Skills/commands hang on new task — shows "Mulling" but no LLM output

[charlesvien]

Context

Reported by @haacked in Slack. This is a recurring issue that started after upgrading to the latest version. Potentially related issue #1617, the LLM appears to not be doing anything at all rather than being slow.

Problem

When running a skill (e.g. /review-code) from a new task screen, the session appears to hang indefinitely. The UI shows "Peeking... Mulling..." status messages and elapsed time, but no "Thinking..." message appears and no LLM output is generated. The chat window seems disconnected from the actual LLM session.

After hitting ESC to interrupt and prompting again (e.g. "It looks like you got stuck"), it recovers and works immediately, the first script runs almost instantly, suggesting the LLM was never actually processing.

Additionally, sub-agents are intermittently not found when running skills.

Steps to reproduce

1. Open a new task
2. Change effort level to xhigh in the dropdown
3. Change mode to Auto (instead of Plan)
4. Run a skill (e.g. /review-code) as the first prompt
5. Observe: no "Thinking..." message appears, only "Mulling..." with elapsed time
6. Wait — it will not produce output on its own
7. Press ESC, then prompt again, it recovers and works

Expected behavior

Skills should begin executing immediately when submitted, showing "Thinking..." and producing output without requiring an ESC + retry cycle.

Additional notes

• Regular prompts (non-skill) with the same effort/mode settings work fine
• The issue is specific to running a skill/command as the first action
• Sub-agents are also intermittently not discovered