overhaul freebuff premium sessions

Author: jahooma

cli/src/components/freebuff-model-selector.tsx

@@ -8,6 +8,8 @@ import {
   FREEBUFF_GEMINI_PRO_MODEL_ID,
   FREEBUFF_KIMI_MODEL_ID,
   FREEBUFF_MODELS,
+  FREEBUFF_PREMIUM_SESSION_LIMIT,
+  isFreebuffPremiumModelId,
   getFreebuffDeploymentAvailabilityLabel,
   isFreebuffModelAvailable,
 } from '@codebuff/common/constants/freebuff-models'
@@ -34,6 +36,10 @@ const FREEBUFF_MODEL_SELECTOR_MODELS = [
   ),
 ]
 
+function formatSessionUnits(units: number): string {
+  return Number.isInteger(units) ? String(units) : units.toFixed(1)
+}
+
 /**
  * Dual-purpose model picker:
  *   - Pre-chat landing (session 'none'): user hasn't joined any queue. Picking
@@ -47,8 +53,7 @@ const FREEBUFF_MODEL_SELECTOR_MODELS = [
  * switch. Mouse clicks are still one-step. On the landing screen, pressing
  * Enter on the already-focused model also commits — there's nothing to lose.
  *
- * Each row shows a live "N ahead" count sourced from the server's
- * `queueDepthByModel` snapshot so the choice is informed.
+ * Each row shows session-budget information so the choice is informed.
  */
 export const FreebuffModelSelector: React.FC = () => {
   const theme = useTheme()
@@ -86,48 +91,10 @@ export const FreebuffModelSelector: React.FC = () => {
     }
   }, [now, selectedModel, session, setSelectedModel])
 
-  // Landing ('none'): depths come from the server snapshot, no "self" to
-  // subtract. In-queue ('queued'): for the user's queue, "ahead" is
-  // `position - 1` (themselves don't count); for every other queue, switching
-  // would land them at the back, so it's that queue's full depth. Null before
-  // any snapshot so the UI doesn't flash misleading zeros — in particular,
-  // landing mode after a session ends initially sets status='none' with no
-  // queueDepthByModel; returning null here keeps the hint blank until the
-  // fetch lands, instead of showing "No wait" on every row.
-  const aheadByModel = useMemo<Record<string, number> | null>(() => {
-    if (session?.status === 'none') {
-      if (!session.queueDepthByModel) return null
-      const depths = session.queueDepthByModel
-      const out: Record<string, number> = {}
-      for (const { id } of FREEBUFF_MODELS) out[id] = depths[id] ?? 0
-      return out
-    }
-    if (session?.status === 'queued') {
-      const depths = session.queueDepthByModel ?? {}
-      const out: Record<string, number> = {}
-      for (const { id } of FREEBUFF_MODELS) {
-        out[id] =
-          id === session.model
-            ? Math.max(0, session.position - 1)
-            : (depths[id] ?? 0)
-      }
-      return out
-    }
-    return null
-  }, [session])
-
-  // Pad the trailing hint ("3 ahead", "No wait", "…") to a fixed width so
-  // buttons don't visibly resize when the queue depth ticks down (12 → 9) or
-  // when the user's selection moves between queues. The tagline is shown
-  // inline with the name now, so it's no longer part of this slot.
+  // Pad the trailing hint to a fixed width so buttons don't visibly resize
+  // when quota snapshots change.
   const hintWidth = useMemo(
-    () =>
-      Math.max(
-        'No wait'.length,
-        '999 ahead'.length,
-        'Used today'.length,
-        'Limit used'.length,
-      ),
+    () => Math.max('Unlimited'.length, '5/5 used'.length, '4.9/5 used'.length),
     [],
   )
 
@@ -246,9 +213,10 @@ export const FreebuffModelSelector: React.FC = () => {
           const isFocused = focusedId === model.id && !isSelected
           const isAvailable = isFreebuffModelAvailable(model.id, new Date(now))
           const rateLimit = rateLimitsByModel?.[model.id]
-          const isQuotaExhausted =
-            rateLimit !== undefined && rateLimit.recentCount >= rateLimit.limit
-          const canJoin = isAvailable && !isQuotaExhausted
+          const isPremium = isFreebuffPremiumModelId(model.id)
+          const hasFullPremiumSession =
+            !rateLimit || rateLimit.recentCount < rateLimit.limit
+          const canJoin = isAvailable && hasFullPremiumSession
           const indicator = isSelected ? '●' : isFocused ? '›' : '○'
           const indicatorColor = isSelected
             ? theme.primary
@@ -263,18 +231,12 @@ export const FreebuffModelSelector: React.FC = () => {
           // anything except re-picking the queue we're already in.
           const interactable =
             !pending && canJoin && model.id !== committedModelId
-          const ahead = aheadByModel?.[model.id]
-          const hint = !isAvailable
-            ? 'Closed'
-            : isQuotaExhausted
-              ? model.id === FREEBUFF_GEMINI_PRO_MODEL_ID
-                ? 'Used today'
-                : 'Limit used'
-              : ahead === undefined
-                ? ''
-                : ahead === 0
-                  ? 'No wait'
-                  : `${ahead} ahead`
+          const quotaHint = rateLimit
+            ? `${formatSessionUnits(rateLimit.recentCount)}/${rateLimit.limit} used`
+            : isPremium
+              ? `0/${FREEBUFF_PREMIUM_SESSION_LIMIT} used`
+              : 'Unlimited'
+          const hint = isAvailable ? quotaHint : 'Closed'
           const hintColor = canJoin ? theme.muted : theme.secondary
 
