Cloud routines (RemoteTrigger) silently fail to execute the configured prompt

Summary

Created a scheduled cloud routine ("Weekly FDSE Update Draft") via the RemoteTrigger API (used by Claude Code's /schedule skill). On every manual test run, the routine's cloud session initializes and reports "Completed," but never appears to execute the configured prompt or call any of its attached MCP connector tools (Google Calendar, Notion, Gmail) — no side effects occur (e.g. no Gmail draft created), and the visible transcript shows only the container-setup steps ("Set up a cloud container" → "Started Claude Code") with nothing after.

Reproduction

1. Created a routine via RemoteTrigger action: create with cron_expression, job_config.ccr.environment_id, job_config.ccr.session_context.events[0].data.message (a user message containing the task prompt), and mcp_connections for Google-Calendar, Notion, and Gmail connectors.
2. Triggered RemoteTrigger action: run three separate times (trigger id trig_01Dyk1jZ75pX2oStpLLwdmhV), including once with message.content as a plain string and once as an Anthropic-style content-block array ([{"type":"text","text":"..."}]).
3. Session IDs from the three runs: cse_011RbZJTgWEmQksGJDecLtsE, cse_01LvFpbvcHVYVeu2KmszWtJU, cse_01N522uqe2xz8mN91czB6Qu3.
4. After each run, checked the target Gmail account for a new draft (task was to create one) — none appeared in any of the three runs.
5. Opened the routine's page in the claude.ai web UI (Code → Routines → routine → Runs panel) to inspect the run transcript. The visible transcript only shows: "Initialized session" → "Set up a cloud container" ✓ → "Clone repository" (skipped) → "Run setup script" (skipped) → "Started Claude Code" ✓ → nothing further (just a stray orange asterisk/logo mark, no message content).
6. The completed run entry in the "Runs" side panel has a > chevron that appears clickable but does nothing when clicked — could not expand to see full run details/logs this way either.

Expected

The routine's cloud agent should receive and act on the configured prompt, calling the attached MCP connector tools as instructed, and the web UI should let me inspect the full run transcript (the chevron should expand run details).

Actual

The prompt appears to never be processed — the session does setup and then ends with nothing further, across 3 separate runs and 2 different message-content formats. The UI provides no way to inspect what actually happened during the run (chevron non-functional).

Environment

• Feature: claude.ai Routines / RemoteTrigger API (cloud scheduled agents), used via Claude Code's built-in /schedule skill.
• Environment used: env_01ScfdEsn6AidKZpFvyHCnZw ("Default", kind anthropic_cloud).
• Model: claude-sonnet-5.

🤖 Reported via Claude Code