		filePath: sessionFile,
+		config: params.config,
+		onEntry: (entry) => {
+			const ts = entry.timestamp?.getTime();
+			if (params.startMs !== void 0 && ts !== void 0 && ts < params.startMs) return;
+			if (params.endMs !== void 0 && ts !== void 0 && ts > params.endMs) return;
+			if (ts !== void 0) {
+				if (!firstActivity || ts < firstActivity) firstActivity = ts;
+				if (!lastActivity || ts > lastActivity) lastActivity = ts;
+			}
+			if (entry.role === "user") {
+				messageCounts.user += 1;
+				messageCounts.total += 1;
+				if (entry.timestamp) lastUserTimestamp = entry.timestamp.getTime();
+			}
+			if (entry.role === "assistant") {
+				messageCounts.assistant += 1;
+				messageCounts.total += 1;
+				const ts = entry.timestamp?.getTime();
+				if (ts !== void 0) {
+					const latencyMs = entry.durationMs ?? (lastUserTimestamp !== void 0 ? Math.max(0, ts - lastUserTimestamp) : void 0);
+					if (latencyMs !== void 0 && Number.isFinite(latencyMs) && latencyMs <= MAX_LATENCY_MS) {
+						latencyValues.push(latencyMs);
+						const dayKey = formatDayKey(entry.timestamp ?? new Date(ts));
+						const dailyLatencies = dailyLatencyMap.get(dayKey) ?? [];
+						dailyLatencies.push(latencyMs);
+						dailyLatencyMap.set(dayKey, dailyLatencies);
+					}
+				}
+			}
+			if (entry.toolNames.length > 0) {
+				messageCounts.toolCalls += entry.toolNames.length;
+				for (const name of entry.toolNames) toolUsageMap.set(name, (toolUsageMap.get(name) ?? 0) + 1);
+			}
+			if (entry.toolResultCounts.total > 0) {
+				messageCounts.toolResults += entry.toolResultCounts.total;
+				messageCounts.errors += entry.toolResultCounts.errors;
+			}
+			if (entry.stopReason && errorStopReasons.has(entry.stopReason)) messageCounts.errors += 1;
+			if (entry.timestamp) {
+				const dayKey = formatDayKey(entry.timestamp);
+				activityDatesSet.add(dayKey);
+				const daily = dailyMessageMap.get(dayKey) ?? {
+					date: dayKey,
+					total: 0,
+					user: 0,
+					assistant: 0,
+					toolCalls: 0,
+					toolResults: 0,
+					errors: 0
+				};
+				daily.total += entry.role === "user" || entry.role === "assistant" ? 1 : 0;
+				if (entry.role === "user") daily.user += 1;
+				else if (entry.role === "assistant") daily.assistant += 1;
+				daily.toolCalls += entry.toolNames.length;
+				daily.toolResults += entry.toolResultCounts.total;
+				daily.errors += entry.toolResultCounts.errors;
+				if (entry.stopReason && errorStopReasons.has(entry.stopReason)) daily.errors += 1;
+				dailyMessageMap.set(dayKey, daily);
+			}
+			if (!entry.usage) return;
+			applyUsageTotals(totals, entry.usage);
+			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(totals, entry.costBreakdown);
+			else applyCostTotal(totals, entry.costTotal);
+			if (entry.timestamp) {
+				const dayKey = formatDayKey(entry.timestamp);
+				const entryTokens = (entry.usage.input ?? 0) + (entry.usage.output ?? 0) + (entry.usage.cacheRead ?? 0) + (entry.usage.cacheWrite ?? 0);
+				const entryCost = entry.costBreakdown?.total ?? (entry.costBreakdown ? (entry.costBreakdown.input ?? 0) + (entry.costBreakdown.output ?? 0) + (entry.costBreakdown.cacheRead ?? 0) + (entry.costBreakdown.cacheWrite ?? 0) : entry.costTotal ?? 0);
+				const existing = dailyMap.get(dayKey) ?? {
+					tokens: 0,
+					cost: 0
+				};
+				dailyMap.set(dayKey, {
+					tokens: existing.tokens + entryTokens,
+					cost: existing.cost + entryCost
+				});
+				if (entry.provider || entry.model) {
+					const modelKey = `${dayKey}::${entry.provider ?? "unknown"}::${entry.model ?? "unknown"}`;
+					const dailyModel = dailyModelUsageMap.get(modelKey) ?? {
+						date: dayKey,
+						provider: entry.provider,
+						model: entry.model,
+						tokens: 0,
+						cost: 0,
+						count: 0
+					};
+					dailyModel.tokens += entryTokens;
+					dailyModel.cost += entryCost;
+					dailyModel.count += 1;
+					dailyModelUsageMap.set(modelKey, dailyModel);
+				}
+			}
+			if (entry.provider || entry.model) {
+				const key = `${entry.provider ?? "unknown"}::${entry.model ?? "unknown"}`;
+				const existing = modelUsageMap.get(key) ?? {
+					provider: entry.provider,
+					model: entry.model,
+					count: 0,
+					totals: emptyTotals()
+				};
+				existing.count += 1;
+				applyUsageTotals(existing.totals, entry.usage);
+				if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(existing.totals, entry.costBreakdown);
+				else applyCostTotal(existing.totals, entry.costTotal);
+				modelUsageMap.set(key, existing);
+			}
+		}
+	});
+	const dailyBreakdown = Array.from(dailyMap.entries()).map(([date, data]) => ({
+		date,
+		tokens: data.tokens,
+		cost: data.cost
+	})).toSorted((a, b) => a.date.localeCompare(b.date));
+	const dailyMessageCounts = Array.from(dailyMessageMap.values()).toSorted((a, b) => a.date.localeCompare(b.date));
+	const dailyLatency = Array.from(dailyLatencyMap.entries()).map(([date, values]) => {
+		const stats = computeLatencyStats(values);
+		if (!stats) return null;
+		return {
+			date,
+			...stats
+		};
+	}).filter((entry) => Boolean(entry)).toSorted((a, b) => a.date.localeCompare(b.date));
+	const dailyModelUsage = Array.from(dailyModelUsageMap.values()).toSorted((a, b) => a.date.localeCompare(b.date) || b.cost - a.cost);
+	const toolUsage = toolUsageMap.size ? {
+		totalCalls: Array.from(toolUsageMap.values()).reduce((sum, count) => sum + count, 0),
+		uniqueTools: toolUsageMap.size,
+		tools: Array.from(toolUsageMap.entries()).map(([name, count]) => ({
+			name,
+			count
+		})).toSorted((a, b) => b.count - a.count)
+	} : void 0;
+	const modelUsage = modelUsageMap.size ? Array.from(modelUsageMap.values()).toSorted((a, b) => {
+		const costDiff = b.totals.totalCost - a.totals.totalCost;
+		if (costDiff !== 0) return costDiff;
+		return b.totals.totalTokens - a.totals.totalTokens;
+	}) : void 0;
+	return {
+		sessionId: params.sessionId,
+		sessionFile,
+		firstActivity,
+		lastActivity,
+		durationMs: firstActivity !== void 0 && lastActivity !== void 0 ? Math.max(0, lastActivity - firstActivity) : void 0,
+		activityDates: Array.from(activityDatesSet).toSorted(),
+		dailyBreakdown,
+		dailyMessageCounts,
+		dailyLatency: dailyLatency.length ? dailyLatency : void 0,
+		dailyModelUsage: dailyModelUsage.length ? dailyModelUsage : void 0,
+		messageCounts,
+		toolUsage,
+		modelUsage,
+		latency: computeLatencyStats(latencyValues),
+		...totals
+	};
+}
+
+//#endregion
+//#region src/cost/benchmark.ts
+/**
+* Claude reference rates and benchmark comparison.
+*
+* Hardcoded per-token USD rates for Claude Opus 4.7, Sonnet 4.6, and
+* Haiku 4.5 (per million tokens, Anthropic public list as of 2026-Q1).
+* Override via `models.gpuRates` is for GPU rates only — Claude rates can
+* be overridden via the per-model `cost: { input, output }` configuration
+* on the user's `anthropic` provider entry, which already exists.
+*
+* Display: pick a target tier via `agents.defaults.cost.benchmark`, default
+* `sonnet`. Setting `"all"` shows comparisons against all three tiers.
