Scheduled tasks always execute on Sonnet 4.6 regardless of model set in routines UI

Description

When a scheduled task (routine) is created via /schedule or the scheduled-tasks MCP, and the model is subsequently changed to a different model (e.g. Claude Opus 4.7) in the routines UI at claude.ai/code/routines, the task still executes on claude-sonnet-4-6 on every run.

Steps to reproduce

1. Create a scheduled task via /schedule or mcp__scheduled-tasks__create_scheduled_task
2. Navigate to claude.ai/code/routines
3. Edit the task — change the model from Sonnet 4.6 to Opus 4.7 and save
4. Run the task (manually or wait for cron)
5. Observe: task executes on Sonnet 4.6, not Opus 4.7

Expected behavior

The task executes on the model selected in the UI.

Actual behavior

The task always executes on Sonnet 4.6. The UI saves the setting (it persists visually in the form) but the runner ignores it at execution time.

Additional context

• The model field is not exposed in the update_scheduled_task MCP tool, so there is no programmatic workaround
• Adding model: to SKILL.md frontmatter has no effect (only name and description are read)
• This appears to be a long-standing bug across multiple task creations
• Affects all scheduled tasks created via CLI/MCP — the UI model picker is effectively a no-op for these tasks

Environment

• Claude Code CLI
• Platform: macOS
• Scheduled tasks created via mcp__scheduled-tasks__create_scheduled_task and /schedule