Merge latest main into recreate/1101-freebuff-commit-attribution

Author: c8dhjp4tyv-bit

.gitignore

@@ -9,6 +9,7 @@ user-emails-*.csv
 .env
 .env.*
 !.env.example
+!.env.local.example
 
 # ctags
 tags
@@ -55,6 +56,8 @@ cli/release-staging/launcher.js
 cli/release-staging/http.js
 freebuff/cli/release/launcher.js
 freebuff/cli/release/http.js
+xcuserdata/
+.claude/settings.local.json
 
 # QA harness fixture: a snapshot of REAL feedback rows, including the Discord
 # and GitHub handles of real people and what they complained about. Useful
@@ -73,6 +76,11 @@ scratch-*
 # Ad-hoc investigation scripts — scratch, never committed
 scripts/_tmp-*
 
+# The intent-batch orchestrator's resume state. It holds a live approval
+# capability (shown once, at the mint) and the cohort's own references, so it
+# is operator-local and never committed.
+scratch/
+
 # Local operator launch manifests contain real advertiser identifiers. The
 # checked-in example is intentionally non-routable.
 ads-pilot-launch.local.json

README.md

@@ -34,19 +34,20 @@ Freebuff includes a curated model catalog. The regular picker currently offers:
 
 | Model                       | Access                  | Best for                                                          |
 | --------------------------- | ----------------------- | ----------------------------------------------------------------- |
-| **GLM 5.3 Flash**           | Full access             | The default in full mode; deepest reasoning, unmetered            |
+| **GLM 5.3 Flash**           | Full and limited access | The default everywhere; deepest reasoning, unmetered              |
+| **DeepSeek V4 Flash 07/31** | Full and limited access | Fast coding and tool use, unmetered                               |
 | **GPT-5.6 Luna**            | Full access             | Strong all-around with native images                              |
-| **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 |
-| **Solar Pro 4**             | Full access             | Limited-time trial; 524K context, text only                       |
+| **MiMo 2.5**                | Full and limited access | Balanced performance with image support                           |
+| **Solar Pro 4**             | Full and limited access | Limited-time trial; 524K context, text only; unmetered at full access |
+| **Muse Spark 1.3**          | Full access             | Meta's agentic coding model; 1M context. Rate limited and shared by every user, so it queues when busy and answers on DeepSeek V4 Flash rather than making you wait |
 
-Most models draw on your normal daily sessions rather than a separate limit. GLM 5.3 Flash and MiMo 2.5 are 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, DeepSeek V4 Flash 07/31, MiMo 2.5 and Solar Pro 4 are unmetered at full access 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:
 
-- **GLM 5.2** is available through earned sessions rather than as an always-unlocked model.
+- **Referrals and bounties** earn extra sessions on top of the free daily allowance.
 - **Gemini 3.1 Flash Lite** powers specialist tasks such as file finding and research rather than appearing in the main picker.
 
 Availability and limits depend on your access tier, product, and current capacity. Freebuff Desktop can also run locally installed Claude Code and Codex agents using your existing provider account; those connected models are separate from Freebuff's included catalog.
@@ -63,15 +64,15 @@ Freebuff uses specialized agents instead of sending every task through one model
 
 ## Free access
 
-Freebuff is available in every country. Supported regions receive full access; other regions and VPN users receive limited access, currently MiMo 2.5 with three one-hour sessions per day, earnable up to seven.
+Freebuff is available in every country. Supported regions receive full access; other regions and VPN users receive limited access to DeepSeek V4 Flash 07/31, GLM 5.3 Flash, MiMo 2.5, and Solar Pro 4. Accounts on Freebucks use the displayed model price and balance. On the legacy session system, limited access includes six one-hour sessions per day, earnable up to seven; GLM uses earned reward sessions instead.
 
 Text ads support the included models. Freebuff shows the applicable session limits and any model-specific data-use notice before you start.
 
 <!-- BEGIN GENERATED FREEBUFF DATA USE -->
 
 **Is my data used to train AI?** Only when a model or feature says data may be used for AI training. Freebuff or the provider may then keep submissions to develop, train, test, evaluate, fine-tune, and improve AI models or products.
 
