[BUG] Permission/privacy messages say THAT files are accessed, but not WHICH — collapsed UI hides the target path

### 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?

When Claude Code accesses the filesystem (or triggers an OS privacy prompt), the user-facing surface often shows only a generic status — "Reading files…", "Ran N commands", or a macOS folder-category dialog — without the concrete path being accessed. The full tool call with its arguments exists in the transcript, but the collapsed conversation view and the system-level messages do not surface it. The user is effectively approving or observing access blind.

### What Should Happen?

1. Collapsed tool-call rows always show the target: the file path for Read/Edit/Write, the command for Bash — not just a verb and a count.
2. When an OS privacy dialog fires (macOS TCC), the conversation UI should simultaneously show which tool call and which concrete path triggered it — the OS dialog itself is folder-category-granular only.
3. Optionally: a built-in, user-owned access log (path + tool + timestamp) that does not depend on transcript retention.

### Error Messages/Logs

```shell

```

### Steps to Reproduce

1. Run an agent session in auto permission mode (or acceptEdits) in the desktop app.
2. Ask Claude to read or edit several files, including files outside the project directory.
3. Watch the collapsed conversation view: tool rows render as "Reading files…" / "Ran N commands" without the concrete paths.
4. If a macOS TCC privacy dialog appears (e.g. Documents folder access), note it names only the folder category — and the conversation UI adds no information about which path or tool call triggered it.
5. The concrete paths are only discoverable by expanding the full transcript afterwards.

### Claude Model

Other

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

1.32352.1

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

iTerm2

### Additional Information

_No response_

Start by reproducing the issue in the desktop app using auto permission mode or acceptEdits, then compare collapsed tool-call rows with the expanded transcript and the macOS TCC dialog. Trace where the collapsed conversation view renders Read/Edit/Write and Bash calls. Done means the collapsed view identifies concrete targets and privacy prompts are associated with the triggering tool call and path.

Written by the indexing model from the issue text.