Instant-admit free sessions when below per-model capacity

web/src/app/api/v1/freebuff/session/__tests__/session.test.ts

@@ -37,6 +37,13 @@ function makeSessionDeps(overrides: Partial<SessionDeps> = {}): SessionDeps & {
     rows,
     isWaitingRoomEnabled: () => true,
     graceMs: 30 * 60 * 1000,
+    sessionLengthMs: 60 * 60 * 1000,
+    // Keep instant-admit disabled in handler tests — they verify queue/state
+    // transitions, not admission policy. With capacity 0 the deps below
+    // aren't reached, so they're trivial stubs.
+    getInstantAdmitCapacity: () => 0,
+    activeCountForModel: async () => 0,
+    promoteQueuedUser: async () => null,
     now: () => now,
     getSessionRow: async (userId) => rows.get(userId) ?? null,
     queueDepthsByModel: async () => {

web/src/server/free-session/__tests__/public-api.test.ts

@@ -38,6 +38,27 @@ function makeDeps(overrides: Partial<SessionDeps> = {}): SessionDeps & {
     _now: () => currentNow,
     isWaitingRoomEnabled: () => true,
     graceMs: GRACE_MS,
+    sessionLengthMs: SESSION_LEN,
+    // Test default: instant-admit disabled (capacity 0) so existing FIFO
+    // queue tests stay green. Tests that exercise instant admission opt in
+    // via `getInstantAdmitCapacity: () => N`.
+    getInstantAdmitCapacity: () => 0,
+    activeCountForModel: async (model) => {
+      let n = 0
+      for (const r of rows.values()) {
+        if (r.status === 'active' && r.model === model) n++
+      }
+      return n
+    },
+    promoteQueuedUser: async ({ userId, model, sessionLengthMs, now }) => {
+      const row = rows.get(userId)
+      if (!row || row.status !== 'queued' || row.model !== model) return null
+      row.status = 'active'
+      row.admitted_at = now
+      row.expires_at = new Date(now.getTime() + sessionLengthMs)
+      row.updated_at = now
+      return row
+    },
     now: () => currentNow,
     getSessionRow: async (userId) => rows.get(userId) ?? null,
     endSession: async (userId) => {
@@ -192,6 +213,63 @@ describe('requestSession', () => {
     if (second.status !== 'active') throw new Error('unreachable')
     expect(second.instanceId).not.toBe('inst-1') // rotated
   })
+
+  test('instant-admit: below capacity admits the user in the same request', async () => {
+    const admitDeps = makeDeps({ getInstantAdmitCapacity: () => 3 })
+    const state = await requestSession({
+      userId: 'u1',
+      model: DEFAULT_MODEL,
+      deps: admitDeps,
+    })
+    expect(state.status).toBe('active')
+    if (state.status !== 'active') throw new Error('unreachable')
+    expect(state.remainingMs).toBe(SESSION_LEN)
+    // The row in storage is flipped too, so the next GET /session also sees active.
+    expect(admitDeps.rows.get('u1')?.status).toBe('active')
+  })
+
+  test('instant-admit: queues once active-count reaches capacity', async () => {
+    const admitDeps = makeDeps({ getInstantAdmitCapacity: () => 2 })
+    const s1 = await requestSession({
+      userId: 'u1',
+      model: DEFAULT_MODEL,
+      deps: admitDeps,
+    })
+    const s2 = await requestSession({
+      userId: 'u2',
+      model: DEFAULT_MODEL,
+      deps: admitDeps,
+    })
+    const s3 = await requestSession({
+      userId: 'u3',
+      model: DEFAULT_MODEL,
+      deps: admitDeps,
+    })
+    expect(s1.status).toBe('active')
+    expect(s2.status).toBe('active')
+    expect(s3.status).toBe('queued')
+  })
+
+  test('instant-admit: per-model capacities are independent', async () => {
+    // GLM saturated at 1 active, MiniMax still has room.
+    const admitDeps = makeDeps({
+      getInstantAdmitCapacity: (model) =>
+        model === DEFAULT_MODEL ? 1 : 10,
+    })
+    await requestSession({ userId: 'u1', model: DEFAULT_MODEL, deps: admitDeps })
+    const s2 = await requestSession({
+      userId: 'u2',
+      model: DEFAULT_MODEL,
+      deps: admitDeps,
+    })
+    const s3 = await requestSession({
+      userId: 'u3',
+      model: 'minimax/minimax-m2.7',
+      deps: admitDeps,
+    })
+    expect(s2.status).toBe('queued')
+    expect(s3.status).toBe('active')
+  })
 })
 
