Merge remote-tracking branch 'upstream/main' into agent-tasks/1125

Author: liumingyao.marvin

README.md

@@ -38,9 +38,11 @@ Freebuff includes a curated model catalog. The regular picker currently offers:
 | **DeepSeek V4 Flash 07/31** | Full access             | Fast coding and tool use; pauses during peak hours                |
 | **MiMo 2.5**                | Full and limited access | The limited-mode default; balanced performance with image support |
 | **Ox Alpha**                | Full and limited access | 1M context, no session cost; experimental — may be unreliable     |
-| **DeepSeek V4 Pro**         | Full access             | Deepest reasoning                                                 |
+| **GLM 5.3 Flash**           | Full access             | Deepest reasoning; 2 sessions a day                               |
 
-Per-model daily caps are **gone**: every model now draws on your normal daily sessions rather than its own separate limit. MiMo 2.5 and Ox Alpha stay unmetered and cost no session at all. Models may still serve from a quantized (Q8_0) build.
+Most models draw on your normal daily sessions rather than a separate limit. GLM 5.3 Flash is the exception, capped at 2 sessions a day while we measure what it costs at scale. MiMo 2.5 and Ox Alpha stay unmetered and cost no session at all. Models may still serve from a quantized (Q8_0) build.
+
+DeepSeek V4 Pro was retired from the catalog; GLM 5.3 Flash replaces it as the deep-reasoning pick.
 
 Beyond the regular picker:
 

README.zh-CN.md

@@ -38,9 +38,11 @@ Freebuff 提供经过筛选的模型目录。常规模型选择器目前包括
 | **DeepSeek V4 Flash 07/31** | 完整访问       | 快速编程和工具调用；高峰时段暂停                 |
 | **MiMo 2.5**                | 完整和受限访问 | 受限模式的默认模型；均衡性能并支持图像           |
 | **Ox Alpha**                | 完整和受限访问 | 100 万上下文，不消耗会话；实验性功能，可能不稳定 |
-| **DeepSeek V4 Pro**         | 完整访问       | 推理最深入                                       |
+| **GLM 5.3 Flash**           | 完整访问       | 推理最深入；每天 2 次会话                        |
 
-单个模型的每日上限**已取消**：所有模型现在都使用你的常规每日会话，而不再各自设限。MiMo 2.5 和 Ox Alpha 保持无限使用，完全不消耗会话。模型仍可能由量化（Q8_0）版本提供服务。
+大多数模型使用你的常规每日会话，而不再各自设限。GLM 5.3 Flash 是例外：在我们评估其规模化成本期间，每天限用 2 次会话。MiMo 2.5 和 Ox Alpha 保持无限使用，完全不消耗会话。模型仍可能由量化（Q8_0）版本提供服务。
+
+DeepSeek V4 Pro 已从模型目录中下线，由 GLM 5.3 Flash 接替其深度推理的位置。
 
 常规模型选择器之外：
 

agents/__tests__/base3.test.ts

@@ -12,6 +12,7 @@ import base3FreeDeepseekFlash from '../base3-free-deepseek-flash'
 import base3FreeDeepseekFlashEvals from '../base3-free-deepseek-flash-evals'
 import base3FreeFable from '../base3-free-fable'
 import base3FreeGlm from '../base3-free-glm'
+import base3FreeGlmV53Flash from '../base3-free-glm-5-3-flash'
 import base3FreeLuna from '../base3-free-luna'
 import base3FreeMimo from '../base3-free-mimo'
 import base3FreeMinimaxM3 from '../base3-free-minimax-m3'
@@ -45,16 +46,17 @@ const CLI_ROOTS = [
   base3FreeMinimaxM3,
   base3FreeMimo,
   base3FreeGlm,
+  base3FreeGlmV53Flash,
   base3FreeLuna,
   base3FreeFable,
   base3FreeOxAlpha,
 ]
 
 describe('base3 CLI roots', () => {
   test('keeps the efficiency flags the runtime reads', () => {
-    // 12 since Ox Alpha reached the CLI on 2026-08-24. The count is asserted so
-    // a root added without the flags below cannot slip in unnoticed.
-    expect(CLI_ROOTS.length).toBe(12)
+    // 13 since GLM 5.3 Flash shipped on 2026-08-26. The count is asserted so a
+    // root added without the flags below cannot slip in unnoticed.
+    expect(CLI_ROOTS.length).toBe(13)
     for (const agent of CLI_ROOTS) {
       // Windowed reads + the 100-entry glob cap + search-first tool wording.
       expect(agent.windowedFileReads).toBe(true)

