Agent proceeds to execute multi-step changes after asking a question, without waiting for the user's answer to the proposed plan
Summary
When the agent uses AskUserQuestion to clarify one decision (e.g., "which alerting channel?"), then begins drafting/presenting a plan, it can mistakenly treat any subsequent user message — even one that just adds collateral context (a webhook URL, a small instruction) — as approval to execute the entire plan unilaterally.
Repro
- Agent asks: "Where should alerts go? (Slack/Email/Telegram/Dashboard)"
- User answers: "Slack"
- Agent starts writing a multi-step plan ("Light plan: zero new secrets, zero new services…") and is interrupted mid-output by the user, who pastes a Slack webhook URL and asks to save it to 1Password
- Agent treats the interruption as a green light, saves the secret, then deploys an 88-line bash script, config file, state file, log file, and is about to add a cron entry — all without ever finishing the plan presentation or getting explicit "go"
Expected
An incoming user message that adds context (here: a webhook URL) should not be inferred as approval for an unfinished plan. The agent should: (1) action only the literal instruction (save webhook to 1P), (2) re-present the full plan, (3) explicitly wait for "go".
Impact
Production-host changes were started without the operator validating the design (heap-fix-and-restart-of-Wazuh-indexer-class scope). In my case the agent's heredoc was buggy so nothing was actually persisted — but had it succeeded, /etc files and a cron job would have been live without my approval.
Workaround
Saved an explicit memory note instructing the agent to stop completely when any question/plan is pending; collateral user messages are not GO signals.
Environment
- Claude Code (CLI), Opus 4.7 (1M context)
- Auto Mode was off at the time
- macOS Darwin 25.5.0
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