Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -16,9 +16,17 @@ All notable changes to DreamCoder will be documented in this file.
- Every request the app itself sends to a provider (through the local proxy, for session titles and for connection tests) is recorded in the diagnostics log with the provider, model, status, timings, the provider's request id and how it ended, including errors in the middle of a stream and clients that disconnect; never the key or the request body (#262).
- **A stalled reply ends after the AI request timeout.** When a provider stops sending anything for longer than the timeout in the network settings, the agent gives up on that reply and starts it again, instead of waiting until you stop it. This now applies to every provider, whether it is reached directly or through the local proxy; keep-alive signals alone do not count as progress (#262).
- Providers in OpenAI format can carry extra request headers and query parameters, such as Azure's `api-version`. The settings form does not offer them yet; for now they are set by editing the provider in `providers.json`. Authentication and the content type still come from the provider settings and override a header of the same name (#262).
- The agent's instructions describe DreamCoder instead of Claude Code: the environment section no longer lists Claude models, Claude Code's products or `/fast`, and says replies are shown in the DreamCoder app; `/help` and feedback point to DreamCoder; the worktree tools, the Read tool and the `update-config` skill describe what the app actually does (#270).
- Plan mode follows one procedure: the agent explores, asks you questions and writes the plan file as it goes, instead of being told in one place not to write any file and in another to launch planning subagents (#270).
- With models other than Claude, notes that arrive in the middle of a turn (messages from other sessions, board notifications, reminders) are sent as their own message instead of being appended to the last tool result (#270).
- The notice for a file that changed after the agent read it no longer claims you or a linter changed it or asks the agent to keep it from you, and shows at most about 2 KB of the change (#270).

### Fixed

- **Saving a memory no longer asks for approval every time.** Writes to the project's memory folder are allowed like the instructions say, also in bypass mode (#270).
- **Stopping a reply keeps the session's context.** Stop ends the agent process only when the reply has not stopped a few seconds after you press Stop, and a restarted agent keeps uploaded images and files, instruction files from subfolders and files restored after compaction, without adding a "Continue from where you left off." message (#270).
- The agent knows the available skills again after compaction, gets only newly added skills when the skill folders change, and does not get the whole list again after a restart (#270).
- **`/goal` keeps working until the goal is met.** Before, the turn ended at the first check that found the goal unmet. Each automatic continuation now tells the agent the goal, what is still missing and how many continuations are left. The turn ends and the goal stays active when the agent needs something only you can provide, or after 10 continuations in a row without a message from you (#270).
- A provider in OpenAI format whose base URL already ends in a version, such as `https://open.bigmodel.cn/api/paas/v4` or Gemini's `.../v1beta/openai`, works: requests, model lists and the connection test no longer add another `/v1`. Base URLs without a version still get `/v1` (#262).
- A streamed reply through the local proxy is no longer cut off when the provider pauses for more than 60 seconds, and stopping a reply or deleting a session now also stops the request to the provider before it starts answering (#262).
- With a manual HTTP proxy set in the network settings, the agent reaches the app's local proxy directly instead of sending those requests to the HTTP proxy (#262).
Expand All @@ -33,6 +41,10 @@ All notable changes to DreamCoder will be documented in this file.
- When the output limit cuts off a tool call in the middle of its arguments, the agent no longer runs the tool with missing arguments; the model is told the call was cut off and retries with smaller pieces (#263).
- **A reply that breaks off in the middle is retried as a stream.** When a provider's stream fails partway, the agent removes the half-finished reply from the conversation, the app and the session history, and asks again with streaming, instead of switching to a slower request without streaming that long answers could not finish within the timeout. Retries before and during a stream share one limit (#263).

### Removed

- The status line setup agent and the `/statusline` command; they set up the terminal's `statusLine` setting, which the app does not run (#270).

## [0.9.0] — 2026-11-17

Sessions now work together: they start and message each other, share a project board, and open side chats. The agent can also make slide decks. These notes cover everything since 0.8.0.
Expand Down
7 changes: 7 additions & 0 deletions docs/features/harness/prompt/context.md
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,7 @@

### T-CTX-02 桌面端设了记忆路径覆盖变量,写记忆文件因此每次都要用户批准,与记忆说明里的「直接写」矛盾

- 状态:已修复(#270):记忆写入豁免覆盖服务器指定的记忆路径。
- 状态:维护者已定(2026-11-18):扩展 CLI 的记忆写入豁免,让它也覆盖服务器指定的记忆路径。
- 类型:问题
- 严重程度:中
Expand All @@ -76,6 +77,7 @@

