Cloud routine (CCR) sessions permanently hang at startup, 100% reproducible over a month

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

I created an automation that looks at my google calendar and mail the upcoming items for the next week.
But it doesn't work. Heres what Claude says.

Environment: env_0134X8nTfAoDHaPiexFcF2XP (Default, anthropic_cloud)
Account: Johan (627f72d0-617e-4050-ab89-7ee057b6f4ea)

Symptom: Every CCR session created under this environment stalls permanently right after the log line "Claude Code process started" — no subsequent tool calls, no assistant turn, no output of any kind. worker_status stays idle indefinitely; the run is eventually marked ROUTINE_RUN_STATUS_ABANDONED.

Impact: A weekly scheduled routine ("Weekly Calendar Digest Email", trig_015apipCSfBfr288MEuKjvW5) has failed on every one of ~11 fires over a month (Aug 20 – Sep 13, 2026), including 4 consecutive real scheduled Sunday fires, with zero successful completions.

Isolation performed (root cause proven, not guessed):

Tested with Google Calendar + Gmail MCP connectors attached — hung.
Tested with mcp_connections fully cleared and a trivial "reply with ping" prompt — hung identically, ruling out connectors/auth.
Verified the same Gmail connector works fine when called directly outside a CCR session (local list_labels call succeeded).
Created a brand-new, never-before-used trigger (trig_011QaCA4pAie7zvATftQvQbQ, minimal prompt "reply with pong", default tool config) and ran it — hung at the exact same point, ruling out anything specific to the original trigger's config or history.

Conclusion: The failure is environment/account-level, not caused by routine configuration, connectors, or prompt content. Something in CCR session startup for this account/environment never hands control to the model after "Claude Code process started."

Affected session IDs:
cse_01TciRG6L4qCe6pXnbpKARMy, cse_015UpoDr2djDLQGE3YnMkYR5, cse_01RW7ZjCFDfMagoK1LebJ5hK, cse_01BB65t9r9LSDgDT8J6G2e2y, cse_01XFAGbP71eDsEAtjiLmr9cA, cse_01Eh4v9z9B2hQDLieh83qFM5, cse_014UFz9YmJZwhcMh3ruxJA2k, cse_013GeQVzs1wGyFuP5j8WHTJp, cse_01MbswZwGW7vCsp5NndaWecL, cse_013xBvAMj9rdSDCuwghVmf9m, cse_01Ke3fyxWeP4bYkVKVfHSizg

### What Should Happen?

Can you provide a solution ?

### Error Messages/Logs

```shell
Environment: env_0134X8nTfAoDHaPiexFcF2XP (Default, anthropic_cloud)
Account: Johan (627f72d0-617e-4050-ab89-7ee057b6f4ea)

Symptom: Every CCR session created under this environment stalls permanently right after the log line "Claude Code process started" — no subsequent tool calls, no assistant turn, no output of any kind. worker_status stays idle indefinitely; the run is eventually marked ROUTINE_RUN_STATUS_ABANDONED.

Impact: A weekly scheduled routine ("Weekly Calendar Digest Email", trig_015apipCSfBfr288MEuKjvW5) has failed on every one of ~11 fires over a month (Aug 20 – Sep 13, 2026), including 4 consecutive real scheduled Sunday fires, with zero successful completions.

Isolation performed (root cause proven, not guessed):

Tested with Google Calendar + Gmail MCP connectors attached — hung.
Tested with mcp_connections fully cleared and a trivial "reply with ping" prompt — hung identically, ruling out connectors/auth.
Verified the same Gmail connector works fine when called directly outside a CCR session (local list_labels call succeeded).
Created a brand-new, never-before-used trigger (trig_011QaCA4pAie7zvATftQvQbQ, minimal prompt "reply with pong", default tool config) and ran it — hung at the exact same point, ruling out anything specific to the original trigger's config or history.

Conclusion: The failure is environment/account-level, not caused by routine configuration, connectors, or prompt content. Something in CCR session startup for this account/environment never hands control to the model after "Claude Code process started."

Affected session IDs:
cse_01TciRG6L4qCe6pXnbpKARMy, cse_015UpoDr2djDLQGE3YnMkYR5, cse_01RW7ZjCFDfMagoK1LebJ5hK, cse_01BB65t9r9LSDgDT8J6G2e2y, cse_01XFAGbP71eDsEAtjiLmr9cA, cse_01Eh4v9z9B2hQDLieh83qFM5, cse_014UFz9YmJZwhcMh3ruxJA2k, cse_013GeQVzs1wGyFuP5j8WHTJp, cse_01MbswZwGW7vCsp5NndaWecL, cse_013xBvAMj9rdSDCuwghVmf9m, cse_01Ke3fyxWeP4bYkVKVfHSizg
```

### Steps to Reproduce

te

### Claude Model

Sonnet (default)

### Is this a regression?

I don't know

### Last Working Version

_No response_

### Claude Code Version

2.110.0

### Platform

Anthropic API

### Operating System

Windows

### Terminal/Shell

Other

### Additional Information

_No response_

Start with the CCR session startup path after the "Claude Code process started" log, using the affected environment and a minimal "reply with ping" trigger without MCP connections. Compare worker status and session logs with a successful direct connector call; done means a CCR routine reaches an assistant turn and is no longer abandoned.

Written by the indexing model from the issue text.