-**How is my data used and stored?** We use prompts, messages, code, files, and repository data to provide the service. We may analyze prompts and messages—including pasted content—to personalize ads, using Freebuff systems and service providers acting on our behalf. Separate uploads and connected repositories are not provided to advertising providers. Where required by law, we provide advertising choices and honor recognized opt-out signals; elsewhere, this processing may be required to use the free service. See the Privacy Policy for retention and details.
+**How is my data used and stored?** We use prompts, messages, agent traces, code, files, and repository data to provide Freebuff. We may analyze prompts and messages to personalize ads. We do not give separately uploaded files or connected repositories to advertising providers. Restricted partners may evaluate connected Cloud repositories or code used with models labeled “May use data for AI training,” but cannot otherwise use, broadly share, or train on it. See the Privacy Policy for retention, eligibility, and advertising choices.
 
 See the [Privacy Policy](https://freebuff.com/privacy-policy) for complete details.
 

README.zh-CN.md

@@ -34,19 +34,20 @@ Freebuff 提供经过筛选的模型目录。常规模型选择器目前包括
 
 | 模型                        | 访问范围       | 适用场景                                         |
 | --------------------------- | -------------- | ------------------------------------------------ |
-| **GLM 5.3 Flash**           | 完整访问       | 完整模式下的默认模型；推理最深入，且不消耗会话   |
+| **GLM 5.3 Flash**           | 完整和受限访问 | 所有模式下的默认模型；推理最深入，不消耗会话         |
+| **DeepSeek V4 Flash 07/31** | 完整和受限访问 | 快速编程和工具调用，不消耗会话                   |
 | **GPT-5.6 Luna**            | 完整访问       | 综合能力强，原生支持图像                         |
-| **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           |
-| **Solar Pro 4**             | 完整访问       | 限时试用；52.4 万上下文，仅支持文本             |
+| **MiMo 2.5**                | 完整和受限访问 | 均衡性能并支持图像                               |
+| **Solar Pro 4**             | 完整和受限访问 | 限时试用；52.4 万上下文，仅支持文本；完整访问下不消耗会话 |
+| **Muse Spark 1.3**          | 完整访问       | Meta 的智能体编码模型；100 万上下文。有速率限制且所有用户共享，繁忙时排队，随后改由 DeepSeek V4 Flash 作答，不会让你一直等 |
 
-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 和 MiMo 2.5 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、DeepSeek V4 Flash 07/31、MiMo 2.5 和 Solar Pro 4 在完整访问下保持无限使用，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：
 
-- **GLM 5.2** 通过获得的会话使用，并非始终解锁。
+- **推荐好友与悬赏任务**可在每日免费额度之外赚取额外会话。
 - **Gemini 3.1 Flash Lite** 用于查找文件和研究等专业任务，不会出现在主模型选择器中。
 
 可用模型和限制取决于你的访问级别、所用产品和当前容量。Freebuff Desktop 还可以通过你现有的提供商账户运行本地安装的 Claude Code 和 Codex 智能体；这些连接的模型不属于 Freebuff 的内置模型目录。
@@ -63,7 +64,7 @@ Freebuff 使用专业化智能体，而不是把所有任务都交给同一个
 
 ## 免费访问
 
-Freebuff 在所有国家和地区均可使用。受支持的地区提供完整访问；其他地区以及使用 VPN 的用户获得受限访问，目前包括 MiMo 2.5，每天可使用六个一小时会话。
+Freebuff 在所有国家和地区均可使用。受支持的地区提供完整访问；其他地区以及使用 VPN 的用户获得受限访问，包括 DeepSeek V4 Flash 07/31、GLM 5.3 Flash、MiMo 2.5 和 Solar Pro 4。使用 Freebucks 的账户按显示的模型价格和余额计费。在旧版会话系统中，受限访问每天提供六个一小时会话，最多可赚取七个；GLM 则使用赚取的奖励会话。
 
