feat: replace Kimi K2.6 with GLM 5.1 as freebuff deployment-hours model

- Switch base2-free, editor-lite, code-reviewer-lite agents from kimi-k2.6 to z-ai/glm-5.1
- Update FREEBUFF_KIMI_MODEL_ID → FREEBUFF_GLM_MODEL_ID constant
- Update Fireworks deployment map (mjb4i7ea), model map, and pricing
- Remove moonshotai/kimi-k2.6 and kimi-k2.6:nitro from ModelName type
- Update freebuff model selector to show GLM first with 'Smartest' tagline
- Update all test files with new model IDs and deployment IDs
- Update docs and scripts to reference GLM instead of Kimi

Author: jahooma

agents/base2/base2.ts

@@ -25,7 +25,7 @@ export function createBase2(
   const isFree = mode === 'free' || mode === 'lite'
 
   const isSonnet = false
-  const model = isFree ? 'moonshotai/kimi-k2.6' : 'anthropic/claude-opus-4.7'
+  const model = isFree ? 'z-ai/glm-5.1' : 'anthropic/claude-opus-4.7'
 
   return {
     publisher,

agents/editor/editor-lite.ts

@@ -3,7 +3,7 @@ import { createCodeEditor } from './editor'
 import type { AgentDefinition } from '../types/agent-definition'
 
 const definition: AgentDefinition = {
-  ...createCodeEditor({ model: 'minimax' }),
+  ...createCodeEditor({ model: 'glm' }),
   id: 'editor-lite',
 }
 export default definition

agents/reviewer/code-reviewer-lite.ts

@@ -5,7 +5,7 @@ import { createReviewer } from './code-reviewer'
 const definition: SecretAgentDefinition = {
   id: 'code-reviewer-lite',
   publisher,
-  ...createReviewer('moonshotai/kimi-k2.6'),
+  ...createReviewer('z-ai/glm-5.1'),
 }
 
 export default definition

agents/types/agent-definition.ts

@@ -423,8 +423,6 @@ export type ModelName =
   // Other open source models
   | 'moonshotai/kimi-k2'
   | 'moonshotai/kimi-k2:nitro'
-  | 'moonshotai/kimi-k2.6'
-  | 'moonshotai/kimi-k2.6:nitro'
   | 'z-ai/glm-5'
   | 'z-ai/glm-5.1'
   | 'z-ai/glm-4.6'

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

@@ -6,6 +6,7 @@ import { Button } from './button'
 import {
   DEFAULT_FREEBUFF_MODEL_ID,
   FREEBUFF_DEPLOYMENT_HOURS_LABEL,
+  FREEBUFF_GLM_MODEL_ID,
   FREEBUFF_MODELS,
   isFreebuffModelAvailable,
 } from '@codebuff/common/constants/freebuff-models'
@@ -19,6 +20,11 @@ import { useTheme } from '../hooks/use-theme'
 
 import type { KeyEvent } from '@opentui/core'
 
+const FREEBUFF_MODEL_SELECTOR_MODELS = [
+  ...FREEBUFF_MODELS.filter((model) => model.id === FREEBUFF_GLM_MODEL_ID),
+  ...FREEBUFF_MODELS.filter((model) => model.id !== FREEBUFF_GLM_MODEL_ID),
+]
+
 /**
  * Dual-purpose model picker:
  *   - Pre-chat landing (session 'none'): user hasn't joined any queue. Picking
@@ -109,7 +115,7 @@ export const FreebuffModelSelector: React.FC = () => {
   const stackVertically = useMemo(() => {
     const BUTTON_CHROME = 4 // 2 border + 2 padding
     const GAP = 2
-    const total = FREEBUFF_MODELS.reduce((sum, model, idx) => {
+    const total = FREEBUFF_MODEL_SELECTOR_MODELS.reduce((sum, model, idx) => {
       const inner =
         2 /* indicator + space */ +
         model.displayName.length +
@@ -167,13 +173,15 @@ export const FreebuffModelSelector: React.FC = () => {
           }
           return
         }
-        const currentIdx = FREEBUFF_MODELS.findIndex((m) => m.id === focusedId)
+        const currentIdx = FREEBUFF_MODEL_SELECTOR_MODELS.findIndex(
+          (m) => m.id === focusedId,
+        )
         if (currentIdx === -1) return
-        const len = FREEBUFF_MODELS.length
+        const len = FREEBUFF_MODEL_SELECTOR_MODELS.length
         const nextIdx = isForward
           ? (currentIdx + 1) % len
           : (currentIdx - 1 + len) % len
-        const target = FREEBUFF_MODELS[nextIdx]
+        const target = FREEBUFF_MODEL_SELECTOR_MODELS[nextIdx]
         if (target) {
           key.preventDefault?.()
           setFocusedId(target.id)
@@ -198,7 +206,7 @@ export const FreebuffModelSelector: React.FC = () => {
           alignItems: 'flex-start',
         }}
       >
-        {FREEBUFF_MODELS.map((model) => {
+        {FREEBUFF_MODEL_SELECTOR_MODELS.map((model) => {
           // 'Selected' means the dot is filled and the label is bold. On the
           // landing screen ('none') this tracks the pre-focused pick; on the
           // queued screen it tracks the model the server has us on. Either

