Sandbox takeover: the saved transcript loses everything the first host delivered before the restart
维护者通常 1 天内回复
@tombeckenham 已经在做这个了。
开始于 2026年11月5日。
评估
这个 Issue 还没有评估数据。
描述
TanStack AI version
@tanstack/ai-sandbox 1.1.0 and main (644a780b0). @tanstack/ai 1.1.0 and main. @tanstack/ai-persistence 1.0.1. @tanstack/ai-claude-code 1.0.1.
Framework/Library version
Server only. chat() driven by sandboxRunDriver in a Node 24 route handler. No client framework is involved.
Describe the bug and the steps to reproduce it
A host restarted during a 40-second Bash call. sandboxRunDriver took over the run on another host and finished it. The saved transcript has the Bash tool result and the final text, but no assistant message with the Bash tool call. The live stream was correct, because the event log already held the first part of the run.
The transcript loses more than that one call. The successor's transcript is missing every chunk that the first host delivered before the restart: assistant text, complete tool calls, and the start of a tool call that was still running. A call that was still running leaves its result as an orphan role: 'tool' message with no call.
Expected: the saved transcript after a takeover is the same as the transcript of a run that did not restart.
Steps:
- Use
claudeCodeTextwithwithPersistenceandwithSandbox(sandbox, { instances, runs, durability }). - Start a run where the agent writes some text and then runs a slow tool.
- Stop the host while the tool runs. Let
sandboxRunDrivertake over the run on another host. - When the run finishes, load the thread from the message store. It holds only the chunks that the successor forwarded.
Suspected cause. On an attach, the harness adapter wraps its replay in alignedIfAttaching. alignToStoredLog suppresses the prefix that the event log already holds and yields only the rest. The wrap is inside the adapter's chatStream, so the suppressed chunks never reach the engine:
- The engine does not see the text of the prefix, so it does not add that text to the transcript.
withSandbox'sonChunkdoes not see theTOOL_CALL_START,TOOL_CALL_ARGSandTOOL_CALL_ENDof the prefix, so the tool-history recorder writes no call. The laterTOOL_CALL_RESULTarrives, and the recorder writes the result alone.
A comment in appendCall says that "a journal replay on takeover re-emits the whole stream". After alignment, that is not true.
Other effects. stripObservedToolCalls removes a result only when it also removed the result's call. So an orphan result stays in the model request of the next turn. claudeCodeText is not affected, because buildPrompt keeps only user and assistant text. I did not check the other harness adapters.
Not checked: whether withPersistence's snapshotStreaming option saves the first host's partial transcript in a form that the successor merges back.
Related: #2035 is a different takeover bug, where the reconnected client receives the rest of the run twice.
Your Minimal, Reproducible Example - (Sandbox Highly Recommended)
https://gist.github.com/adelrodriguez/4e34cfd7224b96dfe05d554e4c0dd1e0
The test runs one chat() with withPersistence(memoryPersistence()) and withSandbox(..., { runs, durability: { adapter: fakeLog(prefix), attach: true } }). Its adapter does what claudeCodeText does: it wraps a fixed replay in alignedIfAttaching. The log already holds the chunks up to the Bash call's TOOL_CALL_END, including a first text message. Place the file in packages/ai-sandbox/tests/ and run vitest run tests/takeover-transcript.repro.test.ts in packages/ai-sandbox. Both tests fail on main:
forwarded to engine: TOOL_CALL_RESULT, TEXT_MESSAGE_START, TEXT_MESSAGE_CONTENT, TEXT_MESSAGE_END, RUN_FINISHED
saved transcript:
{ "role": "user", "content": "Run the slow job" }
{ "role": "tool", "toolCallId": "bash-1", "content": "done" }
{ "role": "assistant", "content": "Finished.", ... }
× keeps the assistant tool call, not only its result
× keeps the text the first host streamed before the restart
AssertionError: expected 'Run the slow job done Finished.' to contain 'Starting the job.'
Tests 2 failed (2)
Do you intend to try to help solve this bug with your own PR?
Maybe, I'll investigate and start debugging
Terms & Code of Conduct
- I agree to follow this project's Code of Conduct
- I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.
🤖 Created with Claude Opus 5.5 in T3 Code
- 主要语言
- TypeScript
- 星标
- 3.3k
- 派生
- 378
- 平均合并
- 1 天 3 小时
- 30 天内合并 PR
- 253
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
TanStack/ai 的其他 Issue
-
openai-base: a Responses max_output_tokens cutoff ends as RUN_ERROR without usage instead of finishReason 'length'可能已有人在做 @tombeckenham 于 8 天前认领。 未关闭has-pr waiting-on: maintainer
难度 2/5 1-3 小时 新手友好度 74/100
TanStack/ai#1916 · 1 条评论 · 已指派 1 人 ·
维护者通常 1 天内回复
-
feat(ai-anthropic): support native on-demand compaction可能已有人在做 @tombeckenham 今天认领。 未关闭waiting-on: maintainer
维护者通常 1 天内回复
-
@tanstack/openai-base: strict structured output skips $defs/definitions, so recursive zod schemas are rejected可能已有人在做 @jherr 今天认领。 未关闭has-pr waiting-on: maintainer
维护者通常 1 天内回复
-
Approval resume validates null-widened history args, so approved tools fail with repairToolInput: false可能已有人在做 @tombeckenham 今天认领。 未关闭waiting-on: maintainer
维护者通常 1 天内回复
-
ai-sandbox: encrypt snapshot blobs at rest with an encrypting blob store wrapper可能已有人在做 @AlemTuzlak 于 1 天前认领。 未关闭waiting-on: maintainer
维护者通常 1 天内回复
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
binding:python bug
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 1 天内回复
-
enhancement
难度 2/5 1-3 小时 新手友好度 78/100
coollabsio/orbit#72 ·
维护者通常 1 天内回复
-
Bug
难度 2/5 1-3 小时 新手友好度 67/100
维护者通常 1 天内回复
-
bug needs-triage potential-regression
难度 2/5 1-3 小时 新手友好度 78/100
aws/aws-sdk-js-v3#8408 ·
维护者通常 1 天内回复