 内置模型由文字广告支持。开始前，Freebuff 会显示适用的会话限制以及模型特定的数据使用提示。
 

agents/__tests__/base-chat.test.ts

@@ -1,3 +1,7 @@
+import {
+  compactionPolicyForModel,
+  DEEPSEEK_FLASH_COMPACTION_POLICY,
+} from '@codebuff/common/constants/compaction-policy'
 import {
   FREEBUFF_DEFAULT_CONTEXT_WINDOW,
   FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID,
@@ -52,6 +56,12 @@ function firstPrunerSpawn(model?: string) {
     agentState: createMockAgentState(100),
     logger: mockLogger,
     model,
+    // What the runtime supplies (run-programmatic-step.ts): the model's pruner
+    // thresholds. base-chat sizes its own budget and reads only the policy.
+    contextPruning: {
+      maxContextLength: 400_000,
+      ...compactionPolicyForModel(model),
+    },
   })
   return generator.next().value as {
     toolName: string
@@ -301,4 +311,29 @@ describe('base-chat pruning triggers', () => {
       60 * 60 * 1000,
     )
   })
+
+  test('non-Flash models keep the 24-hour gap and no token floor', () => {
+    for (const model of [
+      'minimax/minimax-m3',
+      'openai/gpt-5.6-luna',
+      undefined,
+    ]) {
+      const params = firstPrunerSpawn(model).input.params as any
+      expect(params.cacheExpiryMs).toBe(24 * 60 * 60 * 1000)
+      expect(params).not.toHaveProperty('cacheExpiryMinTokens')
+    }
+  })
+
+  test('DeepSeek Flash follows the shared compaction policy: 15 minutes, 40k floor', () => {
+    // Flash's Luminal lane forgets the prefix in ~15 minutes, so the 24-hour
+    // gap only postponed a compaction the cold prefill was going to pay for.
+    const params = firstPrunerSpawn(FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID).input
+      .params as any
+    expect(params).toEqual({
+      maxContextLength: params.maxContextLength,
+      ...DEEPSEEK_FLASH_COMPACTION_POLICY,
+    })
+    expect(params.cacheExpiryMs).toBe(15 * 60 * 1000)
+    expect(params.cacheExpiryMinTokens).toBe(40_000)
+  })
 })

agents/__tests__/base2.test.ts

@@ -7,6 +7,12 @@ import {
   FREEBUFF_MINIMAX_M3_MODEL_ID,
   FREEBUFF_MIMO_V25_MODEL_ID,
 } from '@codebuff/common/constants/freebuff-models'
+import {
+  compactionPolicyForModel,
+  DEEPSEEK_FLASH_COMPACTION_POLICY,
+  DEFAULT_COMPACTION_POLICY,
+} from '@codebuff/common/constants/compaction-policy'
+import { contextPrunerBudgetForModel } from '@codebuff/common/constants/model-config'
 
 import { createBase2 } from '../base2/base2'
 import { createBaseDeep } from '../base2/base-deep'
@@ -148,10 +154,82 @@ describe('production agent step prompts', () => {
     }
   })
 
-  test('plan-only keeps its no-edit constraint in the instructions', () => {
-    const agent = createBase2('default', { planOnly: true })
+  // A user picked plan mode and got the whole feature built and committed. The
+  // instruction paragraph saying not to was there the entire time; what it was
+  // losing to was the rest of the same prompt (a worked example ending in "you
+  // implement the changes using the editor agent") and a toolset that still
+  // held every capability a build turn has. So these assert the capability and
+  // the demonstration, not the paragraph.
+  describe('plan-only', () => {
+    const plan = createBase2('default', { planOnly: true })
+
+    test('holds no tool that writes a file', () => {
+      for (const tool of [
+        'str_replace',
+        'write_file',
+        'propose_str_replace',
+        'propose_write_file',
+      ]) {
+        expect(plan.toolNames).not.toContain(tool)
+      }
+    })
+
+    test('cannot spawn an agent that edits or runs commands', () => {
+      // basher is the shell: `git commit` reached the repository through it
+      for (const agent of [
+        'editor',
+        'editor-multi-prompt',
+        'basher',
+        'tmux-cli',
+      ]) {
+        expect(plan.spawnableAgents).not.toContain(agent)
+      }
+    })
+
+    test('still holds everything it needs to explore', () => {
+      for (const tool of [
+        'read_files',
+        'read_subtree',
+        'list_directory',
+        'glob',
+      ] as const) {
+        expect(plan.toolNames).toContain(tool)
+      }
+      for (const agent of ['file-picker', 'code-searcher', 'researcher-web']) {
+        expect(plan.spawnableAgents).toContain(agent)
+      }
+    })
+
+    test('never demonstrates building in its worked example', () => {
+      const systemPrompt = plan.systemPrompt!
+
+      for (const phrase of [
+        'You implement the changes',
+        'Spawn the editor agent to implement',
+        'Spawn bashers sequentially',
+        'All tests & typechecks pass',
+      ]) {
+        expect(systemPrompt).not.toContain(phrase)
+      }
+      expect(systemPrompt).toContain('You do NOT implement anything')
+    })
+
+    test("refuses to read the user's wording as permission to build", () => {
+      const instructions = plan.instructionsPrompt!
+
+      expect(instructions).toContain('is not permission to leave plan mode')
+      // the old prompt ended with "answer the user's questions or requests
+      // freely", which reads as exactly that permission
+      expect(instructions).not.toContain('requests freely')
+    })
 
