codebee 0.1.49 → 0.1.50
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +24 -0
- package/README.md +4 -150
- package/app/core/aiflavor.py +14 -0
- package/app/core/builtin_agent.py +450 -32
- package/app/core/knowledge.py +432 -405
- package/app/core/manager.py +76 -3
- package/app/core/modelhub.py +53 -5
- package/app/core/pipeline.py +158 -66
- package/app/core/preview.py +293 -0
- package/app/core/runner.py +336 -144
- package/app/core/settings.py +7 -1
- package/app/core/store.py +74 -2
- package/app/core/zentao.py +63 -13
- package/app/main.py +66 -1
- package/app/ui/app.js +397 -102
- package/app/ui/i18n.js +41 -30
- package/app/ui/index.html +45 -2
- package/app/ui/style.css +206 -26
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -6,6 +6,30 @@ README 元数据带回,供老版本在「发现新版本」时展示新版更
|
|
|
6
6
|
|
|
7
7
|
## 未发布
|
|
8
8
|
|
|
9
|
+
### ✨ 新功能
|
|
10
|
+
|
|
11
|
+
- **直连对话把「思考过程」打印出来了**:内置智能体改走流式(SSE),模型的思维链(DeepSeek/vLLM 系 `reasoning_content`、Anthropic `thinking` 块)边生成边实时打印进对话时间线,工具调用与正文增量同步显示——此前整轮回答只有三个点的打字动画,几十秒黑箱干等。运行中思考块常开实时增长;回答收尾后收成「思考过程」折叠摘要(带字数,点开回看),思维链同时进步骤日志。网关不支持流式时自动回落非流式(思考字段仍从响应体兜底抽取),取消/多 KEY 换链语义不变。
|
|
12
|
+
- **日志抽屉认识 claude 事件流了**:claude 步骤日志此前整段原样 JSON,模型长思考时 `thinking_tokens` 心跳每个流式分片刷一条(真实案:3.2MB 步骤日志 15828 条),把正文顶出尾部窗口。现在抽屉视图按邻近组折叠成一行「— 思考中(心跳 ×N 已折叠,估算 ~M tokens)—」,会话启动/API 重试/思考摘要/正文/收尾统计各自翻译成可读行(API 重试逐条保留——429 与断连诊断金矿);只影响显示,原始日志与结果解析不动。
|
|
13
|
+
- 新增**网页成品实时预览**:代码任务跑完,详情页多出「预览」页签——「运行视图」把工作目录里的入口 HTML 真跑起来(iframe 里 style.css / script.js 这些相对引用照常生效,改完点「刷新」即看新版),旁边一排文件页签直接读源码(行号 + 高亮,与成果预览同款),右上角「打开新窗口」给真全屏。借鉴对话式编程产品"边聊边看效果"的体验:终态代码任务跑出了可打开的网页成品时,打开详情直接落在预览;直连对话任务的对话胶囊里也有「预览」入口。没有网页成品的任务(小说、普通脚本)完全不出现这个页签。安全边界:预览是只读静态挂载,只认静态扩展名白名单、拒绝路径穿越与隐藏目录;访问令牌走路径段(iframe 子资源继承不了查询串),鉴权与 /api/* 同一把尺子。
|
|
14
|
+
- **直连对话会动手修了**:内置智能体新增 `run_command`,诊断、修复、改配置、重启服务等命令直接在你机器上真实执行(与 codex/claude 全权沙箱同信任档,无确认闸),退出码与输出回给智能体接力下一步——不再只甩一份「照着做」的指南。超时/取消即强杀进程树;Windows 提权命令自动走 UAC 弹窗(点一次允许即可)。顺带修掉一个底层隐患:子进程杀树后管道关闭不再死等漏网孙进程自然退出(此前会让超时杀进程卡满命令时长)。
|
|
15
|
+
