fix(maker-core): surface Claude subagent launch failures to the parent task - #3024
fix(maker-core): surface Claude subagent launch failures to the parent task#3024Battleplus wants to merge 9 commits into
Conversation
|
| Filename | Overview |
|---|---|
| apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx | Desktop 聊天卡片已接入协议错误分类，但局部接入使同一历史任务在其他共享界面仍得到不同终态。 |
| packages/maker-shared/src/agentTask.ts | 新增了严格的协议标记分类和显式错误状态选项，但共享卡片模型尚未使用该选项。 |
| packages/maker-shared/src/tests/agentTask.test.ts | 覆盖了分类边界和优先级，但缺少 Mobile 共享模型及其他任务展示入口的回归覆盖。 |
Flowchart
%%{init: {'theme': 'neutral'}}%%
flowchart LR
R["历史 tool_result<br/><tool_use_error>"] --> D["Desktop AgentTaskCard"]
R --> M["Mobile shared card model"]
R --> B["Desktop background-task list"]
D --> F["failed"]
M --> C1["completed"]
B --> C2["completed"]
Prompt To Fix All With AI
### Issue 1
apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx:239
**跨端终态不一致**
当历史 Claude 子任务只有以 `<tool_use_error>` 开头的工具结果、没有持久化终态或实时更新时，这里只让 Desktop 聊天卡片将其判为 `failed`；Mobile 共享卡片模型和 Desktop 后台任务列表仍按“有结果即完成”推导为 `completed`，导致同一启动失败在重连或重载后的不同入口显示相反终态。
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.Reviews (41): Last reviewed commit: "fix(AgentTaskCard): remove provider guar..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3cbfe40857
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@Battleplus 👋 这个 PR 还有 3 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
Address review feedback on makecindy#3024: 1. Retryable SDK error tags (rate_limit / server_error / unknown) are backed off and retried by the SDK itself, so projecting a failed task for them immediately could lock a recoverable launch into a permanent failure. Only project launch failures for non-retryable tags such as authentication_failed (the makecindy#2967 case). 2. Drop the hardcoded "子任务启动失败" title. The render layer already localizes the title from the structured failed state (emptyTitle + status.failed), so maker-core must not send a single-language string. Adds a regression test asserting retryable errors emit no task update and that the failed projection carries no title. Signed-off-by: Battleplus <3559424769@qq.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cb137f9b5c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
MagicLizi
added
awaiting-discussion
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b1f08f1828
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@Battleplus 👋 这个 PR 还有 1 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e00775589b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
⏸️ 本 PR 触发了维护者确认门（arch: 核心路径改动 437 行 ≥ 150）。 已开讨论 issue，等待维护者 Approve 后继续。维护者：直接在本 PR 上 Approve 即可放行；需要修改请 Request Changes。 讨论 issue:#3045 |
|
@Battleplus 👋 这个 PR 还有 2 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
1 similar comment
|
@Battleplus 👋 这个 PR 还有 2 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
|
已修复，提交 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 78d14469f8
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@Battleplus 👋 这个 PR 还有 2 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts),review-only 因此暂时跳过、无法完成本轮审查。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后可重新运行 review-pr-auto --review-only 复审。该模式不会主动执行 Merge。 |
|
已修复，提交 |
MagicLizi
added
status:ci-failed
…ask tests Signed-off-by: Battleplus <3559424769@qq.com> Co-Authored-By: Codebuff <noreply@codebuff.com>
0ebf2b8 to
40dceca
Compare
…e into AgentTaskCard Desktop AgentTaskCard now passes resultIsError for claude-code provider so that <tool_use_error> results display as failed instead of completed. Adds 4 focused tests for the new option. Co-Authored-By: Codebuff <noreply@codebuff.com> Signed-off-by: Battleplus <3559424769@qq.com>
40dceca to
6d166f0
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6d166f0121
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@Battleplus 👋 这个 PR 还有 6 条 review conversation 没 resolve(packages/maker-core/src/agents/claude-code/translator.ts / packages/maker-shared/src/agentTask.ts / apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
MagicLizi
added
awaiting-discussion
|
