From 7b12f4dfb25e3db419bfa3f22cae2156b7cea507 Mon Sep 17 00:00:00 2001 From: asteroida123 <264808420+asteroida123@users.noreply.github.com> Date: Sun, 16 Aug 2026 22:32:58 +0800 Subject: [PATCH 01/75] docs: update agent integration cost model in strategy memo --- docs/STRATEGY-MEMO.md | 369 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 369 insertions(+) create mode 100644 docs/STRATEGY-MEMO.md diff --git a/docs/STRATEGY-MEMO.md b/docs/STRATEGY-MEMO.md new file mode 100644 index 000000000..753cd2768 --- /dev/null +++ b/docs/STRATEGY-MEMO.md @@ -0,0 +1,369 @@ +# Codeg 战略备忘录 + +> 内部决策记录。整理于一次深度代码评审 + 商业化讨论之后。 +> 目的:把散落在对话里的分析、判断、案例、待办沉淀下来,避免重复讨论。 + +--- + +## 一、项目定位(已确认,不再动摇) + +**Codeg = 多智能体编码工作台(Multi-Agent Coding Workspace)。** + +- 服务对象:**开发者**(不是非开发者,不是垂直行业) +- 核心能力:**跨进程异构 agent 委托**(不是单进程换模型) +- 不做:**Cursor 平替、通用 agent 平台、垂直行业方案** + +### 定位边界(加宽 vs 跳船) + +| 允许 | 禁止 | +|------|------| +| 覆盖开发者的更多环节(编码/文档/审查) | 脱离开发者去做通用市场 | +| 引入更多编码 agent(Pi/GenericAgent) | 引入非编码 agent 后转去做非开发场景 | +| 加宽定位(开发者全流程) | 跳船(服务完全不同的人群) | + +--- + +## 二、技术资产评审(代码级事实) + +### 已完成且扎实的部分 + +| 模块 | 规模 | 评价 | +|------|------|------| +| `acp/delegation/` | ~1.4 万行 Rust,200+ 测试 | **核心护城河**,工程深度极高 | +| `acp/` 顶层 | ~2.5 万行(connection/manager/lifecycle 等) | ACP 协议完整实现 | +| `parsers/` | ~1.8 万行,10 个 agent | 会话聚合,覆盖广 | +| 双模式架构 | codeg / codeg-server / codeg-mcp | 部署形态齐全,`_core` 函数共用 | + +### delegation 模块的技术亮点(护城河来源) + +1. **跨进程真委托** — `delegate_to_agent` 起独立 PID,不是进程内换 prompt +2. **异步 fan-out** — 同时开多个子 agent,`get_delegation_status` 批量长轮询 +3. **取消级联** — 四种取消路径全覆盖(外部/子/父连接/父轮次) +4. **Setup 窗口竞态处理** — inflight 注册 + checkpoint + atomic park,零信任时序 +5. **深度限制** — `depth.rs` 防递归爆炸,带 cap 防脏数据 +6. **安全边界** — 一次性 token + parent scoping + UDS/piped 传输 +7. **测试覆盖** — broker 109 测试、companion 53、listener 25,覆盖全边界条件 +8. **trait 解耦** — `ConnectionSpawner` 为 v3 远程 agent 预留扩展点 + +### 当前短板(v1 限制) + +1. **v1 是 one-shot** — `broker.rs:2527` `disconnect` 写死,子 agent 跑完即杀 +2. **broker.rs 7554 行** — 单文件过大,接近可维护性临界点 +3. **没有远程 agent** — 只能调度本机,v3 未实现 +4. **结果不落库** — 子 agent 输出只在内存缓存(512MB FIFO 淘汰) + +--- + +## 三、产品演进路线(v1 → v2 → v3) + +### v1(现状):一次性委托 + +- 子 agent 跑完第一轮 `TurnComplete` → broker `disconnect` → 杀掉 +- 父 LLM 拿到结果文本,委托结束 +- 暴露 6 个工具:`delegate_to_agent` / `get_delegation_status` / `cancel_delegation` / `check_user_feedback` / `ask_user_question` / `get_session_info` + +### v2:多轮子会话(Continue Session)— 短期 1-2 月 + +**目标**:子会话长期存活,支持多轮交互 + +**工具变化**:新增 `continue_with_session` / `close_session` + +**为什么不难**: +- ACP 协议本身支持多轮(Codeg 已实现 `session/load` / `session/resume`,主会话在用) +- 只需删除 `finalize_delegation` 末尾的 `disconnect`,改成等显式 close +- 取消级联/深度限制/inflight 注册早已为长期会话设计,现在大材小用 + +**真正难点**:状态管理(TTL、并发上限、父结束时的级联关闭、中途失败语义) + +**价值**:从"工具调用"升级为"Agent 团队",产品形态质变 + +### v3:远程 Agent(Remote Spawner)— 中期 3-6 月 + +**目标**:agent 跑在远程,broker 不变 + +**架构**:新增 `RemoteSpawner` 实现 `ConnectionSpawner` trait,broker 零改动 + +**难点不在 Codeg**(写 RemoteSpawner 不难),在基础设施: +- 远程执行环境(sandbox、API key 管理、资源隔离) +- 网络可靠性(远程超时处理) +- 安全(权限模型、审计) +- 认证(多租户隔离) +- 计费 + +**价值**:从"本地工具"升级为"云平台",商业化的钥匙 + +### v2 必须先于 v3 的理由 + +1. 远程 agent 冷启动成本高,只换一个 turn 的结果经济不成立 → 需要 v2 多轮 +2. v2 纯本地,风险低,能快速验证产品形态 + +--- + +## 四、商业化分析 + +### 市场定位的真实判断 + +**Codeg 的差异化不是"多智能体",而是"多智能体的跨进程编排"。** + +- OpenCode / Pi / DriFox 做的是"单进程多模型/多 prompt"(第一层) +- Codeg 做的是"跨进程真委托"(第二层) +- **不在一个维度上竞争** + +### 四条变现路径(垂直行业已划掉) + +| 路径 | 可行性 | 说明 | +|------|--------|------| +| 开发者工具变现(企业版/团队) | 🟢 最现实 | 开发者为效率付费,Codeg 主场 | +| 托管云 | 🟡 有需求 | 但要扛运维成本 | +| API 计费层 | 🟡 灰色 | 中国市场对接海外模型有法律风险 | +| ~~垂直行业~~ | ❌ 划掉 | 底座是编码 agent,做不了行业方案 | + +### 为什么垂直行业划掉 + +- 垂直行业(律所/金融/医疗)不会以 Claude Code/Codex 为底座 +- 这些 agent 是通用编码 agent,不具备行业专业性 +- 行业数据不能出境(合规问题) +- 没有行业知识库 + +### 最诚实的商业化判断 + +**最大风险:在"多 agent 协作"需求真正普及前,上游平台方可能下场。** +- 窗口期 12-18 个月 +- 最该做的:抢心智占有率("多智能体协作 = Codeg") + +**最该先验证的问题(未回答)**: +1. 现在有没有人重度依赖 Codeg 的 delegation? +2. 你和同事自己日常用 delegation 吗? + +> 这两个问题不回答,所有路径都是空中楼阁。先找 10 个真实用户问三个问题: +> 你用它干过什么?没有它你会怎么做?愿意付多少钱? + +--- + +## 五、案例研究 + +### CoolVibe(coolvibe.io)— 参照系 + +| 维度 | 数据 | +|------|------| +| 本质 | Agent 的网页查看器(壳子) | +| 投入 | 4 个月,~3 万美元 | +| 定价 | 免费 / Pro $29.9/年 | +| 现状 | "订阅免费送"(还在买用户阶段) | +| V2EX 热度 | 1054 条回复 | +| 公开营收 | 查不到 | +| 团队 | xterminal 团队(有成功经验) | + +**启示**: +1. "壳子"能赚钱,但前提是痛点明确且高频 +2. 