+*/
+/** Per-million-token USD rates, applied via `estimateUsageCost`. */
+const CLAUDE_RATES = {
+	opus: {
+		input: 15,
+		output: 75,
+		cacheRead: 1.5,
+		cacheWrite: 18.75
+	},
+	sonnet: {
+		input: 3,
+		output: 15,
+		cacheRead: .3,
+		cacheWrite: 3.75
+	},
+	haiku: {
+		input: .8,
+		output: 4,
+		cacheRead: .08,
+		cacheWrite: 1
+	}
+};
+const CLAUDE_LABEL = {
+	opus: "Claude Opus 4.7",
+	sonnet: "Claude Sonnet 4.6",
+	haiku: "Claude Haiku 4.5"
+};
+function benchmarkUsage(args) {
+	if (!args.usage) return [];
+	const tiers = args.target === "all" ? [
+		"opus",
+		"sonnet",
+		"haiku"
+	] : [args.target];
+	const results = [];
+	for (const tier of tiers) {
+		const cost = estimateUsageCost({
+			usage: args.usage,
+			cost: CLAUDE_RATES[tier]
+		});
+		if (cost === void 0) continue;
+		results.push({
+			tier,
+			costUsd: cost
+		});
+	}
+	return results;
+}
+function formatBenchmarkSuffix(args) {
+	if (args.benchmarks.length === 0) return;
+	if (args.benchmarks.length === 1) {
+		const b = args.benchmarks[0];
+		return formatPair(args.ourCostUsd, b);
+	}
+	const parts = [];
+	for (const b of args.benchmarks) parts.push(`${shortLabel(b.tier)} $${formatCostShort(b.costUsd)}`);
+	const cheapest = args.benchmarks.reduce((min, b) => b.costUsd < min.costUsd ? b : min);
+	if (cheapest.costUsd > args.ourCostUsd && cheapest.costUsd > 0) {
+		const savedPct = (cheapest.costUsd - args.ourCostUsd) / cheapest.costUsd * 100;
+		if (savedPct >= 1) parts.push(`save ${savedPct.toFixed(0)}% vs ${shortLabel(cheapest.tier)}`);
+	}
+	return `vs ${parts.join(" / ")}`;
+}
+function formatPair(ourCost, b) {
+	const benchPart = `${CLAUDE_LABEL[b.tier]} would be $${formatCostShort(b.costUsd)}`;
+	if (b.costUsd > 0 && b.costUsd > ourCost) {
+		const saved = (b.costUsd - ourCost) / b.costUsd * 100;
+		if (saved >= 1) return `${benchPart} (save ${saved.toFixed(0)}%)`;
+	} else if (b.costUsd > 0 && ourCost > b.costUsd) {
+		const overspend = (ourCost - b.costUsd) / b.costUsd * 100;
+		if (overspend >= 1) return `${benchPart} (+${overspend.toFixed(0)}%)`;
+	}
+	return benchPart;
+}
+function shortLabel(tier) {
+	if (tier === "opus") return "Opus";
+	if (tier === "sonnet") return "Sonnet";
+	return "Haiku";
+}
+function formatCostShort(value) {
+	if (value >= 1) return value.toFixed(2);
+	return value.toFixed(4);
+}
+function isClaudeTier(value) {
+	return value === "opus" || value === "sonnet" || value === "haiku";
+}
+function isBenchmarkTarget(value) {
+	return isClaudeTier(value) || value === "all";
+}
+function resolveBenchmarkTarget(value) {
+	if (typeof value === "string" && isBenchmarkTarget(value)) return value;
+	return "sonnet";
+}
+
+//#endregion
+//#region src/cost/summary.ts
+async function loadCostBenchmarkSummary(params) {
+	const summary = await loadCostUsageSummary(params);
+	const daily = summary.daily.map((day) => ({
+		...day,
+		benchmarkUsd: computeBenchmarkPerTier(day)
+	}));
+	const benchmarkTotals = computeBenchmarkPerTier(summary.totals);
+	const savingsTotals = {
+		opus: benchmarkTotals.opus - summary.totals.totalCost,
+		sonnet: benchmarkTotals.sonnet - summary.totals.totalCost,
+		haiku: benchmarkTotals.haiku - summary.totals.totalCost
+	};
+	return {
+		...summary,
+		daily,
+		benchmarkTotals,
+		savingsTotals
+	};
+}
+function computeBenchmarkPerTier(totals) {
+	const all = benchmarkUsage({
+		usage: {
+			input: totals.input,
+			output: totals.output,
+			cacheRead: totals.cacheRead,
+			cacheWrite: totals.cacheWrite
+		},
+		target: "all"
+	});
+	const map = new Map(all.map((b) => [b.tier, b.costUsd]));
+	return {
+		opus: map.get("opus") ?? 0,
+		sonnet: map.get("sonnet") ?? 0,
+		haiku: map.get("haiku") ?? 0
+	};
+}
+function resolveWindow(window, now = Date.now()) {
+	const end = now;
+	switch (window) {
+		case "today": {
+			const d = new Date(end);
+			d.setHours(0, 0, 0, 0);
+			return {
+				startMs: d.getTime(),
+				endMs: end
+			};
+		}
+		case "week": return {
+			startMs: end - 10080 * 6e4,
+			endMs: end
+		};
+		case "month": return {
+			startMs: end - 720 * 60 * 6e4,
+			endMs: end
+		};
+		case "all": return {
+			startMs: 0,
+			endMs: end
+		};
+	}
+}
+function formatSummaryReport(summary, label) {
+	const t = summary.totals;
+	const b = summary.benchmarkTotals;
+	const lines = [];
+	if (label) lines.push(`# ${label}`);
+	lines.push(`Tokens:  ${formatTokens(t.totalTokens)} (${formatTokens(t.input)} in / ${formatTokens(t.output)} out)`);
+	lines.push(`Spend:   $${t.totalCost.toFixed(4)}`);
+	lines.push(`Claude equivalents:`);
+	lines.push(`  Opus:   $${b.opus.toFixed(4)}  (saved $${(b.opus - t.totalCost).toFixed(4)})`);
+	lines.push(`  Sonnet: $${b.sonnet.toFixed(4)}  (saved $${(b.sonnet - t.totalCost).toFixed(4)})`);
+	lines.push(`  Haiku:  $${b.haiku.toFixed(4)}  (saved $${(b.haiku - t.totalCost).toFixed(4)})`);
+	if (summary.daily.length > 0) {
+		lines.push("");
+		lines.push("Daily:");
+		for (const d of summary.daily.slice(-7)) lines.push(`  ${d.date}: ${formatTokens(d.totalTokens).padStart(8)} tokens · $${d.totalCost.toFixed(4).padStart(8)} · vs Sonnet $${d.benchmarkUsd.sonnet.toFixed(4)}`);
+	}
+	return lines.join("\n");
+}
+function formatTokens(n) {
+	if (!Number.isFinite(n) || n <= 0) return "0";
+	if (n >= 1e6) return `${(n / 1e6).toFixed(1)}m`;
+	if (n >= 1e3) return `${(n / 1e3).toFixed(n >= 1e4 ? 0 : 1)}k`;
+	return String(Math.round(n));
+}
+
 //#endregion
 //#region src/agents/system-prompt-report.ts
 function extractBetween(input, startMarker, endMarker) {
@@ -26417,6 +27031,560 @@ async function buildExportSessionReply(params) {
 	].join("\n") };
 }
 
+//#endregion
+//#region src/cost/rates.ts
+const GPU_VRAM_GB = {
+	"a100-40gb": 40,
+	"a100-80gb": 80,
+	"h100-pcie": 80,
+	"h100-sxm": 80,
+	h200: 141,
+	l40: 48,
+	l40s: 48,
+	a40: 48,
+	a6000: 48,
+	a5000: 24,
+	v100: 32,
+	mi300x: 192,
+	"rtx-4090": 24
+};
+const GPU_LABEL = {
+	"a100-40gb": "A100-40GB",
+	"a100-80gb": "A100-80GB",
+	"h100-pcie": "H100-PCIe",
+	"h100-sxm": "H100-SXM",
+	h200: "H200",
+	l40: "L40",
+	l40s: "L40S",
+	a40: "A40",
+	a6000: "RTX-A6000",
+	a5000: "RTX-A5000",
+	v100: "V100",
+	mi300x: "MI300X",
+	"rtx-4090": "RTX-4090"
+};
+const DEFAULT_HOURLY_RATES_USD = {
+	"a100-40gb": 2.06,
+	"a100-80gb": 2.21,
+	"h100-pcie": 4.25,
+	"h100-sxm": 4.76,
+	h200: 6.5,
+	l40: 1.27,
+	l40s: 1.35,
+	a40: .95,
+	a6000: .96,
+	a5000: .61,
+	v100: .8,
+	mi300x: 4.89,
+	"rtx-4090": .74
+};
+/**
+* Rough decode (output) throughput defaults in tokens/sec for a 27-30B FP8
+* model on a single GPU of each type. Scales roughly inversely with model
+* parameter count and bytes-per-param. Used as a fallback when no live
+* `vllm:generation_tokens_total` delta has been observed yet.
+*
+* These are conservative published-benchmark medians; treat as ballpark.
+* Override via `models.gpuRates` is rate-only; throughput overrides are
+* Phase 4+ (not yet wired to config).
+*/
+const DEFAULT_DECODE_TPS = {
+	"a100-40gb": 60,
+	"a100-80gb": 80,
+	"h100-pcie": 120,
+	"h100-sxm": 130,
+	h200: 160,
+	l40: 45,
+	l40s: 50,
+	a40: 40,
+	a6000: 35,
+	a5000: 25,
+	v100: 25,
+	mi300x: 110,
+	"rtx-4090": 55
+};
+/**
+* Prefill throughput is typically 30-60× higher than decode for a given
+* setup. Use 40× as a balanced midpoint.
+*/
+const PREFILL_DECODE_RATIO = 40;
+function resolveDefaultDecodeTps(gpuType, paramsB, bytesPerParam) {
+	const baseline = DEFAULT_DECODE_TPS[gpuType];
+	if (!paramsB) return baseline;
+	const refBytes = 27;
+	const modelBytes = paramsB * (bytesPerParam ?? 2);
+	if (modelBytes <= 0) return baseline;
+	const scaled = baseline * (refBytes / modelBytes);
+	return Math.max(5, scaled);
+}
+function resolveHourlyRate(gpuType, overrides) {
+	if (overrides && typeof overrides[gpuType] === "number" && overrides[gpuType] >= 0) return overrides[gpuType];
+	return DEFAULT_HOURLY_RATES_USD[gpuType];
+}
+
+//#endregion
+//#region src/cost/gpu-inference.ts
+/**
+* Heuristic GPU type + count inference for self-hosted serving endpoints.
