Scheduled tasks ignore model selector — always run on Sonnet

Description

Scheduled tasks in the Claude Code desktop app ignore the model selected in the task editor UI and always execute on Sonnet, regardless of what's configured.

Steps to Reproduce

1. Create or edit a scheduled task in the Claude Code desktop app
2. Set the model selector to Opus 4.6 (1M context) (visible in the bottom-right of the prompt editor)
3. Save the task
4. Let it run on its cron schedule

Expected Behavior

The task should execute using the selected model (Opus 4.6).

Actual Behavior

The task runs on Sonnet. Output quality is consistent with Sonnet, not Opus. Confirmed across multiple runs over several weeks.

The model selector appears in the UI and persists visually, but the scheduled task runner does not respect it.

Environment

• Claude Code desktop app (macOS, Darwin 25.3.0)
• Task: cos-daily — runs at 5:37 AM daily
• Model selector shows: Opus 4.6 (1M context)
• Actual execution model: Sonnet (based on output quality and error patterns)

Workaround

Currently using Agent(model: "opus") subagent calls within the prompt to force Opus execution for quality-sensitive work. The outer shell still runs on Sonnet but delegates immediately.

Additional Context

This appears to be the same limitation reported for OpenClaw skills, where model: claude-opus-4-6 in SKILL.md is also ignored. Would be helpful to know if this is a known platform constraint or a bug.