fix(agents): drop messaging-tool metadata from sensitive-stop return
Remove didSendViaMessagingTool, messagingToolSentTexts, messagingToolSentMediaUrls, messagingToolSentTargets, and successfulCronAdds from the sensitive stop_reason return path. Propagating these fields causes the safety-filter refusal payload to be suppressed when a messaging tool has already sent to the same target, leaving the user with no safety explanation. This aligns the sensitive error path with other early-return error branches (role_ordering, image_size) that also omit messaging-tool metadata.
This commit is contained in:
parent
ff954a711b
commit
c621974a9b
@ -1348,11 +1348,6 @@ export async function runEmbeddedPiAgent(
|
||||
systemPromptReport: attempt.systemPromptReport,
|
||||
error: { kind: "sensitive", message: errorText },
|
||||
},
|
||||
didSendViaMessagingTool: attempt.didSendViaMessagingTool,
|
||||
messagingToolSentTexts: attempt.messagingToolSentTexts,
|
||||
messagingToolSentMediaUrls: attempt.messagingToolSentMediaUrls,
|
||||
messagingToolSentTargets: attempt.messagingToolSentTargets,
|
||||
successfulCronAdds: attempt.successfulCronAdds,
|
||||
};
|
||||
}
|
||||
// Handle role ordering errors with a user-friendly message
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user