--dangerously-skip-permissions does not propagate to context: fork sub-agents

Bug description

When a session is started with claude -p --dangerously-skip-permissions, skills that use context: fork in their frontmatter do not inherit the --dangerously-skip-permissions flag. Commands outside the skill's allowed-tools still trigger permission confirmation prompts in the forked sub-agent.

The documentation states:

If the parent uses bypassPermissions, this takes precedence and cannot be overridden

This implies that --dangerously-skip-permissions should propagate to forked sub-agents, but the actual behavior contradicts this.

Steps to reproduce

1. Create a skill with context: fork in its frontmatter and an allowed-tools list:

``yaml
---
name: my-skill
context: fork
allowed-tools: Bash(git status:*), Read
---
``

1. Run claude -p "/my-skill" --dangerously-skip-permissions
2. Inside the skill, attempt to run a command not in allowed-tools (e.g., ps)
3. A permission confirmation prompt appears

Expected behavior

--dangerously-skip-permissions propagates to context: fork sub-agents, and all commands execute without permission prompts.

Actual behavior

The forked sub-agent applies its own allowed-tools permission checks as if --dangerously-skip-permissions was not set. Commands outside allowed-tools trigger confirmation prompts.

Workaround

Extract the skill body (stripping the YAML frontmatter) and pass it directly as a prompt to claude -p, bypassing the context: fork frontmatter interpretation entirely:

SKILL_FILE="skills/my-skill/SKILL.md"
FRONTMATTER_END=$(awk '/^---$/{n++; if(n==2){print NR; exit}}' "$SKILL_FILE")
SKILL_BODY=$(tail -n +"$((FRONTMATTER_END + 1))" "$SKILL_FILE")
claude -p "$SKILL_BODY" --dangerously-skip-permissions

Environment

• Claude Code CLI (latest version as of Feb 2026)
• macOS
• Skills using context: fork with allowed-tools