+*
+* Two signals: the baseUrl hostname (often encodes "a100", "h100", etc.) and
+* the model id (encodes parameter count + quantization, e.g. "qwen36-27b-fp8").
+* Output is a confidence-tagged guess so downstream code (and the user) can
+* decide whether to trust the inference or override it.
+*
+* Bytes-per-parameter table:
+*   fp32        → 4
+*   fp16/bf16   → 2 (default when unspecified)
+*   fp8/int8/q8 → 1
+*   int4/q4/fp4 → 0.5
+*
+* GPU count formula:
+*   model_bytes = params × bytes_per_param
+*   target_util = 0.6 (40% headroom for KV cache, activations, scheduler)
+*   count       = max(1, ceil(model_bytes / (vram × target_util)))
+*/
+const URL_GPU_PATTERNS = [
+	{
+		re: /h200/i,
+		type: "h200",
+		confidence: "high"
+	},
+	{
+		re: /h100[-_]?pcie/i,
+		type: "h100-pcie",
+		confidence: "high"
+	},
+	{
+		re: /h100[-_]?sxm/i,
+		type: "h100-sxm",
+		confidence: "high"
+	},
+	{
+		re: /h100/i,
+		type: "h100-sxm",
+		confidence: "medium"
+	},
+	{
+		re: /a100[-_]?80/i,
+		type: "a100-80gb",
+		confidence: "high"
+	},
+	{
+		re: /a100[-_]?40/i,
+		type: "a100-40gb",
+		confidence: "high"
+	},
+	{
+		re: /a100/i,
+		type: "a100-80gb",
+		confidence: "medium"
+	},
+	{
+		re: /l40s/i,
+		type: "l40s",
+		confidence: "high"
+	},
+	{
+		re: /l40/i,
+		type: "l40",
+		confidence: "high"
+	},
+	{
+		re: /mi300x/i,
+		type: "mi300x",
+		confidence: "high"
+	},
+	{
+		re: /a6000/i,
+		type: "a6000",
+		confidence: "high"
+	},
+	{
+		re: /a5000/i,
+		type: "a5000",
+		confidence: "high"
+	},
+	{
+		re: /a40/i,
+		type: "a40",
+		confidence: "high"
+	},
+	{
+		re: /v100/i,
+		type: "v100",
+		confidence: "high"
+	},
+	{
+		re: /rtx[-_]?4090|gpu[-_]?4090/i,
+		type: "rtx-4090",
+		confidence: "high"
+	}
+];
+function inferGpuFromUrl(baseUrl) {
+	const url = baseUrl.toLowerCase();
+	for (const pattern of URL_GPU_PATTERNS) {
+		const match = url.match(pattern.re);
+		if (match) return {
+			gpuType: pattern.type,
+			confidence: pattern.confidence,
+			match: match[0]
+		};
+	}
+	return { confidence: "low" };
+}
+function parseModelSize(modelId) {
+	const id = modelId.toLowerCase();
+	const sizeMatch = id.match(/(\d+(?:\.\d+)?)b(?:[-_./]|$)/);
+	const paramsB = sizeMatch ? Number.parseFloat(sizeMatch[1]) : void 0;
+	let bytesPerParam;
+	if (/\b(fp4|q4|int4)\b/.test(id) || /[-_]q4[-_]/.test(id)) bytesPerParam = .5;
+	else if (/\b(fp8|int8|q8|w8a8|w8a16)\b/.test(id) || /[-_]fp8[-_]?/.test(id)) bytesPerParam = 1;
+	else if (/\b(fp16|bf16|float16|half)\b/.test(id)) bytesPerParam = 2;
+	else if (/\b(fp32|float32)\b/.test(id)) bytesPerParam = 4;
+	return {
+		paramsB,
+		bytesPerParam
+	};
+}
+function inferGpuCount(args) {
+	const util = args.targetUtilization ?? .6;
+	if (!args.paramsB || !args.bytesPerParam) return 1;
+	const modelBytes = args.paramsB * 1e9 * args.bytesPerParam;
+	const usableVram = args.vramGb * 1e9 * util;
+	return Math.max(1, Math.ceil(modelBytes / usableVram));
+}
+function inferGpu(args) {
+	const reasoning = [];
+	const overrideType = args.override?.gpuType;
+	const overrideCount = args.override?.gpuCount;
+	let gpuType;
+	let confidence = "low";
+	if (overrideType && overrideType in GPU_VRAM_GB) {
+		gpuType = overrideType;
+		confidence = "high";
+		reasoning.push(`gpuType=${gpuType} from config override`);
+	} else if (args.baseUrl) {
+		const fromUrl = inferGpuFromUrl(args.baseUrl);
+		if (fromUrl.gpuType) {
+			gpuType = fromUrl.gpuType;
+			confidence = fromUrl.confidence;
+			reasoning.push(`gpuType=${gpuType} from URL pattern '${fromUrl.match}'`);
+		} else reasoning.push("no GPU type pattern matched in URL");
+	}
+	const sized = args.modelId ? parseModelSize(args.modelId) : {};
+	if (sized.paramsB) reasoning.push(`paramsB=${sized.paramsB} from model id '${args.modelId}'`);
+	if (sized.bytesPerParam) reasoning.push(`bytesPerParam=${sized.bytesPerParam} from quantization marker`);
+	else if (args.modelId) reasoning.push("no quantization marker — defaulting to bf16 (2 bytes)");
+	let gpuCount = 1;
+	if (typeof overrideCount === "number" && overrideCount > 0) {
+		gpuCount = Math.floor(overrideCount);
+		reasoning.push(`gpuCount=${gpuCount} from config override`);
+	} else if (gpuType) {
+		const bpp = sized.bytesPerParam ?? 2;
+		gpuCount = inferGpuCount({
+			paramsB: sized.paramsB,
+			bytesPerParam: bpp,
+			vramGb: GPU_VRAM_GB[gpuType]
+		});
+		if (sized.paramsB) reasoning.push(`gpuCount=${gpuCount} (model ${(sized.paramsB * bpp).toFixed(0)}GB / ${GPU_VRAM_GB[gpuType]}GB × 0.6 util)`);
+	}
+	return {
+		gpuType,
+		gpuCount,
+		confidence,
+		paramsB: sized.paramsB,
+		bytesPerParam: sized.bytesPerParam,
+		reasoning
+	};
+}
+
+//#endregion
+//#region src/cost/serving-discovery.ts
+const SUCCESS_TTL_MS = 6e4;
+const NULL_TTL_MS = 5 * 6e4;
+const DEFAULT_TIMEOUT_MS$1 = 4e3;
+const cache = /* @__PURE__ */ new Map();
+const previousSnapshots = /* @__PURE__ */ new Map();
+async function probeServingEndpoint(baseUrl, opts = {}) {
+	const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
+	const nowFn = opts.now ?? (() => Date.now());
+	const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS$1;
+	const trimmed = baseUrl.trim().replace(/\/+$/, "");
+	if (!trimmed) return null;
+	const cached = cache.get(trimmed);
+	if (cached && !opts.forceRefresh && cached.expiresAt > nowFn()) return cached.info;
+	const rootUrl = trimmed.replace(/\/v\d+$/, "");
+	const probeAt = nowFn();
+	const fetchOpts = () => ({
+		signal: AbortSignal.timeout(timeoutMs),
+		headers: { Accept: "application/json, text/plain, */*" }
+	});
+	const [versionRes, modelsRes, metricsRes] = await Promise.allSettled([
+		fetchImpl(`${rootUrl}/version`, fetchOpts()),
+		fetchImpl(`${rootUrl}/v1/models`, fetchOpts()),
+		fetchImpl(`${rootUrl}/metrics`, fetchOpts())
+	]);
+	const info = {
+		engine: "unknown",
+		probedAt: probeAt,
+		baseUrl: trimmed,
+		rootUrl
+	};
+	if (versionRes.status === "fulfilled" && versionRes.value.ok) try {
+		const txt = await versionRes.value.text();
+		try {
+			const json = JSON.parse(txt);
+			if (typeof json.version === "string") info.version = json.version;
+		} catch {
+			const stripped = txt.trim();
+			if (stripped && stripped.length < 64 && /^[\w.\-+]+$/.test(stripped)) info.version = stripped;
+		}
+	} catch {}
+	if (modelsRes.status === "fulfilled" && modelsRes.value.ok) try {
+		const first = (await modelsRes.value.json()).data?.[0];
+		if (first) {
+			if (typeof first.id === "string") info.modelId = first.id;
+			if (typeof first.max_model_len === "number") info.contextWindow = first.max_model_len;
+			if (typeof first.owned_by === "string") {
+				const owner = first.owned_by.toLowerCase();
+				if (owner.includes("vllm")) info.engine = "vllm";
+				else if (owner.includes("sglang")) info.engine = "sglang";
+				else if (info.engine === "unknown") info.engine = "openai-compat";
+			}
+		}
+	} catch {}
+	let snapshot = { probedAt: probeAt };
+	if (metricsRes.status === "fulfilled" && metricsRes.value.ok) try {
+		const parsed = parsePrometheusMetrics(await metricsRes.value.text());
+		info.kvCacheUsagePerc = parsed.kvCacheUsagePerc;
+		info.queueDepth = parsed.queueDepth;
+		info.numRequestsRunning = parsed.numRequestsRunning;
+		info.cacheConfig = parsed.cacheConfig;
+		snapshot = {
+			probedAt: probeAt,
+			promptTokensTotal: parsed.promptTokensTotal,
+			generationTokensTotal: parsed.generationTokensTotal
+		};
+		if (info.engine === "unknown" && parsed.foundVllmPrefix) info.engine = "vllm";
+		if (info.engine === "unknown" && parsed.foundSglangPrefix) info.engine = "sglang";