-    expect(agent.instructionsPrompt).toContain('Do not make file changes')
+    test('leaves the build modes alone', () => {
+      const build = createBase2('default')
+
+      expect(build.toolNames).toContain('write_file')
+      expect(build.spawnableAgents).toContain('editor')
+      expect(build.spawnableAgents).toContain('basher')
+    })
   })
 })
 
@@ -329,13 +407,26 @@ describe('base2 optional tools', () => {
 })
 
 describe('base2 context pruning', () => {
+  /** What the runtime hands a serialized generator for `model`
+   *  (run-programmatic-step.ts): the model plus its pruner thresholds. */
+  const stepContextFor = (model: string) => ({
+    model,
+    contextPruning: {
+      maxContextLength: contextPrunerBudgetForModel(model as any),
+      ...compactionPolicyForModel(model),
+    },
+  })
+
   const getContextPrunerParams = (
     mode: Parameters<typeof createBase2>[0],
     options?: Parameters<typeof createBase2>[1],
     params?: Record<string, unknown>,
   ) => {
     const base2 = createBase2(mode, options)
-    const generator = base2.handleSteps!({ params } as any)
+    const generator = base2.handleSteps!({
+      params,
+      ...stepContextFor(base2.model),
+    } as any)
     const step = generator.next().value as any
     return step.input.params
   }
@@ -350,7 +441,10 @@ describe('base2 context pruning', () => {
     const isolatedHandleSteps = new Function(
       `return (${handleStepsString})`,
     )() as NonNullable<typeof base2.handleSteps>
-    const generator = isolatedHandleSteps({ params: undefined } as any)
+    const generator = isolatedHandleSteps({
+      params: undefined,
+      ...stepContextFor(base2.model),
+    } as any)
     const step = generator.next().value as any
     return step.input.params
   }
@@ -384,14 +478,39 @@ describe('base2 context pruning', () => {
   test('free non-MiniMax/Kimi models default context pruning to 400k tokens', () => {
     expect(
       getContextPrunerParams('free', {
-        model: FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID,
+        model: FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
       }),
     ).toEqual({
       maxContextLength: 400_000,
       cacheExpiryMs: 30 * 60 * 1000,
     })
   })
 
+  test('DeepSeek Flash prunes under the Flash policy; other models keep 30 minutes, no floor', () => {
+    // Flash's numbers arrive through the runtime's `contextPruning`; the model
+    // id is the one literal in the generator, so pin it to the constant.
+    expect(FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID).toBe(
+      'deepseek/deepseek-v4-flash',
+    )
+    const flashParams = {
+      maxContextLength: 400_000,
+      ...DEEPSEEK_FLASH_COMPACTION_POLICY,
+    }
+    const flash = { model: FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID }
+    expect(getContextPrunerParams('free', flash)).toEqual(flashParams)
+    expect(getSerializedContextPrunerParams('default', flash)).toEqual(
+      flashParams,
+    )
+    // base2 is the kill-switch fallback, so it deliberately does NOT take the
+    // default policy's hour/140k for the rest.
+    expect(DEFAULT_COMPACTION_POLICY.cacheExpiryMs).toBe(60 * 60 * 1000)
+    expect(
+      getContextPrunerParams('free', {
+        model: FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
+      }),
+    ).toEqual({ maxContextLength: 400_000, cacheExpiryMs: 30 * 60 * 1000 })
+  })
+
   test('free mode preserves explicit context pruning params', () => {
     const base2 = createBase2('free')
     const generator = base2.handleSteps!({