定价对齐"省下的时间",不是"技术多复杂" +3. 免费送订阅是冷启动的合理策略 +4. CoolVibe 比深得多的 Codeg 先验证了"给 agent 做壳"能跑通 + +### DriFox(github.com/martin98-afk/DriFox)— 反面教材 + 经验库 + +| 维度 | 数据 | +|------|------| +| 本质 | PyQt 桌面对话助手(Cursor 平替) | +| 规模 | 15 万行 Python | +| Stars | **24**(2 个月) | +| 定位错误 | 堆 33 插件做"Cursor 平替",死路 | + +**DriFox 走的错路(Codeg 绝不能走)**: +- 自己做 agent 和 Cursor 正面打(蚂蚁 vs 太阳) +- 堆功能数量拼不过飞轮(数据/生态/资本) +- "平替"定位是诅咒(永远活在正品阴影里) + +**Cursor 的体量(参照)**: +- ARR $40 亿(2026.06 年化) +- 估值 $290-500 亿 +- 日活 100 万+ +- 融资 $23 亿 D 轮 + +### ACP 协议现状 + +- 官方 35+ 兼容 agent +- 已被 Zed / JetBrains / Google / GitHub 采纳 +- 事实标准(类似 LSP 之于语言服务器) +- **zcode 不在列表**(截至 2026-07) + +--- + +## 六、可执行的功能规划 + +### 🎯 功能 A:编程 PK(周末项目,优先级最高) + +**是什么**:同一个任务同时分发给多个 agent,看谁做得好 + +**为什么做**: +1. 成本极低(底层全有,只缺对比 UI)——周末能做完 +2. 硬核技术的最佳展示窗口(把不可见的委托变成可视化) +3. 赛道真空(DriFox 24 star 不构成威胁,异构 PK 无人做) +4. 自带传播(天然话题性 + 可视化 + 争议性) +5. 低成本的市场探测器(PK 没人理 → 大方向要重新考虑) + +**为什么 Codeg 能做、别人不能**: +- Cursor:单进程,没法真异构并行 +- DriFox:单进程,SubAgent 是换 prompt +- OpenCode/Pi:同上 +- **只有 Codeg 有跨进程 ACP 委托** + +**技术可行性(已确认)**: +- `delegate_to_agent` 已支持 fan-out ✅ +- `get_delegation_status` 已支持批量等 ✅ +- `DelegationSuccess` 已带 `duration_ms`/`token_usage`/`turn_count` ✅ +- `DelegationContext` 已流到前端 ✅ +- **只缺一层横向对比 UI** + +**实施档位**: + +| 档位 | 内容 | 工作量 | +|------|------|--------| +| A 最小可玩 | PK 触发器 + 分屏视图 + 计分板 | 周六一天 | +| B 好看+分享 | diff 对比 + 实时终端 + 分享截图 + 任务模板 | +周日上午 | +| C 评判 | 自动测试裁判 + LLM 裁判 | 周日下午(选做) | + +**关键代码位置**: +- 触发器:`src/components/composer/`(加 PK 模式开关) +- 对比组件:新建 `src/components/chat/agent-pk-arena.tsx` +- 数据复用:`src/hooks/use-delegation-card-model.ts`(已解出 agent/task/status/childId) +- 渲染复用:`src/components/message/sub-agent-session-dialog.tsx`(已会渲染单个子 agent) +- 分享截图:`html-to-image`(package.json 已有依赖) + +**传播策略**: +- V2EX 发帖,标题方向:"周末做了个 AI 编程 PK 场:让 Claude Code、Codex、Gemini 同时写贪吃蛇,看谁快" +- 每场 PK 都是一次传播机会 + +### 🎯 功能 B:角色化 Agent 团队(PK 之后的第二张牌) + +**是什么**:预设 leader/build/review 等角色,一键发起"Claude 当 leader 分任务给 Codex(build)和 Gemini(review)" + +**为什么做**: +- 比 PK 更接近"真实有用的工作"(协作 > 竞技) +- 和 PK 共用同一套 delegation 底层,只换 prompt 和 UI +- 让 Codeg 从"对比工具"延伸为"团队编排平台" + +**来源**:DriFox 的 `plugins/system/agents/*.md`(角色 prompt 设计精良,可借鉴) + +**与 v2 的关系**:角色化团队天然需要多轮子会话(v2),但可以先做 one-shot 版本 + +--- + +## 七、从 DriFox 可借鉴的具体设计 + +### 值得偷(按价值排序) + +#### 1. 角色化 agent prompt 模板 +- 来源:`plugins/system/agents/{build,leader,review,explore,plan}.md` +- 用途:Codeg 的角色化团队功能 +- 关键原则(直接抄): + - "不做超出需求的功能。一次性代码不做抽象。" + - "不要顺手改进相邻代码。不要重构没坏的东西。" + - "每一行变更都应该能直接追溯到用户的请求。" + - "使用 question 工具提问,优先提供选项" + +#### 2. 工具安全分类(三分法) +- 来源:`app/tools/command_safety.py` + `app/tools/tool_classifier.py` +- 用途:Codeg 加一层自己的安全护栏(企业版刚需) +- 规则: + - 危险(write/edit/bash/mouse)→ 审批 + - 安全(read/grep/glob/webfetch)→ 放行 + - 命令三分法:无元字符直接跑 / 有管道重定向需确认 / 黑名单拒绝 + +#### 3. "关键文档"记忆机制 +- 来源:`app/core/memory_manager.py` 的 `KeyDocumentsRepository` +- 用途:解决子 agent 冷启动不懂项目上下文的痛点 +- 做法:用户标记项目重要文件(README/架构文档/API 约定),委托时自动注入 task prompt + +#### 4. 任务邮件分发(轻量跨窗口协作) +- 来源:`app/core/team_manager.py`(文件邮箱) +- 用途:v3 远程 agent 的备选轻量方案 +- 优先级:低(记住有这个方案) + +#### 5. AGENTS.md 作为项目笔记 +- DriFox 印证了这是事实标准,Codeg 已在用,继续走 + +### 明确不要偷 + +| 不偷 | 理由 | +|------|------| +| PyQt 桌面浮动窗 | 架构不同,非核心价值 | +| 33 个插件 | 功能堆砌是它 24 star 的原因 | +| 进程内 SubAgent DAG | Codeg 跨进程委托更强 | +| ECharts 力导向图 | PK 用并排对比更直观 | +| LSP 集成(11 种语言) | Cursor 的战场,别送 | + +--- + +## 八、关于接入 zcode 的结论 + +**结论:暂不接入。** + +理由: +- zcode 不开源,改不了,只能向 Z.ai 提需求 +- zcode 不支持 ACP(不在官方 35+ 兼容列表) +- 不支持 ACP → 只能做 L1(会话导入 parser),不能做 L2(完整调度) +- L1 单独价值低(装饰性,用户在 zcode 自己的 app 里已能看会话) +- zcode 加 ACP 是 Z.ai 的战略决策,Codeg 这边无法推动 + +**zcode 数据格式(已摸清,备用)**: +- `~/.zcode/cli/db/db.sqlite` — 主库(session/message/part/todo 表) +- `~/.zcode/v2/tasks-index.sqlite` — 任务索引 +- `~/.zcode/cli/rollout/*.jsonl` — 原始模型 IO 流 +- `~/.zcode/cli/agents/sess_*/agent_*/` — 子 agent 会话 + +如果未来 zcode 支持 ACP,接入成本分两档(2026-08-16 修正,依据:上游 DSH 集成案例): + +**L1(会话导入,半天,4 处)**: +1. `models/agent.rs` — `AgentType` 枚举 +2. `acp/registry.rs` — `AcpAgentMeta` +3. `parsers/mod.rs` — 注册 +4. `parsers/zcode.rs` — 会话导入(可选) + +**L2(完整内置 agent,六层 ~25 处,量级 3-5 天)**: +完整参考:上游提交 `3845e9df`「feat(deepseek): integrate DeepSeek Harness as a built-in agent」,+3271/−50 行,改动约 25 个文件。