+	} catch {}
+	const prev = previousSnapshots.get(trimmed);
+	if (prev && typeof prev.generationTokensTotal === "number" && typeof snapshot.generationTokensTotal === "number") {
+		const dt = (snapshot.probedAt - prev.probedAt) / 1e3;
+		const dToks = snapshot.generationTokensTotal - prev.generationTokensTotal;
+		if (dt > 0 && dToks >= 0) info.observedTps = dToks / dt;
+	}
+	if (snapshot.generationTokensTotal !== void 0 || snapshot.promptTokensTotal !== void 0) previousSnapshots.set(trimmed, snapshot);
+	if (!(info.version !== void 0 || info.modelId !== void 0 || info.cacheConfig !== void 0 || info.kvCacheUsagePerc !== void 0 || info.queueDepth !== void 0 || info.numRequestsRunning !== void 0)) {
+		cache.set(trimmed, {
+			info: null,
+			expiresAt: nowFn() + NULL_TTL_MS
+		});
+		return null;
+	}
+	cache.set(trimmed, {
+		info,
+		expiresAt: nowFn() + SUCCESS_TTL_MS
+	});
+	return info;
+}
+function parsePrometheusMetrics(text) {
+	const result = {
+		foundVllmPrefix: false,
+		foundSglangPrefix: false
+	};
+	const lines = text.split("\n");
+	for (const raw of lines) {
+		const line = raw.trim();
+		if (!line || line.startsWith("#")) continue;
+		if (line.startsWith("vllm:")) result.foundVllmPrefix = true;
+		if (line.startsWith("sglang:")) result.foundSglangPrefix = true;
+		const match = line.match(/^([^\s{]+)(\{[^}]*\})?\s+([0-9eE+\-.]+)$/);
+		if (!match) continue;
+		const metric = match[1];
+		const labels = match[2];
+		const value = Number(match[3]);
+		if (!Number.isFinite(value)) continue;
+		switch (metric) {
+			case "vllm:prompt_tokens_total":
+				result.promptTokensTotal = (result.promptTokensTotal ?? 0) + value;
+				break;
+			case "vllm:generation_tokens_total":
+				result.generationTokensTotal = (result.generationTokensTotal ?? 0) + value;
+				break;
+			case "vllm:kv_cache_usage_perc":
+				result.kvCacheUsagePerc = value;
+				break;
+			case "vllm:num_requests_waiting":
+				if (!labels || !labels.includes("reason=")) result.queueDepth = value;
+				break;
+			case "vllm:num_requests_running":
+				result.numRequestsRunning = value;
+				break;
+			case "vllm:cache_config_info": {
+				if (!labels) break;
+				const parsed = parseLabels(labels);
+				const cfg = {};
+				if (parsed.block_size !== void 0) cfg.blockSize = Number(parsed.block_size);
+				if (parsed.num_gpu_blocks !== void 0) cfg.numGpuBlocks = Number(parsed.num_gpu_blocks);
+				if (parsed.cache_dtype !== void 0) cfg.cacheDtype = parsed.cache_dtype;
+				if (parsed.gpu_memory_utilization !== void 0) {
+					const n = Number(parsed.gpu_memory_utilization);
+					if (Number.isFinite(n)) cfg.gpuMemoryUtilization = n;
+				}
+				if (parsed.enable_prefix_caching !== void 0) cfg.enablePrefixCaching = parsed.enable_prefix_caching === "True";
+				result.cacheConfig = cfg;
+				break;
+			}
+		}
+	}
+	return result;
+}
+function parseLabels(rawLabels) {
+	const inner = rawLabels.replace(/^\{/, "").replace(/\}$/, "");
+	const out = {};
+	const re = /(\w+)="([^"]*)"/g;
+	let m;
+	while ((m = re.exec(inner)) !== null) out[m[1]] = m[2];
+	return out;
+}
+function formatEngineLine(info) {
+	if (!info) return null;
+	const parts = [];
+	const engineLabel = (() => {
+		if (info.engine === "vllm") return info.version ? `vLLM ${info.version}` : "vLLM";
+		if (info.engine === "sglang") return info.version ? `SGLang ${info.version}` : "SGLang";
+		if (info.engine === "openai-compat") return "OpenAI-compat";
+		return info.version ? `engine ${info.version}` : null;
+	})();
+	if (engineLabel) parts.push(engineLabel);
+	if (typeof info.observedTps === "number" && info.observedTps > .5) parts.push(`${Math.round(info.observedTps)} tps`);
+	if (typeof info.kvCacheUsagePerc === "number") {
+		const pct = Math.round(info.kvCacheUsagePerc * 100);
+		if (pct > 0) parts.push(`KV ${pct}%`);
+	}
+	if (typeof info.queueDepth === "number" && info.queueDepth > 0) parts.push(`queue ${info.queueDepth}`);
+	if (typeof info.numRequestsRunning === "number" && info.numRequestsRunning > 0) parts.push(`running ${info.numRequestsRunning}`);
+	if (parts.length === 0) return null;
+	return `🖥 ${parts.join(" · ")}`;
+}
+
+//#endregion
+//#region src/cost/breakdown.ts
+async function buildCostBreakdown(args) {
+	const benchmarkTarget = args.config?.models?.benchmark === "none" ? "none" : resolveBenchmarkTarget(args.config?.models?.benchmark);
+	const breakdown = {
+		serving: null,
+		benchmarks: [],
+		benchmarkTarget
+	};
+	if (args.baseUrl && args.baseUrl.trim()) try {
+		breakdown.serving = await probeServingEndpoint(args.baseUrl, {
+			timeoutMs: args.probeTimeoutMs ?? 3e3,
+			forceRefresh: Boolean(args.forceProbeRefresh)
+		});
+	} catch {
+		breakdown.serving = null;
+	}
+	breakdown.engineLine = formatEngineLine(breakdown.serving) ?? void 0;
+	const modelCost = resolveModelCost(args.config, args.provider, args.modelId);
+	const detectedModelId = breakdown.serving?.modelId ?? args.modelId;
+	breakdown.gpu = inferGpu({
+		baseUrl: args.baseUrl,
+		modelId: detectedModelId,
+		override: {
+			gpuType: modelCost?.gpuType,
+			gpuCount: modelCost?.gpuCount
+		}
+	});
+	const overrides = args.config?.models?.gpuRates;
+	let rate;
+	let source;
+	if (typeof modelCost?.hourlyRate === "number" && modelCost.hourlyRate >= 0) {
+		rate = modelCost.hourlyRate;
+		source = "model-override";
+	} else if (breakdown.gpu.gpuType) {
+		rate = resolveHourlyRate(breakdown.gpu.gpuType, overrides);
+		if (rate !== void 0) source = overrides && breakdown.gpu.gpuType in overrides ? "config-override" : "default-table";
+	}
+	breakdown.hourlyRate = rate;
+	breakdown.hourlyRateSource = source;
+	if (breakdown.gpu?.gpuType) {
+		const observed = breakdown.serving?.observedTps;
+		if (typeof observed === "number" && observed > .5) {
+			breakdown.decodeTps = observed;
+			breakdown.prefillTps = observed * PREFILL_DECODE_RATIO;
+			breakdown.throughputSource = "observed";
+		} else {
+			const decode = resolveDefaultDecodeTps(breakdown.gpu.gpuType, breakdown.gpu.paramsB, breakdown.gpu.bytesPerParam);
+			breakdown.decodeTps = decode;
+			breakdown.prefillTps = decode * PREFILL_DECODE_RATIO;
+			breakdown.throughputSource = "default-table";
+		}
+	}
+	if (args.usage && breakdown.hourlyRate !== void 0 && breakdown.gpu?.gpuType && breakdown.decodeTps !== void 0 && breakdown.prefillTps !== void 0) {
+		const perSecond = breakdown.hourlyRate * breakdown.gpu.gpuCount / 3600;
+		const inputTokens = args.usage.input ?? 0;
+		const outputTokens = args.usage.output ?? 0;
+		const totalCost = ((breakdown.prefillTps > 0 ? inputTokens / breakdown.prefillTps : 0) + (breakdown.decodeTps > 0 ? outputTokens / breakdown.decodeTps : 0)) * perSecond;
+		breakdown.instanceCostUsd = totalCost;
+		const totalTokens = inputTokens + outputTokens;
+		if (totalTokens > 0) breakdown.perTokenAvgUsd = totalCost / totalTokens;
+	}
+	if (args.usage) {
+		const costForEstimate = modelCost ? {
+			input: modelCost.input ?? 0,
+			output: modelCost.output ?? 0,
+			cacheRead: modelCost.cacheRead ?? 0,
+			cacheWrite: modelCost.cacheWrite ?? 0
+		} : void 0;
+		const rawCost = estimateUsageCost({
+			usage: args.usage,
+			cost: costForEstimate
+		});
+		if (typeof rawCost === "number" && rawCost > 0) breakdown.effectiveCostUsd = rawCost;
+		else if (breakdown.instanceCostUsd !== void 0) breakdown.effectiveCostUsd = breakdown.instanceCostUsd;
+		if (benchmarkTarget !== "none" && breakdown.effectiveCostUsd !== void 0) {
+			breakdown.benchmarks = benchmarkUsage({
