[BUG] Opus over-investigates simple debug tasks, consuming disproportionate quota

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

What happened
I asked Claude (Opus 4.6) to debug two straightforward firmware issues:

A Modbus exception code 2 error shown by a PC debug tool
A current calibration offset bug
Both problems could be diagnosed with 2-3 targeted tool calls (one API query to see which register block was failing, one version check, one look at the calibration formula). Instead, Claude:

Spawned an Explore agent with "very thorough" breadth that read the entire register map, all Modbus handlers, EEPROM layout, PC tool session code, HTTP API handlers, Go register tables, and calibration code across both firmware and PC tool
Then proceeded to re-read many of the same files itself (uartjob.c, tables.go, session.go, devices.go, server.go, mcb_config.c, carrier_config.h, DataProcess.c, current_calibration.c)
Ran a full git show of a large commit diff
Read multiple grep results across the codebase
Only then made the actual diagnostic API call (curl /api/state) that immediately revealed the answer
This consumed roughly 40% of my daily quota on what should have been a quick diagnosis.

What Should Happen?

Expected behavior
For a targeted debug question ("tool shows error X, what's wrong"), Claude should:

Get the observable symptom first (one API call or one file read)
Form a hypothesis from the symptom
Read only the specific code needed to confirm
Fix
Not exhaustively read every related file in the project before looking at the actual error.

Error Messages/Logs

Steps to Reproduce

1. Open a medium-sized embedded project (firmware + PC tool, ~50 source files) in Claude Code desktop app
2. Ask Claude (Opus 4.6, auto mode) a targeted debug question: "my debug tool keeps showing Modbus exception code 2, and current calibration is off — what's wrong?"
3. Provide context that a live serial connection is available for querying
4. Observe that Claude spawns an Explore agent with "very thorough" breadth, reads 15+ files exhaustively, then re-reads many of the same files, before making the one API call that immediately reveals the answer
5. ~40% of daily quota consumed on a 2-3 tool call diagnosis

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

1.46388.4

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Environment
Claude Code desktop app (Code tab), Windows 11
Model: Opus 4.6 (claude-opus-4-6)
The project is an embedded firmware + Go PC tool (~medium codebase)
Auto mode was active

Additional context
I had previously told Claude not to overthink in past sessions, but the behavior persisted. The Explore agent call with "very thorough" search breadth was the biggest offender — it returned a massive report that consumed significant context, and most of it was unnecessary for the diagnosis.

I wanted him(opus 4.6) to report this bug, but instead, he committed my previous work. His intelligence has really taken a nosedive.
Opus 5 suffers from severe hallucinations, so I reverted to Opus 4.8—but that has bad hallucinations too. Now I’m using Opus 4.6, yet its intelligence is lacking and the hallucinations are still terrible.
GPT is far better than you guys these days, and I have a subscription to it. Two weeks ago, I really loved using Claude Code, but now I have no desire to use it; the hallucinations are so bad that I can't get any real work done—it’s mostly just a waste of time.