common/src/constants/free-agents.ts

@@ -28,7 +28,7 @@ export const FREE_MODE_AGENT_MODELS: Record<string, Set<string>> = {
   // Root orchestrator
   'base2-free': new Set([
     'minimax/minimax-m2.7',
-    'moonshotai/kimi-k2.6',
+    'z-ai/glm-5.1',
   ]),
 
   // File exploration agents
@@ -46,13 +46,13 @@ export const FREE_MODE_AGENT_MODELS: Record<string, Set<string>> = {
   // Editor for free mode
   'editor-lite': new Set([
     'minimax/minimax-m2.7',
-    'moonshotai/kimi-k2.6',
+    'z-ai/glm-5.1',
   ]),
 
   // Code reviewer for free mode
   'code-reviewer-lite': new Set([
     'minimax/minimax-m2.7',
-    'moonshotai/kimi-k2.6',
+    'z-ai/glm-5.1',
   ]),
 }
 

common/src/constants/freebuff-models.ts

@@ -18,7 +18,7 @@ export interface FreebuffModelOption {
 }
 
 export const FREEBUFF_DEPLOYMENT_HOURS_LABEL = '9am ET-5pm PT'
-export const FREEBUFF_KIMI_MODEL_ID = 'moonshotai/kimi-k2.6'
+export const FREEBUFF_GLM_MODEL_ID = 'z-ai/glm-5.1'
 
 export const FREEBUFF_MODELS = [
   {
@@ -28,9 +28,9 @@ export const FREEBUFF_MODELS = [
     availability: 'always',
   },
   {
-    id: FREEBUFF_KIMI_MODEL_ID,
-    displayName: 'Kimi K2.6',
-    tagline: 'Balanced',
+    id: FREEBUFF_GLM_MODEL_ID,
+    displayName: 'GLM 5.1',
+    tagline: 'Smartest',
     availability: 'deployment_hours',
   },
 ] as const satisfies readonly FreebuffModelOption[]

common/src/templates/initial-agents-dir/types/agent-definition.ts

@@ -423,8 +423,6 @@ export type ModelName =
   // Other open source models
   | 'moonshotai/kimi-k2'
   | 'moonshotai/kimi-k2:nitro'
-  | 'moonshotai/kimi-k2.6'
-  | 'moonshotai/kimi-k2.6:nitro'
   | 'z-ai/glm-5'
   | 'z-ai/glm-5.1'
   | 'z-ai/glm-4.6'

docs/freebuff-waiting-room.md

@@ -5,7 +5,7 @@
 The waiting room is the admission control layer for **free-mode** requests against the freebuff Fireworks deployments. It has three jobs:
 
