Hey, thanks for the detailed report. This is a known issue. We’ve had a few reports that in Plan mode, tasks are generated incorrectly, especially with non-Anthropic models.
Related threads for reference:
Your note that Opus 4.6 works but Auto, Composer 1.5, and Codex 5.2 don’t matches what other users have reported. Anthropic models seem to handle the task format correctly, while others don’t.
As a temporary workaround:
- Use Opus 4.6, or another Anthropic model like Sonnet 4.6, specifically for Plan mode when you need tasks and the sub-agent workflow
- For everything else, you can still use Auto
The team is aware, and we don’t have an ETA for a fix yet. Let me know if using Opus for Plan mode is enough for your current workflow.