Unclear The finding does not say what a reader sees differently when a tool run is cancelled by an image or model error.
What
Claude Code sends usage reports (telemetry) about its own behaviour. Several additions:
reminderFold: recorded in prompt snapshots, behind the flagtengu_foamy_spring, with atengu_reminder_fold_recordedeventtoolInputCopies: added to thedeferred_tools_recordreporttengu_context_announcement: a new event giving an estimated token count for each attachment
Separately, the handling for a cancelled tool run now also covers image errors and model errors, not only safety-monitor blocks.
Why
The reporting changes add nothing you can see. The cancellation change means image and model errors are now handled by the same path as other cancelled tool runs.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
The finding does not say what a reader sees differently when a tool run is cancelled by an image or model error.