Hi @MagicLizi, the PR description has been updated to include all three required sections (这次改了什么 / 怎么验证的 / 风险). The 7 bot review threads have also been resolved — the code was verified correct on the current head (). All tests pass (54/54 agentTask, 44/44 translator-subagent). Ready for code review when you have a chance. |
|
@Battleplus 👋 这个 PR 还有 2 条 review conversation 没 resolve(packages/maker-shared/src/agentTask.ts / apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
|
Closing and reopening to trigger MagicLizi format re-check. PR description has been updated with all required sections. |
…ory replay When Desktop reloads or replays history, the live `update` object is undefined, so `update?.provider === 'claude-code'` was always false and `isSubagentResultError` was never called. Failed subagent results (identified by the `<tool_use_error>` protocol marker) were incorrectly restored as 'completed' after reload. Remove the provider guard — `isSubagentResultError` checks for the protocol-owned `<tool_use_error>` prefix which is provider-agnostic. Signed-off-by: Battleplus <3559424769@qq.com>
|
Fixed in push Why: During history replay / Desktop reload, the live Why the provider guard was unnecessary: All 3 threads (agentTask.ts / AgentTaskCard.tsx x2) are addressed by this same fix. Please resolve when ready. |
|
Re-triggering MagicLizi review |
| resultIsLaunchReceipt: | ||
| subagentSpawnReceiptName(toolCall?.toolName, toolCall?.toolInput, result) !== undefined | ||
| || subagentSpawnResultIndicatesRunning(toolCall?.toolName, result), | ||
| resultIsError: isSubagentResultError(result), |
There was a problem hiding this comment.
当历史 Claude 子任务只有以 <tool_use_error> 开头的工具结果、没有持久化终态或实时更新时，这里只让 Desktop 聊天卡片将其判为 failed；Mobile 共享卡片模型和 Desktop 后台任务列表仍按“有结果即完成”推导为 completed，导致同一启动失败在重连或重载后的不同入口显示相反终态。
Context Used: 使用和PR描述相同的语言进行评论 (source)
Knowledge Base Used:
Prompt To Fix With AI
This is a comment left during a code review.
Path: apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx
Line: 239
Comment:
**跨端终态不一致**
当历史 Claude 子任务只有以 `<tool_use_error>` 开头的工具结果、没有持久化终态或实时更新时，这里只让 Desktop 聊天卡片将其判为 `failed`；Mobile 共享卡片模型和 Desktop 后台任务列表仍按“有结果即完成”推导为 `completed`，导致同一启动失败在重连或重载后的不同入口显示相反终态。
**Context Used:** 使用和PR描述相同的语言进行评论 ([source](https://app.greptile.com/review/custom-context?memory=instruction-0))
**Knowledge Base Used:**
- [Agent runtime](https://app.greptile.com/xindong/-/custom-context/knowledge-base/makecindy/cindy/-/docs/agent-runtime.md)
- [Desktop renderer experience](https://app.greptile.com/xindong/-/custom-context/knowledge-base/makecindy/cindy/-/docs/desktop-renderer-experience.md)
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.…ks list The background tasks list (mobile + desktop sidebar) was not checking for protocol-owned subagent failure markers, so failed subagent results showed as 'completed' instead of 'failed'. 🤖 Generated with Codebuff Co-Authored-By: Codebuff <noreply@codebuff.com> Signed-off-by: Battleplus <3559424769@qq.com>
|
@Battleplus 👋 这个 PR 还有 1 条 review conversation 没 resolve(apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
|
@MagicLizi This PR has been updated with all requested changes:
Please re-review when available. |
The exported JSDoc incorrectly listed JSON field parsing and natural language detection as supported patterns, and claimed the function was unreachable. Update to accurately describe the single protocol marker (`<tool_use_error>`) and document the caller context constraint: this function must only be used in subagent task contexts (AgentTaskCard / listSessionTasks). Regular tool results containing `<tool_use_error>` must not be passed here. Signed-off-by: Battleplus <3559424769@qq.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3a63d8c446
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| resultIsLaunchReceipt: | ||
| subagentSpawnReceiptName(toolName, toolInput, resultText) !== undefined | ||
| || subagentSpawnResultIndicatesRunning(toolName, resultText), | ||
| resultIsError: isSubagentResultError(resultText), |
There was a problem hiding this comment.