### T-CTX-03 压缩后技能清单不再发送,模型不知道有哪些技能

- 状态:已修复(#270):已发送的技能从对话历史推导,压缩后在压缩附件里补发完整清单,重启后不重复发送。
- 类型:问题
- 严重程度:中
- 体量:技能清单首轮约 1 500 字符起(6 个内置技能 1 410 + 头 65 + 包装),上限 8 000(200K 窗口);压缩后这部分缺失
Expand All @@ -91,6 +93,7 @@

### T-CTX-04 第一条 user 消息的外层包装说「可能无关、不要回应」,内层说「必须严格遵守」,记忆索引也被归为「必须遵守的指令」

- 状态:已修复(#270):外层改为「自动附上的上下文,不是用户的话」;前言改为用户消息优先;`MEMORY.md` 移出前言,单独说明。`# claudeMd` 等标题改名留给 #278。
- 类型:问题
- 严重程度:中
- 体量:固定包装 290 字符 + 内层前言 193 字符 + 每个文件块头约 65 字符加路径,每个请求都发
Expand All @@ -105,6 +108,7 @@

### T-CTX-05 环境段对非 Claude 模型发 552 字符的 Claude 产品与模型信息,宿主写成 Claude Code,没有 DreamCoder 桌面端的说明

- 状态:已修复(#270):同 T-SYS-04。
- 类型:问题
- 严重程度:中
- 体量:环境段 823 字符(本仓库路径下实测),其中 3 行 552 字符为 Claude 专属;每个请求都发
Expand All @@ -120,6 +124,7 @@

### T-CTX-06 「会话内只算一次」的上下文实际是每个 CLI 进程算一次;桌面端切换 effort、模型或进出 bypass 模式都会重启 CLI,git 状态、指令文件、日期全部重算,前缀缓存失配,钩子输出和嵌套 CLAUDE.md 重复追加

- 状态:部分修复(#270):恢复时按历史去重嵌套 CLAUDE.md,SessionStart(resume) 输出与上一份相同时不再追加。git 状态、指令文件、日期改为写进历史的附件另开 issue。
- 类型:问题
- 严重程度:中
- 体量:重启后第一个请求的整段 system 与历史全部按未命中计费(system 约 2.6 万字符加历史,体量见 comparison/request-cache.md Q7);钩子输出视钩子而定,每次重启追加一份
Expand Down Expand Up @@ -166,6 +171,7 @@

### T-CTX-09 `update-config` 技能在清单里写「Claude Code harness」「not Claude」,并让模型用本场景不存在的 Config 工具

- 状态:已修复(#270):描述和正文改为 DreamCoder 与「模型」的说法,删掉 Config 工具那句。
- 类型:问题
- 严重程度:低
- 体量:清单条目截到 250 字符;6 个内置技能清单共 1 410 字符,首轮一次
Expand Down Expand Up @@ -255,6 +261,7 @@

### T-CTX-16 子 agent 继承的上下文不成套:拿到 `MEMORY.md` 索引但没有记忆说明,每个子 agent 再收一份技能清单

- 状态:已修复(#270):子 agent 的指令文件不再带 `MEMORY.md`;技能清单部分未改。
- 类型:优化
- 严重程度:低
- 体量:general-purpose 子 agent 每次启动多约 1 500 字符起的技能清单;`MEMORY.md` 视索引
Expand Down
8 changes: 8 additions & 0 deletions docs/features/harness/prompt/reminders.md
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@

### T-INJ-02 CLI 重启后,除排队消息外的所有中途注入文字都从历史里消失,技能清单整份重发,前缀从第一个附件处失配

- 状态:已修复(#270):transcript 写入所有附件;用户停止后只有这一轮没有因中断结束时才结束 CLI。
- 状态:维护者已定(2026-11-18):transcript 保留所有附件;用户停止后,只有这一轮没有停下时才结束 CLI。
- 类型:问题
- 严重程度:高
Expand Down Expand Up @@ -110,6 +111,7 @@

### T-INJ-06 文件被外部改动的通知:归因写死为「用户或 linter」,要求不告诉用户,每个文件最多 8K 字符

- 状态:已修复(#270):归因改为中性(自己运行的命令、用户或其他会话),去掉「不要告诉用户」,片段上限 2 KB 并提示用 Read 看全文。「本轮自己用 Bash 改的文件不发通知」未做。
- 类型:问题
- 严重程度:中
- 体量:338 字符包装加 diff 片段(上限 8,192 字节,`src/tools/FileEditTool/utils.ts:355`,`395-405`)。每个读过、之后被 Edit/Write 以外的方式改动的文件一次。
Expand All @@ -133,6 +135,7 @@