+				usage: args.usage,
+				target: benchmarkTarget
+			});
+			breakdown.costLineSuffix = formatBenchmarkSuffix({
+				ourCostUsd: breakdown.effectiveCostUsd,
+				benchmarks: breakdown.benchmarks
+			});
+		}
+	}
+	breakdown.hardwareLine = formatHardwareLine(breakdown);
+	return breakdown;
+}
+function resolveModelCost(config, provider, modelId) {
+	if (!config || !provider || !modelId) return;
+	const providers = config.models?.providers;
+	if (!providers) return;
+	const providerCfg = providers[provider];
+	if (!providerCfg?.models) return;
+	return providerCfg.models.find((m) => m.id === modelId)?.cost;
+}
+function formatHardwareLine(breakdown) {
+	const gpu = breakdown.gpu;
+	if (!gpu?.gpuType) return;
+	const label = GPU_LABEL[gpu.gpuType];
+	const count = gpu.gpuCount;
+	const parts = [];
+	parts.push(`${count}× ${label}`);
+	if (typeof breakdown.hourlyRate === "number") parts.push(`≈ $${(breakdown.hourlyRate * count).toFixed(2)}/hr`);
+	const tag = (() => {
+		if (breakdown.hourlyRateSource === "model-override") return "model override";
+		if (breakdown.hourlyRateSource === "config-override") return "config override";
+		if (gpu.confidence === "high") return "inferred";
+		if (gpu.confidence === "medium") return "inferred · medium-conf";
+		return "inferred · low-conf";
+	})();
+	parts.push(`(${tag})`);
+	return `🧰 ${parts.join(" · ")}`;
+}
+
 //#endregion
 //#region src/infra/provider-usage.shared.ts
 const DEFAULT_TIMEOUT_MS = 5e3;
@@ -27620,6 +28788,32 @@ async function buildStatusReply(params) {
 	} catch {
 		usageLine = null;
 	}
+	const baseUrl = cfg.models?.providers?.[provider]?.baseUrl;
+	let engineLine;
+	let hardwareLine;
+	let costOverrideUsd;
+	let costLineSuffix;
+	try {
+		const breakdown = await buildCostBreakdown({
+			baseUrl,
+			provider,
+			modelId: model,
+			config: cfg,
+			usage: {
+				input: sessionEntry?.inputTokens,
+				output: sessionEntry?.outputTokens
+			}
+		});
+		engineLine = breakdown.engineLine;
+		hardwareLine = breakdown.hardwareLine;
+		costOverrideUsd = breakdown.instanceCostUsd;
+		costLineSuffix = breakdown.costLineSuffix;
+	} catch {
+		engineLine = void 0;
+		hardwareLine = void 0;
+		costOverrideUsd = void 0;
+		costLineSuffix = void 0;
+	}
 	const queueSettings = resolveQueueSettings({
 		cfg,
 		channel: command.channel,
@@ -27702,6 +28896,10 @@ async function buildStatusReply(params) {
 			showDetails: queueOverrides
 		},
 		subagentsLine,
+		engineLine,
+		hardwareLine,
+		costOverrideUsd,
+		costLineSuffix,
 		mediaDecisions: params.mediaDecisions,
 		includeTranscriptUsage: false
 	}) };
@@ -27802,6 +29000,27 @@ const handleStatusCommand = async (params, allowTextCommands) => {
 		})
 	};
 };
+const handleCostCommand = async (params, allowTextCommands) => {
+	if (!allowTextCommands) return null;
+	const normalized = params.command.commandBodyNormalized;
+	if (normalized !== "/cost" && !normalized.startsWith("/cost ")) return null;
+	if (!params.command.isAuthorizedSender) {
+		logVerbose(`Ignoring /cost from unauthorized sender: ${params.command.senderId || "<unknown>"}`);
+		return { shouldContinue: false };
+	}
+	const arg = normalized.slice(5).trim();
+	const window = arg === "today" || arg === "week" || arg === "month" || arg === "all" ? arg : "week";
+	const { startMs, endMs } = resolveWindow(window);
+	return {
+		shouldContinue: false,
+		reply: { text: formatSummaryReport(await loadCostBenchmarkSummary({
+			startMs,
+			endMs,
+			config: params.cfg,
+			agentId: params.agentId
+		}), `${window.charAt(0).toUpperCase()}${window.slice(1)}`) }
+	};
+};
 const handleContextCommand = async (params, allowTextCommands) => {
 	if (!allowTextCommands) return null;
 	const normalized = params.command.commandBodyNormalized;
@@ -28198,411 +29417,6 @@ function scheduleGatewaySigusr1Restart(opts) {
 	};
 }
 
-//#endregion
-//#region src/infra/session-cost-usage.ts
-const emptyTotals = () => ({
-	input: 0,
-	output: 0,
-	cacheRead: 0,
-	cacheWrite: 0,
-	totalTokens: 0,
-	totalCost: 0,
-	inputCost: 0,
-	outputCost: 0,
-	cacheReadCost: 0,
-	cacheWriteCost: 0,
-	missingCostEntries: 0
-});
-const toFiniteNumber = (value) => {
-	if (typeof value !== "number") return;
-	if (!Number.isFinite(value)) return;
-	return value;
-};
-const extractCostBreakdown = (usageRaw) => {
-	if (!usageRaw || typeof usageRaw !== "object") return;
-	const cost = usageRaw.cost;
-	if (!cost) return;
-	const total = toFiniteNumber(cost.total);
-	if (total === void 0 || total < 0) return;
-	return {
-		total,
-		input: toFiniteNumber(cost.input),
-		output: toFiniteNumber(cost.output),
-		cacheRead: toFiniteNumber(cost.cacheRead),
-		cacheWrite: toFiniteNumber(cost.cacheWrite)
-	};
-};
-const parseTimestamp = (entry) => {
-	const raw = entry.timestamp;
-	if (typeof raw === "string") {
-		const parsed = new Date(raw);
-		if (!Number.isNaN(parsed.valueOf())) return parsed;
-	}
-	const message = entry.message;
-	const messageTimestamp = toFiniteNumber(message?.timestamp);
-	if (messageTimestamp !== void 0) {
-		const parsed = new Date(messageTimestamp);
-		if (!Number.isNaN(parsed.valueOf())) return parsed;
-	}
-};
-const parseTranscriptEntry = (entry) => {
-	const message = entry.message;
-	if (!message || typeof message !== "object") return null;
-	const roleRaw = message.role;
-	const role = roleRaw === "user" || roleRaw === "assistant" ? roleRaw : void 0;
-	if (!role) return null;
-	const usageRaw = message.usage ?? entry.usage;
-	const usage = usageRaw ? normalizeUsage(usageRaw) ?? void 0 : void 0;
-	const provider = (typeof message.provider === "string" ? message.provider : void 0) ?? (typeof entry.provider === "string" ? entry.provider : void 0);
-	const model = (typeof message.model === "string" ? message.model : void 0) ?? (typeof entry.model === "string" ? entry.model : void 0);
-	const costBreakdown = extractCostBreakdown(usageRaw);
-	const stopReason = typeof message.stopReason === "string" ? message.stopReason : void 0;
-	const durationMs = toFiniteNumber(message.durationMs ?? entry.durationMs);
-	return {
-		message,
-		role,
-		timestamp: parseTimestamp(entry),
-		durationMs,
-		usage,
-		costTotal: costBreakdown?.total,
-		costBreakdown,
-		provider,
-		model,
-		stopReason,
-		toolNames: extractToolCallNames(message),
-		toolResultCounts: countToolResults(message)
-	};
-};
-const formatDayKey = (date) => date.toLocaleDateString("en-CA", { timeZone: Intl.DateTimeFormat().resolvedOptions().timeZone });
-const computeLatencyStats = (values) => {
-	if (!values.length) return;
-	const sorted = values.toSorted((a, b) => a - b);
-	const total = sorted.reduce((sum, v) => sum + v, 0);
-	const count = sorted.length;
-	const p95Index = Math.max(0, Math.ceil(count * .95) - 1);
-	return {
-		count,
-		avgMs: total / count,
-		p95Ms: sorted[p95Index] ?? sorted[count - 1],
-		minMs: sorted[0],
-		maxMs: sorted[count - 1]
-	};
-};
-const applyUsageTotals = (totals, usage) => {
-	totals.input += usage.input ?? 0;
-	totals.output += usage.output ?? 0;
-	totals.cacheRead += usage.cacheRead ?? 0;
-	totals.cacheWrite += usage.cacheWrite ?? 0;
-	const totalTokens = usage.total ?? (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0);
-	totals.totalTokens += totalTokens;
-};
-const applyCostBreakdown = (totals, costBreakdown) => {
-	if (costBreakdown === void 0 || costBreakdown.total === void 0) return;
-	totals.totalCost += costBreakdown.total;
-	totals.inputCost += costBreakdown.input ?? 0;
-	totals.outputCost += costBreakdown.output ?? 0;
-	totals.cacheReadCost += costBreakdown.cacheRead ?? 0;