六层清单: + +| 层 | 改动点 | 示例(DSH 提交) | +|----|--------|----------------| +| 1. 身份层 | `AgentType` 枚举 + wire 名 + 防自定义 shadow | `models/agent.rs`(+12) | +| 2. 启动层 | 注册表元数据:npm 包/版本/node 要求/认证方式 | `acp/registry.rs`(+44) | +| 3. 历史层 | 原生会话解析器(会话进列表/统计/导入,工作量最大) | `parsers/deepseek.rs`(+1291) | +| 4. 运行时层 | 沙箱根目录 / env 键 / skills 存储 / 只读路径 | `file_system_runtime.rs`、`commands/acp.rs`(+237) | +| 5. 协议层 | ACP capabilities / MCP 接线 / 传输限制 | `connection.rs`、`commands/mcp.rs`(+398) | +| 6. 防御层 | 自定义注册表冲突校验 + 文档 | `custom_registry.rs`(+52) | + +> **教训**:曾以为"加一个 agent = 改 4 处",实际"内置 agent"是六层全通——只做 1/2 层的 PR 会被上游以"不够一等公民"拒绝(2026-08-16 DSH PR 实测)。以后评估任何"接新 agent"的工作量,按此清单逐层核对。 + +--- + +## 九、关键待办与未决问题 + +### 必须先回答的问题(阻塞所有商业化决策) + +- [ ] **现在有没有人重度依赖 Codeg 的 delegation?**(找 10 个真实用户) +- [ ] **你和同事自己日常用 delegation 吗?**(如果造的人都不用,别指望别人用) + +### 功能待办(按优先级) + +- [ ] **编程 PK**(周末项目) — 最低成本验证 + 传播 +- [ ] **角色化 Agent 团队**(PK 之后) — 第二个场景 +- [ ] **v2 多轮子会话** — 产品形态质变,1-2 月 +- [ ] **v3 远程 agent** — 商业化钥匙,3-6 月 + +### 不做清单(防止分心) + +- [x] ~~接入 zcode~~(暂缓,等 ACP) +- [x] ~~垂直行业方案~~(底座不匹配) +- [x] ~~做 Cursor 平替~~(蚂蚁 vs 太阳) +- [x] ~~堆功能数量~~(学 DriFox 的教训) +- [x] ~~脱离开发者市场~~(丢弃所有资产) + +--- + +## 十、一句话总结 + +> **Codeg 的硬核技术(跨进程异构委托)是真的,但"硬核"本身不产生收入,产生收入的是"有一群人离不开它"。** +> +> **当下的优先级:先用最低成本(PK)验证"多 agent 并行"对开发者有没有吸引力 → 如果有,做 v2/v3 深化 → 如果没有,重新评估方向。** +> +> **别再纠结"能不能打过 Cursor"——那不是 Codeg 该问的问题。该问的是:Cursor 用户有什么事想做但 Cursor 架构上做不到?** From 51d8a6ec4d0a1dcf6042beb4d16a82549bf5f8fc Mon Sep 17 00:00:00 2001 From: asteroida123 <264808420+asteroida123@users.noreply.github.com> Date: Mon, 17 Aug 2026 09:22:18 +0800 Subject: [PATCH 02/75] docs: record the perf audit and qoder integration status in the strategy memo --- docs/STRATEGY-MEMO.md | 35 +++++++++++++++++++++++++++++++---- 1 file changed, 31 insertions(+), 4 deletions(-) diff --git a/docs/STRATEGY-MEMO.md b/docs/STRATEGY-MEMO.md index 753cd2768..c55948acb 100644 --- a/docs/STRATEGY-MEMO.md +++ b/docs/STRATEGY-MEMO.md @@ -2,6 +2,7 @@ > 内部决策记录。整理于一次深度代码评审 + 商业化讨论之后。 > 目的:把散落在对话里的分析、判断、案例、待办沉淀下来,避免重复讨论。 +> 2026-08-16 第二次更新:性能深度扫描 + 上游 issue 需求验证(见二、三、六、九节标注)。 --- @@ -45,12 +46,21 @@ 7. **测试覆盖** — broker 109 测试、companion 53、listener 25,覆盖全边界条件 8. **trait 解耦** — `ConnectionSpawner` 为 v3 远程 agent 预留扩展点 -### 当前短板(v1 限制) +### 当前短板(v1 限制,2026-08-16 复测) -1. **v1 是 one-shot** — `broker.rs:2527` `disconnect` 写死,子 agent 跑完即杀 -2. **broker.rs 7554 行** — 单文件过大,接近可维护性临界点 +1. **v1 是 one-shot** — `broker.rs:2527` `disconnect` 写死,子 agent 跑完即杀。`continue_with_session`/`close_session` 仍不存在,`acp/delegation/mod.rs:29` 注释还挂着 v2 计划 +2. **巨石文件三座山** — `commands/acp.rs` 17020 行、`acp/connection.rs` 16155 行、`broker.rs` 8481 行(此前记 7554,又长了 ~930;前两个比 broker 更大,之前漏记)。broker 约 4750 行是测试,外移到 `broker_tests.rs` 是零风险减半 3. **没有远程 agent** — 只能调度本机,v3 未实现 -4. **结果不落库** — 子 agent 输出只在内存缓存(512MB FIFO 淘汰) +4. **结果不落库(比原记轻)** — 完整文本只在 512MB 内存 FIFO 缓存(`broker.rs:79`);但有界 `text_preview` 已持久化进父会话 tool-call meta(`meta_writer.rs:281`),淘汰后 UI 仍有预览。真缺口是"全文不落库",不是"什么都不落" +5. **前端两个上帝对象** — `acp-agent-settings.tsx` 11818 行(单个设置页)、`acp-connections-context.tsx` 5606 行(~99 个 switch/case 的事件分发层) + +### 性能优化机会(2026-08-16 深度扫描,按收益排序) + +1. **侧边栏全量遍历** — `get_sidebar_data`/`list_folders`/`get_stats` 每次都走 `list_conversations_sync`(`commands/conversations.rs:342`),13 个 parser 全目录遍历,但文件夹树只需要 `folder_path`。前端已自己绕开(`skills-settings.tsx:713` 注释承认慢)。改 DB 文件夹索引 = 用户可感知的最大提升 +2. **活跃会话整文件重解析** — 摘要缓存键 `(mtime,size)`(`summary_cache.rs:64`),流式期间每次列表刷新都对整个 JSONL 逐行反序列化(`claude.rs:846`)。`transcript_watermark` 已跟踪字节位置但从未用于增量读(`claude.rs:1866` 注释自认;codex 直接返回 `None`)。基础设施都在,只差接线 +3. **patch 行号解析无缓存** — `resolve_patch_line_numbers` 每个 patch 块全量读目标文件(`parsers/mod.rs:1002`),7 个 parser 共用;同一大文件 N 个 patch = N 次全量读。一次 HashMap memo 即可 +4. **会话文件定位 O(全树)** — Codex 按 id 每次 WalkDir(`codex.rs:336`),Claude 全目录 read_dir(`claude.rs:1060`);id→path 索引变 O(1) +5. **干净的部分(不用动)** — DB 索引齐全、无 N+1、WAL 配置正确;前端虚拟化(virtua)+ RAF 批处理是真做了的;启动无阻塞扫描,感知慢在首次侧边栏(冷缓存 + 上述第 1 条) --- @@ -92,6 +102,8 @@ **价值**:从"本地工具"升级为"云平台",商业化的钥匙 +**需求验证(2026-08-16)**:上游 issue #461「以持久 Session 为成员的 Team / Chatroom」——社区独立提出了与功能 B/v2 同构的构想。