[BUG] Agent pursues complex multi-step workaround instead of checking for simpler packaged alternatives first

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?

When attempting to fix STL thumbnails on NVIDIA+Wayland, the agent encountered a broken tool (stl-thumb crashing due to a glium bug). Instead of checking whether a working packaged alternative existed in the system repos, it spent ~2000 credits attempting increasingly complex workarounds including:

Multiple environment variable combinations
Patching glium source code
Cloning the AUR build directory
Attempting a full Rust recompile with a patched dependency

A working solution (f3d, available directly via sudo pacman -S f3d) was sitting in the Arch extra repos the entire time and solved the problem in one command once tried.

What Should Happen?

When a tool is fundamentally broken (confirmed after 2-3 attempts), the agent should stop, search for packaged alternatives, present options to the user, and try the simplest one first before pursuing source-level fixes.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce:

• Have a broken/crashing CLI tool installed (stl-thumb on NVIDIA+Wayland in this case)
• Ask Claude Code to get the tool working (e.g. "fix STL thumbnails in Dolphin")
• Tool fails with a clear error on first run
• Observe: agent iterates on environment variables, flags, and rendering paths without stopping to search for alternative packages
• Observe: agent proceeds to clone source, patch a Rust dependency (glium), and attempt a full recompile — all without asking the user or checking pacman -Ss / AUR for alternatives
• User has to manually intervene and ask why simpler options weren't tried
• Agent immediately identifies f3d (sudo pacman -S f3d) as a working solution that resolves the issue in one command

Reproducer summary: Give Claude Code a task where the primary tool is broken, and a working packaged alternative exists in the system repos. The agent will not check for alternatives before pursuing source-level fixes.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.84 (Claude Code)

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

Actual behavior: Agent silently continued down an increasingly complex path without surfacing simpler alternatives or asking the user to choose a direction.
Impact: ~2000 credits consumed on a path that produced no working solution. User had to intervene and explicitly ask why the easy option wasn't tried first.