           const borderColor = isSelected

cli/src/components/status-bar.tsx

@@ -66,6 +66,9 @@ const formatSessionRemaining = (ms: number): string => {
   return minutes === 0 ? `${hours}h left` : `${hours}h ${minutes}m left`
 }
 
+const formatSessionUnits = (units: number): string =>
+  Number.isInteger(units) ? String(units) : units.toFixed(1)
+
 interface StatusBarProps {
   timerStartTime: number | null
   isAtBottom: boolean
@@ -131,7 +134,8 @@ export const StatusBar = ({
 
       case 'clipboard':
         // Use green color for feedback success messages
-        const isFeedbackSuccess = statusIndicatorState.message.includes('Feedback sent')
+        const isFeedbackSuccess =
+          statusIndicatorState.message.includes('Feedback sent')
         return (
           <span fg={isFeedbackSuccess ? theme.success : theme.primary}>
             {statusIndicatorState.message}
@@ -142,12 +146,7 @@ export const StatusBar = ({
         return <span fg={theme.success}>Reconnected</span>
 
       case 'retrying':
-        return (
-          <ShimmerText
-            text="retrying..."
-            primaryColor={theme.warning}
-          />
-        )
+        return <ShimmerText text="retrying..." primaryColor={theme.warning} />
 
       case 'connecting':
         return <ShimmerText text="connecting..." />
@@ -180,9 +179,16 @@ export const StatusBar = ({
             freebuffSession?.status === 'active'
               ? getFreebuffModel(freebuffSession.model).displayName
               : null
+          const quotaText =
+            freebuffSession?.status === 'active' && freebuffSession.rateLimit
+              ? `Premium ${formatSessionUnits(freebuffSession.rateLimit.recentCount)}/${freebuffSession.rateLimit.limit} used · `
+              : freebuffSession?.status === 'active'
+                ? 'Unlimited · '
+                : ''
           return (
             <span fg={isUrgent ? theme.warning : theme.secondary}>
-              {modelName ? `${modelName} · ` : ''}Free session ·{' '}
+              {modelName ? `${modelName} · ` : ''}
+              {quotaText}Free session ·{' '}
               {formatSessionRemaining(sessionProgress.remainingMs)}
             </span>
           )
@@ -260,12 +266,18 @@ export const StatusBar = ({
         }}
       >
         <text style={{ wrapMode: 'none' }}>{elapsedTimeContent}</text>
-        {onStop && (statusIndicatorState.kind === 'waiting' || statusIndicatorState.kind === 'streaming') && (
-          <StatusActionButton onClick={onStop}>■ Esc</StatusActionButton>
-        )}
-        {onEndSession && statusIndicatorState.kind === 'idle' && freebuffSession?.status === 'active' && (
-          <StatusActionButton onClick={onEndSession}>✕ End session</StatusActionButton>
-        )}
+        {onStop &&
+          (statusIndicatorState.kind === 'waiting' ||
+            statusIndicatorState.kind === 'streaming') && (
+            <StatusActionButton onClick={onStop}>■ Esc</StatusActionButton>
+          )}
+        {onEndSession &&
+          statusIndicatorState.kind === 'idle' &&
+          freebuffSession?.status === 'active' && (
+            <StatusActionButton onClick={onEndSession}>
+              ✕ End session
+            </StatusActionButton>
+          )}
         {sessionProgress !== null &&
           sessionProgress.remainingMs < COUNTDOWN_VISIBLE_MS &&
           statusIndicatorState.kind !== 'idle' && (

cli/src/components/waiting-room-screen.tsx

@@ -56,6 +56,9 @@ const formatRetryAfter = (ms: number): string => {
   return rem === 0 ? `${hours}h` : `${hours}h ${rem}m`
 }
 