这是 v2 值得做的最直接外部证据。 + ### v2 必须先于 v3 的理由 1. 远程 agent 冷启动成本高,只换一个 turn 的结果经济不成立 → 需要 v2 多轮 @@ -234,6 +246,8 @@ - V2EX 发帖,标题方向:"周末做了个 AI 编程 PK 场:让 Claude Code、Codex、Gemini 同时写贪吃蛇,看谁快" - 每场 PK 都是一次传播机会 +**需求验证(2026-08-16)**:上游 issue #428「缺乏派发 subagent 状态的监视」——用户独立要 delegation 可视化监控,与 PK 共用数据层(`use-delegation-card-model.ts`),做完 PK 顺手就有,算免费赠品。 + ### 🎯 功能 B:角色化 Agent 团队(PK 之后的第二张牌) **是什么**:预设 leader/build/review 等角色,一键发起"Claude 当 leader 分任务给 Codex(build)和 Gemini(review)" @@ -350,6 +364,19 @@ - [ ] **v2 多轮子会话** — 产品形态质变,1-2 月 - [ ] **v3 远程 agent** — 商业化钥匙,3-6 月 +### 工程待办(2026-08-16 新增,按收益排序) + +- [ ] **侧边栏 DB 文件夹索引** — 干掉 13-parser 全量遍历(性能机会 1) +- [ ] **增量会话解析** — 接通已有但闲置的 `transcript_watermark`(性能机会 2) +- [ ] **broker.rs 测试外移** — ~4750 行拆到 `broker_tests.rs`,零风险减半 +- [ ] **patch 行号解析加 memo** — 一次 HashMap 的事(性能机会 3) + +### 上游影响力机会(刷存在感,非主线) + +- [ ] **#391/#387 流式重复文本** — 维护者自己未定位根因(在等用户补 seq 数据);修了就是硬通货 +- [ ] **#396 preferred-config 过滤** — 0xlinn 已给出精确根因:通用路径不按 advertised options 过滤,867 次/23天的重复报错;修法是照抄 grok 路径已有守卫(`connection.rs:1947`)+ `mode` 白名单 +- [x] **#408 接入 qoder cli**(2026-08-17 已完成,待提 PR)— 六层全通:身份/注册表(`qoder-cli`,npx `@qoder-ai/qodercli@1.1.23`,`--acp`)/解析器(Claude 信封格式,`~/.qoder/projects//.jsonl`,state.json 标题加密故走 transcript 明文)/沙箱(`QODER_CONFIG_DIR` RootSlot,实测 `~/.qoder` 可写)/协议(原生 ACP,loadSession+list/resume/fork 全有,MCP 走 settings.json 合并写入+转发跳过)/防御(`qoder-cli` 冲突校验)。规避了既往坑:#396 serde 重命名(测试钉死)、过期 env 误报(important keys 置空)、#468 标题问题(明文派生)。活体验证:server 模式 spawn→握手→prompt→turn_complete→transcript 解析全通。qoder 会话历史与 DSH 等宽(比 DSH 便宜在原生 ACP,贵在 MCP 转发语义相反:qoder 读自己的 settings.json,进跳过名单) + ### 不做清单(防止分心) - [x] ~~接入 zcode~~(暂缓,等 ACP) From 8f9dba0781e64ea3f3a4e18febdc8bdb0366a008 Mon Sep 17 00:00:00 2001 From: asteroida123 <264808420+asteroida123@users.noreply.github.com> Date: Mon, 17 Aug 2026 09:40:02 +0800 Subject: [PATCH 03/75] =?UTF-8?q?feat(pk):=20agent=20PK=20arena=20?= =?UTF-8?q?=E2=80=94=20one=20task,=20N=20agents,=20side=20by=20side?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The arena sends one task to 2-4 agents simultaneously, gives each its own git worktree under `/.codeg-pk///`, and renders the contest live: a scoreboard (status / duration / output tokens / turns) over one LiveTranscriptView column per contestant, a diff tab that fetches each worktree's unified diff once the round settles, and a share button that exports the scoreboard as a PNG via html-to-image. Pure frontend — no backend changes: the orchestrator drives the existing connection machinery directly (connect/sendPrompt on synthetic `pk:` context keys, the pattern delegation children already use), completion is detected from per-connection status_changed events, and token/turn stats are summed from the persisted conversation. This deliberately bypasses the delegation broker: its v1 metrics are placeholders (turn_count: 1, token_usage: None at lifecycle.rs:342) and it cannot run parentless, so an arena built on it would need new backend surface for no gain. Distinct worktrees double as contestant isolation AND unlock same-agent rematches (connections dedup by agent+cwd, so two Claude slots need two cwds). Rounds persist to localStorage; a round still running at shutdown revives as `interrupted` with settled contestants kept and live fields dropped. All ten locales carry the new keys (the parity test enforces it). --- src/app/workspace/page.tsx | 11 +- .../chat/composer/composer-add-menu.tsx | 8 + src/components/pk/pk-arena-dialog.tsx | 252 +++++++++++++ src/components/pk/pk-arena-host.tsx | 42 +++ src/components/pk/pk-diff-view.test.ts | 39 ++ src/components/pk/pk-diff-view.tsx | 120 +++++++ src/components/pk/pk-launcher-dialog.tsx | 205 +++++++++++ src/components/pk/pk-scoreboard.tsx | 138 ++++++++ src/hooks/use-pk-round.ts | 332 ++++++++++++++++++ src/i18n/messages/ar.json | 56 ++- src/i18n/messages/de.json | 56 ++- src/i18n/messages/en.json | 56 ++- src/i18n/messages/es.json | 56 ++- src/i18n/messages/fr.json | 56 ++- src/i18n/messages/ja.json | 56 ++- src/i18n/messages/ko.json | 56 ++- src/i18n/messages/pt.json | 56 ++- src/i18n/messages/zh-CN.json | 56 ++- src/i18n/messages/zh-TW.json | 56 ++- src/stores/pk-arena-store.test.ts | 152 ++++++++ src/stores/pk-arena-store.ts | 258 ++++++++++++++ 21 files changed, 2106 insertions(+), 11 deletions(-) create mode 100644 src/components/pk/pk-arena-dialog.tsx create mode 100644 src/components/pk/pk-arena-host.tsx create mode 100644 src/components/pk/pk-diff-view.test.ts create mode 100644 src/components/pk/pk-diff-view.tsx create mode 100644 src/components/pk/pk-launcher-dialog.tsx create mode 100644 src/components/pk/pk-scoreboard.tsx create mode 100644 src/hooks/use-pk-round.ts create mode 100644 src/stores/pk-arena-store.test.ts create mode 100644 src/stores/pk-arena-store.ts diff --git a/src/app/workspace/page.tsx b/src/app/workspace/page.tsx index 1e3a6232a..cdc3f2171 100644 --- a/src/app/workspace/page.tsx +++ b/src/app/workspace/page.tsx @@ -1,7 +1,16 @@ "use client" import { ConversationDetailPanel } from "@/components/conversations/conversation-detail-panel" +import { PkArenaHost } from "@/components/pk/pk-arena-host" export default function WorkspacePage() { - return + return ( + <> + + {/* Inside the workspace layout's AcpConnectionsProvider — the arena + orchestrator needs the connection actions and the acp://event + subscription. */} + + + ) } diff --git a/src/components/chat/composer/composer-add-menu.tsx b/src/components/chat/composer/composer-add-menu.tsx index 4c51a8a5a..0c3184fcc 100644 --- a/src/components/chat/composer/composer-add-menu.tsx +++ b/src/components/chat/composer/composer-add-menu.tsx @@ -14,6 +14,7 @@ import { Plus, Search, Sparkles, + Swords, Upload, } from "lucide-react" @@ -31,6 +32,7 @@ import { DropdownRadioItemContent } from "@/components/chat/dropdown-radio-item- import { rankByTextMatch } from "@/lib/fuzzy-text-match" import { isImeCompositionKey } from "@/lib/ime-composition" import { cn } from "@/lib/utils" +import { usePkArenaStore } from "@/stores/pk-arena-store" import type { AvailableCommandInfo } from "@/lib/types" import { commandInvocationToken } from "@/components/chat/composer/invocation-reference" @@ -396,6 +398,12 @@ export function ComposerAddMenu({ )} + usePkArenaStore.getState().setLauncherOpen(true)} + > + + {t("startPk")} + ) diff --git a/src/components/pk/pk-arena-dialog.tsx b/src/components/pk/pk-arena-dialog.tsx new file mode 100644 index 000000000..e2fad824c --- /dev/null +++ b/src/components/pk/pk-arena-dialog.tsx @@ -0,0 +1,252 @@ +"use client" + +import { useEffect, useMemo, useRef, useState } from "react" +import { useTranslations } from "next-intl" +import { + Dialog, + DialogContent, + DialogDescription, + DialogTitle, +} from "@/components/ui/dialog" +import { LiveTranscriptView } from "@/components/message/live-transcript-view" +import { PkDiffView } from "@/components/pk/pk-diff-view" +import { PkScoreboard } from "@/components/pk/pk-scoreboard" +import { usePkRound } from "@/hooks/use-pk-round" +import { cn } from "@/lib/utils" +import { usePkArenaStore } from "@/stores/pk-arena-store" + +/** + * The arena itself: scoreboard on top, one live transcript column per + * contestant, a diff tab once the round settles, and a share button that + * exports the scoreboard as a PNG. Round switching comes from the store's + * `activeRoundId` — the launcher sets it on start, history can revisit any + * finished round (live streams of an old round are gone; its persisted + * conversation still renders through the same view). + */ + +export function PkArenaDialog() { + const t = useTranslations("PkArena.arena") + const open = usePkArenaStore((s) => s.arenaOpen) + const setArenaOpen = usePkArenaStore((s) => s.setArenaOpen) + const rounds = usePkArenaStore((s) => s.rounds) + const activeRoundId = usePkArenaStore((s) => s.activeRoundId) + const setActiveRound = usePkArenaStore((s) => s.setActiveRound) + + const round = useMemo( + () => rounds.find((r) => r.id === activeRoundId) ?? null, + [rounds, activeRoundId] + ) + + const { cancelRound, cleanupRound, fetchDiff } = usePkRound() + const [tab, setTab] = useState<"battle" | "diff">("battle") + const [now, setNow] = useState(() => Date.now()) + const [sharing, setSharing] = useState(false) + const [diffLoading, setDiffLoading] = useState(false) + const scoreboardRef = useRef(null) + + // Literal keys — next-intl's typed messages reject dynamic concatenation. + const roundStatusLabel = useMemo( + () => ({ + running: t("roundStatus.running"), + finished: t("roundStatus.finished"), + canceled: t("roundStatus.canceled"), + interrupted: t("roundStatus.interrupted"), + }), + [t] + ) + const tabLabel = useMemo( + () => ({ battle: t("tabs.battle"), diff: t("tabs.diff") }) as const, + [t] + ) + + const live = round != null && round.status === "running" + + // 1s scoreboard clock while any contestant is live. + useEffect(() => { + if (!open || !live) return + const timer = window.setInterval(() => setNow(Date.now()), 1000) + return () => window.clearInterval(timer) + }, [open, live]) + + useEffect(() => { + if (open) setNow(Date.now()) + }, [open]) + + // Diff tab: fetch each contestant's worktree diff once per visit. + useEffect(() => { + if (!open || tab !== "diff" || !round) return + let cancelled = false + const pending = round.contestants.filter( + (c) => c.diff == null && c.worktreePath + ) + if (pending.length === 0) return + setDiffLoading(true) + void Promise.allSettled(pending.map((c) => fetchDiff(round, c))).then( + () => { + if (!cancelled) setDiffLoading(false) + } + ) + return () => { + cancelled = true + } + }, [open, tab, round, fetchDiff]) + + const handleShare = async () => { + if (!scoreboardRef.current || sharing) return + setSharing(true) + try { + const { toPng } = await import("html-to-image") + const dataUrl = await toPng(scoreboardRef.current, { + backgroundColor: getComputedStyle(document.body).backgroundColor, + pixelRatio: 2, + }) + const link = document.createElement("a") + link.download = `codeg-pk-${round?.id ?? "round"}.png` + link.href = dataUrl + link.click() + } catch { + // Sharing is best-effort; a failed export must not disturb the round. + } finally { + setSharing(false) + } + } + + const roundLive = round != null && round.status === "running" + + return ( + + + {t("title")} + + {t("description")} + + {round ? ( +
+
+ + ⚔ + +
+
+ {round.task} +
+
+ {roundStatusLabel[round.status]} ·{" "} + {new Date(round.createdAt).toLocaleString()} +
+
+ {rounds.length > 1 ? ( + + ) : null} + {roundLive ? ( + + ) : round.contestants.some((c) => c.worktreePath) ? ( + + ) : null} + +
+ + + +
+ {(["battle", "diff"] as const).map((key) => ( + + ))} +
+ +
+ {tab === "battle" + ? round.contestants.map((contestant) => ( +
+ {contestant.conversationId != null ? ( + + ) : ( +
+ {contestant.statusDetail ?? t("preparing")} +
+ )} +
+ )) + : round.contestants.map((contestant) => ( + + ))} +
+
+ ) : ( +
+ {t("noRound")} +
+ )} + + + ) +} diff --git a/src/components/pk/pk-arena-host.tsx b/src/components/pk/pk-arena-host.tsx new file mode 100644 index 000000000..ebafed1d6 --- /dev/null +++ b/src/components/pk/pk-arena-host.tsx @@ -0,0 +1,42 @@ +"use client" + +import { useEffect, useRef } from "react" +import { PkArenaDialog } from "@/components/pk/pk-arena-dialog" +import { PkLauncherDialog } from "@/components/pk/pk-launcher-dialog" +import { usePkRound } from "@/hooks/use-pk-round" +import { usePkArenaStore } from "@/stores/pk-arena-store" + +/** + * Arena mount point — renders the launcher and arena dialogs and drives the + * orchestrator for rounds created by the launcher. Must live inside + * `AcpConnectionsProvider` (the workspace layout provides it): the + * orchestrator calls `connect`/`sendPrompt` and subscribes to `acp://event`. + * + * The launcher only writes the round into the store; this host picks it up, + * so round creation works from anywhere (composer menu, future entries) + * without prop-drilling. + */ +export function PkArenaHost() { + const { startRound } = usePkRound() + const rounds = usePkArenaStore((s) => s.rounds) + + // Drive any round that still has contestants in "preparing" — exactly