Unclear Whether the revised text is used depends on a setting whose default is not stated.
What
EndConversation is a tool the model can use to end a conversation. This release prepares a revised version of its prompt:
- The EndConversation tool description and its confirmation re-read prompt appear in the bundle with a revised-prompt flag check. The text is largely about background forks and self-harm exclusions.
- The EndConversation hint in the list of deferred tools (tools loaded only when needed) can now use alternative wording: "for sustained user abuse directed at the assistant, after a warning, or when the user requests it". The default wording is unchanged.
- New stored feature value slots appear:
endConversationRevisedPrompt,connectorUploadFlagWaitandconnectorUploadFlagKeptFor. - Two query paths now pass a
sendSideQueryBetaoption through, defaulting to a helper call.
The stock ToolSearch description captured under this release is 953 characters, unchanged from the previous release.
Why
The default prompt is unchanged. The revised wording, which describes when the model may end a conversation, applies only where the flag check turns it on.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
Whether the revised text is used depends on a setting whose default is not stated.