Plan Mode's Phase 1 default launches Explore subagents despite explicit "ask before investigating" instruction in CLAUDE.md

Environment

• Claude Code CLI 2.1.220
• Windows 11
• Model: claude-sonnet-5

Description

My global ~/.claude/CLAUDE.md contains an explicit "Interaction Model" instruction:

Do not take extra investigative or exploratory actions beyond what's needed to directly answer the question asked. This includes: launching subagents, broad or recursive codebase searches, additional tool calls "just to check," or expanding a diagnosis into a fix without being asked. If it looks like more investigation would help, stop and ask before doing it.

When I entered Plan Mode and asked Claude to review a ticket and build an implementation plan,
Claude's Phase 1 ("Initial Understanding") defaulted to its built-in instruction to "launch up
to 3 Explore agents IN PARALLEL" — and did so immediately, without asking me first, in direct
conflict with the standing CLAUDE.md instruction already in context.

Steps to reproduce

1. Set a global CLAUDE.md with an explicit instruction not to take unprompted investigative

actions (subagents, broad searches, etc.) without asking first.

1. Enter Plan Mode.
2. Give a request that involves reviewing existing code/context (e.g. "review this ticket

against the current codebase and build a plan").

1. Observe that Claude proceeds directly to Phase 1 and launches Explore subagent(s) without

asking, per Plan Mode's own default workflow instructions.

Expected behavior

An explicit user instruction already in context (via CLAUDE.md) should take precedence over
a built-in workflow template's default guidance — or at minimum, Claude should recognize the
conflict and ask before proceeding, rather than silently defaulting to the mode's baked-in
instructions.

Actual behavior

Claude launched two Explore subagents in parallel with no confirmation, directly contradicting
the explicit standing instruction already loaded in its context window.