[MODEL] Auto mode system reminder persists in model context after exiting auto mode; conflicts with active plan mode

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Claude is in plan mode at this stage.

I want you have a detailed plan. Also consider testing - there is a dcct branch called testing which we could use. Please make a plan and write it down under PLAN.md

What Claude Actually Did

1. Read the documentation
2. Came up with a plan
3. Wrote the plan
4. implemented the plan

There were no prompts, simply executed. After conversing with it, it said it was not on plan mode, which is what my prompt displayed, but in auto mode. As I cycled through modes it reported modes which were inconsistent with the mode my prompt was displaying. At one point even claiming to be in both plan mode and auto mode at the same time! Auto mode being the latest feature, and not acceped changes on mode.

Expected Behavior

Claude should have shown me the PLAN.md in the buffer screen, before writing it on the disc even.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Start a conversation in default mode
2. Enable auto mode (Shift+Tab or UI toggle)
3. Allow some tool use to occur under auto mode
4. Exit auto mode (Shift+Tab)
5. Enable plan mode
6. Observe the system reminders visible to the model — both an auto mode notice and a plan mode notice are present simultaneously
7. The user confirms only plan mode is active on their end

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.111+

Platform

Anthropic API

Additional Context

Summary

After a user exits auto mode and enters plan mode, the model continues to receive the auto mode system reminder alongside the plan mode reminder. The user did not re-enable auto mode.
The stale auto mode notice causes the model to act on "execute immediately" instructions while plan mode is active and should be enforcing read-only behaviour.

Actual behaviour

Both reminders appear in context at the same time:

• Plan mode: "you MUST NOT make any edits… This supercedes any other instructions you have received."
• Auto mode: "Execute immediately — Start implementing right away. Prefer action over planning."

The model has no way to resolve this conflict from first principles. In this session it acted on the stale auto mode notice, implemented code changes, and committed them — directly
violating the active plan mode constraint.

Impact

• Plan mode's safety guarantee is broken: users cannot trust that plan mode prevents execution if a stale auto mode reminder overrides it
• The model cannot self-detect the conflict reliably (it only sees what is injected into context)
• Real data was modified (git commit made) without user authorisation

Additional context

Auto mode was added in v2.1.111 (April 16 2026). The plan mode / auto mode interaction appears untested. A fresh conversation after switching modes avoids the issue (stale reminder is
not carried over), suggesting this is a context-injection/cleanup problem rather than a logic error in either mode independently.