### 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? ## Summary When dispatching a custom sub-agent via the Task/Agent tool, the sub-agent prompts for Edit permission on every individual Edit tool call, reg

[BUG] Sub-agent permission mode inheritance broken — permissionMode frontmatter and parent acceptEdits both ineffective for Edit tool calls

Resolved 💬 3 comments Opened May 7, 2026 by TinoSilfver Closed Jun 5, 2026

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?

Summary

When dispatching a custom sub-agent via the Task/Agent tool, the sub-agent
prompts for Edit permission on every individual Edit tool call, regardless of:

Parent session's permission mode being set to "Edit automatically" (acceptEdits)

via the in-product UI selector

.claude/settings.json containing "defaultMode": "acceptEdits" under

permissions

.claude/settings.json containing "Edit(**)" and "Write(**)" in

permissions.allow

The sub-agent's own frontmatter declaring permissionMode: acceptEdits

The sub-agent's Edit calls each surface a "Do you want to make this edit?"
permission prompt. A pipeline run dispatching one Doc Agent that performs
~10 file edits results in ~10-12 manual permission prompts even with all of
the above settings in place. Confirmed across four consecutive pipeline runs
(2026-05-06 → 2026-05-08).

Expected behaviour

Per official documentation
(https://code.claude.com/docs/en/sub-agents.md, "Permission Modes" section,
line 388 at time of writing):

"If the parent uses bypassPermissions or acceptEdits, this takes precedence and cannot be overridden."

Either:
(a) Parent session's acceptEdits mode propagates to dispatched sub-agents and
eliminates Edit-tool prompts, OR
(b) Sub-agent's own permissionMode: acceptEdits frontmatter takes effect when
parent inheritance does not.

Actual behaviour

Neither (a) nor (b) eliminates the Edit prompts. Every Edit call by the
sub-agent surfaces a permission prompt to the human user.

Reproduction

Configure .claude/settings.json:

```json
{
"permissions": {
"defaultMode": "acceptEdits",
"allow": ["Edit()", "Write()", "Read(**)"]
}
}

What Should Happen?

Expected behaviour

Per official documentation
(https://code.claude.com/docs/en/sub-agents.md, "Permission Modes" section,
line 388 at time of writing):

"If the parent uses bypassPermissions or acceptEdits, this takes precedence and cannot be overridden."

Error Messages/Logs

Steps to Reproduce

Reproduction

Configure .claude/settings.json:

```json
{
"permissions": {
"defaultMode": "acceptEdits",
"allow": ["Edit()", "Write()", "Read(**)"]
}
}

Create a custom sub-agent at .claude/agents/test-agent.md:

---
name: test-agent
description: Test agent for permission inheritance reproduction.
tools: Read, Write, Edit, Glob, Grep
permissionMode: acceptEdits
model: sonnet
---
You are a test agent. When dispatched, edit three files in the repo
(any three .md files), adding a single comment line to each.

In the parent Claude Code session, ensure the in-product mode selector

shows "Edit automatically" (acceptEdits).

Dispatch the sub-agent via the Agent/Task tool, instructing it to perform

the three edits.

Observe: every Edit call surfaces a "Do you want to make this edit?"

prompt. The parent session's mode selector does not affect the sub-agent.
The sub-agent's own permissionMode: acceptEdits does not affect itself.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.132 (VS Code extension, win32-x64)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

<head>

Environment<o:p></o:p>

<ul style='margin-top:0cm' type=disc>
<li class=MsoNormal style='mso-list:l4 level1 lfo1;tab-stops:list 36.0pt'>Claude Code version: 2.1.132 (VS
Code extension, win32-x64)<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l4 level1 lfo1;tab-stops:list 36.0pt'>Platform: Windows 11 Pro 10.0.26200
(PowerShell 7+)<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l4 level1 lfo1;tab-stops:list 36.0pt'>Project: ClubEvents (custom multi-agent
SDD pipeline; ~9 custom sub-agents in .claude/agents/)<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l4 level1 lfo1;tab-stops:list 36.0pt'>Settings: .claude/settings.json includes defaultMode:
acceptEdits and Edit()/Write() allow rules.<o:p></o:p></li>
</ul>

Real-world impact<o:p></o:p>

In our project we run
a /feature-request orchestrator that dispatches up to 8 sub-agents per pipeline
run (Spec, Tech, Code, Doc, Compliance, Gap, Final Verdict, Task). Measured
friction across four consecutive pipeline runs:<o:p></o:p>

Across ~10 pipeline
runs per month this is ~100-150 unnecessary manual approval prompts. Each
prompt requires human attention, breaks orchestration flow, and accumulates
non-trivial wall-clock and cognitive cost.<o:p></o:p>

Related issues<o:p></o:p>

<ul style='margin-top:0cm' type=disc>
<li class=MsoNormal style='mso-list:l3 level1 lfo2;tab-stops:list 36.0pt'>#28584 — "[BUG] Subagents prompt for permission on every tool
call starting v2.1.56" — closed; this issue may be a regression of
the same root cause with the addition
of permissionMode frontmatter being equally ineffective<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l3 level1 lfo2;tab-stops:list 36.0pt'>#18950 — "[BUG] Skills/subagents do not inherit user-level
permissions from settings.json" — open; describes the inheritance
gap; this issue extends it to the
per-agent permissionMode frontmatter<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l3 level1 lfo2;tab-stops:list 36.0pt'>#40241 — "[BUG] --dangerously-skip-permissions does not
propagate to subagents" — closed as duplicate of #42696; same family
of inheritance bugs<o:p></o:p></li>
</ul>

