[BUG]Cloud routine "run now" consistently stalls before init — session-start hook never resolves

Preflight Checklist

• [x] I have searched existing issues 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?

Summary

Manually firing a scheduled cloud routine (RemoteTrigger action: "run") reliably stalls the resulting session before it ever reaches model init. Reproduced 4/4 times, including with a minimal zero-connector, one-word-prompt configuration, ruling out prompt content and MCP connectors as the cause.

Environment

• Feature: Cloud routines (claude.ai/code/routines), created via the RemoteTrigger API (action: "create") from a Claude Code CLI session.
• Environment: Default (env_01Xa3JcWJDBted2mfTDdoLrQ, kind anthropic_cloud)
• Model: claude-sonnet-5

Steps to reproduce

1. Create a routine with job_config.ccr pointing at the Default environment, a claude-sonnet-5 session, and one user-role event as the initial prompt. (Reproduced both with 4 MCP connectors attached — Gmail, Slack, Google Calendar, Notion — and with zero connectors attached.)
2. Call RemoteTrigger action: "run" on the trigger.
3. Poll RemoteTrigger action: "get_run_log" on the returned session_id.

Observed behavior

The log always stops at the same point, across all 4 attempts:
env[info]: Allocating sandbox
env[info]: Environment runner started
env[info]: No sources configured
env[info]: No setup script configured
env[info]: Starting Claude Code
env[info]: Claude Code process started
followed only by non-transcript bookkeeping events (system/hook_started, system/hook_response, system/commands_changed, autocompact_state, active_goal). No init: line ever appears, no assistant turn ever starts, and list_runs reports worker_status: "idle" indefinitely (observed idle 8+ minutes before giving up). This happened on 3 separate full-config runs and 1 minimal (no MCP connectors, one-word prompt) diagnostic run.

Expected behavior

The session should reach init: model=... cwd=... and begin processing the initial prompt shortly after "Claude Code process started," as it does when a user manually types a message into the same stalled session (observed: typing "continue" into a stalled session did successfully reach init — but as a brand-new session with no memory of the original prompt, since persist_session: false, so the original task was lost, not resumed).

Notes

• This may be specific to sessions started via the trigger's run action / initial events array vs. a user-typed message — since manual typing into the same stalled session did unstick it (but into a fresh context).
• Have not yet been able to confirm whether the actual cron-scheduled fire (as opposed to manual "run now") hits the same code path / same bug.
• Session IDs from reproduction, if useful for backend log lookup: cse_01VyVvsjANKwnoMZXhghfJP3, cse_013JhabXiBYSzz6N3Vmg5wRa, cse_01AVpE1TDevU1B68SNn7m9i4 (full-config runs), cse_01Bfe91zhfhUXWoicVE89xoN (minimal diagnostic run).

What Should Happen?

Expected behavior
The session should reach init: model=... cwd=... and begin processing the initial prompt shortly after "Claude Code process started," as it does when a user manually types a message into the same stalled session (observed: typing "continue" into a stalled session did successfully reach init — but as a brand-new session with no memory of the original prompt, since persist_session: false, so the original task was lost, not resumed).

Error Messages/Logs

This may be specific to sessions started via the trigger's `run` action / initial `events` array vs. a user-typed message — since manual typing into the same stalled session did unstick it (but into a fresh context).
- Have not yet been able to confirm whether the actual cron-scheduled fire (as opposed to manual "run now") hits the same code path / same bug.
- Session IDs from reproduction, if useful for backend log lookup: `cse_01VyVvsjANKwnoMZXhghfJP3`, `cse_013JhabXiBYSzz6N3Vmg5wRa`, `cse_01AVpE1TDevU1B68SNn7m9i4` (full-config runs), `cse_01Bfe91zhfhUXWoicVE89xoN` (minimal diagnostic run).

Steps to Reproduce

1. Create a routine with job_config.ccr pointing at the Default environment, a claude-sonnet-5 session, and one user-role event as the initial prompt. (Reproduced both with 4 MCP connectors attached — Gmail, Slack, Google Calendar, Notion — and with zero connectors attached.)
2. Call RemoteTrigger action: "run" on the trigger.
3. Poll RemoteTrigger action: "get_run_log" on the returned session_id.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.241

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

PowerShell

Additional Information

_No response_