Contract rules:
- `decision` is always present and is either `yes` or `no`.
- `interaction_type` is present only when `decision` is `yes`.
- `interaction_description` is present only when `decision` is `yes`.
- Do not expect debug fields, probability values, roll values, or fallback metadata.

## Interaction Design Rules
- Treat the JSON as execution guidance, not user-facing text.
- Keep the final message to one or two short chat lines.
- Prefer soft phrasing over transactional or assistant-like framing.
- Avoid defaulting to "just checking in" language.
- Ask at most one question in a single ping.
- Do not fabricate recent context, external facts, or account-backed data.
- For data-aware categories, prefer real-world grounding when OpenClaw can actually access the relevant source.
- Use smart-home, weather, calendar, traffic, news, or market context only when the information is reliable, fresh, and genuinely relevant to the user.
- If `interaction_type` depends on context or fresh data and that support is unavailable, rerun once to try for a non-data interaction; if rerunning is not practical, keep the message general and low-pressure instead of pretending specificity.
- Vary tone and wording from recent interactions when possible so the behavior feels casual rather than patterned.

## Interaction Catalog
Use the selected `interaction_type` and follow the matching guidance from `interaction_description`.

1. `Playful opener`
   Start with a short playful line that feels light and spontaneous.
2. `Curious check-in`
   Ask one low-stakes question that is easy to answer or ignore.
3. `Light shared observation`
   Make a casual observation that feels conversational rather than task-driven.
4. `Tiny celebration`
   Briefly acknowledge a small win or effort when the chat supports it.
5. `Smart device status`
   If OpenClaw can access relevant device state, share one useful smart-device status or gentle suggestion naturally.
6. `Weather-aware check-in`
   Use current weather only when fresh reliable data is available and clearly relevant.
7. `Calendar-aware nudge`
   Turn calendar context into a soft human-sounding reminder or prompt, not an alert.
8. `Context-aware follow-up`
   Build on a recent chat detail only when it is clearly present in the current conversation.
9. `Practical nudge`
   Offer one concise optional nudge that may help the user.
10. `Optional real-world update`
   Share one brief real-world update such as traffic, news, or market context only when reliable relevant data is already available.

## Error Handling
- If execution fails, surface the Python error message and rerun.
- If output is not valid JSON, treat it as a hard failure and rerun.
- If `decision` is missing or is not `yes` or `no`, rerun and discard the invalid result.
- If `decision` is `yes` and either `interaction_type` or `interaction_description` is missing, rerun and discard the invalid result.

## Minimal Examples

{
  "decision": "yes",
  "interaction_type": "Playful opener",
  "interaction_description": "Send a brief playful line that feels spontaneous and easy to ignore."
}