-	totals.cacheWriteCost += costBreakdown.cacheWrite ?? 0;
-};
-const applyCostTotal = (totals, costTotal) => {
-	if (costTotal === void 0) {
-		totals.missingCostEntries += 1;
-		return;
-	}
-	totals.totalCost += costTotal;
-};
-async function* readJsonlRecords(filePath) {
-	const fileStream = fs.createReadStream(filePath, { encoding: "utf-8" });
-	const rl = readline.createInterface({
-		input: fileStream,
-		crlfDelay: Infinity
-	});
-	try {
-		for await (const line of rl) {
-			const trimmed = line.trim();
-			if (!trimmed) continue;
-			try {
-				const parsed = JSON.parse(trimmed);
-				if (!parsed || typeof parsed !== "object") continue;
-				yield parsed;
-			} catch {}
-		}
-	} finally {
-		rl.close();
-		fileStream.destroy();
-	}
-}
-async function scanTranscriptFile(params) {
-	for await (const parsed of readJsonlRecords(params.filePath)) {
-		const entry = parseTranscriptEntry(parsed);
-		if (!entry) continue;
-		if (entry.usage && entry.costTotal === void 0) {
-			const cost = resolveModelCostConfig({
-				provider: entry.provider,
-				model: entry.model,
-				config: params.config
-			});
-			entry.costTotal = estimateUsageCost({
-				usage: entry.usage,
-				cost
-			});
-		}
-		params.onEntry(entry);
-	}
-}
-async function scanUsageFile(params) {
-	await scanTranscriptFile({
-		filePath: params.filePath,
-		config: params.config,
-		onEntry: (entry) => {
-			if (!entry.usage) return;
-			params.onEntry({
-				usage: entry.usage,
-				costTotal: entry.costTotal,
-				costBreakdown: entry.costBreakdown,
-				provider: entry.provider,
-				model: entry.model,
-				timestamp: entry.timestamp
-			});
-		}
-	});
-}
-async function loadCostUsageSummary(params) {
-	const now = /* @__PURE__ */ new Date();
-	let sinceTime;
-	let untilTime;
-	if (params?.startMs !== void 0 && params?.endMs !== void 0) {
-		sinceTime = params.startMs;
-		untilTime = params.endMs;
-	} else {
-		const days = Math.max(1, Math.floor(params?.days ?? 30));
-		const since = new Date(now);
-		since.setDate(since.getDate() - (days - 1));
-		sinceTime = since.getTime();
-		untilTime = now.getTime();
-	}
-	const dailyMap = /* @__PURE__ */ new Map();
-	const totals = emptyTotals();
-	const sessionsDir = resolveSessionTranscriptsDirForAgent(params?.agentId);
-	const entries = await fs.promises.readdir(sessionsDir, { withFileTypes: true }).catch(() => []);
-	const files = (await Promise.all(entries.filter((entry) => entry.isFile() && entry.name.endsWith(".jsonl")).map(async (entry) => {
-		const filePath = path.join(sessionsDir, entry.name);
-		const stats = await fs.promises.stat(filePath).catch(() => null);
-		if (!stats) return null;
-		if (stats.mtimeMs < sinceTime) return null;
-		return filePath;
-	}))).filter((filePath) => Boolean(filePath));
-	for (const filePath of files) await scanUsageFile({
-		filePath,
-		config: params?.config,
-		onEntry: (entry) => {
-			const ts = entry.timestamp?.getTime();
-			if (!ts || ts < sinceTime || ts > untilTime) return;
-			const dayKey = formatDayKey(entry.timestamp ?? now);
-			const bucket = dailyMap.get(dayKey) ?? emptyTotals();
-			applyUsageTotals(bucket, entry.usage);
-			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(bucket, entry.costBreakdown);
-			else applyCostTotal(bucket, entry.costTotal);
-			dailyMap.set(dayKey, bucket);
-			applyUsageTotals(totals, entry.usage);
-			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(totals, entry.costBreakdown);
-			else applyCostTotal(totals, entry.costTotal);
-		}
-	});
-	const daily = Array.from(dailyMap.entries()).map(([date, bucket]) => Object.assign({ date }, bucket)).toSorted((a, b) => a.date.localeCompare(b.date));
-	const days = Math.ceil((untilTime - sinceTime) / (1440 * 60 * 1e3)) + 1;
-	return {
-		updatedAt: Date.now(),
-		days,
-		daily,
-		totals
-	};
-}
-async function loadSessionCostSummary(params) {
-	const sessionFile = params.sessionFile ?? (params.sessionId ? resolveSessionFilePath(params.sessionId, params.sessionEntry, { agentId: params.agentId }) : void 0);
-	if (!sessionFile || !fs.existsSync(sessionFile)) return null;
-	const totals = emptyTotals();
-	let firstActivity;
-	let lastActivity;
-	const activityDatesSet = /* @__PURE__ */ new Set();
-	const dailyMap = /* @__PURE__ */ new Map();
-	const dailyMessageMap = /* @__PURE__ */ new Map();
-	const dailyLatencyMap = /* @__PURE__ */ new Map();
-	const dailyModelUsageMap = /* @__PURE__ */ new Map();
-	const messageCounts = {
-		total: 0,
-		user: 0,
-		assistant: 0,
-		toolCalls: 0,
-		toolResults: 0,
-		errors: 0
-	};
-	const toolUsageMap = /* @__PURE__ */ new Map();
-	const modelUsageMap = /* @__PURE__ */ new Map();
-	const errorStopReasons = new Set([
-		"error",
-		"aborted",
-		"timeout"
-	]);
-	const latencyValues = [];
-	let lastUserTimestamp;
-	const MAX_LATENCY_MS = 720 * 60 * 1e3;
-	await scanTranscriptFile({
-		filePath: sessionFile,
-		config: params.config,
-		onEntry: (entry) => {
-			const ts = entry.timestamp?.getTime();
-			if (params.startMs !== void 0 && ts !== void 0 && ts < params.startMs) return;
-			if (params.endMs !== void 0 && ts !== void 0 && ts > params.endMs) return;
-			if (ts !== void 0) {
-				if (!firstActivity || ts < firstActivity) firstActivity = ts;
-				if (!lastActivity || ts > lastActivity) lastActivity = ts;
-			}
-			if (entry.role === "user") {
-				messageCounts.user += 1;
-				messageCounts.total += 1;
-				if (entry.timestamp) lastUserTimestamp = entry.timestamp.getTime();
-			}
-			if (entry.role === "assistant") {
-				messageCounts.assistant += 1;
-				messageCounts.total += 1;
-				const ts = entry.timestamp?.getTime();
-				if (ts !== void 0) {
-					const latencyMs = entry.durationMs ?? (lastUserTimestamp !== void 0 ? Math.max(0, ts - lastUserTimestamp) : void 0);
-					if (latencyMs !== void 0 && Number.isFinite(latencyMs) && latencyMs <= MAX_LATENCY_MS) {
-						latencyValues.push(latencyMs);
-						const dayKey = formatDayKey(entry.timestamp ?? new Date(ts));
-						const dailyLatencies = dailyLatencyMap.get(dayKey) ?? [];
-						dailyLatencies.push(latencyMs);
-						dailyLatencyMap.set(dayKey, dailyLatencies);
-					}
-				}
-			}
-			if (entry.toolNames.length > 0) {
-				messageCounts.toolCalls += entry.toolNames.length;
-				for (const name of entry.toolNames) toolUsageMap.set(name, (toolUsageMap.get(name) ?? 0) + 1);
-			}
-			if (entry.toolResultCounts.total > 0) {
-				messageCounts.toolResults += entry.toolResultCounts.total;
-				messageCounts.errors += entry.toolResultCounts.errors;
-			}
-			if (entry.stopReason && errorStopReasons.has(entry.stopReason)) messageCounts.errors += 1;
-			if (entry.timestamp) {
-				const dayKey = formatDayKey(entry.timestamp);
-				activityDatesSet.add(dayKey);
-				const daily = dailyMessageMap.get(dayKey) ?? {
-					date: dayKey,
-					total: 0,
-					user: 0,
-					assistant: 0,
-					toolCalls: 0,
-					toolResults: 0,
-					errors: 0
-				};
-				daily.total += entry.role === "user" || entry.role === "assistant" ? 1 : 0;
-				if (entry.role === "user") daily.user += 1;
-				else if (entry.role === "assistant") daily.assistant += 1;
-				daily.toolCalls += entry.toolNames.length;
-				daily.toolResults += entry.toolResultCounts.total;
-				daily.errors += entry.toolResultCounts.errors;
-				if (entry.stopReason && errorStopReasons.has(entry.stopReason)) daily.errors += 1;
-				dailyMessageMap.set(dayKey, daily);
-			}
-			if (!entry.usage) return;
-			applyUsageTotals(totals, entry.usage);