+const formatSessionUnits = (units: number): string =>
+  Number.isInteger(units) ? String(units) : units.toFixed(1)
+
 const PRIVACY_SIGNAL_LABELS: Partial<Record<FreebuffIpPrivacySignal, string>> =
   {
     anonymous: 'anonymized network',
@@ -260,17 +263,16 @@ export const WaitingRoomScreen: React.FC<WaitingRoomScreenProps> = ({
                   <span>Elapsed </span>
                   {formatElapsed(elapsedMs)}
                 </text>
-                {/* Per-model session quota (e.g. Kimi K2.6 caps at 5/12h). Only
-                    rendered for rate-limited models so the Minimax queue stays
-                    clutter-free. */}
+                {/* Premium session quota. Minimax is unlimited, so it has no
+                    rateLimit payload and skips this line. */}
                 {session.rateLimit && (
                   <text style={{ fg: theme.muted, alignSelf: 'flex-start' }}>
-                    <span>Sessions </span>
+                    <span>Premium sessions </span>
                     <span fg={theme.foreground}>
-                      {session.rateLimit.recentCount} /{' '}
+                      {formatSessionUnits(session.rateLimit.recentCount)} /{' '}
                       {session.rateLimit.limit}
                     </span>
-                    <span> used in last {session.rateLimit.windowHours}h</span>
+                    <span> used today</span>
                   </text>
                 )}
               </box>
@@ -343,8 +345,8 @@ export const WaitingRoomScreen: React.FC<WaitingRoomScreenProps> = ({
             </>
           )}
 
-          {/* Per-model session quota exhausted (e.g. 5+ Kimi sessions in the
-              last 12h). Terminal for this run — the user can exit and come
+          {/* Shared premium-session quota exhausted. Terminal for this run —
+              the user can exit and come
               back once the oldest session in the window rolls off. */}
           {session?.status === 'rate_limited' && (
             <>
@@ -354,10 +356,9 @@ export const WaitingRoomScreen: React.FC<WaitingRoomScreenProps> = ({
               <text style={{ fg: theme.muted, wrapMode: 'word' }}>
                 You've used{' '}
                 <span fg={theme.foreground}>
-                  {session.recentCount} of {session.limit}
+                  {formatSessionUnits(session.recentCount)} of {session.limit}
                 </span>{' '}
-                hour-long sessions on {session.model} in the last{' '}
-                {session.windowHours}h. Try again in{' '}
+                premium sessions today. Try again in{' '}
                 <span fg={theme.foreground}>
                   {formatRetryAfter(session.retryAfterMs)}
                 </span>

common/src/constants/freebuff-models.ts

@@ -25,6 +25,8 @@ export const FREEBUFF_GEMINI_PRO_MODEL_ID = 'google/gemini-3.1-pro-preview'
 export const FREEBUFF_GLM_MODEL_ID = 'z-ai/glm-5.1'
 export const FREEBUFF_KIMI_MODEL_ID = 'moonshotai/kimi-k2.6'
 export const FREEBUFF_MINIMAX_MODEL_ID = 'minimax/minimax-m2.7'
+export const FREEBUFF_PREMIUM_SESSION_LIMIT = 5
+export const FREEBUFF_PREMIUM_SESSION_WINDOW_HOURS = 20
 const FREEBUFF_EASTERN_TIMEZONE = 'America/New_York'
 const FREEBUFF_PACIFIC_TIMEZONE = 'America/Los_Angeles'
 
@@ -45,13 +47,13 @@ export const FREEBUFF_MODELS = [
   {
     id: FREEBUFF_GEMINI_PRO_MODEL_ID,
     displayName: 'Gemini 3.1 Pro',
-    tagline: 'Deepest, 1/day',
+    tagline: 'Deepest',
     availability: 'always',
   },
   {
     id: FREEBUFF_MINIMAX_MODEL_ID,
     displayName: 'MiniMax M2.7',
-    tagline: 'Fastest',
+    tagline: 'Fastest, unlimited',
     availability: 'always',
   },
   {
@@ -71,6 +73,12 @@ export const LEGACY_FREEBUFF_MODELS = [
   },
 ] as const satisfies readonly FreebuffModelOption[]
 
+export const FREEBUFF_PREMIUM_MODEL_IDS = [
+  FREEBUFF_GEMINI_PRO_MODEL_ID,
+  FREEBUFF_KIMI_MODEL_ID,
+  FREEBUFF_GLM_MODEL_ID,
+] as const
+
 export const SUPPORTED_FREEBUFF_MODELS = [
   ...FREEBUFF_MODELS,
   ...LEGACY_FREEBUFF_MODELS,
@@ -79,6 +87,7 @@ export const SUPPORTED_FREEBUFF_MODELS = [
 export type FreebuffModelId = (typeof FREEBUFF_MODELS)[number]['id']
 export type SupportedFreebuffModelId =
   (typeof SUPPORTED_FREEBUFF_MODELS)[number]['id']
+export type FreebuffPremiumModelId = (typeof FREEBUFF_PREMIUM_MODEL_IDS)[number]
 
