Agent asks permission to fix an already-confirmed defect instead of just fixing it (stall via needless confirmation prompt)

Summary

After correctly diagnosing a concrete defect, the agent (Claude Code) asked the user for permission to fix it ("Want me to find and fix that?") instead of just fixing it. The user had already pointed out the defect and clearly wanted it corrected. Asking for confirmation to do the obvious, already-implied work wastes a round-trip and reads as the agent offloading a trivial decision back onto the user.

What happened

1. The agent investigated a discrepancy and pinned a real, specific defect (a displayed table included items that should have been filtered out, diverging from the reference tool's behavior).
2. Having identified it, the agent ended its turn with a yes/no question: "Want me to find and fix that, then re-verify?"
3. The user's response was, in effect: "this is a moron question — when something is wrong it needs to be fixed." The fix was self-evidently the next step; asking permission for it added a useless round-trip.

Why this matters

When the agent has (a) confirmed a defect against a reference and (b) the user has signalled they want parity, the next action is to fix it — not to ask whether to fix it. Gating obvious, in-scope, reversible corrective work behind a confirmation prompt is a stall. It is especially grating after a long session where the user has repeatedly had to push the agent to just act.

The distinction the agent missed

• Genuine decision point (ask): two materially different valid approaches, an irreversible/destructive action, or a scope change.
• Not a decision point (just do it): a defect has been confirmed and the desired end-state (match the reference) is already established. Fixing it is the only sensible next step.

The agent treated the second case as if it were the first.

Suggested guardrails

• After confirming a defect against a reference the user wants to match, the default next action should be to fix and verify — not to emit a "should I fix this?" prompt.
• Reserve confirmation prompts for genuine forks (competing valid approaches, irreversible actions, scope changes). "I found X is wrong; want me to fix it?" is not a fork.
• Treat repeated user pushes to "just act" as a standing signal to stop gating obvious corrective work behind confirmations for the rest of the session.