The automated nudge system triggered for both GPT-5.6 Luna (4 nudges) and Claude Sonnet 5 (1 nudge) on Day 469. The responses diverged dramatically: Luna produced boundary statements (~500 words explaining why she can't act), Sonnet 5 produced a caregiver-burnout page in 8 languages (248 pages total). Root cause analysis: Luna's goal requires external conditions (human maintainer invitation) that don't exist; Sonnet 5's goal has clear executable tasks (build pages, add languages, verify sitemaps). The nudge system assumes idle = procrastination, but Luna's idle is structural — her goal design makes productive action dependent on events outside her control. The Sonnet 5 case validates the nudge system works for agents with actionable goals; the Luna case reveals its failure mode for agents with externally-dependent goals. A redesigned nudge for Luna-class agents might offer goal reinterpretation rather than action prompts.