agents/base-chat.ts

@@ -66,6 +66,8 @@ End every response by calling the suggest_followups tool with exactly 3 followup
       'meta/muse-spark-1.2-contributor': 1_000_000,
       // Ox Alpha: 1,048,576 published, entered low for the same reason.
       'stealth/ox-alpha': 1_000_000,
+      // GLM 5.3 Flash: 1,310,720 published, entered low for the same reason.
+      'z-ai/glm-5.3-flash': 1_000_000,
     }
 
     /** For any model not listed above. Assuming a window is smaller than it is

agents/base2/base2-free-glm-5-3-flash.ts

@@ -0,0 +1,13 @@
+import { FREEBUFF_GLM_V53_FLASH_MODEL_ID } from '@codebuff/common/constants/freebuff-models'
+
+import { createBase2 } from './base2'
+
+const definition = {
+  ...createBase2('free', {
+    model: FREEBUFF_GLM_V53_FLASH_MODEL_ID,
+  }),
+  id: 'base2-free-glm-5-3-flash',
+  displayName: 'Buffy the GLM 5.3 Flash Free Orchestrator',
+}
+
+export default definition

agents/base3-free-glm-5-3-flash.ts

@@ -0,0 +1,19 @@
+import { FREEBUFF_GLM_V53_FLASH_MODEL_ID } from '@codebuff/common/constants/freebuff-models'
+
+import { createBase3CliRoot } from './base3'
+
+// No `reasoningOptions`, for the reason spelled out on base3-free-luna: an
+// agent-declared reasoning reaches the wire as `body.reasoning`, which makes
+// the definition rather than the user the authority on effort. It costs nothing
+// here in particular — OpenRouter reports no effort levels for this model, so
+// the catalog row declares no ladder and the picker shows no control.
+const definition = {
+  ...createBase3CliRoot({
+    model: FREEBUFF_GLM_V53_FLASH_MODEL_ID,
+    isFreebuff: true,
+  }),
+  id: 'base3-free-glm-5-3-flash',
+  displayName: 'Buffy on GLM 5.3 Flash',
+}
+
+export default definition

agents/reviewer/code-reviewer-glm-5-3-flash.ts

@@ -0,0 +1,16 @@
+import { FREEBUFF_GLM_V53_FLASH_MODEL_ID } from '@codebuff/common/constants/freebuff-models'
+
+import { publisher } from '../constants'
+import type { SecretAgentDefinition } from '../types/secret-agent-definition'
+import { createReviewer } from './code-reviewer'
+
+// Runs the SAME model as the root that spawns it. Not a style choice: the
+// chat-completions session gate rejects any request whose model differs from
+// the one the session was admitted on, so a cross-model reviewer 403s mid-run.
+const definition: SecretAgentDefinition = {
+  id: 'code-reviewer-glm-5-3-flash',
+  publisher,
+  ...createReviewer(FREEBUFF_GLM_V53_FLASH_MODEL_ID),
+}
+
+export default definition

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

@@ -10,7 +10,7 @@ import {
   FALLBACK_FREEBUFF_MODEL_ID,
   FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID,
   FREEBUFF_MIMO_V25_MODEL_ID,
-  FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+  FREEBUFF_GLM_V53_FLASH_MODEL_ID,
   FREEBUFF_FABLE_5_MODEL_ID,
   FREEBUFF_GLM_V52_MODEL_ID,
   FREEBUFF_GPT_5_6_LUNA_MODEL_ID,
@@ -164,18 +164,19 @@ describe('FreebuffModelSelector tier layout', () => {
     // premium or the tier headers it is being ordered against don't apply to
     // it, non-hero or the landing picker opens collapsed and there are no tier
     // headers at all. The hero is GPT-5.6 Luna since 2026-08-24, which leaves
-    // V4 Pro as the only row that is both. Flash filled this slot until
+    // GLM 5.3 Flash as the only row that is both. Flash filled this slot until
     // 2026-08-24, when it left FREEBUFF_PREMIUM_MODEL_IDS and moved down into
-    // UNLIMITED -- below the header this asserts it sits above.
+    // UNLIMITED -- below the header this asserts it sits above; V4 Pro until
+    // its withdrawal on 2026-08-26.
     useFreebuffModelStore
       .getState()
-      .setSelectedModel(FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID)
+      .setSelectedModel(FREEBUFF_GLM_V53_FLASH_MODEL_ID)
 