Suggested
resolution<o:p></o:p>

Either: (a) Restore
the documented inheritance semantics — parent acceptEdits propagates to
dispatched sub-agents, OR (b) Fix permissionMode: acceptEdits in sub-agent
frontmatter to take effect when present (since the parent inheritance is
currently broken, this is the only project-tier workaround), OR (c) Provide an
explicit project-tier permission-policy mechanism for sub-agents (e.g.,
.claude/agents-settings.json or a subagents block in .claude/settings.json).<o:p></o:p>

The current state
forces project owners to either (i) accept ~10-15 manual prompts per pipeline
run, or (ii) avoid sub-agent dispatches entirely (which defeats the purpose of
the Agent/Task tool for orchestrator patterns).<o:p></o:p>

Compensation
request<o:p></o:p>

Investigating this
issue has consumed substantial time and tokens on my side as a paying user:<o:p></o:p>

<ul style='margin-top:0cm' type=disc>
<li class=MsoNormal style='mso-list:l2 level1 lfo3;tab-stops:list 36.0pt'>Four consecutive pipeline runs (2026-05-06
→ 2026-05-08) each surfacing manual Edit-prompts at the Doc Agent stage —
~30-50 manual approvals in total across the four runs.<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l2 level1 lfo3;tab-stops:list 36.0pt'>Process Rules Batches 2-4 alone consumed
approximately 6.7M effective billable token-equivalents end-to-end (1.29M

2.18M + 3.22M); the STORY-T03 feature run added another ~13M. Full

pipeline measurement memos with per-stage breakdown available on request.<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l2 level1 lfo3;tab-stops:list 36.0pt'>A dedicated between-package research
session (using the claude-code-guide sub-agent with WebFetch +
WebSearch) to triangulate the documented behaviour against the observed
behaviour, including reading the official sub-agents documentation and four
GitHub issues (#28584, #18950, #40241, #42696).<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l2 level1 lfo3;tab-stops:list 36.0pt'>Authoring this bug report itself across
multiple iterative drafts.<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l2 level1 lfo3;tab-stops:list 36.0pt'>The Polku 1 mitigation attempt
(per-agent permissionMode: acceptEdits applied to six sub-agents
on 2026-05-07, commit 0d00752) and the follow-up Paketti 4 pipeline
run dedicated to validating that Polku 1 (per-agent frontmatter) does not
work — an additional pipeline that exists solely because of this bug.<o:p></o:p></li>
</ul>

Given that:<o:p></o:p>

<ol style='margin-top:0cm' start=1 type=1>
<li class=MsoNormal style='mso-list:l0 level1 lfo4;tab-stops:list 36.0pt'>The bug is documented in multiple prior
issues (#28584 closed without real fix, #18950 still open) without
resolution from Anthropic.<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l0 level1 lfo4;tab-stops:list 36.0pt'>The official documentation explicitly
promises behaviour the product does not deliver.<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l0 level1 lfo4;tab-stops:list 36.0pt'>The investigation cost has been borne
entirely by paying users without a functional escalation channel.<o:p></o:p></li>
</ol>

I request Claude
usage credit compensation for the documented investigation effort and token
consumption.<o:p></o:p>

If GitHub issues are
not the canonical compensation channel, please redirect or copy this request to
the appropriate Anthropic support channel — I will also be sending a parallel
message to support@anthropic.com referencing this issue for compensation handling.<o:p></o:p>

Validation status<o:p></o:p>

Polku 1 (per-agent
permissionMode: acceptEdits frontmatter) applied to six sub-agents on
2026-05-07 (commit 0d00752). Validation pipeline run Paketti 4 /
process-rules-batch-4-measurement-automation completed 2026-05-08:<o:p></o:p>

<ul style='margin-top:0cm' type=disc>
<li class=MsoNormal style='mso-list:l1 level1 lfo5;tab-stops:list 36.0pt'>Stage 5 sub-agent Edit-prompts: 5-6
surfaced<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l1 level1 lfo5;tab-stops:list 36.0pt'>Stage 5 Edit operations performed: 4-7<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l1 level1 lfo5;tab-stops:list 36.0pt'>Prompt-per-Edit ratio: ~1:1 (unchanged
from pre-Polku-1 baseline)<o:p></o:p></li>
<li class=MsoNormal style='mso-list:l1 level1 lfo5;tab-stops:list 36.0pt'>Conclusion: Polku 1 ineffective. Per-agent permissionMode:
acceptEdits frontmatter does not bypass the underlying bug; sub-agent
dispatches continue to prompt on every Edit tool call.<o:p></o:p></li>
</ul>

This validates that
the bug spans both inheritance paths — parent session propagation AND per-agent
frontmatter declaration — confirming the issue is deeper than #18950 and #28584
individually describe.<o:p></o:p>

</body>

</html>

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