### T-INJ-07 `/goal` 续跑时模型只收到评估器的一句理由,没有目标原文、预算和无进展检查,也没有续跑上限

- 状态:已修复(#270):续跑消息写明是 /goal 自动续跑、目标原文、评估器理由;同一条用户消息内最多连续 10 次(计数放在查询循环里,压缩不清零),第 10 次要求汇报。另修了评审时发现的问题:评估器判未完成时同时要求停止,`executeHooks` 把停止要求和阻断信息分两次产出,结果 /goal 第一次判未完成就结束这一轮,从未自动续跑;现在 /goal 的 hook 判未完成时只返回阻断信息。评估器另有第三种答案「卡在用户这里」(`blocked: true`):这一轮结束、不续跑,goal 保留。
- 类型:问题
- 严重程度:中
- 体量:每次续跑 55 字符前缀加评估器理由;`/goal` 设定时 `<command-name>/goal</command-name>…<local-command-stdout>Goal set: …</local-command-stdout>` 一次。
Expand Down Expand Up @@ -161,6 +164,7 @@

### T-INJ-09 plan 模式首条提醒 4.7K 字符,要求最多 3 个 Explore 和至少 1 个 Plan 子 agent

- 状态:已修复(#270):随 T-TOOL-09 删除 5 阶段流程,不再要求 Explore/Plan 子 agent。
- 类型:优化
- 严重程度:低
- 体量:完整版 4,719 字符(不含计划文件路径),之后每 5 个用户轮发简版 312 字符,每第 5 次再发完整版(`attachments.ts:256-259`,`1187-1253`);子 agent 版 955 字符;重新进入 990 字符;退出 137 字符。
Expand All @@ -175,6 +179,7 @@

### T-INJ-10 技能目录有任何变化就整份重发技能清单,而不是只发新增的

- 状态:已修复(#270):同 T-CTX-03,目录变化只发新增技能。
- 类型:问题
- 严重程度:低
- 体量:每次变化一份完整清单(上限同 T-INJ-02),头部固定 103 字符
Expand All @@ -201,6 +206,7 @@

### T-INJ-12 跨会话和看板消息在一轮中途送达时被并进无关的工具结果,每条都重复约 300 字符的固定说明

- 状态:已修复(#270):非 Claude 模型的中途文字保留为独立文本块(Claude 模型照旧拼接);信封里的权限说明缩短为一句,还原改为按标签解析。
- 类型:优化
- 严重程度:低
- 体量:信封加固定说明,session 消息 584 字符,brief 808,report 561;看板按用途 440~772(实测,名称和 id 用短样例)。每条消息一次。
Expand All @@ -216,6 +222,7 @@

### T-INJ-13 压缩摘要让模型去读 JSONL 格式的完整 transcript,文件通常超过 Read 的 256KB 上限

- 状态:已修复(#270):改为用 Grep 搜索 transcript。
- 类型:问题
- 严重程度:低
- 体量:每次压缩一次,325 字符包装(自动压缩时 606,含「Continue the conversation ... without asking」那段)
Expand Down Expand Up @@ -264,6 +271,7 @@

### T-INJ-17 恢复会话时插入「Continue from where you left off.」和假的 assistant「No response requested.」,但不会自动续跑

- 状态:已修复(#270):恢复时不再插入续跑消息和假回复;只有设置了自动续跑时才补发。
- 类型:问题
- 严重程度:低
- 体量:33 + 23 字符,每次 CLI 重启一组(与 T-INJ-02 叠加)
Expand Down
6 changes: 6 additions & 0 deletions docs/features/harness/prompt/system.md
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,7 @@

### T-SYS-03 外部构建没有「完成前先验证」和「如实汇报结果」,这两条只对 ant 打开

- 状态:已修复(#270):system 提示词和工具描述里 `USER_TYPE === 'ant'` 的文字分支已删除;ant 专用的行为代码未动。
- 状态:维护者已定(2026-11-18):这些条文不对外部构建打开,`USER_TYPE === 'ant'` 的文字分支直接删除。
- 类型:缺失
- 严重程度:高
Expand All @@ -130,6 +131,7 @@

### T-SYS-04 Environment 段带三行 Claude 专属文字(540 字符),其中两行在任何模型下都与 DreamCoder 不符,且自相矛盾

- 状态:已修复(#270):三行 Claude 文字对所有模型删除;由桌面服务器启动时(`DREAMCODER_DESKTOP_HOST=1`)加一行 DreamCoder 宿主说明。
- 类型:问题
- 严重程度:中
- 体量:正文 540 字符,含列表符与换行 552,每个请求。
Expand All @@ -145,6 +147,7 @@

### T-SYS-05 memory 说明让模型「直接用 Write 写」,但桌面端 memory 目录在配置目录里,每次写都要用户批准,bypass 模式也一样

- 状态:已修复(#270):记忆写入豁免覆盖服务器指定的记忆路径(`src/utils/permissions/filesystem.ts`)。
- 状态:维护者已定(2026-11-18):扩展 CLI 的记忆写入豁免,让它也覆盖服务器指定的记忆路径(同 T-CTX-02)。
- 类型:问题
- 严重程度:中
Expand Down Expand Up @@ -219,6 +222,7 @@