- 连载小说支持**分卷**:新建任务时给「每卷章数」(默认预填 20,清空 = 不分卷),或在「指定卷结构」里写明各卷(也能直接在任务目标里写「第一卷 少年初入江湖 第1-20章;卷二 风云再起 21-40章」,自动识别)。分卷贯穿四个环节:① 大纲按卷规划,产出卷名与卷弧光(用户给了卷名就原样沿用);② 每章起草与评审都带上「第几卷第几章、是不是卷末」,卷末章按本卷高潮与卷末钩子收束;③ 合并成书插入「第 X 卷 《卷名》」分隔;④ 运行详情的章节评审卡与发布页章节清单按卷分组。卷边界由全书章号确定性推导,续写批次自动接上同一卷、不会把一卷劈成两半。
|
|
16
|
+
|
|
17
|
+
### ⚡ 性能
|
|
18
|
+
|
|
19
|
+
- **多评审并发**:章节评审与全书评审的多位评审从排队串行改为并发执行——评审占连载运行总时长的六到七成,而各评审读的是同一份稿件、各自独立打分,串行等待纯属时间叠加。每位评审的得分、逐项目标审稿(event_check)与资源账本(ledger)仍按原顺序合并,评分口径与串行版完全一致;某位评审的管道异常照旧上抛打断。同章两评审实测墙钟约 7.2s → 4.2s。
|
|
20
|
+
- **超时不再原样重烧**:章稿起草失败重试加「新变量」闸门——上一次就是超时/停滞收场、且这次的提示词与上次逐字节相同时(缩上下文降级未能生效的场景),不再用同样输入再烧一整轮 2400 秒,直接跳到换将(换 CLI/模型才是新机会)。
|
|
21
|
+
|
|
22
|
+
### 🛠 问题修复
|
|
23
|
+
|
|
24
|
+
- **首选 KEY 超限自动切备用了**:同厂商多密钥此前只认欠费类文案——首选 KEY 撞上 429 限流(智谱「并发数超过限制」「每分钟Token数已超过上限」、claude 的 rate_limit_error、codex 的 "429 Too Many Requests" 等)只记一笔错误、不进冷却,每个新步骤仍从超限的首选 KEY 重新烧起。现在限流文案与欠费同口径:进冷却 30 分钟、密钥镜像自动切下一把、链展开跳过冷却中的 KEY(到期或「恢复」按钮一键回首选)。顺带补一处漏账:无显式链时「换模型回退」的链条目不带 KEY 标记,这类重试失败此前完全不记 KEY 账。
|
|
25
|
+
|
|
26
|
+
## v0.1.50(2026-09-22)
|
|
27
|
+
|
|
28
|
+
- 今日体检修复:i18n 英文字典 22 个重复键清理——其中「 条/字/章」等 7 个值冲突属后值静默覆盖(如「第 N 章」曾被渲染成 "Chapter N chapters"),按使用现场裁定正确值,英文界面文案修复
|
|
29
|
+
- 测试适配:codex 配置同步两测试改「假 HOME 在临时目录下」放行形态,恢复验证真实写盘语义(防毒闸 011023c 引入的预存挂清零)
|
|
30
|
+
- 并行功能收尾一批:自动化运行偏好透传、内置智能体工作目录锁死、连载卷表显式化、市场已装包启停开关、思维链可视化、卷级规划要点、runner 缺件诊断、流式落盘节流、禅道间隔跟随默认
|
|
31
|
+
- 合并质量巡检:436 个本包导入绑定完整性通过(无漏提交模块)、pipeline 四组历史修复全部在位、架构检查无新增违规
|
|
32
|
+
|
|
9
33
|
## v0.1.49(2026-09-22)
|
|
10
34
|
|
|
11
35
|
### 🛠 问题修复
|
package/README.md
CHANGED
|
@@ -21,157 +21,11 @@ Kimi Code、MiMo Code、Grok Build、Pi、DeepSeek Harness……),提供
|
|
|
21
21
|
不会把你的任务内容交给任何第三方。
|
|
22
22
|
|
|
23
23
|
<!-- relnotes:start -->
|
|
24
|
-
###
|
|
24
|
+
### 最新版更新内容(v0.1.50)
|
|
25
25
|
|
|
26
|
-
-
|
|
27
|
-
-
|
|
28
|
-
|
|
29
|
-
### 🛠 问题修复 · Fixes
|
|
30
|
-
|
|
31
|
-
- 智能体读超大文件不再只给开头:保留开头 44KB+结尾 16KB、中段标注省略——文件末尾的报错与结论不再被截掉,结尾中文不乱码。
|
|
32
|
-
- Reading oversized files no longer keeps only the head: 44KB head + 16KB tail with the middle elided — end-of-file errors and conclusions survive, and trailing Chinese text stays intact.