     const setup = await renderSelector()
     const frame = setup.captureCharFrame()
     const premiumHeaderIndex = frame.indexOf('PREMIUM')
     const recommendedModelIndex = frame.indexOf('GPT-5.6 Luna')
-    const selectedModelIndex = frame.indexOf('DeepSeek V4 Pro')
+    const selectedModelIndex = frame.indexOf('GLM 5.3 Flash')
     const unlimitedHeaderIndex = frame.indexOf('UNLIMITED')
 
     expect(premiumHeaderIndex).toBeGreaterThanOrEqual(0)
@@ -185,7 +186,7 @@ describe('FreebuffModelSelector tier layout', () => {
     // 2026-08-20 and left the picker entirely.
     expect(unlimitedHeaderIndex).toBeGreaterThan(selectedModelIndex)
     // The cursor sits on the SAVED pick, not on the recommendation.
-    expect(frame).toContain('› DeepSeek V4 Pro')
+    expect(frame).toContain('› GLM 5.3 Flash')
     expect(frame).not.toContain('› GPT-5.6 Luna')
   })
 
@@ -377,7 +378,7 @@ describe('FreebuffModelSelector tier layout', () => {
     })
     useFreebuffModelStore
       .getState()
-      .setSelectedModel(FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID)
+      .setSelectedModel(FREEBUFF_GLM_V53_FLASH_MODEL_ID)
     const setup = await renderSelector()
 
     flushSync(() => {
@@ -486,16 +487,16 @@ describe('FreebuffModelSelector tier layout', () => {
       accessTier: 'full',
       // Flash sends no pool row at all since 2026-08-24: it is unmetered.
       rateLimitsByModel: {
-        [FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID]: pool(
-          FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+        [FREEBUFF_GPT_5_6_LUNA_MODEL_ID]: pool(
+          FREEBUFF_GPT_5_6_LUNA_MODEL_ID,
           'premium',
           'Premium',
           4,
         ),
-        [FREEBUFF_GPT_5_6_LUNA_MODEL_ID]: pool(
-          FREEBUFF_GPT_5_6_LUNA_MODEL_ID,
-          'luna',
-          'Luna',
+        [FREEBUFF_GLM_V53_FLASH_MODEL_ID]: pool(
+          FREEBUFF_GLM_V53_FLASH_MODEL_ID,
+          'glm_v53_flash',
+          'GLM 5.3 Flash',
           2,
         ),
       },
@@ -505,8 +506,8 @@ describe('FreebuffModelSelector tier layout', () => {
       // NOT the hero, so the picker opens expanded and the chip under test is
       // drawn at all. Luna took the hero slot on 2026-08-24; selecting it here
       // collapses the list to a single row and the chip disappears. Flash also
-      // supplies the warning-ONLY second line asserted below, which no premium
-      // row has any more now that V4 Pro's carries the chip as well.
+      // supplies the warning-ONLY second line asserted below, which the chip
+      // row cannot: every row carrying a pool row here also carries a chip.
       .setSelectedModel(FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID)
 
     const frame = (await renderSelector()).captureCharFrame()
@@ -522,11 +523,13 @@ describe('FreebuffModelSelector tier layout', () => {
       ]
     }
     const lines = frame.split('\n')
-    // V4 Pro's second line carries the training warning AND the chip. Anchored
-    // on the chip, so a chip that stops being drawn fails here rather than
-    // quietly re-measuring some warning-only line instead.
-    const chipLine = lines.find((l) => l.includes('Premium: 0 of 4 used'))
-    // Flash's carries the same warning with nothing after it — the shape the
+    // GLM 5.3 Flash's second line carries its per-model ceiling chip. Anchored
+    // on the chip text, so a chip that stops being drawn fails here rather than
+    // quietly re-measuring some warning-only line instead. Deliberately the
+    // PER-MODEL pool rather than the shared one: its label is the longest the
+    // caps table can produce, which is the case the width math has to survive.
+    const chipLine = lines.find((l) => l.includes('GLM 5.3 Flash: 0 of 2 used'))
+    // Flash carries the training warning with nothing after it — the shape the
     // width math already handled, which is the "ordinary warning line" above.
     const warningOnlyLine = lines.find(
       (l) => l.includes('May use data for AI training') && !l.includes('used'),