### T-SYS-10 `/help` 指向 Claude Code,反馈句只剩半句

- 状态:已修复(#270):`/help` 说明和反馈地址改为 DreamCoder 的。
- 类型:问题
- 严重程度:低
- 体量:163 字符,每个请求。
Expand All @@ -230,6 +234,7 @@

### T-SYS-11 关于输出怎么呈现的描述与桌面端不符

- 状态:已修复(#270):去掉等宽字体、CommonMark、可点击链接的承诺和 Anthropic 示例;冒号规则本身留给 T-SYS-06。
- 类型:问题
- 严重程度:低
- 体量:约 450 字符,每个请求。
Expand All @@ -244,6 +249,7 @@

### T-SYS-12 statusline-setup 子 agent 的 7K 正文针对 DreamCoder 没有的状态栏

- 状态:已修复(#270):statusline-setup 子 agent 和 `/statusline` 命令已删除。
- 类型:问题
- 严重程度:低
- 体量:正文 6 993 字符,合计约 7.9K,每次调用发一次。
Expand Down
6 changes: 6 additions & 0 deletions docs/features/harness/prompt/tools.md
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,7 @@

### T-TOOL-04 Read 描述说模型能看图、能读 PDF,并称「Claude Code is a multimodal LLM」;经代理时图片和 PDF 都到不了模型

- 状态:已修复(#270):经代理时描述写明图片和 PDF 内容到不了模型;直连时保留图片说明,去掉「Claude Code」。代理侧见 P-REQ-03/04。
- 类型:问题
- 严重程度:中
- 体量:图片句 165 字符、PDF 句 253 字符、截图句 323 字符,每个请求都发。
Expand Down Expand Up @@ -184,6 +185,7 @@

### T-TOOL-09 计划模式有三份流程说明,EnterPlanMode 的结果说「不要写任何文件」,计划模式提醒又要求写计划文件

- 状态:已修复(#270):只保留计划模式提醒里的 interview 流程;EnterPlanMode 描述删掉「What Happens」,结果只允许写计划文件。
- 类型:问题
- 严重程度:中
- 体量:EnterPlanMode 描述「What Happens in Plan Mode」370 字符(每个请求);EnterPlanMode 结果 472 字符(每次进入);计划模式提醒(T-INJ 范围)。
Expand Down Expand Up @@ -213,6 +215,7 @@

### T-TOOL-11 AskUserQuestion 的 schema 暴露三个内部字段,以及桌面端不渲染的 `preview`;`preview` 的描述指向描述里不存在的段落

- 状态:已修复(#270):发给模型的 schema 去掉 `answers`、`annotations`、`metadata` 和选项的 `preview`,内部 schema 不变。
- 类型:问题
- 严重程度:中
- 体量:schema 2,966 字符,其中 `answers`/`annotations`/`metadata` 955 字符、`preview` 241 字符;工具标 D,本场景每个请求都发。
Expand Down Expand Up @@ -241,6 +244,7 @@

### T-TOOL-13 EnterWorktree/ExitWorktree 与桌面端自己的 worktree 流程重复,文字里有 hook、tmux、退出时提示等桌面端没有的行为

- 状态:已修复(#270):描述去掉 hook、tmux 和退出提示,写明桌面端面板仍显示原目录。核实结果:会话中途进入 worktree 后桌面端面板不跟随(CLI 只在 init 时报告 cwd),未修。
- 状态:维护者已定(2026-11-18):保留这两个工具;按桌面端的实际行为改描述,并核实桌面端面板是否跟随 CLI 进入 worktree。
- 类型:问题
- 严重程度:中
Expand Down Expand Up @@ -297,6 +301,7 @@

### T-TOOL-17 Agent 描述里约 1.2K 字符的示例和 agent 条目对本场景无用:示例 agent 不存在,statusline-setup 配置的是桌面端没有的界面

- 状态:部分修复(#270):statusline-setup 已从内置列表删除(同 T-SYS-12);示例段和 build-validator 句未改。
- 类型:优化
- 严重程度:低
- 体量:示例段 1,002 字符、statusline-setup 一行约 110 字符、「build-validator / test-runner」一句约 150 字符,每个请求都发。
Expand Down Expand Up @@ -341,6 +346,7 @@

### T-TOOL-21 缺少「编辑后不必重读验证」的说明;Edit 的「最小 old_string」提示只给 ant

- 状态:部分修复(#270):ant 专用的「最小 old_string」提示已删除;「编辑后不必重读」未定。
- 状态:维护者已定(2026-11-18):「最小 old_string」这条 ant 专用提示随 ant 分支一起删除;「编辑后不必重读」未定。
- 类型:缺失
- 严重程度:低
Expand Down
Loading
Loading