Three tools with clear, domain-specific descriptions (50-150 chars each) and well-formed input schemas. All parameters are typed and constrained. However, output schemas are not documented, LLMs cannot predict what fields to expect from responses. Tool names follow verb_noun convention (get_*). No error handling guidance or recovery hints. Descriptions lack WHEN-to-use context and prerequisites. No parameter descriptions beyond the schema itself.

• Document output schemas for all three tools. Example for get_attention_state: {"type": "object", "properties": {"mood": {"type": "string", "enum": ["focused", "frazzled", "away", "calm"]}, "switches_per_hr": {"type": "number"}, ...}}
• Add parameter descriptions in the schema. For offset_days: 'Number of days in the past (0 = today, 1 = yesterday, etc.). Defaults to 0 for today's metrics.'
• Add readOnlyHint: true to all three tools' input schemas to signal they are safe to retry and do not modify state.
• Expand tool descriptions with WHEN-to-use guidance. E.g., 'Call this BEFORE interrupting the user with questions or notifications. Returns one of: focused (batch non-urgent asks), frazzled (consolidate into one message), away (queue results), calm (normal interaction).'
• Add error handling guidance: 'If AttentionOS collector is not running, this tool will return an error. Ensure the desktop app is active and the local database (~/.attentionos/attn.db) is accessible.'