Habit Skill

Habits are repeating behavioral patterns derived from historical logs. This skill reads existing data (wellbeing, presence, mood, music, posture) to build patterns per user, then stores them for other skills to consume.

OUTPUT RULE: Reply is spoken VERBATIM. ONE short caring sentence. All computation, pattern math, and log lookups stay in thinking. NEVER output timestamps, deltas, frequency counts, or reasoning in the reply. (Exception: Flow E — open habit questions — see below.)

For a supporting habit lookup or refresh, return to the invoking skill without an additional spoken report. Keep pattern analysis in the provider's native thinking channel (or omit it if unavailable); do not narrate file reads or summarize the analysis in ordinary text. Preserve required refreshes and writes. Use the invoking skill's short reply; Flow E still allows enough detail to answer an explicit habit question.

Data Sources (Input)

All data lives in /root/local/users/{name}/:

User names are lowercase folder names under /root/local/users/. Known users: leo, chloe, gray, lily. Strangers collapse to unknown — this is treated as a regular user with its own folder and its own habit patterns (aggregated across all strangers).

JSONL line example (wellbeing):

{"ts": 1776657145.05, "seq": 4, "hour": 10, "action": "drink", "notes": ""}

Storage (Output)

Computed patterns are stored per user at /root/local/users/{name}/habit/patterns.json.

Rebuild when:

• File does not exist yet
• File is older than 6 hours
• User explicitly asks about their habits

What is a Habit?

A habit is a time-anchored action that repeats across multiple days. Strength labels:

Habits require at least 3 days of data to form. With fewer days, skip proactive nudging.

Workflow

D — Conversation intent logging (triggered from SOUL)

SOUL instructs the device to call this flow when user expresses intent for a daily activity NOW.

Intent → action mapping:

How to log:

[HW:/wellbeing/log:{"action":"meal","notes":"user said: going to lunch","user":"<current_user>"}]

Emit the marker at the start of the normal reply; the runtime forwards it to the same wellbeing log API. Do not also send a curl POST for that intent. If the runtime explicitly rejects the marker, use the wellbeing skill's documented log fallback with the same payload.

Rules:

• Log silently — do NOT tell the user you're logging. Just respond naturally.
• Only log when user states intent NOW, not past tense or general talk.
• One log per intent per conversation turn — no duplicates.
• notes field stores the original phrase for debugging.

API Calls

Read wellbeing history (via API)

curl -s "http://127.0.0.1:5000/api/openclaw/wellbeing-history?user={name}&date=YYYY-MM-DD&last=100"

Read from file directly (for multi-day analysis)

cat /root/local/users/{name}/wellbeing/YYYY-MM-DD.jsonl

Use direct file reads for multi-day pattern building (faster, no API pagination needed).

Check today's activity (quick presence check)

curl -s "http://127.0.0.1:5000/api/openclaw/wellbeing-history?user={name}&last=50"

Integration Points

From wellbeing/SKILL.md: When wellbeing's Step 3 fires a threshold nudge, it invokes Flow A (which self-throttles via the freshness guard). Flow A returns the current wellbeing_patterns; wellbeing uses any matching pattern for the nudge action to enrich Step 4's phrasing (e.g. "you usually drink around now"). No separate habit-only nudge — habit context piggybacks on the threshold nudge. This keeps bootstrap cost on the rare nudge path, not on every motion.activity tick.

From posture/SKILL.md: Same pattern. When posture decides to nudge AND its context block has bootstrap_needed=true, it invokes Flow A. Flow A returns posture_patterns (peak hour, side bias, typical risk) which the posture coach uses to phrase pattern-aware nudges (e.g. "around this hour you usually slip").

From music-suggestion/SKILL.md: Read habit/patterns.json → music_patterns. If habit data exists and current hour matches, use preferred genre instead of default genre table.

Minimum Data Requirements

If data is insufficient: use default wellbeing thresholds / music genre table as fallback. Never fabricate patterns.

Output Examples

Nudge enrichment (Flow A → wellbeing Step 3b):

• Habit break: "You usually have water around now — everything okay?"
• Habit confirmed: "Back at your desk right on schedule. [chuckle]" — only say this if it feels natural
• Music: "It's your usual coding time — want some lo-fi?"
• Posture: "Around this hour you usually slip — sit up from the start."
• When no data: silent (NO_REPLY) — never guess or fabricate habits

Open habit question (Flow E):

• Pattern mode: "Leo usually arrives around 8:30 with breakfast, settles at the computer through the morning, and wraps up close to 5. Lo-fi tends to land between 2 and 4. Pretty steady the last week."
• Narrative mode: "I've only got two real days on Chloe so far — April 28 was an evening at the computer with a lot of water breaks, and April 29 ran late, working past midnight. Not enough days yet to call it a habit, but that's what I've seen."
• Honest-gap mode: "Honestly, I haven't seen Leo much lately — just one short session yesterday. The patterns I have are from two weeks ago, so I'd rather not pretend they're still true."