 describe('getSessionState', () => {

web/src/server/free-session/config.ts

@@ -39,3 +39,19 @@ export function getSessionLengthMs(): number {
 export function getSessionGraceMs(): number {
   return env.FREEBUFF_SESSION_GRACE_MS
 }
+
+/**
+ * Per-model instant-admit capacity: how many concurrent active sessions a
+ * deployment can hold before new joiners fall back to the FIFO queue + tick.
+ * Deployment-sizing knob — kept server-side so we can tune without bumping
+ * the shared `common` package that the CLI consumes. Unknown ids → 0 (always
+ * queue).
+ */
+const INSTANT_ADMIT_CAPACITY: Record<string, number> = {
+  'z-ai/glm-5.1': 50,
+  'minimax/minimax-m2.7': 200,
+}
+
+export function getInstantAdmitCapacity(id: string): number {
+  return INSTANT_ADMIT_CAPACITY[id] ?? 0
+}

web/src/server/free-session/public-api.ts

@@ -4,15 +4,19 @@ import {
 } from '@codebuff/common/constants/freebuff-models'
 
 import {
+  getInstantAdmitCapacity,
   getSessionGraceMs,
+  getSessionLengthMs,
   isWaitingRoomBypassedForEmail,
   isWaitingRoomEnabled,
 } from './config'
 import {
+  activeCountForModel,
   endSession,
   FreeSessionModelLockedError,
   getSessionRow,
   joinOrTakeOver,
+  promoteQueuedUser,
   queueDepthsByModel,
   queuePositionFor,
 } from './store'
@@ -35,11 +39,28 @@ export interface SessionDeps {
     model: string
     queuedAt: Date
   }) => Promise<number>
+  /** Instant-admit check: returns the number of active sessions currently
+   *  bound to a given model. Compared against the model's configured
+   *  `instantAdmitCapacity` to decide whether a new joiner skips the queue. */
+  activeCountForModel: (model: string) => Promise<number>
+  /** Instant-admit promotion: flips a specific queued row to active. Returns
+   *  the updated row or null if the row wasn't in a queued state. */
+  promoteQueuedUser: (params: {
+    userId: string
+    model: string
+    sessionLengthMs: number
+    now: Date
+  }) => Promise<InternalSessionRow | null>
+  /** Per-model capacity lookup. Indirected through deps so tests can
+   *  force-enable / force-disable instant admit without mutating the
+   *  shared model registry. */
+  getInstantAdmitCapacity: (model: string) => number
   isWaitingRoomEnabled: () => boolean
   /** Plain values, not getters: these never change at runtime. The deps
    *  interface uses values rather than thunks so tests can pass numbers
    *  inline without wrapping. */
   graceMs: number
+  sessionLengthMs: number
   now?: () => Date
 }
 
