[BUG] [Cowork] Subagent invocations rejected with no permission prompt shown

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) 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?

## What happened

Four `Agent` (subagent) invocations were rejected. The model received:

> The user doesn't want to proceed with this tool use. The tool use was rejected

**No permission dialog ever appeared.** The first two happened with
auto-approve enabled. I then switched to manual approve specifically to catch
the prompt, and still saw nothing — the call was rejected without any dialog.

## Sequence

| # | Subagent task | Result |
|---|---|---|
| 1 | Web research — enterprise server hardware | Rejected. Retry succeeded. |
| 2 | Web research — datacenter power/cooling | Rejected. Retry succeeded. |
| 3 | Analysis task, longest prompt of the three | Rejected |
| 4 | Same prompt again, manual approve enabled | Rejected. Abandoned. |

## The part that seems most diagnostic

**The same work succeeded when done inline in the main conversation instead of
in a subagent.** So whatever rejected it was not rejecting the content — it
looks specific to the subagent invocation path.

Two prompts succeeded on retry while a third failed twice, so it does not look
like simple flakiness either. The one that never succeeded had the longest
prompt and contained the most unusual strings (numeric format names, hardware
model numbers, acronyms) — that is a guess, not a diagnosis, but it is the main
thing separating it from the two that eventually worked.

## Why this is disruptive

From the model's side, "rejected" is indistinguishable from "the user declined."
Claude correctly stopped and waited each time. Since no dialog had appeared, I
had no idea anything was pending — so each occurrence was several minutes of
silence before I noticed and prompted it again.

An error that distinguishes *user declined* from *request never reached the
user* would remove most of the friction, even if the underlying delivery
problem takes longer to fix.

## Prior reports

Searched before filing. #21474, #16393 and #18885 describe the same symptom but
are all closed as duplicates, and all appear to be Claude Code CLI rather than
Cowork. #18885 describes a plausible mechanism — subagents running in a mode
where they cannot prompt, so anything outside the allowlist is auto-denied
silently — which matches what I saw.

Environment

- Cowork mode, Claude desktop app (macOS)
- Version: Claude 1.24012.9 (03c61d) 2026-07-24T04:59:17.000Z
- Mac Mini
- Model: claude-opus-5
- Date: 2026-07-25
- Session: `8bf98dc5-2e65-451b-823b-22820ebe98ed` /
`5475c59a-0d9c-48aa-9d58-23e47999f2d5`
- Context: long session — subagent web research, editing a local git repo via a
connected folder, running Python, committing

### What Should Happen?

Show permission dialog (manual approvals) or auto-approve (if selected).

### Error Messages/Logs

```shell

```

### Steps to Reproduce

See above.

### Claude Model

Opus

### Is this a regression?

I don't know

### Last Working Version

_No response_

### Claude Code Version

Claude 1.24012.9 (03c61d) 2026-07-24T04:59:17.000Z

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Terminal.app (macOS)

### Additional Information

_No response_

No source files, tests, or entry points are identified. Start by reproducing the rejected subagent invocation in Cowork with manual approval enabled, then compare the subagent and inline execution paths; done means rejected requests show a permission dialog or a distinct delivery error.

Written by the indexing model from the issue text.