[BUG] Scheduled Claude Routine stalls on deferred tool schema loading — requires manual input to continue
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?
When a Claude Routine runs on a schedule, it stalls at ToolSearch calls used to load deferred MCP tool schemas. The session waits indefinitely for a human response before it can continue. In today's run (April 18, 2026), the routine started at 8:41 AM but only completed at 10:12 AM — a 1.5-hour delay — because three separate ToolSearch blocking points required manual "Tool loaded." responses.
What Should Happen?
Expected behavior: Scheduled (unattended) routines should be able to load tool schemas automatically without requiring human input.
Connectors in use: Gmail, Google Drive, Slack
Workaround attempted: Added read-only MCP tools to permissions.allow in .claude/settings.json, but this does not address the schema-loading block.
Error Messages/Logs
Steps to Reproduce
- Create a Claude Routine with Gmail, Google Drive, and Slack connectors enabled.
- In the routine's system prompt, reference MCP tools from those connectors.
- Schedule the routine to run at a fixed time (e.g. 8:35 AM).
- When the routine runs, it issues ToolSearch calls to load deferred MCP tool schemas.
- The session blocks at each ToolSearch call, waiting for a human "continue" response.
- Without manual intervention, the session never proceeds — it stalls indefinitely.
Note: In attended use the session works fine if a human responds at each pause.
The issue is that scheduled/unattended runs cannot provide that response.
Claude Model
Opus
Is this a regression?
No, this never worked
Last Working Version
_No response_
Claude Code Version
2.1.114 (Claude Code)
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Other
Additional Information
Claude Routines (claude.ai) — no terminal
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