 /** What new freebuff users see selected in the picker. May not be currently
  *  available (Kimi is closed outside deployment hours); callers that need an
@@ -113,6 +122,13 @@ export function isSupportedFreebuffModelId(
   return SUPPORTED_FREEBUFF_MODELS.some((m) => m.id === id)
 }
 
+export function isFreebuffPremiumModelId(
+  id: string | null | undefined,
+): id is FreebuffPremiumModelId {
+  if (!id) return false
+  return FREEBUFF_PREMIUM_MODEL_IDS.some((modelId) => modelId === id)
+}
+
 export function resolveSupportedFreebuffModel(
   id: string | null | undefined,
 ): SupportedFreebuffModelId {

common/src/types/freebuff-session.ts

@@ -7,11 +7,12 @@
  */
 
 /**
- * Per-model usage counter surfaced to the CLI so the waiting-room UI can
- * render "N of M sessions used" alongside queue/active state. Present when
- * the joined model has a rate limit applied. `recentCount` is the number of
- * admissions inside `windowHours` at the time the response was produced —
- * see also the standalone `rate_limited` status for the reject path.
+ * Usage counter surfaced to the CLI so the waiting-room UI can render
+ * "N of M sessions used" alongside queue/active state. Present when the
+ * joined model consumes premium Freebuff sessions. `recentCount` is the
+ * rounded session units inside `windowHours` at the time the response was
+ * produced — see also the standalone `rate_limited` status for the reject
+ * path.
  */
 export interface FreebuffSessionRateLimit {
   model: string
@@ -60,9 +61,9 @@ export type FreebuffSessionServerResponse =
        *  committing the user to a queue. Present on GET responses; not
        *  returned from POST (POST never produces `none`). */
       queueDepthByModel?: Record<string, number>
-      /** Current quota snapshots for rate-limited models, keyed by model id.
-       *  Lets the picker show exhausted daily/session caps before the user
-       *  commits to a queue. */
+      /** Current quota snapshots for premium models, keyed by model id. Lets
+       *  the picker show daily premium-session usage before the user commits
+       *  to a queue. */
       rateLimitsByModel?: FreebuffSessionRateLimitByModel
     }
   | {
@@ -80,9 +81,7 @@ export type FreebuffSessionServerResponse =
       queueDepthByModel: Record<string, number>
       estimatedWaitMs: number
       queuedAt: string
-      /** Rate-limit quota for rate-limited models. Absent
-       *  for unlimited models or when the status was produced outside the
-       *  rate-limit check path (e.g. pure read via GET). */
+      /** Premium-session quota for this model. Absent for unlimited models. */
       rateLimit?: FreebuffSessionRateLimit
       rateLimitsByModel?: FreebuffSessionRateLimitByModel
     }
@@ -94,9 +93,7 @@ export type FreebuffSessionServerResponse =
       admittedAt: string
       expiresAt: string
       remainingMs: number
-      /** Rate-limit quota for rate-limited models. Absent
-       *  for unlimited models or when the status was produced outside the
-       *  rate-limit check path (e.g. pure read via GET). */
+      /** Premium-session quota for this model. Absent for unlimited models. */
       rateLimit?: FreebuffSessionRateLimit
       rateLimitsByModel?: FreebuffSessionRateLimitByModel
     }
@@ -161,21 +158,20 @@ export type FreebuffSessionServerResponse =
       status: 'banned'
     }
   | {
-      /** User has used up their per-model admission quota in the rolling
-       *  window. Returned from POST
-       *  /session before the user is placed in the queue. `retryAfterMs` is
-       *  the time until the oldest admission inside the window falls off
-       *  and one quota slot opens up — clients should show the user when
-       *  they can try again. Terminal for the CLI's current poll session;
+      /** User has used up their shared premium-session quota in the rolling
+       *  window. Returned from POST /session before the user is placed in the
+       *  queue. `retryAfterMs` is the time until enough session units fall out
+       *  of the window to open one quota slot — clients should show the user
+       *  when they can try again. Terminal for the CLI's current poll session;
        *  the user can exit and come back later. */
       status: 'rate_limited'
       /** The freebuff model the user tried to join. */
       model: string
-      /** Max admissions permitted per window (e.g. 5). */
+      /** Max premium session units permitted per window (e.g. 5). */
       limit: number
       /** Rolling window size in hours (e.g. 20). */
       windowHours: number
-      /** Admission count inside the window at check time — will be ≥ limit. */
+      /** Premium session units inside the window at check time — will be ≥ limit. */
       recentCount: number
       /** Milliseconds from now until the oldest admission in the window
        *  exits and the user regains one quota slot. */

packages/internal/src/db/migrations/0048_magical_firebird.sql

@@ -0,0 +1 @@
+ALTER TABLE "free_session_admit" ADD COLUMN "session_units" numeric(3, 1) DEFAULT '1.0' NOT NULL;