[BUG] Pending permission/approval state not always rendered in session view ("waiting on nothing" — agent announces action, UI shows neither prompt nor progress)

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?

TITLE: Pending permission/approval state not always rendered in session view ("waiting on nothing" — agent announces action, UI shows neither prompt nor progress)

Environment: Claude Code desktop app, Windows 10 Pro 19045.

Summary: In interactive sessions we hit states where the agent is waiting on something (an approval, or a next-turn signal) but the UI displays nothing actionable: no prompt card, no spinner, last message may even end with the agent announcing "starting now". A user nudge (any message) resumes it. Related observation from the agent side: sessions report "no permission prompt encountered" while the human WAS clicking approval prompts — tool calls just block until approved, invisibly to the model, so agent-side reports cannot be trusted for authorization auditing either.

Steps to reproduce (not fully deterministic):

1. Give a session instructions ending in a multi-step plan; approvals granted along the way.
2. Occasionally the session ends a turn announcing imminent action; no tool call follows, no pending prompt is rendered, yet a pending "continue" state seems to exist (a subsequent short user message like "go" — which appeared to be already queued/expected — resumes work immediately).
3. Separately: approval prompt cards sometimes appear only after a delay, or an expired/stale prompt is shown for an already-dead run.

Expected: any state where the runtime waits on user input should be visibly rendered (prompt card or explicit "waiting" indicator); a turn that ends without pending input should be clearly final.

Impact: users can't distinguish "working", "waiting invisibly", and "done"; automation supervision requires constant human polling.

What Should Happen?

Expected: any state where the runtime waits on user input should be visibly rendered (prompt card or explicit "waiting" indicator); a turn that ends without pending input should be clearly final.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce (not fully deterministic):

1. Give a session instructions ending in a multi-step plan; approvals granted along the way.
2. Occasionally the session ends a turn announcing imminent action; no tool call follows, no pending prompt is rendered, yet a pending "continue" state seems to exist (a subsequent short user message like "go" — which appeared to be already queued/expected — resumes work immediately).
3. Separately: approval prompt cards sometimes appear only after a delay, or an expired/stale prompt is shown for an already-dead run.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code desktop app, Windows 10 Pro 19045

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_