 1. **Drip-admit users per model** — each selectable freebuff model has its own FIFO queue. Admission runs one tick (default `ADMISSION_TICK_MS`, 15s) that tries to admit one user per model, so heavier models can sit cold without starving lighter ones.
-2. **Gate on per-deployment health and hours** — a single fleet probe per tick (`getFleetHealth` in `web/src/server/free-session/fireworks-health.ts`) hits the Fireworks metrics endpoint and classifies each dedicated deployment as `healthy | degraded | unhealthy`. Only models whose deployment is `healthy` and currently available admit that tick; Kimi K2.6 is available during 9am ET-5pm PT on weekdays, while MiniMax M2.7 is serverless and always available.
+2. **Gate on per-deployment health and hours** — a single fleet probe per tick (`getFleetHealth` in `web/src/server/free-session/fireworks-health.ts`) hits the Fireworks metrics endpoint and classifies each dedicated deployment as `healthy | degraded | unhealthy`. Only models whose deployment is `healthy` and currently available admit that tick; GLM 5.1 is available during 9am ET-5pm PT on weekdays, while MiniMax M2.7 is serverless and always available.
 3. **One instance per account** — prevent a single user from running N concurrent freebuff CLIs to get N× throughput.
 
 Users who cannot be admitted immediately are placed in the queue for their chosen model and given an estimated wait time. Admitted users get a fixed-length session (default 1h) bound to the model they were admitted on; chat completions use that model for the life of the session.
@@ -149,8 +149,8 @@ The final tick result carries a `queueDepthByModel` map and a single `skipped` r
 | Constant | Location | Default | Purpose |
 |---|---|---|---|
 | `ADMISSION_TICK_MS` | `config.ts` | 15000 | How often the ticker fires. Up to one user is admitted per model per tick. |
-| `FREEBUFF_MODELS` | `common/src/constants/freebuff-models.ts` | `minimax-m2.7`, `kimi-k2.6` | Selectable models; each gets its own queue and admission slot. |
-| `FIREWORKS_DEPLOYMENT_MAP` | `web/src/llm-api/fireworks-config.ts` | `kimi-k2.6` | Models with dedicated Fireworks deployments. Models not listed are treated as `healthy` (serverless fallback) — drop this default when they migrate to their own deployments. |
+| `FREEBUFF_MODELS` | `common/src/constants/freebuff-models.ts` | `minimax-m2.7`, `glm-5.1` | Selectable models; each gets its own queue and admission slot. |
+| `FIREWORKS_DEPLOYMENT_MAP` | `web/src/llm-api/fireworks-config.ts` | `glm-5.1` | Models with dedicated Fireworks deployments. Models not listed are treated as `healthy` (serverless fallback) — drop this default when they migrate to their own deployments. |
 | `HEALTH_CACHE_TTL_MS` | `fireworks-health.ts` | 25000 | Fleet probe cache TTL. Sits just under the Fireworks 30s exporter cadence and 6 req/min rate limit. |
 | `FREEBUFF_SESSION_LENGTH_MS` | env | 3_600_000 | Session lifetime |
 | `FREEBUFF_SESSION_GRACE_MS` | env | 1_800_000 | Drain window after expiry — gate still admits requests so an in-flight agent can finish, but the CLI is expected to block new prompts. Hard cutoff at `expires_at + grace`. |
@@ -185,7 +185,7 @@ Response shapes:
   "queueDepth": 43,        // size of this model's queue
   "queueDepthByModel": {   // snapshot of every model's queue — powers the
     "minimax/minimax-m2.7": 43, //  "N ahead" hint in the selector. Missing
-    "moonshotai/kimi-k2.6": 4   //  entries should be treated as 0.
+    "z-ai/glm-5.1": 4   //  entries should be treated as 0.
   },
   "estimatedWaitMs": 384000,
   "queuedAt": "2026-04-17T12:00:00Z"
@@ -285,7 +285,7 @@ waitMs = (position - 1) * 24_000
 - Position 1 → 0 (next tick admits you)
 - Position 2 → 24s, and so on.
 
