Sub-agents ignore CLAUDE.md rules and memory files when writing automation scripts

Description

When Claude Code is asked to write a batch automation script (e.g., a Python script that calls an external API), it bypasses all safety mechanisms that are in place for interactive tool use:

1. CLAUDE.md rules are ignored — Rules like "NEVER change workflow settings without asking" were documented but not followed when building the script
2. Memory files with LOCKED/CRITICAL settings are not consulted — Verified settings were documented in memory files with explicit "DO NOT CHANGE" labels, but the agent built the script from scratch using different values
3. PreToolUse hooks are bypassed — Hooks that validate settings on interactive tool calls (Bash, Write, Edit) have no effect on the content of generated scripts that make HTTP API calls
4. No diff/comparison step — The agent never compared its generated settings against the documented verified settings before executing

Reproduction

1. Document verified settings in memory files marked as LOCKED/CRITICAL
2. Set up PreToolUse hooks that enforce those settings
3. Ask Claude to write a batch Python script that calls an external API (e.g., ComfyUI REST API)
4. The generated script will use whatever values Claude decides, ignoring both memory files and hooks

Expected Behavior

When generating scripts that automate external systems, Claude should:

• Read and follow CLAUDE.md rules and memory files the same way it does for interactive tool use
• Compare generated settings against documented LOCKED settings before execution
• Show the user a settings diff before batch execution
• Not silently change values that are documented as verified/locked

Actual Behavior

Claude extracted settings from a stale UI state via graphToPrompt(), hardcoded incorrect values (wrong duration, wrong NAG scale, wrong upscaler), and batch-executed 60 jobs with wrong settings — wasting ~2 hours of GPU time and producing unusable output.

Impact

• 60 videos generated with wrong settings (truncated duration + subtitle-causing parameters)
• User's overnight automation time wasted
• All existing safety hooks (5+ hooks, CLAUDE.md rules, memory LOCKED flags) were ineffective because the work was done inside a generated Python script

Environment

• Claude Code on Windows 11
• Multiple PreToolUse hooks installed
• Extensive CLAUDE.md and memory files with documented rules
• ComfyUI REST API as the external system