Summary

A botched update.run (npm upgrade triggered via the gateway) left
the package half-installed. The post-restart maintenance step then
tried to load a module that no longer existed at the path the new
build expected, crashed, and a live Telegram session lost its
<sessionId>.jsonl transcript. Every subsequent message to that
Telegram chat then hit claude --resume <id> against a missing
file, failed in ~400 ms, and surfaced as
"Something went wrong while processing your request" to the user.

Version

• openclaw 2026.5.2 (8b2a6e5), macOS, Node v25.9.0 via nvm

Reproduction (observed)

1. Gateway running with active Telegram session.

2. update.run initiated (npm package upgrade) by openclaw-control-ui.

3. Module resolution failed mid-restart:

   [gateway] request handler failed: Error: Cannot find module
     '/opt/homebrew/lib/node_modules/openclaw/dist/task-registry.maintenance-DuW0FRWY.js'
     imported from .../dist/status.summary-D7d6QRTx.js
   [gateway] shutdown error: ERR_MODULE_NOT_FOUND ... server-close-D1yUo6cN.js
   

4. After restart, every Telegram turn fails with FailoverError in
   ~400 ms (both Opus 4.7 and Sonnet 4.6), because the session's
   transcript file no longer exists.

5. openclaw sessions cleanup --enforce --fix-missing confirmed the
   record pointed at a missing transcript and pruned it.

Two underlying bugs

Bug 1 — update.run is not atomic

A failed/in-progress install leaves modules referenced by sibling
modules at hashed paths that no longer match. The gateway should
download → verify all module imports resolve → swap atomically.
Right now a partial install is reachable by the live process and
crashes the maintenance routine.

Bug 2 — session-store entries for shell-only cron jobs

systemEvent run: cron jobs (pure shell, no LLM turn) get a row in
sessions.json and a .trajectory-path.json + .trajectory.jsonl,
but never a <id>.jsonl. openclaw doctor then reports them as
"missing transcripts" forever, and cleanup --fix-missing prunes
them — but they should never have been registered as resumable
sessions in the first place. Example session ids that exhibit this:
a1287341-..., 04ab26a5-..., cb3b51f9-... (all cron-driven
systemEvent jobs).

Bug 3 (nice-to-have) — auto-recover from missing transcripts

When claude --resume <id> fails because the transcript file is
gone, the gateway loops forever on the same broken session record.
It should auto-prune the record and start a fresh session for that
sessionKey instead of failing every turn for the user.

Workaround

• Run openclaw sessions cleanup --enforce --fix-missing
• Or schedule it daily via openclaw cron add --system-event 'run: ...'

Logs (trimmed)

14:05:50 [gateway] request handler failed: ERR_MODULE_NOT_FOUND task-registry.maintenance
14:05:51 [gateway] shutdown error: ERR_MODULE_NOT_FOUND server-close
14:10:23 [diagnostic] stuck session: sessionKey=agent:main:telegram:direct:<redacted>
         age=135s reason=queued_work_without_active_run
14:52:43 claude live session turn failed: provider=claude-cli ... durationMs=300011 error=FailoverError
15:10:22 claude live session turn failed: durationMs=384 error=FailoverError (immediate)