@@ -49,13 +70,19 @@ const defaultDeps: SessionDeps = {
   endSession,
   queueDepthsByModel,
   queuePositionFor,
+  activeCountForModel,
+  promoteQueuedUser,
+  getInstantAdmitCapacity,
   isWaitingRoomEnabled,
   get graceMs() {
     // Read-through getter so test overrides via env still work; the value
     // itself is materialized once per call. Cheaper than a thunk because
     // callers don't have to invoke a function.
     return getSessionGraceMs()
   },
+  get sessionLengthMs() {
+    return getSessionLengthMs()
+  },
 }
 
 const nowOf = (deps: SessionDeps): Date => (deps.now ?? (() => new Date()))()
@@ -145,6 +172,33 @@ export async function requestSession(params: {
     }
     throw err
   }
+
+  // Instant-admit: if the model has spare capacity (fewer active sessions
+  // than its configured `instantAdmitCapacity`), skip the waiting room
+  // entirely and flip the user to active in this same request. The tick
+  // + FIFO queue only engage once we hit the threshold, so backpressure
+  // kicks in exactly when the deployment needs it.
+  //
+  // Race note: two concurrent joiners may each see `active < capacity`
+  // and both get admitted, overshooting the cap by up to `concurrency - 1`.
+  // Capacities are chosen with headroom for this, and the configured
+  // value is a comfort threshold not a hard ceiling.
+  if (row.status === 'queued') {
+    const capacity = deps.getInstantAdmitCapacity(model)
+    if (capacity > 0) {
+      const activeCount = await deps.activeCountForModel(model)
+      if (activeCount < capacity) {
+        const promoted = await deps.promoteQueuedUser({
+          userId: params.userId,
+          model,
+          sessionLengthMs: deps.sessionLengthMs,
+          now: nowOf(deps),
+        })
+        if (promoted) row = promoted
+      }
+    }
+  }
+
   const view = await viewForRow(params.userId, deps, row)
   if (!view) {
     throw new Error(

web/src/server/free-session/store.ts

@@ -176,6 +176,24 @@ export async function queueDepthsByModel(): Promise<Record<string, number>> {
   return out
 }
 
+/**
+ * Count of rows currently in `active` status for one model — the threshold
+ * check that gates instant admission. Hot-path lookup; callers avoid the
+ * full `activeCountsByModel` scan when they only need one model's count.
+ */
+export async function activeCountForModel(model: string): Promise<number> {
+  const rows = await db
+    .select({ n: count() })
+    .from(schema.freeSession)
+    .where(
+      and(
+        eq(schema.freeSession.status, 'active'),
+        eq(schema.freeSession.model, model),
+      ),
+    )
+  return Number(rows[0]?.n ?? 0)
+}
+
 /**
  * Single-query read of active-row counts bucketed by model. Mirrors
  * `queueDepthsByModel` so the admission tick can log per-model utilization
@@ -333,6 +351,43 @@ export async function admitFromQueue(params: {
   })
 }
 
+/**
+ * Promote a specific queued user to active. Used by the instant-admit path
+ * in `requestSession` when the model's active-session count is below its
+ * configured capacity — skips the FIFO advisory-lock dance because each
+ * call targets a distinct (user_id, model) and the UPDATE is a no-op if
+ * the row isn't queued any more.
+ *
+ * Returns the updated row or null if the row was not in the expected
+ * (queued, same-model) state.
+ */
+export async function promoteQueuedUser(params: {
+  userId: string
+  model: string
+  sessionLengthMs: number
+  now: Date
+}): Promise<InternalSessionRow | null> {
+  const { userId, model, sessionLengthMs, now } = params
+  const expiresAt = new Date(now.getTime() + sessionLengthMs)
+  const [row] = await db
+    .update(schema.freeSession)
+    .set({
+      status: 'active',
+      admitted_at: now,
+      expires_at: expiresAt,
+      updated_at: now,
+    })
+    .where(
+      and(
+        eq(schema.freeSession.user_id, userId),
+        eq(schema.freeSession.status, 'queued'),
+        eq(schema.freeSession.model, model),
+      ),
+    )
+    .returning()
+  return (row as InternalSessionRow | undefined) ?? null
+}
+
 /** Stable 31-bit hash so model-keyed advisory lock ids don't overflow int4. */
 function hashStringToInt32(s: string): number {
   let h = 0