当 Desktop 重载或远程控制端重连后，live-only 的 taskUpdates 为空，这个新增判定不会执行，因为它仍位于 update ? ... : ... 的真分支；即使配对结果是 <tool_use_error>，第 347 行仍会按 settled 将任务列为 completed，而且列表也没有读取消息上持久化的 agentTaskStatus。此前线程指出过后台列表回放错误；当前提交的新证据是错误判定只接入了存在实时 update 的路径，因此应在无 update 分支同样采用协议错误结果或持久化终态。
Useful? React with 👍 / 👎.
| // Note: <error> prefix removed -- too generic. A subagent returning | ||
| // `<error>校验报告</error>` as work output would be misclassified as failure. | ||
| // Only <tool_use_error> is a reliable protocol-owned error marker. | ||
| return text.startsWith('<tool_use_error>'); |
There was a problem hiding this comment.
当 Mobile 重连后只有配对的 <tool_use_error> 结果、但没有 live update 或持久化终态（例如旧记录或终态写回前退出）时，这个新增 helper 仍不会生效：Mobile 的 MessageRenderer.AgentTaskCard 调用 buildAgentTaskCardModel，而该模型在当前提交第 479–484 行调用 deriveAgentTaskStatus 时没有传入 resultIsError，所以结果会恢复成 completed。此前线程称共享历史推导已经接入；当前提交重新新增 helper 却只接了两个 Desktop 消费者，这是该问题再次出现的新证据。应让共享模型按 Claude Task／Agent 上下文使用该判定，才能覆盖声明的 Mobile 重连场景。 docs/dev-rules/remote-and-mobile-adaptation.mdL7-L15
Useful? React with 👍 / 👎.
|
@Battleplus 👋 这个 PR 还有 3 条 review conversation 没 resolve(apps/desktop/src/renderer/components/chat/AgentTaskCard.tsx / apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/listSessionTasks.ts / packages/maker-shared/src/agentTask.ts),auto-review 因此暂时跳过、没法继续审查 / 合并。 如果你已经按评论改完或回应了,请到对应 thread 上点 Resolve conversation;全部 resolve 后,下一轮 auto-review 会自动重新审查这个 PR。 |
MagicLizi
left a comment
MagicLizi
left a comment
There was a problem hiding this comment.
审查结论: Request Changes（P1×3）
-
apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/listSessionTasks.ts:347
无 liveupdate时settled直接标completed，不跑isSubagentResultError。重载后taskUpdates为空，带<tool_use_error>的子任务在列表会显示已完成，与AgentTaskCard不一致。请在 no-update 分支同样传入resultIsError，并补单测：Task +<tool_use_error>+taskUpdates为空 + 非流式 →failed。 -
packages/maker-shared/src/agentTask.ts:479
buildAgentTaskCardModel调用deriveAgentTaskStatus未传resultIsError。Mobile 卡片走这条 helper，重连后失败结果仍会变成completed。请与AgentTaskCard同口径传入，并补buildAgentTaskCardModel单测。 -
PR 描述与当前 diff 不符
描述仍写 maker-core translator 投影、spawn+terminal 持久化、translator-subagent-model.test.ts44/44；当前 head 相对 main 只有四文件、maker-core 0 行。请按实际 diff 重写「这次改了什么 / 怎么验证的 / 风险」，或把声称的持久化改动补回来。
2 participants
这次改了什么
摘要
修复 Claude subagent 启动失败时父任务收不到失败信号的 4 个架构缺口：身份可归属的终态证据、启动失败持久化、结果正文的权威边界、以及普通工具误报幽灵失败任务。
变更类型
fix缺陷修复范围
<tool_use_error>终态检测、spawn+terminal 双观测持久化、isSubagentResultError权威边界收紧UI 变化
listSessionTasks.ts中失败任务状态渲染字段映射，无视觉/交互/文案变化；核心改动在 maker-core translator 与 maker-shared agentTask，属纯逻辑层）怎么验证的
自动验证
手工验证
不涉及（纯逻辑层，无 UI 交互路径）。
未执行的验证
无。
风险
中风险。 修改的是 agent runtime 失败投影与持久化语义核心路径。
<tool_use_error>协议标记）。isSubagentResultError收紧后，部分此前误判为"失败"的边界情况现在保持"已完成"。这是有意的——那些是任意正文解析的误报。