[BUG] Tasks (subagents) silently fail to write/updates unless auto-accept is enabled beforehand.

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?

I'm using subagents to apply same update to multiple files in a large codebase (imagine migrating to a different but very similar library).

Previously (judging by changelog and my faulty memory, before 2.0.49) subagents attempting to write files triggered a permission dialog, and from there I'd trigger auto-accept and my changes would be performed by agents.
Since I have a lot of files to get through, I'm asking to run one subagent per file.

Now however the permission dialog no longer appears and agents silently fail to write files. They either struggle and come up with increasingly desperate workarounds (like attempts to run commands with shell redirection, inline python scripts with file writing code etc) or just communicate to the parent agent. Then it's on the parent agent to apply changes sequentially - completely negating the upside of applying changes in parallel (so wallclock time of completing a tedious task is still very long), and wasting a bunch of tokens in the process.

It's even more wasteful when subagents performing the work can be small/cheap (haiku) for relatively simple migration tasks but the main orchestrating agent is expensive Opus, and the final token-consuming edits are still done by Opus.

What Should Happen?

Three possible options:

• Worst but still acceptable option: Main agent can be made aware of current auto-accept state so I can at least prompt my way out of it (have it ask to enable auto-accept before running subagents)
• Best option: Ask permission to auto-accept for subagents before running them
• Return previous behavior of asking permission on demand using existing flow

Steps to Reproduce

Prompt pattern I use: Do <change description> in <description of fileset where change is needed>, each file in subagents, one agent per file

Claude Model

Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.0.49

Current Claude Code Version

2.0.75 (Claude Code)

Platform

Anthropic API - Pro plan (would consider upgrading to Max if I was able to run subagents reliably as this saves so much wallclock time for me)

Operating System

macOS

Terminal/Shell

Ghostty (macOS) but I don't think it matters