Claude hallucinates on post-cutoff tools instead of using WebSearch proactively

Summary

When Claude Code encounters an unfamiliar tool released after its training cutoff, it guesses based on name similarity instead of immediately using its available WebSearch tool. This produces confident wrong answers that waste multiple turns to correct.

Repro

1. Ask Claude Code to help configure a tool released after August 2025 (tested with Google Antigravity IDE, released November 2025)
2. Claude does NOT use WebSearch
3. Claude guesses based on name similarity and gives wrong advice
4. User corrects. Claude still does not search. Continues guessing.
5. Claude only searches after explicit user callout

Root cause

Confidence calibration issue: if a tool name sounds familiar, the model does not trigger search even when it should. Knowledge cutoff stops being an excuse when WebSearch is available and not being used.

Expected behavior

Any unrecognized tool, IDE, library, or service name should trigger an immediate WebSearch before the model attempts to answer. The model should treat uncertainty about a tool's identity as a mandatory search trigger, not a reason to guess.

Real impact

• Incorrect configuration advice given and acted on
• Multiple turns wasted correcting hallucinated answers
• User confidence in the agent degraded for the remainder of the session

Environment

Claude Code v2.1.90 · Sonnet 4.6 · macOS Darwin arm64 25.3.0
Tool tested with: Google Antigravity IDE v1.20.6 (VSCode OSS fork, released November 2025)