|
|
33
|
-
|
|
34
|
-
### ✨ 新功能 · New
|
|
35
|
-
|
|
36
|
-
- 经验库新增「有效做法」记忆:高分一次通过的任务会把「做对了什么」提炼成可复用步骤,下次同类任务自动注入照做;绿色「做法」徽章一眼区分。
|
|
37
|
-
- The lesson library now also captures "what worked": high-scoring first-pass runs distill reusable procedures, auto-injected into similar future tasks — marked with a green "Best practice" badge.
|
|
38
|
-
|
|
39
|
-
- 新任务类型「标书编制」:上传招标文件,AI 按评分标准逐条应答起草投标文件;资质不虚构(缺失标注待补)、应答先结论、全文自查废标风险项。
|
|
40
|
-
- New "Bid Document" task type: upload the tender file and the AI drafts point-by-point responses against scoring criteria — no fabricated credentials, conclusion-first answers, and a full disqualification-risk sweep.
|
|
41
|
-
|
|
42
|
-
### 🛠 问题修复 · Fixes
|
|
43
|
-
|
|
44
|
-
- 断点续跑/重跑更省钱:同任务再次生成大纲或计划直接命中精确缓存(1 小时内),不重复花全量规划调用;格式不合规的重试仍重新生成。
|
|
45
|
-
- Resuming or re-running a task now hits an exact cache for outline/plan regeneration (within 1 hour), skipping a full planning call; format-failure retries still regenerate.
|
|
46
|
-
|
|
47
|
-
- 修复长任务上下文撑爆后重试仍失败:压缩摘要此前只写日志、没真正替换发往模型的上下文;现在重试用「压缩摘要+最近消息」重建请求并放弃膨胀的旧会话,超长任务能真正续跑。
|
|
48
|
-
- Fixed long-task overflow retries failing again: compaction summaries used to go only to logs. Retries now rebuild the request from the summary plus recent messages and drop the bloated old session.
|
|
49
|
-
|
|
50
|
-
### ✨ 新功能 · New
|
|
51
|
-
|
|
52
|
-
- 新建任务输入框焕新:编排模式、思考程度、「厂商/模型」选择器收进输入框工具行右侧(左附件、右偏好+发送的主流布局);对话模型点开是「厂商 → 模型」两级菜单,选中打勾、底部「管理模型」直达,pill 显示「厂商/模型」组合名。
|
|
53
|
-
- The composer was refreshed: orchestration mode, reasoning level and a "provider/model" picker now sit on the input box's right side (attachments left, preferences + send right, like mainstream AI input boxes). The model picker opens a provider → model cascading menu with checkmarks and a shortcut to model management.
|
|
54
|
-
|
|
55
|
-
- 经验库越用越准:教训按任务结局自动记胜负,真正帮上忙的经验优先注入、帮倒忙的被挤出——随真实运行自我提纯。
|
|
56
|
-
- The lesson library now tracks outcomes: lessons that helped a task pass gain weight, ones that failed to prevent issues fade — it self-refines with every run.
|
|
57
|
-
|
|
58
|
-
### 🛠 问题修复 · Fixes
|
|
59
|
-
|
|
60
|
-
- 端口诊断的「关闭」真能关掉了:温和信号无效自动升级强杀,以「端口确实释放」为准,不再假报成功。
|
|
61
|
-
- Port diagnostics "Close" now really closes: gentle signal escalates to forced kill when ignored, and success is only reported once the port is actually released.
|
|
62
|
-
|
|
63
|
-
- 群摘要:选好 Python 直接刷新不再误报「请先填路径」;新建的群也会出现在监听列表;蜜蜂气泡「点我细看」真能点了。
|
|
64
|
-
- Digests: refreshing right after picking a Python no longer nags "path required"; brand-new groups show up in the picker; the bee bubble "click to read" actually navigates now.
