The shell commands the agent ran, read from the Bash tool calls of the
conversation.
A phrase in the transcript is not a command that ran: the skill text the
agent read and its own explanation of what it did not do both quote
commands. Only a tool-call part naming the Bash tool counts, which is
what the adapter emits with output: "messages".
The shell commands the agent ran, read from the
Bashtool calls of the conversation.A phrase in the transcript is not a command that ran: the skill text the agent read and its own explanation of what it did not do both quote commands. Only a
tool-callpart naming theBashtool counts, which is what the adapter emits withoutput: "messages".