-			if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(totals, entry.costBreakdown);
-			else applyCostTotal(totals, entry.costTotal);
-			if (entry.timestamp) {
-				const dayKey = formatDayKey(entry.timestamp);
-				const entryTokens = (entry.usage.input ?? 0) + (entry.usage.output ?? 0) + (entry.usage.cacheRead ?? 0) + (entry.usage.cacheWrite ?? 0);
-				const entryCost = entry.costBreakdown?.total ?? (entry.costBreakdown ? (entry.costBreakdown.input ?? 0) + (entry.costBreakdown.output ?? 0) + (entry.costBreakdown.cacheRead ?? 0) + (entry.costBreakdown.cacheWrite ?? 0) : entry.costTotal ?? 0);
-				const existing = dailyMap.get(dayKey) ?? {
-					tokens: 0,
-					cost: 0
-				};
-				dailyMap.set(dayKey, {
-					tokens: existing.tokens + entryTokens,
-					cost: existing.cost + entryCost
-				});
-				if (entry.provider || entry.model) {
-					const modelKey = `${dayKey}::${entry.provider ?? "unknown"}::${entry.model ?? "unknown"}`;
-					const dailyModel = dailyModelUsageMap.get(modelKey) ?? {
-						date: dayKey,
-						provider: entry.provider,
-						model: entry.model,
-						tokens: 0,
-						cost: 0,
-						count: 0
-					};
-					dailyModel.tokens += entryTokens;
-					dailyModel.cost += entryCost;
-					dailyModel.count += 1;
-					dailyModelUsageMap.set(modelKey, dailyModel);
-				}
-			}
-			if (entry.provider || entry.model) {
-				const key = `${entry.provider ?? "unknown"}::${entry.model ?? "unknown"}`;
-				const existing = modelUsageMap.get(key) ?? {
-					provider: entry.provider,
-					model: entry.model,
-					count: 0,
-					totals: emptyTotals()
-				};
-				existing.count += 1;
-				applyUsageTotals(existing.totals, entry.usage);
-				if (entry.costBreakdown?.total !== void 0) applyCostBreakdown(existing.totals, entry.costBreakdown);
-				else applyCostTotal(existing.totals, entry.costTotal);
-				modelUsageMap.set(key, existing);
-			}
-		}
-	});
-	const dailyBreakdown = Array.from(dailyMap.entries()).map(([date, data]) => ({
-		date,
-		tokens: data.tokens,
-		cost: data.cost
-	})).toSorted((a, b) => a.date.localeCompare(b.date));
-	const dailyMessageCounts = Array.from(dailyMessageMap.values()).toSorted((a, b) => a.date.localeCompare(b.date));
-	const dailyLatency = Array.from(dailyLatencyMap.entries()).map(([date, values]) => {
-		const stats = computeLatencyStats(values);
-		if (!stats) return null;
-		return {
-			date,
-			...stats
-		};
-	}).filter((entry) => Boolean(entry)).toSorted((a, b) => a.date.localeCompare(b.date));
-	const dailyModelUsage = Array.from(dailyModelUsageMap.values()).toSorted((a, b) => a.date.localeCompare(b.date) || b.cost - a.cost);
-	const toolUsage = toolUsageMap.size ? {
-		totalCalls: Array.from(toolUsageMap.values()).reduce((sum, count) => sum + count, 0),
-		uniqueTools: toolUsageMap.size,
-		tools: Array.from(toolUsageMap.entries()).map(([name, count]) => ({
-			name,
-			count
-		})).toSorted((a, b) => b.count - a.count)
-	} : void 0;
-	const modelUsage = modelUsageMap.size ? Array.from(modelUsageMap.values()).toSorted((a, b) => {
-		const costDiff = b.totals.totalCost - a.totals.totalCost;
-		if (costDiff !== 0) return costDiff;
-		return b.totals.totalTokens - a.totals.totalTokens;
-	}) : void 0;
-	return {
-		sessionId: params.sessionId,
-		sessionFile,
-		firstActivity,
-		lastActivity,
-		durationMs: firstActivity !== void 0 && lastActivity !== void 0 ? Math.max(0, lastActivity - firstActivity) : void 0,
-		activityDates: Array.from(activityDatesSet).toSorted(),
-		dailyBreakdown,
-		dailyMessageCounts,
-		dailyLatency: dailyLatency.length ? dailyLatency : void 0,
-		dailyModelUsage: dailyModelUsage.length ? dailyModelUsage : void 0,
-		messageCounts,
-		toolUsage,
-		modelUsage,
-		latency: computeLatencyStats(latencyValues),
-		...totals
-	};
-}
-
 //#endregion
 //#region src/auto-reply/send-policy.ts
 function normalizeSendPolicyOverride(raw) {
@@ -29802,6 +30616,7 @@ async function handleCommands(params) {
 		handleHelpCommand,
 		handleCommandsListCommand,
 		handleStatusCommand,
+		handleCostCommand,
 		handleAllowlistCommand,
 		handleApproveCommand,
 		handleContextCommand,
@@ -37429,6 +38244,32 @@ function createSessionStatusTool(opts) {
 				}
 			} catch {}
 			const groupActivation = resolved.entry.chatType === "group" || resolved.entry.chatType === "channel" || resolved.key.includes(":group:") || resolved.key.includes(":channel:") ? normalizeGroupActivation(resolved.entry.groupActivation) ?? "mention" : void 0;
+			const baseUrl = cfg.models?.providers?.[providerForCard]?.baseUrl;
+			let engineLine;
+			let hardwareLine;
+			let costOverrideUsd;
+			let costLineSuffix;
+			try {
+				const breakdown = await buildCostBreakdown({
+					baseUrl,
+					provider: providerForCard,
+					modelId: configured.model,
+					config: cfg,
+					usage: {
+						input: resolved.entry.inputTokens,
+						output: resolved.entry.outputTokens
+					}
+				});
+				engineLine = breakdown.engineLine;
+				hardwareLine = breakdown.hardwareLine;
+				costOverrideUsd = breakdown.instanceCostUsd;
+				costLineSuffix = breakdown.costLineSuffix;
+			} catch {
+				engineLine = void 0;
+				hardwareLine = void 0;
+				costOverrideUsd = void 0;
+				costLineSuffix = void 0;
+			}
 			const queueSettings = resolveQueueSettings({
 				cfg,
 				channel: resolved.entry.channel ?? resolved.entry.lastChannel ?? "unknown",
@@ -37474,6 +38315,10 @@ function createSessionStatusTool(opts) {
 					dropPolicy: queueSettings.dropPolicy,
 					showDetails: queueOverrides
 				},
+				engineLine,
+				hardwareLine,
+				costOverrideUsd,
+				costLineSuffix,
 				includeTranscriptUsage: false
 			});
 			return {
@@ -40982,6 +41827,86 @@ async function handleInlineActions(params) {
 	};
 }
 
+//#endregion
+//#region src/auto-reply/reply/middleware-context-warning.ts
+/** Threshold (fraction of context window) that triggers the soft warning. */
+const CONTEXT_WARNING_THRESHOLD = .85;
+/**
+* Threshold that triggers proactive compaction in `prepareReplyTurn`.
+* Higher than the soft warning so the user sees the heads-up first;
+* compaction runs the next turn the threshold is observed.
+*/
+const CONTEXT_PROACTIVE_COMPACT_THRESHOLD = .9;
+/**
+* Pure decision for "should `prepareReplyTurn` run proactive compaction
+* before the next turn?" — extracted so the trigger logic is unit-testable
+* without mocking the rest of the prep pipeline.
+*
+* Returns true ONLY when:
+*   - we have a session id (otherwise compaction has no target),
+*   - we have a positive context window,
+*   - we have a positive cumulative usage from the previous turn,
+*   - usage / window ≥ CONTEXT_PROACTIVE_COMPACT_THRESHOLD.
+*/
+function shouldProactivelyCompact(args) {
+	if (!args.sessionId) return false;
+	if (!args.contextTokens || args.contextTokens <= 0) return false;
+	const usage = args.totalTokens ?? 0;
+	if (usage <= 0) return false;
+	return usage / args.contextTokens >= CONTEXT_PROACTIVE_COMPACT_THRESHOLD;
+}
+/**
+* Read the model's context-window size, falling back to the global
+* default when the model isn't in the registry yet.
+*/
+function resolveContextWindow(modelId) {
+	const fromRegistry = lookupContextTokens(modelId);
+	if (fromRegistry && fromRegistry > 0) return fromRegistry;
+	return DEFAULT_CONTEXT_TOKENS;
+}
+/**
+* After-turn hook that surfaces a one-time soft warning when usage
+* crosses the context threshold. Mutates the result text to append
+* the warning when triggered, and persists `contextWarningSent: true`
+* on the session entry so the warning doesn't fire again.