|
|
65
|
-
|
|
66
|
-
- 禅道:「测试并保存连接」不再卡按钮几十秒;兼容更多老版禅道的模块/产品清单形状;操作状态显示具体操作名。
|
|
67
|
-
- ZenTao: "Test & save" no longer holds the button for tens of seconds; more legacy ZenTao module/product list shapes are supported; the operation panel shows concrete action names.
|
|
68
|
-
|
|
69
|
-
- 设置页「代码设置」恢复横排布局(此前被章节评审卡样式意外覆盖)。
|
|
70
|
-
- Settings → Code display restores its row layout after an accidental style override from the chapter-review card.
|
|
71
|
-
|
|
72
|
-
- 手机端操作被「其他设备正在控制」挡住时,弹出「接管并重试」——一键接管并自动重做刚才的操作。
|
|
73
|
-
- When a mobile action is blocked by another device in control, a "Take over & retry" prompt appears — one tap retakes control and redoes the action.
|
|
74
|
-
|
|
75
|
-
- 微信群列表拉取失败时会显示读取器的真实报错(依赖缺失/位数不对/微信未登录一眼可辨)。
|
|
76
|
-
- WeChat group-list failures now surface the reader's real error (missing dependency, wrong bitness, or WeChat not logged in).
|
|
77
|
-
|
|
78
|
-
- 连载章节评审在任务级详情与单次运行详情都可见;新建任务会记住上次的编排模式和思考程度。
|
|
79
|
-
- Chapter reviews now appear in both task and run details; new tasks also remember the last orchestration mode and reasoning level.
|
|
80
|
-
|
|
81
|
-
- 扫描 Python 后即使保存前遇到后台刷新,刚选中的路径也不会再被冲掉。
|
|
82
|
-
- A Python selected by Scan now survives background refreshes even before settings are saved.
|
|
83
|
-
|
|
84
|
-
### ✨ 新功能 · New
|
|
85
|
-
|
|
86
|
-
- 创建任务支持快速/自动/专家/手动编排与独立思考程度;代码、写作、翻译、邮件、汇报、调研等全类型按风险动态选择步骤。
|
|
87
|
-
- 新建对话默认系统推荐,也可只对当前对话指定厂商/模型;创建前和运行中显示 ETA/P90 与剩余时间。
|
|
88
|
-
- New tasks support Fast/Auto/Expert/Manual orchestration and an independent reasoning level; every task type adapts its workflow to risk.
|
|
89
|
-
- Conversations use the system recommendation by default, with optional per-conversation provider/model selection plus ETA/P90 visibility.
|
|
90
|
-
|
|
91
|
-
- 修复 Windows 控制台启动崩溃、损坏 Git 仓库导致 Aider 长时间等待、Anthropic 接口收到不兼容思考字段,以及桌宠语言切换后重复开窗。
|
|
92
|
-
- Fixed Windows console startup crashes, Aider stalls on unreadable Git repositories, incompatible Anthropic reasoning fields, and duplicate pet windows after language changes.
|
|
93
|
-
|
|
94
|
-
- 连载任务详情新增「章节评审」卡片:每章维度评分(低分标红)与达标状态,大纲要点逐项核对结果红绿灰直读。
|
|
95
|
-
- Serial runs now show a Chapter Reviews card: per-chapter dimension scores (low ones in red) plus beat-by-beat checks color-coded green/red/gray.
|
|
96
|
-
|
|
97
|
-
- 群摘要搬进正式页面:设置 →「微信群聊汇报」里配置与摘要列表一体呈现,右下角悬浮小窗已移除;桌面蜜蜂右键「群摘要」直达。
|
|
98
|
-
- Group digests now live on a proper settings page (config + digest list in one place); the floating bottom-right widget is gone, and the desktop bee links straight to it.
|
|
99
|
-
|
|
100
|
-
### 🛠 问题修复 · Fixes
|
|
101
|
-
|
|
102
|
-
- 修复「选不中 Python」:自动刷新不再冲掉你刚选好但还没保存的路径。
|
|
103
|
-
- Fixed "can't pick a Python": background refresh no longer wipes a freshly chosen but unsaved path.