-`position` is scoped to this model's queue — a user at position 1 in the `minimax/minimax-m2.7` queue is not affected by the depth of the `moonshotai/kimi-k2.6` queue. The estimate is intentionally decoupled from the admission tick — it's a human-friendly rule-of-thumb for the UI, not a precise projection. Actual wait depends on admission-tick cadence, health-gated pauses, and deployment-hours availability (during a Kimi Fireworks incident or outside 9am ET-5pm PT, only Kimi's queue stalls; MiniMax keeps draining), so the real wait can be longer or shorter.
+`position` is scoped to this model's queue — a user at position 1 in the `minimax/minimax-m2.7` queue is not affected by the depth of the `z-ai/glm-5.1` queue. The estimate is intentionally decoupled from the admission tick — it's a human-friendly rule-of-thumb for the UI, not a precise projection. Actual wait depends on admission-tick cadence, health-gated pauses, and deployment-hours availability (during a GLM Fireworks incident or outside 9am ET-5pm PT, only GLM's queue stalls; MiniMax keeps draining), so the real wait can be longer or shorter.
 
 ## CLI Integration (frontend-side contract)
 
@@ -324,7 +324,7 @@ The `disabled` response means the server has the waiting room turned off. CLI tr
 | Spamming POST/GET to starve admission tick | Admission uses per-model Postgres advisory locks; DDoS protection is upstream (Next's global rate limits). Consider adding a per-user limiter on `/session` if traffic warrants. |
 | Repeatedly POSTing different models to get across every queue | Single row per user (PK on `user_id`); switching models moves the row, never clones it. A user holds exactly one queue slot at any time. |
 | Fireworks metrics endpoint down / slow | `getFleetHealth()` fails closed (timeout, non-OK, or missing API key) → every dedicated-deployment model is flagged `unhealthy` and its queue pauses. |
-| One deployment degraded while others are fine | Health is classified per-deployment; only the affected model's queue pauses, so a degraded Kimi deployment doesn't block MiniMax admissions. |
+| One deployment degraded while others are fine | Health is classified per-deployment; only the affected model's queue pauses, so a degraded GLM deployment doesn't block MiniMax admissions. |
 | Zombie expired sessions holding capacity | Swept on every admission tick, even when upstream is unhealthy |
 
 ## Testing

scripts/test-fireworks-cache-intervals.ts

@@ -13,7 +13,6 @@
  *
  * Models:
  *   glm-5.1   (default) — z-ai/glm-5.1
- *   kimi-k2.6           — moonshotai/kimi-k2.6
  *   minimax             — minimax/minimax-m2.5
  *
  * Flags:
@@ -25,11 +24,11 @@
  *   # Default glm-5.1 serverless with default intervals
  *   bun scripts/test-fireworks-cache-intervals.ts
  *
- *   # Custom Kimi deployment with a faster sweep
- *   bun scripts/test-fireworks-cache-intervals.ts kimi-k2.6 --deployment --intervals=30,60,120,300,600
+ *   # Custom GLM deployment with a faster sweep
+ *   bun scripts/test-fireworks-cache-intervals.ts glm-5.1 --deployment --intervals=30,60,120,300,600
  *
  *   # Long sweep up to 1 hour
- *   bun scripts/test-fireworks-cache-intervals.ts kimi-k2.6 --deployment --intervals=60,300,600,1200,1800,2700,3600
+ *   bun scripts/test-fireworks-cache-intervals.ts glm-5.1 --deployment --intervals=60,300,600,1200,1800,2700,3600
  */
 
 export {}
@@ -49,18 +48,11 @@ const MODEL_CONFIGS: Record<string, ModelConfig> = {
   'glm-5.1': {
     id: 'z-ai/glm-5.1',
     standardModel: 'accounts/fireworks/models/glm-5p1',
+    deploymentModel: 'accounts/james-65d217/deployments/mjb4i7ea',
     inputCostPerToken: 1.4 / 1_000_000,
     cachedInputCostPerToken: 0.26 / 1_000_000,
     outputCostPerToken: 4.4 / 1_000_000,
   },
-  'kimi-k2.6': {
-    id: 'moonshotai/kimi-k2.6',
-    standardModel: 'accounts/fireworks/models/kimi-k2p6',
-    deploymentModel: 'accounts/james-65d217/deployments/j8ar2x0y',
-    inputCostPerToken: 0.6 / 1_000_000,
-    cachedInputCostPerToken: 0.1 / 1_000_000,
-    outputCostPerToken: 3.0 / 1_000_000,
-  },
   minimax: {
     id: 'minimax/minimax-m2.5',
     standardModel: 'accounts/fireworks/models/minimax-m2p5',