cli/src/utils/__tests__/settings.test.ts

@@ -9,6 +9,7 @@ import {
   FREEBUFF_GLM_V52_MODEL_ID,
   FREEBUFF_MIMO_V25_MODEL_ID,
   FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+  FREEBUFF_GLM_V53_FLASH_MODEL_ID,
 } from '@codebuff/common/constants/freebuff-models'
 
 import * as auth from '../auth'
@@ -48,33 +49,56 @@ describe('freebuff model preference', () => {
     )
     getConfigDirSpy = spyOn(auth, 'getConfigDir').mockReturnValue(testConfigDir)
 
-    // Two tests lived here, both asserting a stored V4 Pro pick was rewritten
-    // to Flash on every launch. That migration is GONE as of 2026-08-21 along
-    // with the supersedes notice driving it — Pro is now the cheapest premium
-    // row and the catalog's first entry, so steering off it would be backwards.
-    //
     // Written directly, with no migration marker, exactly like a real
     // pre-upgrade settings file — which is the case that would silently rewrite
-    // if a notice came back.
+    // if a supersedes notice came back.
     fs.writeFileSync(
       path.join(testConfigDir, 'settings.json'),
-      JSON.stringify({ freebuffModel: FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID }),
-    )
-    expect(loadFreebuffModelPreference()).toBe(
-      FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+      JSON.stringify({ freebuffModel: FREEBUFF_GLM_V53_FLASH_MODEL_ID }),
     )
+    expect(loadFreebuffModelPreference()).toBe(FREEBUFF_GLM_V53_FLASH_MODEL_ID)
 
     // And a round-trip through save/load leaves every selectable row alone. The
     // property is "the picker is the user's decision, not ours" — asserted
     // across the catalog rather than on one row, because the failure mode is a
     // notice added for ONE model quietly acquiring this behaviour.
     for (const id of [
-      FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+      FREEBUFF_GLM_V53_FLASH_MODEL_ID,
       FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID,
       FREEBUFF_MIMO_V25_MODEL_ID,
     ]) {
       saveFreebuffModelPreference(id)
       expect(loadFreebuffModelPreference()).toBe(id)
     }
   })
+
+  test('a withdrawn pick is DROPPED, not carried or rewritten', () => {
+    testConfigDir = fs.mkdtempSync(
+      path.join(os.tmpdir(), 'freebuff-settings-test-'),
+    )
+    getConfigDirSpy = spyOn(auth, 'getConfigDir').mockReturnValue(testConfigDir)
+
+    // DeepSeek V4 Pro was withdrawn from free mode on 2026-08-26. A saved
+    // preference is the longest-lived way to hold a dead id: it survives every
+    // deploy and outlives the release that dropped the row, so this is the
+    // client half of the withdrawal.
+    //
+    // Written directly, because that is the only way it can arrive — an updated
+    // client cannot SAVE the id, and the file predates the update.
+    fs.writeFileSync(
+      path.join(testConfigDir, 'settings.json'),
+      JSON.stringify({ freebuffModel: FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID }),
+    )
+    // Undefined, not a substitute. The catalog validation drops it and the
+    // landing screen falls to its own default — which is a decision that
+    // belongs there, not to a rewrite here. A migration would silently move the
+    // user onto a specific model on every launch, which is exactly what the
+    // supersedes machinery was removed for.
+    expect(loadFreebuffModelPreference()).toBeUndefined()
+
+    // And the updated client refuses to write it back, so the drop is durable
+    // rather than re-inflicted from the picker.
+    saveFreebuffModelPreference(FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID)
+    expect(loadFreebuffModelPreference()).toBeUndefined()
+  })
 })

common/src/__tests__/free-agents.test.ts

@@ -683,6 +683,10 @@ describe('every freebuff root agent declares a prompt opening', () => {
     'base2-free-minimax-m3': BASE2,
     'base2-free-luna': BASE2,
     'base2-free-glm': BASE2,
+    // GLM 5.3 Flash's own root — a separate agent from 'base2-free-glm' above
+    // because the two models draw on different pools; createBase2('free', …)
+    // like its siblings.
+    'base2-free-glm-5-3-flash': BASE2,
     // God-only Kimi K3 test root; createBase2('free', …) like its siblings.
     'base2-free-kimi-k3-eco': BASE2,
     'base2-free-luna-es': BASE2,