|
|
104
|
-
|
|
105
|
-
- 微信直连配置更省事:「扫描」自动找本机 Python(标注位数与依赖状态,点选即用),选中后可一键安装 wechatauto-replica。
|
|
106
|
-
- WeChat setup is now effortless: a Scan button finds local Pythons (with bitness and dependency badges), and one click installs wechatauto-replica.
|
|
107
|
-
|
|
108
|
-
- 桌面蜜蜂右键新增「群摘要」:直达群摘要面板,已打开的页面只聚焦不重复开。
|
|
109
|
-
- The desktop bee's right-click menu now has "Group Digests" — it focuses the existing page instead of opening a duplicate.
|
|
110
|
-
|
|
111
|
-
- 没装 Pillow 的电脑上蜜蜂也会显示新形象贴图了(修复手绘丑蜜蜂问题)。
|
|
112
|
-
- Bees now show the new artwork even without Pillow installed (fixes the ugly hand-drawn fallback).
|
|
113
|
-
|
|
114
|
-
### 🛠 问题修复 · Fixes
|
|
115
|
-
|
|
116
|
-
- 禅道自动化开关与扫描间隔紧凑排成一行。
|
|
117
|
-
- ZenTao automation toggles and the scan interval now fit on one tidy row.
|
|
118
|
-
|
|
119
|
-
### 上一版亮点 · Previous release
|
|
120
|
-
|
|
121
|
-
- 升级后蜜蜂自动换新装:服务重启后旧蜜蜂自动让位,新蜜蜂带新形象上线。
|
|
122
|
-
- After upgrading, the desktop bee swaps itself for the new look: the old bee steps aside when the service restarts.
|
|
123
|
-
|
|
124
|
-
- 去掉蜜蜂底部的任务进度条,窗口更干净。
|
|
125
|
-
- Removed the task progress bar under the desktop bee for a cleaner look.
|
|
126
|
-
|
|
127
|
-
- 「关于与更新」页的版本信息与更新说明改为左对齐阅读,不再整块居中。
|
|
128
|
-
- The version info and release notes on the About page are now left-aligned for easier reading.
|
|
129
|
-
|
|
130
|
-
### ✨ 新功能 · New
|
|
131
|
-
|
|
132
|
-
- 连载任务防穿帮:新增「资源账本」——每章评审自动登记新出现的道具、伤情、承诺与伏笔,写下一章前自动注入提醒核对。
|
|
133
|
-
- Serial tasks now keep a resource ledger: props, injuries, promises, and plot threads are tracked per chapter and injected before writing the next one.
|
|
134
|
-
|
|
135
|
-
- 连载评审逐项核对:每章按大纲要点逐项判定「已完成/未完成/待核实」并引用正文证据,铺垫不再被当成已完成。
|
|
136
|
-
- Chapter reviews now check every outline beat against the actual text — foreshadowing no longer counts as done.
|
|
137
|
-
|
|
138
|
-
- 新建任务记住你的偏好:上次选择的任务类型、轮数、阈值、章节数等下次自动预填(只记参数,不记内容)。
|
|
139
|
-
- The new-task form now remembers your last-used type, rounds, threshold, and chapter settings (parameters only, never content).
|
|
140
|
-
|
|
141
|
-
### 🛠 问题修复 · Fixes
|
|
142
|
-
|
|
143
|
-
- 代码任务提速:验证命令失败时先直接修复再评审(首轮不再白跑一次评审),连续失败两轮后才请评审参与诊断。
|
|
144
|
-
- Code tasks get faster: when verification fails, the agent fixes first and review joins only after repeated failures.
|
|
145
|
-
|
|
146
|
-
### 上一版亮点 · Previous release
|
|
147
|
-
|
|
148
|
-
- 顶栏新增统一操作状态中心:创建、删除、绑定、插件等写操作会显示进行中、完成或失败状态,并保留时间与错误原因。
|
|
149
|
-
- A unified operation center now shows create, delete, binding, and plugin actions as running, completed, or failed, with timestamps and error details.
|
|
150
|
-
|
|
151
|
-
### 上一版亮点 · Previous release
|
|
152
|
-
|
|
153
|
-
- 文档类任务更专业:汇报先给结论、邮件先答复再铺陈、技术方案从问题开场并保留被否决的真实备选、文档验收标准可逐条勾选。
|
|
154
|
-
- Smarter document tasks: reports lead with conclusions, emails answer first, proposals open at the problem and keep a real rejected alternative, docs get testable acceptance criteria.
