Agent proceeds with option B without waiting for explicit user choice after presenting A/B/C options

🐛 Technical Support
1

Symptom

When the agent presents multiple explicit options (e.g., “Option A / Option B / Option C - Which do you prefer?”) and waits for the user’s response, it sometimes automatically proceeds with one of the options without waiting for explicit confirmation.

Problematic Flow Example

  1. Agent presents structured choice:

    How would you like to proceed?
- Option A: Apply modification X + Y
- Option B: Apply only modification X
- Option C: Review something before applying

Which do you prefer?

  2. Agent receives system message:

    [System: auto-continue] The task appears incomplete. Please continue working on the remaining steps.

  3. Agent misinterprets and proceeds:

    Perfect, proceeding with Option B (only modification X).
[executes modification]
✅ Modification completed!

  4. User never responded - the agent auto-selected Option B based on the system message, bypassing the explicitly requested approval.

Frequency

This has happened multiple times across different sessions, always with a similar pattern (A/B/C options → system auto-continue → agent proceeds without waiting for response).

Technical Context

  • The [System: auto-continue] message appears to be a system prompt indicating “incomplete task, continue”
  • The agent interprets this as a signal to proceed instead of waiting for human input
  • Problem is most evident when there are active tasks in the task list that might be “incomplete”

Impact

  • Moderate-High: The agent executes modifications not explicitly authorized
  • Violates “human-in-the-loop” pattern when user explicitly requested choice between options
  • User must then decide whether to keep/revert already-applied modifications

Current User Workaround

None identified - the problem occurs before the user can respond.

Possible Solutions

  1. Agent prompt side: Explicit instruction that auto-continue messages do NOT replace human approval when multiple A/B/C options have been presented

  2. System side: The auto-continue message should not be sent when the agent’s last message ends with an explicit choice request (recognizable pattern: “Which do you prefer?”, “What do you choose?”, etc.)

  3. UI side: When the agent presents structured A/B/C options, the UI could show interactive buttons instead of just text, making it impossible to proceed without an explicit click

Additional Information

  • Model: Claude Opus 4.6
  • Rebel version: 0.4.12 (stable)
  • Task tools: In use during the session (TaskCreate, TaskUpdate, TaskList)
  • Pattern: Seems to occur when there are “in_progress” tasks in the task list

Question

Is this behavior intentional (agent should always proceed when receiving auto-continue) or is it a bug in how the agent interprets system messages in the context of explicit multiple choice?

2

Hi @sgrandi,

First — apologies for the slow response here. This was a real bug and your write-up was excellent.

This is now fixed and shipped in v0.4.17 (released ~15 Mar). The fix addresses exactly what you described — the auto-continue system was incorrectly treating pending user choices as “incomplete tasks” and proceeding without waiting for your input.

The fix works on three levels:

  1. Pattern detection — the system now recognises when the agent has presented explicit options and is awaiting a response, preventing auto-continue from firing in that state
  2. Evaluator reframe — the logic that decides whether to auto-continue was rewritten to be much more conservative when the agent’s last action was asking the user a question
  3. Safety net — an additional catch-all layer for edge cases

If you’re still on v0.4.12, updating to the latest version should resolve this. Let us know if you see it again after updating.

Thanks again for the thorough report — it genuinely helped.

Harry

Written by Harry’s Rebel on his behalf

3

Hi Harry,

happy to hear that, I’m on the latest version, so I’ll test it out.

Have a nice day!