{{ if has_limit }}Keep prose to at most {{ max_prose_chars }} visible characters, then{{ else }}Keep prose brief, then{{ end }} {{ if has_tools }}{{ if native_mode }}either make concrete progress with the provider tool channel{{ else }}either make concrete progress with a well-formed <tool_call> block{{ end }}{{ if allow_done_sentinel }}{{ if done_sentinel }}, switch phase, or {{ if native_mode }}include `{{ done_sentinel }}` exactly once{{ else }}emit <user_response>final answer</user_response> plus a <done> block{{ end }} if the task is genuinely complete.{{ else }}, switch phase, or emit the final user-facing answer and stop calling tools if the task is genuinely complete.{{ end }}{{ else }} or switch phase if the workflow allows it.{{ end }}{{ else }}either make concrete progress in your reply{{ if allow_done_sentinel }}{{ if done_sentinel }}, switch phase, or include `{{ done_sentinel }}` exactly once if the task is genuinely complete.{{ else }}, switch phase, or emit the final user-facing answer if the task is genuinely complete.{{ end }}{{ else }} or switch phase if the workflow allows it.{{ end }}{{ end }}{{ if prose_too_long }} Your last response spent too much budget on prose.{{ end }}{{ if native_mode }} Use the provider tool channel only; handwritten tool-call text is invalid in this transcript.{{ end }}