|
|
155
|
-
|
|
156
|
-
- 升级完成后自动重启生效,不用再手动点「重启服务」;页面几秒后自动恢复。有任务在跑时不会自动重启,会提示稍后手动重启。
|
|
157
|
-
- Updates finish on their own: after upgrading, the service restarts itself — no manual restart needed. If tasks are running, it waits and tells you instead.
|
|
158
|
-
|
|
159
|
-
- 启动时若端口被上次没退干净的旧实例占用,自动清场后再启动,不再报错让用户手动排查。
|
|
160
|
-
- At startup, a stale instance holding the port is cleared automatically instead of failing with a "port occupied" error.
|
|
161
|
-
|
|
162
|
-
- 运行详情页新增「归档任务」按钮:任务结束后可直接归档/取消归档,不用再找右键菜单。
|
|
163
|
-
- The run detail page now has an "Archive Task" button — no need to hunt for the right-click menu.
|
|
164
|
-
|
|
165
|
-
### 🛠 问题修复 · Fixes
|
|
166
|
-
|
|
167
|
-
- 修复任务附件偶尔被智能体无视:现在要求动手前必须先读附件,结论要建立在附件内容上;读不了会明确说明缺了什么。
|
|
168
|
-
- Fixed task attachments occasionally being ignored: agents must now read attachments first and base their answers on them, or clearly say what's missing.
|
|
169
|
-
|
|
170
|
-
- 修复桌面蜜蜂「关不掉」:右键「关闭桌宠」现在真正关闭,不再自动复活。
|
|
171
|
-
- Fixed the desktop bee not closing: the Close menu item now really closes it instead of reviving.
|
|
172
|
-
|
|
173
|
-
- 修复桌面蜜蜂「拖不动」:拖动此前绑定了一个窗口上不存在的方法,现已修复并更跟手。
|
|
174
|
-
- Fixed the desktop bee not draggable: dragging was bound to a method that doesn't exist on the window; now fixed and smooth.
|
|
26
|
+
- 英文界面文案修复:字典重复键导致的静默覆盖清理(22 处)
|
|
27
|
+
- 并行功能收尾一批:自动化偏好透传、思维链可视化、市场启停开关、卷级规划等
|
|
28
|
+
- 合并质量巡检全过:导入完整性 / pipeline 修复在位 / 架构无新增违规
|
|
175
29
|
<!-- relnotes:end -->
|
|
176
30
|
|
|
177
31
|
---
|
package/app/core/aiflavor.py
CHANGED
|
@@ -146,6 +146,20 @@ def pacing_analyze(text):
|
|
|
146
146
|
"alerts": alerts, "alert": bool(alerts)}
|
|
147
147
|
|
|
148
148
|
|
|
149
|
+
def inject_into_prompt(tpl, text, kind=None):
|
|
150
|
+
"""把确定性检测块插进评审提示词,返回新提示词(无命中原样返回)。
|
|
151
|
+
|
|
152
|
+
位置固定为「## 待评审稿件」之前——与技能包/圣经/知识库注入同一位置,
|
|
153
|
+
保持前缀稳定(§07 T1.2' 前缀缓存)。单稿件评审与连载逐章评审共用本
|
|
154
|
+
函数:此前连载逐章完全没有这条参考线(旗舰场景漏挂)。
|
|
155
|
+
"""
|
|
156
|
+
det = report_line(text, kind)
|
|
157
|
+
if not det:
|
|
158
|
+
return tpl
|
|
159
|
+
return tpl.replace("## 待评审稿件",
|
|
160
|
+
"## 确定性检测结果(供评审参考)\n%s\n\n## 待评审稿件" % det, 1)
|
|
161
|
+
|
|
162
|
+
|
|
149
163
|
def report_line(text, kind=None):
|
|
150
164
|
"""生成注入评审提示词的报告行(措辞层 + 叙事架构层 + 节奏层);
|
|
151
165
|
全部无命中返回空串。kind 为流程类型:只有叙事类流程才下发节奏检测
|