What kind of decision the agent made, in the vocabulary the corpus uses.
These are deliberately the decision kinds, not the nextStep.type values — taskComplete is a
decision but not a step type, and Retry/ForEach/While are control flow the corpus does not
assert on. Anything the normalizer cannot place lands in other, which is a legitimate
observation rather than an error.
Type Declaration
"action"
"subAgent"
"chat"
"taskComplete"
"payloadChange"
"other"
"unparseable"
Output existed but could not be read as a decision — the malformed class.
"empty"
No usable output at all (empty content, no tool calls).
What kind of decision the agent made, in the vocabulary the corpus uses.
These are deliberately the decision kinds, not the
nextStep.typevalues —taskCompleteis a decision but not a step type, andRetry/ForEach/Whileare control flow the corpus does not assert on. Anything the normalizer cannot place lands inother, which is a legitimate observation rather than an error.