[MODEL] Explore subagent being used to solve the problem

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

Other unexpected behavior

What You Asked Claude to Do

There is a nuanced but commonly occurring issue with Opus 4.6 misusing the intended purpose of the Explore subagent. Opus often asks the Explore agent to solve the actual problem instead of identifying areas of code. (i.e. Attempting to offload higher-level thinking to Explore)

What Claude Actually Did

Here is an example from our project:

1. The user asks Claude if there is a bug in how image paths are being captured and then saved to the SQLite database if the application were to be running on Windows
2. Claude doesn't know where this code exists, so instead of searching the codebase itself, it launches an Explore agent. (✅ This is good. An agent should be used here.)
3. Claude prompts the Explore agent as follows:

I need to understand how image file paths are gathered in this codebase. Please explore these areas thoroughly:

1. Backend: Look at the /api/pipeline/add-images endpoint in collect/backend/services/pipeline/route.py. How does it call pipeline_db.add_images()? What does it return to the frontend?
2. Frontend: AddImagesTab — Read collect/frontend/src/components/add-images/AddImagesTab.tsx completely. How does it handle the "Process" button and what data does it receive from the API after adding?

_**3. How would this returned data (specifically the full image path) interact with a Windows file system?

1. How would these paths differ if pipeline_db.add_images() were run within a Windows environment?**_

Be THOROUGH in your analysis. I will need to make edits.

1. Claude receives the output from Explore and does not read any files on it's own, relying on the determination that Explore returned to it.

Expected Behavior

Claude should NOT have relied on an Explore agent (running the Haiku model) to diagnose the potential issue. Instead, it should have simply asked Explore to... explore... the codebase, find where the filepaths are gathered, and return the information.

Then, Claude should have made its _own_ determination if the logic would cause issues on various OS's.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Ask Claude a question (of any complexity) about your codebase that does not require it to modify code in the same run.
2. Examine the prompt it sends to the Explore agent.
3. Note if this prompt instructs the agent to do more than simply exploring the codebase and identifying relevant patterns.

Claude Model

Opus

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.62

Platform

Anthropic API

Additional Context

I have attempted to mitigate this issue by overriding the Explore agent and providing modified instructions for execution, but none have seemed to work right. Opus does not seem to pick up on the nuance of not instructing Explore to solve the problem it was handed.