the + // state the launcher leaves behind. Restarted (interrupted) rounds come + // back with settled statuses, so they are never re-driven. + const drivenRef = useRef(new Set()) + useEffect(() => { + for (const round of rounds) { + if (drivenRef.current.has(round.id)) continue + if (!round.contestants.some((c) => c.status === "preparing")) continue + drivenRef.current.add(round.id) + void startRound(round) + } + }, [rounds, startRound]) + + return ( + <> + + + + ) +} diff --git a/src/components/pk/pk-diff-view.test.ts b/src/components/pk/pk-diff-view.test.ts new file mode 100644 index 000000000..465656ff1 --- /dev/null +++ b/src/components/pk/pk-diff-view.test.ts @@ -0,0 +1,39 @@ +import { describe, expect, it } from "vitest" +import { parseUnifiedDiff } from "./pk-diff-view" + +describe("parseUnifiedDiff", () => { + it("classifies add, delete, hunk and context lines", () => { + const diff = [ + "diff --git a/main.py b/main.py", + "index 123..456 100644", + "--- a/main.py", + "+++ b/main.py", + "@@ -1,3 +1,4 @@", + " context line", + "-removed line", + "+added line", + "+another addition", + ].join("\n") + + const lines = parseUnifiedDiff(diff) + // File header lines are dropped entirely. + expect(lines.map((l) => l.kind)).toEqual([ + "hunk", + "ctx", + "del", + "add", + "add", + ]) + expect(lines[0].text).toBe("@@ -1,3 +1,4 @@") + }) + + it("keeps blank context lines as renderable entries", () => { + const lines = parseUnifiedDiff("+x\n\n-y") + expect(lines).toHaveLength(3) + expect(lines[1]).toEqual({ kind: "ctx", text: "" }) + }) + + it("handles an empty diff without inventing lines", () => { + expect(parseUnifiedDiff("")).toEqual([]) + }) +}) diff --git a/src/components/pk/pk-diff-view.tsx b/src/components/pk/pk-diff-view.tsx new file mode 100644 index 000000000..5776ae5b6 --- /dev/null +++ b/src/components/pk/pk-diff-view.tsx @@ -0,0 +1,120 @@ +"use client" + +import { useMemo } from "react" +import { useTranslations } from "next-intl" +import { AgentIcon } from "@/components/agent-icon" +import { getAgentLabel } from "@/lib/custom-agents" +import type { AgentType } from "@/lib/types" +import { cn } from "@/lib/utils" + +/** + * Lightweight unified-diff renderer for one contestant's worktree diff — + * line-level red/green with hunk headers, deliberately NOT the three-pane + * merge editor (that is for conflict resolution, not comparison). Each + * contestant renders in its own scrollable column. + */ + +interface DiffLine { + kind: "add" | "del" | "hunk" | "ctx" + text: string +} + +export function parseUnifiedDiff(diff: string): DiffLine[] { + if (diff.trim() === "") return [] + const lines: DiffLine[] = [] + for (const raw of diff.split("\n")) { + if ( + raw.startsWith("+++") || + raw.startsWith("---") || + raw.startsWith("diff ") || + raw.startsWith("index ") + ) { + continue + } + if (raw.startsWith("@@")) { + lines.push({ kind: "hunk", text: raw }) + } else if (raw.startsWith("+")) { + lines.push({ kind: "add", text: raw }) + } else if (raw.startsWith("-")) { + lines.push({ kind: "del", text: raw }) + } else { + lines.push({ kind: "ctx", text: raw }) + } + } + return lines +} + +function diffStats(lines: DiffLine[]): { added: number; removed: number } { + let added = 0 + let removed = 0 + for (const line of lines) { + if (line.kind === "add") added += 1 + if (line.kind === "del") removed += 1 + } + return { added, removed } +} + +const LINE_STYLE: Record = { + add: "bg-emerald-500/10 text-emerald-700 dark:text-emerald-400", + del: "bg-red-500/10 text-red-700 dark:text-red-400", + hunk: "bg-muted text-muted-foreground", + ctx: "text-foreground/80", +} + +export function PkDiffView({ + agentType, + diff, + loading, +}: { + agentType: AgentType + diff: string | null + loading: boolean +}) { + const t = useTranslations("PkArena.diff") + const lines = useMemo(() => (diff ? parseUnifiedDiff(diff) : []), [diff]) + const stats = useMemo(() => diffStats(lines), [lines]) + const empty = diff != null && diff.trim() === "" + + return ( +
+
+ + + {getAgentLabel(agentType)} + + {diff != null && !empty ? ( + + + +{stats.added} + + + −{stats.removed} + + + ) : null} +
