【Bug】hiclaw-v1.1.1，worker的Matrix accessToken被file-sync同步覆盖后导致Worker永久断联，仅能通过容器重启恢复 || [Bug] hiclaw-v1.1.1, the worker's Matrix accessToken is overwritten by file-sync synchronization, causing the worker to be permanently disconnected, and can only be restored by restarting the container

[Thassarian-Hitagi]

Bug Description

只要worker重启过一次（token更新），后续worker只要执行文件同步（file-sync），就会100%断连。

。

Steps to Reproduce

1.当worker只要重启过一次，再根据任务要求执行file-sync的时候，会把minio中的openclaw.json（包含初始accessToken，此时已失效）同步到本地。
2.文件同步完成后，Worker仍可以正常发送至少1条消息（Matrix长连接会话令牌缓存有效期30120秒，此期间不需要重新校验Token）
3.等待30120秒（到Matrix客户端默认心跳周期），Matrix连接被服务器主动断开，返回401/403认证错误
4.Matrix插件自动重连，一直使用已经被覆盖的无效旧Token，重连永久失败，Worker完全断联，仅能通过重启容器恢复

AI Analysis

根本原因：
1.Hiclaw Gateway内置工作区状态持久化机制：worker的本地文件自动同步到minio时，不包含工作区的openclaw.json （没问题，防止worker自己不小心改坏）
2.minio中的openclaw.json，文件中channels.matrix.accessToken是最初创建的时候生成的。
3.worker容器重启时，matrix分配的accessToken会更新

确认容器重启会重新生成token、确认minio中token与实际生效的不一致

Relevant Logs

Component

OpenClaw Worker

Version / Commit

No response

---

Bug Description

As long as the worker is restarted once (token update), subsequent workers will be 100% disconnected as long as they perform file synchronization (file-sync).

.

Steps to Reproduce

1. When the worker only needs to be restarted once and then executes file-sync according to the task requirements, the openclaw.json in minio (including the initial accessToken, which is invalid at this time) will be synchronized to the local.
2. After file synchronization is completed, Worker can still send at least 1 message normally (Matrix long connection session token cache is valid for 30~120 seconds, and there is no need to re-verify the token during this period)
3. Wait for 30~120 seconds (to the default heartbeat period of the Matrix client). The Matrix connection is actively disconnected by the server and a 401/403 authentication error is returned.
4. The Matrix plug-in automatically reconnects and always uses the invalid old token that has been overwritten. The reconnection fails permanently and the worker is completely disconnected and can only be restored by restarting the container.

AI Analysis

Root cause:

1. Hiclaw Gateway has a built-in workspace state persistence mechanism: when the worker's local files are automatically synchronized to minio, the openclaw.json of the workspace is not included (no problem, to prevent the worker from accidentally changing it)
2. In openclaw.json in minio, the channels.matrix.accessToken in the file is generated when it is initially created.
3. When the worker container is restarted, the accessToken assigned by matrix will be updated.

Confirm that restarting the container will regenerate the token, and confirm that the token in minio is inconsistent with the actual one.

Relevant Logs

Component

OpenClaw Worker

Version/Commit

No response

[Thassarian-Hitagi]

临时处理：每次worker重启、稳定之后，让worker把自己的openclaw.json推送到minio可以缓解这个问题。

另外doubao2.0pro的agent说，配置环境变量+ 修改本地的extensions/matrix/src/connection.ts + extensions/matrix/src/index.ts 能修正本地问题，不过这个感觉很不不靠谱，还没尝试过。

根治，好像需要各位大佬去改hiclaw/openclaw-base镜像？

[shiyiyue1102]

Matrix accessToken 被 file-sync 覆盖后 Worker 断联，这个比较关键。@Sunrisea 麻烦重点看下凭据文件和同步目录的边界，避免用户只能靠重启恢复。

---

The Worker is disconnected after the Matrix accessToken is overwritten by file-sync. This is critical. @Sunrisea Please focus on checking the boundaries between the credential file and the synchronization directory to prevent users from having to rely on restarting to recover.

[Sunrisea]

看了下，复现路径没问题，根因就是：worker 重启会重新 login 拿新 token 写回本地 openclaw.json，但本地往 MinIO 同步时 openclaw.json 是被 exclude 的，所以 MinIO 一直是最早那个旧 token，file-sync 再把它拉回来盖掉本地的新的。

但是我看代码这块其实 #519（v1.0.9）就把 file-sync 从直接覆盖改成本地优先合并了，#715（v1.1.1）又把 accessToken 单独拎出来强制以本地为准。正常情况下 v1.1.1 上不管是 file-sync 还是后台兜底同步，都应该保住本地新 token 才对。

所以这里有个点想让你确认下。worker 实际跑的那份 hiclaw-sync.sh 不在 worker 镜像里，是放在这个 worker 的 MinIO 工作区（agents//skills/file-sync/scripts/hiclaw-sync.sh），由 controller/Manager 推上去的，worker 镜像只是个 wrapper 调它。我怀疑你这边 MinIO 里那份脚本还是老版本（#519 之前那种直接 mc mirror --overwrite、不走合并的），所以保护逻辑根本没生效——这种一般是各组件版本没拉齐，或者存量 worker 的脚本副本没跟着升级刷新。当然这只是猜测，得你这边核实下。

麻烦帮忙看两个东西：

1. controller / Manager / worker 三个镜像各自的 tag，确认是不是都到 v1.1.1+ 了，别只看一个整体版本号；
2. 出问题的 worker 上 cat 一下 ~/hiclaw-fs/agents//skills/file-sync/scripts/hiclaw-sync.sh，看开头有没有 source merge-openclaw-config.sh、有没有先把本地备份到 /tmp/openclaw-local-sync.json 再 merge 那段。有就是新版，只有一行 mc mirror --overwrite 就是老版。

如果 controller/Manager 版本没跟上，建议整套升到最新，让 controller reconcile / Manager 的 upgrade-builtins 把存量 worker 的脚本刷成新版，然后把受影响的 worker 重启一次（触发重新 login）就能恢复。MinIO 里那个旧 openclaw.json 不用管，新版合并会以本地为准。

把 accessToken 拆到 credentials 那套更彻底，但工作量比较大，建议先走镜像升级止血。

要不核对下版本和脚本？

[Thassarian-Hitagi]

目前是v1.1.2版本，包括woker的镜像也是1.1.2
截止目前这个问题还是存在的

[Thassarian-Hitagi]

仔细看了下manager自身是1.1.1，有点怪，会跟这个有关吗？有办法单独调整manager的镜像吗。。

[Thassarian-Hitagi]

合并脚本检查.txt

这两个脚本都存在，而且确实有token本地优先的逻辑，不过没生效。
/opt/hiclaw/scripts/lib/merge-openclaw-config.sh

skills/file-sync/scripts/hiclaw-sync.sh
本地脚本详情见附件。

同步文件前，worker状态正常时：

worker执行一次文件同步后：

[RerankerGuo]

Opened PR #983 for this: #983

The likely failure window is that hiclaw-sync bulk-mirrors the remote worker directory over the local workspace before merging openclaw.json, so OpenClaw can briefly see the stale MinIO Matrix accessToken. The PR changes file-sync to fetch remote openclaw.json into /tmp, exclude openclaw.json from the bulk mirror, and then merge the temporary remote copy into the local config afterward.

Verified locally with:

• bash -n manager/agent/worker-agent/skills/file-sync/scripts/hiclaw-sync.sh tests/test-17-worker-config-verify.sh
• git diff --check
• static check that the pushed file-sync script contains --exclude "openclaw.json"