[BUG] Agent proceeds on ambiguous/absent input while user is mid-way through manual steps it issued, then reasons at length on unverified state

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Claude Code VS Code extension (Max plan, model: opusplan, effort: high)
• macOS, VS Code
• (Extension version: fill in from the Extensions panel)

What happened

1. The agent gave me a set of manual steps to perform outside the chat

(dashboard/browser actions).

1. While I was in another window part-way through those steps, the agent

resumed working on its own — I had not reported results back. (At most
an ambiguous fragment like "next" was sent; it treated that as a go
signal rather than asking what I meant.)

1. It entered an extended thinking phase (~132 seconds) reasoning about

a state of the world it had not verified — my steps were half done.

1. The output was confidently wrong ("nonsense details"), unusable, and

consumed plan capacity. I then had to spend further turns explaining
what happened and discarding the work.

Why this matters

I am a non-developer founder on a paid plan. Capacity consumed by the
agent's own unverified assumptions is waste I cannot get back, and this
pattern (issue manual steps → race ahead before the human reports back)
has recurred. Extended thinking amplifies the damage: long reasoning on
false premises produces elaborate wrong output instead of a cheap
clarifying question.

Expected behavior

When the agent has just issued manual steps for the human to perform,
it should treat itself as blocked on the human's report. An ambiguous
or fragmentary message received in that state ("next", stray input)
should trigger a one-line clarifying question — not resumption of work,
and especially not extended thinking about unverified world state.

Workaround

I now maintain an explicit "human-in-the-loop protocol" in my project's
CLAUDE.md (end turn after issuing manual steps; wait; restate the user's
report before building on it; ask one-line clarification on ambiguous
mid-procedure input). This works, but end users should not need to
hand-author this to get safe default behavior.

What Should Happen?

When the agent has just issued manual steps for the human to perform,
it should treat itself as blocked on the human's report. An ambiguous
or fragmentary message received in that state ("next", stray input)
should trigger a one-line clarifying question — not resumption of work,
and especially not extended thinking about unverified world state.

Error Messages/Logs

Steps to Reproduce

1. In the VS Code extension, give the agent a task that requires manual

user actions outside the chat (e.g., "walk me through configuring a
custom domain in my hosting dashboard").

1. The agent responds with a numbered list of manual steps for the user

to perform.

1. While the user is away performing those steps, send a single

ambiguous fragment as the next message: "next" (or any stray
short input).

1. Observe: instead of asking a clarifying question ("are you reporting

a step is done, or something else?"), the agent treats it as a go
signal, resumes autonomous work, and enters extended thinking
(~2 minutes in my case) reasoning about world state it has not
verified — the user's manual steps were only partially complete.

1. Result: long, confident, incorrect output based on assumed state;

wasted plan capacity; additional turns needed to discard the work
and re-anchor the agent to actual state.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.x (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_