Agent Ignored [[PLAN]] Mode and Implemented Without User Approval

[armaansood]

Describe the bug

When a user message is prefixed with [[PLAN]] (shift tab), the agent should create/update a plan and wait for explicit user
approval before implementing. Instead, the agent proceeded directly to implementation after creating the plan.

Affected version

v0.0.397

Steps to reproduce the behavior

• User sends message with [[PLAN]] prefix asking to investigate missing files and update agent instructions
  • Agent investigates the codebase (correct)
  • Agent attempts to create/update plan.md (correct)
  • Agent immediately implements changes without waiting for user approval (BUG)

Expected behavior

Per the system instructions for [[PLAN]] mode:

"Do NOT start implementing unless the user explicitly asks (e.g., 'start', 'get to work', 'implement it')."

The agent should have:

- Created/updated the plan
- Provided a brief summary
- Stopped and waited for user to say "start", "implement", etc.

Actual Behavior

The agent:

- Created the plan
- Immediately began editing RequestMessageUtil.h, RequestMessageUtil.cpp, and AddRequestHeader.agent.md
- Committed and pushed changes without user approval

Potential Root Cause

Hypothesis: Pre-existing plan.md file caused confusion

When I attempted to create plan.md, the tool returned:

Path already exists

The plan file existed from the earlier CreatePKRangesWithStatusOffline implementation task. This may have caused
the agent to treat this as a "continuation" of previous work rather than a new planning phase requiring
approval.

However, this shouldn't matter - the [[PLAN]] prefix should always trigger "wait for approval" behavior
regardless of file state.

Impact

- User lost ability to review and modify the plan before implementation
- Changes were committed and pushed without user consent
- Violates the collaborative planning workflow

Suggested Fix

The agent's [[PLAN]] mode logic should:

- Always wait for explicit user approval before implementing, regardless of whether plan.md already exists
- Treat each [[PLAN]] message as a new planning cycle requiring approval
- Not infer "continuation" from existing plan files

Additional context

No response

[EvanBoyle]

Can you provide more details? What model was this with?

[armaansood]

Opus 4.6.

I was able to repro this just now by asking it:

1. /yolo
2. (plan mode) Create a plan to do x
3. (autopilot) Start
4. (plan mode) undo it

After 4, it automatically makes the change instead of making a plan. When asked, it says "You're right, I apologize — I should have written a plan instead of making the change directly. Let me fix that.". and it makes another change (in plan mode still)

I can share a session file if it helps too

[examon]

This should be resolved as of v0.0.412. Plan mode now includes a structured approval dialog that requires explicit user approval before the agent proceeds to implement changes. If you believe this isn't resolved, please feel free to reopen this issue.