hidden-thinking#

The scanner detects sessions where model thinking is hidden. Claude Code’s default settings replace thinking text with summaries, leaving empty thinking blocks in the transcript. The scanner records whether the newest thinking block is empty and opens an issue when it is.

#How it works

The note task finds the single newest thinking block across all scanned sessions and records whether its text is empty. The note pins the session line of the block. It also freezes the observation context, including the block’s timestamp, the model, and the showThinkingSummaries setting at observation time. The snapshot survives after the session file is deleted or the setting is changed.

The issue task reads the newest observation. When the block was empty, it opens an issue. The issue description explains the cause and the fix, which is setting showThinkingSummaries to true in ~/.claude/settings.json. It also lists known cases the setting does not cover.

A closed issue reopens when a later scan observes a new empty thinking block. This covers the case where the setting regresses after you fix it.

#What it writes

#Running

The scanner is in the default group, so it runs when you run gage scan without selecting scanners. To run it alone:

gage scan -s hidden-thinking