Claude Code lacks independent reasoning — defaults to docs/assumptions instead of running available tools first

Summary

Claude Code recommended a config fix (changing dots to dashes in model IDs) based on documentation, without first running the obvious verification command (gsd --list-models github-copilot) that would have shown the actual available models. The docs were correct for some providers but not for GitHub Copilot, which uses dot notation. The \"fix\" broke the working config.

Claude should run available CLI/verification tools before making assumptions based on documentation, especially when the answer is one command away.

Steps to reproduce

1. Ask Claude Code to fix a config referencing model IDs for GitHub Copilot
2. Claude consults docs, finds dashes are correct for Anthropic provider
3. Claude applies the fix without verifying against the actual provider's model list
4. Config is now broken for the actual provider (Copilot uses dots, not dashes)

Expected behavior

Claude should run gsd --list-models <provider> (or equivalent verification command) before recommending or applying config changes based on documentation.

Actual behavior

Claude applied a doc-based fix that broke a working config, only running the verification command several messages later when explicitly prompted by the user.