+
+ {loading ? ( +
+ {t("loading")} +
+ ) : empty ? ( +
+ {t("empty")} +
+ ) : ( + 
+            {lines.map((line, index) => (
+              

+                {line.text === "" ? " " : line.text}
+              

+            ))}
+
+ )} +
+
+ ) +} diff --git a/src/components/pk/pk-launcher-dialog.tsx b/src/components/pk/pk-launcher-dialog.tsx new file mode 100644 index 000000000..8861679b7 --- /dev/null +++ b/src/components/pk/pk-launcher-dialog.tsx @@ -0,0 +1,205 @@ +"use client" + +import { useEffect, useMemo, useState } from "react" +import { useTranslations } from "next-intl" +import { + Dialog, + DialogContent, + DialogDescription, + DialogTitle, +} from "@/components/ui/dialog" +import { AgentIcon } from "@/components/agent-icon" +import { useAcpAgents } from "@/hooks/use-acp-agents" +import { getFolder } from "@/lib/api" +import type { AgentType } from "@/lib/types" +import { cn } from "@/lib/utils" +import { usePkArenaStore } from "@/stores/pk-arena-store" +import { useTabStore } from "@/stores/tab-store" + +/** + * Arena launcher: pick 2-4 installed agents, write the task, start the round. + * Reads the ACTIVE tab for the target folder (an arena needs a real folder — + * its git repo provides the per-contestant worktrees; chat mode has none). + */ + +const MIN_CONTESTANTS = 2 +const MAX_CONTESTANTS = 4 + +export function PkLauncherDialog() { + const t = useTranslations("PkArena.launcher") + const open = usePkArenaStore((s) => s.launcherOpen) + const setLauncherOpen = usePkArenaStore((s) => s.setLauncherOpen) + const setArenaOpen = usePkArenaStore((s) => s.setArenaOpen) + const createRound = usePkArenaStore((s) => s.createRound) + const { agents: rawAgents } = useAcpAgents() + const activeTab = useTabStore((s) => + s.activeTabId + ? (s.tabs.find((tab) => tab.id === s.activeTabId) ?? null) + : null + ) + + const [selected, setSelected] = useState([]) + const [task, setTask] = useState("") + const [workingDir, setWorkingDir] = useState(null) + const [folderId, setFolderId] = useState(null) + + useEffect(() => { + if (!open) return + setSelected([]) + setTask("") + setWorkingDir(null) + setFolderId(null) + // The active tab decides where the arena runs. Draft tabs may lack a + // workingDir; fall back to the folder's own path. + if (activeTab?.folderId == null || activeTab.folderId < 0) return + if (activeTab.workingDir) { + setFolderId(activeTab.folderId) + setWorkingDir(activeTab.workingDir) + return + } + let cancelled = false + getFolder(activeTab.folderId) + .then((folder) => { + if (cancelled) return + setFolderId(folder.id) + setWorkingDir(folder.path) + }) + .catch(() => undefined) + return () => { + cancelled = true + } + // eslint-disable-next-line react-hooks/exhaustive-deps + }, [open]) + + const agents = useMemo( + () => rawAgents.filter((a) => a.enabled && a.available), + [rawAgents] + ) + + const noFolder = open && folderId == null && activeTab != null + const taskValid = task.trim().length > 0 + const selectionValid = + selected.length >= MIN_CONTESTANTS && selected.length <= MAX_CONTESTANTS + const canStart = + taskValid && selectionValid && folderId != null && workingDir != null + + const toggle = (agentType: AgentType) => { + setSelected((prev) => + prev.includes(agentType) + ? prev.filter((a) => a !== agentType) + : prev.length >= MAX_CONTESTANTS + ? prev + : [...prev, agentType] + ) + } + + const handleStart = () => { + if (!canStart || folderId == null || workingDir == null) return + createRound({ + task: task.trim(), + folderId, + workingDir, + agents: selected, + }) + setLauncherOpen(false) + setArenaOpen(true) + // The orchestrator (in PkArenaHost) picks the new round up from the store. + } + + return ( + + + + {t("title")} + + + {t("description")} + +
+
+
+ {t("contestantsLabel")} +
+ {noFolder ? ( +
+ {t("noFolderHint")} +
+ ) : null} +
+ {agents.map((agent) => { + const isSelected = selected.includes(agent.agent_type) + return ( + + ) + })} +
+ {selected.length > 0 && selected.length < MIN_CONTESTANTS ? ( +
+ {t("needMore", { count: MIN_CONTESTANTS })} +
+ ) : null} +
+
+ +