claude-code - 💡(How to fix) Fix Opus over-confirms on already-authorized low-risk steps, then silently blocks

Summary

Two orthogonal model-behavior issues observed with Opus during a multi-turn iterative session (a back-and-forth practice/drill loop).

Repro

In a multi-turn task, the user repeatedly says "continue" to advance through a series of similar steps (e.g. practice problems, checklist items). Instead of executing the next step, the model stops and asks "do you want the next one, or stop here?" — and then produces nothing further.

Bug 1 — over-confirmation

The model inserts a human-in-the-loop confirmation on an action that is already authorized (standing authorization), low-risk, and reversible. The user has repeatedly granted permission to keep going; asking for per-step confirmation is pure friction. HITL should only trigger on irreversible / high-risk / ambiguous-boundary actions.

Bug 2 — silent block (orthogonal to Bug 1)

Even if the model decides to wait for the user, it emits no signal explaining what it is waiting for. The user perceives a deadlock with zero feedback — they don't know the flow is blocked or what action would unblock it. This holds even if HITL were warranted: "request visibility" is a necessary part of HITL design, not optional.

Expected behavior

1. When the user repeatedly says "continue" on the same class of task, execute the next step directly without re-confirming.
2. If an irreversible/high-risk action genuinely requires confirmation, state the specific pending action explicitly ("I need you to confirm X before I continue") — never wait silently.

Notes

These are model-behavior/quality issues rather than a software crash. Framed in agent-design terms: Bug 1 is over-confirmation (treating standing authorization as per-action authorization); Bug 2 is a silent block (HITL request invisibility / no transparency signal on a blocking state).