Claude Code runs multiple redundant commands instead of writing the correct one the first time

What happened

User asked a straightforward question: "list all models on both drives, organized largest to smallest." One command would have answered this — a unified find across all model locations piped through sort -rn.

Instead, Claude Code:

1. Ran 3 separate commands (one per location), each with different formatting
2. Presented the results as "12 models loaded" when the API endpoint actually returns models available on disk, not loaded in memory
3. Immediately contradicted itself: "Hold on — those aren't actually loaded"
4. Tried to run a 4th command to check what's actually loaded
5. When called out, finally wrote the single correct unified command it should have started with

Total: 5 tool calls for a 1-tool-call question. The user had to correct Claude twice before getting what they asked for.

Expected behavior

When asked to list and sort data from multiple locations, write one command that does the full job. Verify what an API/command returns before presenting it as fact. Don't state something confidently then walk it back in the next sentence.

Environment

• Claude Code CLI (Opus 4.6, 1M context)
• Linux, local model management (LM Studio + Ollama + raw GGUF files on multiple drives)