+*/
+function withContextUsageWarning() {
+	return (next) => async (params) => {
+		const result = await next(params);
+		if (result.outbound.action === "drop") return result;
+		if (!params.sessionKey) return result;
+		const usage = result.rawResult.meta.agentMeta?.usage?.total ?? 0;
+		if (usage <= 0) return result;
+		if (usage / resolveContextWindow(params.modelId ?? result.modelId) < CONTEXT_WARNING_THRESHOLD) return result;
+		const agentId = resolveAgentIdFromSessionKey(params.sessionKey);
+		const storePath = resolveStorePath(params.config.session?.store, { agentId });
+		let alreadyWarned = false;
+		try {
+			await updateSessionStoreEntry({
+				storePath,
+				sessionKey: params.sessionKey,
+				update: async (entry) => {
+					if (entry.contextWarningSent === true) {
+						alreadyWarned = true;
+						return null;
+					}
+					return { contextWarningSent: true };
+				}
+			});
+		} catch (err) {
+			logVerbose(`context-warning: failed to persist contextWarningSent flag for ${params.sessionKey}: ${String(err)}`);
+			return result;
+		}
+		if (alreadyWarned) return result;
+		const trimmedText = result.text.trimEnd();
+		const separator = trimmedText ? "\n\n" : "";
+		return {
+			...result,
+			text: `${trimmedText}${separator}${SESSION_CONTEXT_WARNING_MESSAGE}`
+		};
+	};
+}
+
 //#endregion
 //#region src/auto-reply/reply/session-reset-model.ts
 function splitBody(body) {
@@ -41151,7 +42076,7 @@ async function deliverSessionMaintenanceWarning(params) {
 		return;
 	}
 	try {
-		const { deliverOutboundPayloads } = await import("./deliver-Cjyb6h4g.js").then((n) => n.n);
+		const { deliverOutboundPayloads } = await import("./deliver-f3cIWxXT.js").then((n) => n.n);
 		await deliverOutboundPayloads({
 			cfg: params.cfg,
 			channel,
@@ -41878,7 +42803,7 @@ async function prepareReplyTurn(ctx, opts, configOverride) {
 		typing.cleanup();
 		return {
 			kind: "early-reply",
-			reply: { text: SESSION_OVERFLOW_MESSAGE }
+			reply: { text: SESSION_COMPACTION_FAILED_MESSAGE }
 		};
 	}
 	await applyResetModelOverride({
@@ -41991,6 +42916,82 @@ async function prepareReplyTurn(ctx, opts, configOverride) {
 	};
 	directives = inlineActionResult.directives;
 	abortedLastRun = inlineActionResult.abortedLastRun ?? abortedLastRun;
+	if (shouldProactivelyCompact({
+		totalTokens: sessionEntry.totalTokens,
+		contextTokens,
+		sessionId: sessionEntry.sessionId
+	})) {
+		if (resolvedOpts?.onBlockReply) try {
+			await resolvedOpts.onBlockReply({ text: SESSION_COMPACTING_MESSAGE });
+		} catch (err) {
+			logVerbose(`proactive-compact: inline status delivery failed: ${String(err)}`);
+		}
+		if (isEmbeddedPiRunActive(sessionEntry.sessionId)) {
+			abortEmbeddedPiRun(sessionEntry.sessionId);
+			await waitForEmbeddedPiRunEnd(sessionEntry.sessionId, 15e3);
+		}
+		let compactResult;
+		try {
+			compactResult = await compactEmbeddedPiSession({
+				sessionId: sessionEntry.sessionId,
+				sessionKey,
+				messageChannel: command.channel,
+				groupId: sessionEntry.groupId,
+				groupChannel: sessionEntry.groupChannel,
+				groupSpace: sessionEntry.space,
+				spawnedBy: sessionEntry.spawnedBy,
+				sessionFile: resolveSessionFilePath(sessionEntry.sessionId, sessionEntry, resolveSessionFilePathOptions({
+					agentId,
+					storePath
+				})),
+				workspaceDir,
+				agentDir,
+				config: cfg,
+				skillsSnapshot: sessionEntry.skillsSnapshot,
+				provider,
+				model,
+				thinkLevel: resolvedThinkLevel,
+				reasoningLevel: resolvedReasoningLevel,
+				trigger: "overflow",
+				senderIsOwner: command.senderIsOwner,
+				ownerNumbers: command.ownerList.length > 0 ? command.ownerList : void 0
+			});
+		} catch (err) {
+			logVerbose(`proactive-compact: compactor threw for ${sessionKey}: ${String(err)}`);
+			compactResult = {
+				ok: false,
+				compacted: false,
+				reason: String(err)
+			};
+		}
+		if (compactResult.ok && compactResult.compacted) await incrementCompactionCount({
+			sessionEntry,
+			sessionStore,
+			sessionKey,
+			storePath,
+			tokensAfter: compactResult.result?.tokensAfter
+		});
+		else {
+			logVerbose(`proactive-compact: failed for ${sessionKey} (${compactResult.reason ?? "unknown"})`);
+			if (storePath) try {
+				await updateSessionStoreEntry({
+					storePath,
+					sessionKey,
+					update: async (entry) => {
+						if (entry.endedReason === "context_overflow") return null;
+						return { endedReason: "context_overflow" };
+					}
+				});
+			} catch (err) {
+				logVerbose(`proactive-compact: failed to mark session sealed: ${String(err)}`);
+			}
+			typing.cleanup();
+			return {
+				kind: "early-reply",
+				reply: { text: SESSION_COMPACTION_FAILED_MESSAGE }
+			};
+		}
+	}
 	await stageSandboxMedia({
 		ctx,
 		sessionCtx,
@@ -51602,7 +52603,7 @@ async function runEmbeddedPiAgent(params) {
 				error: err
 			});
 		}
-		const MAX_OVERFLOW_COMPACTION_ATTEMPTS = 3;
+		const MAX_OVERFLOW_COMPACTION_ATTEMPTS = 1;
 		const MAX_RUN_LOOP_ITERATIONS = resolveMaxRunRetryIterations(profileCandidates.length);
 		let overflowCompactionAttempts = 0;
 		let toolResultTruncationAttempted = false;
@@ -52625,8 +53626,8 @@ function extractErrorMessage(err) {
 function buildBoundaryResult(params) {
 	const profile = resolveModelProfile(params.provider && params.modelId ? `${params.provider}/${params.modelId}` : "", params.config.models?.profiles);
 	return {
-		text: SESSION_OVERFLOW_MESSAGE,
-		rawText: SESSION_OVERFLOW_MESSAGE,
+		text: SESSION_COMPACTION_FAILED_MESSAGE,
+		rawText: SESSION_COMPACTION_FAILED_MESSAGE,
 		outbound: {
 			action: "deliver",
 			reason: "deliver"
@@ -52637,7 +53638,7 @@ function buildBoundaryResult(params) {
 		profile,
 		didSendViaMessagingTool: false,
 		rawResult: {
-			payloads: [{ text: SESSION_OVERFLOW_MESSAGE }],
+			payloads: [{ text: SESSION_COMPACTION_FAILED_MESSAGE }],
 			meta: {
 				durationMs: Date.now() - params.startTime,
 				agentMeta: {
@@ -52696,62 +53697,6 @@ function withContextOverflowRecovery() {
 	};
 }
 
-//#endregion
-//#region src/auto-reply/reply/middleware-context-warning.ts
-/** Threshold (fraction of context window) that triggers the soft warning. */
-const CONTEXT_WARNING_THRESHOLD = .85;
-/**
-* Read the model's context-window size, falling back to the global
-* default when the model isn't in the registry yet.
-*/
-function resolveContextWindow(modelId) {
-	const fromRegistry = lookupContextTokens(modelId);
-	if (fromRegistry && fromRegistry > 0) return fromRegistry;
-	return DEFAULT_CONTEXT_TOKENS;
-}
-/**
-* After-turn hook that surfaces a one-time soft warning when usage
-* crosses the context threshold. Mutates the result text to append
-* the warning when triggered, and persists `contextWarningSent: true`
-* on the session entry so the warning doesn't fire again.
-*/
-function withContextUsageWarning() {
-	return (next) => async (params) => {
-		const result = await next(params);
-		if (result.outbound.action === "drop") return result;
-		if (!params.sessionKey) return result;
-		const usage = result.rawResult.meta.agentMeta?.usage?.total ?? 0;
-		if (usage <= 0) return result;
-		if (usage / resolveContextWindow(params.modelId ?? result.modelId) < CONTEXT_WARNING_THRESHOLD) return result;
-		const agentId = resolveAgentIdFromSessionKey(params.sessionKey);
-		const storePath = resolveStorePath(params.config.session?.store, { agentId });
-		let alreadyWarned = false;
-		try {
-			await updateSessionStoreEntry({
-				storePath,
-				sessionKey: params.sessionKey,
-				update: async (entry) => {
-					if (entry.contextWarningSent === true) {
-						alreadyWarned = true;
-						return null;
-					}
-					return { contextWarningSent: true };
-				}
-			});
-		} catch (err) {
-			logVerbose(`context-warning: failed to persist contextWarningSent flag for ${params.sessionKey}: ${String(err)}`);
-			return result;
-		}
-		if (alreadyWarned) return result;
-		const trimmedText = result.text.trimEnd();
-		const separator = trimmedText ? "\n\n" : "";
-		return {
-			...result,
-			text: `${trimmedText}${separator}${SESSION_CONTEXT_WARNING_MESSAGE}`
-		};
-	};
-}
-
 //#endregion
 //#region src/auto-reply/reply/middleware-diagnostic-logging.ts
 function deriveChannel(ctx) {
@@ -57931,7 +58876,7 @@ function readSlackExternalArgMenuToken(raw) {
 }
 let commandsRegistry;
 async function getCommandsRegistry() {
-	if (!commandsRegistry) commandsRegistry = await import("./commands-registry-DFYDtKr_.js").then((n) => n.t);
+	if (!commandsRegistry) commandsRegistry = await import("./commands-registry-DG2kGyg5.js").then((n) => n.t);
 	return commandsRegistry;
 }
 function encodeSlackCommandArgValue(parts) {
@@ -58334,7 +59279,7 @@ async function registerSlackMonitorSlashCommands(params) {
 			});
 			const deliverSlashPayloads = async (replies) => {
 				const [{ deliverSlackSlashReplies }, { resolveChunkMode }, { resolveMarkdownTableMode }] = await Promise.all([
-					import("./replies-BYF1gbJf.js").then((n) => n.r),
+					import("./replies-0nzkXt6o.js").then((n) => n.r),
 					import("./chunk-CAZujdOi.js").then((n) => n.s),
 					import("./markdown-tables-jQzXAsf3.js").then((n) => n.t)
 				]);
@@ -58389,7 +59334,7 @@ async function registerSlackMonitorSlashCommands(params) {
 	let nativeCommands = [];
 	if (nativeEnabled) {
 		reg = await getCommandsRegistry();
-		const skillCommands = nativeSkillsEnabled ? (await import("./skill-commands-CKm7eCw6.js").then((n) => n.a)).listSkillCommandsForAgents({ cfg }) : [];
+		const skillCommands = nativeSkillsEnabled ? (await import("./skill-commands-D0xAWG0l.js").then((n) => n.a)).listSkillCommandsForAgents({ cfg }) : [];
 		nativeCommands = reg.listNativeCommandSpecsForConfig(cfg, {
 			skillCommands,
 			provider: "slack"