Says it will ask user, but then doesn't

[ekr1]

Issue

Very often, the agent will generate a response "...long explanation ending in choices. I will ask the user to select one of the choices". But then it proceeds to pick a choice itself and continues the loop.

Here is an example:

...
**Suggested implementation**

Want me to:
- **(A)** write the Python script (dry-run first, prints planned edits per 
file), or
- **(B)** also run it and apply?

Recommend A first: generate script, dry-run, review diff of name-matching 
(esp. kebabcase corners + which env files), then apply. Confirm which envs to
touch (prod only, or all).

Which — A or B, and prod-only or all envs?
56k ◇ 0.0% ↑ 882 ↓ 616k ◇ 0.0% ↑↓

I'll build the script, dry-run first.
61k ◇ 0.0% ↑ 1.8k ↓ 679k ◇ 0.0% ↑↓

Tool Call: Local • Command

Command:
cat > /tmp/apply_limits.py << 'PYEOF'
...

As you can see, it came up with a choice, and a recommendation, but then the agent loop immediately continued, I was not actually having a chance to enter anything.

I assume this could be solved by fine-tuning the system prompt for the agent behaviour somehow. Maybe caveman mode somehow influences the text emitted by the LLM so cecli's agent code that would have returned control to the user doesn't pick up on it?

Version and model info

cecli v1.1.0
Models claude-opus-4.8 (main) • claude-haiku-4.5 (weak)
Settings agent (edit format) • 2k think tokens • infinite output
Environment .git (176 files) • map (4096 tokens, auto refresh)
├─ Subagents memorizer • worker
├─ Skills caveman