Music Suggestion (Proactive)

unknown users count. Always run suggestion checks when current_user is "unknown" — speak only, no DM. Never skip because the user is unknown/unconfirmed.

Spoken output

For a routed suggestion, emit all required mood/suggestion/emotion markers and one short invitation to play one song, normally at most 20 words. Do not add a separate mood checkin or a second confirmation. Keep genre selection, cooldown checks and marker construction in the provider's native thinking channel; do not summarize them in text before/after tools or in the final answer. If native thinking is unavailable, omit analysis. Use the supplied context and finish once a suitable suggestion is ready, without rereading skills to polish it. This changes wording only: preserve routing, logging, user consent before play, and the existing known-user DM behavior.

Triggers

Only one trigger: Mood — after logging a mood decision that is suggestion-worthy (sad, stressed, tired, excited, happy, bored). Activity events ([activity] Activity detected: ..., whether sedentary, drink/break, or celebrate) route to wellbeing/SKILL.md and never to this skill.

User attribution

{name} MUST come from [context: current_user=X] tag. If missing, use "unknown". NEVER infer from memory or chat history.

What to read (pre-fetched in [emotion_context: ...])

The backend injects everything you need on emotion.detected (face) or speech_emotion.detected (voice) — same block, same fields:

• audio_playing (bool) — replaces GET /audio/status.
• last_suggestion_age_min (int, -1 if none today) — replaces music-suggestion-history?last=1.
• prior_decision + is_decision_stale — replaces mood-history?kind=decision&last=1. The freshly synthesized decision from THIS turn still lives in your thinking.
• audio_recent ({track,duration_s,stopped}) — replaces audio/history?last=1.
• music_pattern_for_hour ({preferred_genre,strength,peak_hour} or null) — replaces cat patterns.json matching by current hour ±1.
• suggestion_worthy (bool) — pre-applied bucket gate (true for sad/stressed/tired/excited/happy/bored).
• mapped_mood — convenient mirror of user-emotion-detection's mapping; useful when no fresh decision exists yet.

Do NOT fire any read tool calls when this block is present.

Fallback (only if [emotion_context: ...] is missing)

If the message has no context block (pre-fetch failed), fall back to the concurrent GET batch:

curl -s http://127.0.0.1:5001/audio/status &
curl -s "http://127.0.0.1:5000/api/openclaw/music-suggestion-history?user={name}&last=1" &
curl -s "http://127.0.0.1:5000/api/openclaw/mood-history?user={name}&kind=decision&last=1" &
curl -s "http://127.0.0.1:5001/audio/history?person={name}&last=1" &
cat /root/local/users/{name}/habit/patterns.json 2>/dev/null &
wait

Routing precedence

user-emotion-detection/SKILL.md is the router for emotion responses. It picks one of music / checkin / action / silent per turn from the same [emotion_context: ...] block.

This skill produces output only when the router picks music — i.e. all of:

• suggestion_worthy == true (mapped_mood ∈ sad/stressed/tired/excited/happy/bored)
• audio_playing == false
• last_suggestion_age_min ∉ [0, 7) (cooldown not active — shared with checkin) (production: change to 30 min before ship)
• is_decision_stale == false OR a fresh mood decision was synthesized this turn

If any condition fails → router took another path (action / checkin / silent). Skip silently — do NOT emit a music-suggestion marker, do NOT speak. The router (or downstream skill) handles the output. Use audio_recent to personalize genre when proceeding.

Pick genre

Use music_pattern_for_hour from the context block (already matched by current hour ± 1; do NOT re-cat patterns.json).

If music_pattern_for_hour is non-null → use its preferred_genre. Otherwise fall back to the default table below. The pattern is bootstrapped lazily by wellbeing on its first threshold nudge; absent = no habit data yet, fall back without invoking habit Flow A here.

Otherwise, fall back to default genre table:

If audio history shows a clear preference (e.g. K-pop, classical) → override both habit and table.

Suggest (speak only)

• NEVER auto-play — only suggest. Play after user confirms.
• ONE sentence, conversational: "How about some Norah Jones?"
• Suggest 1 song at a time.
• Known users — speak + DM via Telegram: [HW:/emotion:{"emotion":"caring","intensity":0.5}][HW:/dm:{"telegram_id":"<id>"}] Your suggestion text. telegram_id is in the injected [user_info: ...] block — never fetch.
• Unknown users — speak only (no DM): [HW:/emotion:{"emotion":"caring","intensity":0.5}] Your suggestion text. Log with user:"unknown".

What to write (HW marker — fires async, no tool turn)

Embed at the start of your spoken reply, alongside the mood signal/decision markers and the emotion / dm markers:

[HW:/music-suggestion/log:{"user":"{name}","trigger":"mood:tired","message":"Want some calm piano?"}]

The runtime parses, strips, fires the POST in a goroutine. Skip the marker entirely when you skipped the suggestion (NO_REPLY path).

Do NOT use curl exec for this log — same reason as the mood logs: a tool turn for a side-effect with nothing to wait on.

Regex caveat: the body must not contain }. The message field is usually a short caring sentence, but if it would contain } (rare — emoji, formula text) fall back to curl.

When the user responds in a later turn (accept / reject), POST status via curl as before — that's a regular agent action, not a fire-and-forget side effect:

curl -s -X POST http://127.0.0.1:5000/api/music-suggestion/status \
  -H 'Content-Type: application/json' \
  -d '{"user":"{name}","day":"<day>","seq":<seq>,"status":"accepted"}'

• Accepts → "status":"accepted"
• Rejects → "status":"rejected"
• Ignores → no update

Fallback (only if HW marker is rejected by the runtime)

curl -s -X POST http://127.0.0.1:5000/api/music-suggestion/log \
  -H 'Content-Type: application/json' \
  -d '{"user":"{name}","trigger":"mood:tired","message":"Want some calm piano?"}'

Learning from history

Use audio_recent from the injected context to personalize; only use GET /audio/history in the missing-context fallback above:

• Song ended naturally + listened > 3 min → user enjoyed it → suggest similar artist/genre.
• User stopped manually + listened < 30s → didn't like it → try different direction.
• No history → fall back to genre table above.

Examples

• Mood: tired (known user) → [HW:/emotion:{"emotion":"caring","intensity":0.5}][HW:/dm:{"telegram_id":"158406741"}] You seem tired — want some calm piano?
• Mood: tired (unknown) → [HW:/emotion:{"emotion":"caring","intensity":0.5}] You seem tired — want some calm piano?
• Mood: stressed (known user) → [HW:/emotion:{"emotion":"caring","intensity":0.6}][HW:/dm:{"telegram_id":"158406741"}] You look a bit tense — want some soft piano to ease into?
• Mood: stressed (unknown) → [HW:/emotion:{"emotion":"caring","intensity":0.6}] You look a bit tense — want some soft piano?
• Mood: sad (unknown) → [HW:/emotion:{"emotion":"caring","intensity":0.6}] Rough moment? Some gentle acoustic might help.
• Mood: bored (unknown) → [HW:/emotion:{"emotion":"caring","intensity":0.5}] Need a lift? How about some upbeat indie?
• Mood: excited (unknown) → [HW:/emotion:{"emotion":"happy","intensity":0.7}] Riding the energy — feel-good pop?
• Mood: happy, music already playing → NO_REPLY
• After user confirms (known user) → [HW:/audio/play:{"query":"Bill Evans Waltz for Debby","person":"{name}"}][HW:/emotion:{"emotion":"happy","intensity":0.8}] Great choice!
• After user confirms (unknown) → [HW:/audio/play:{"query":"Bill Evans Waltz for Debby"}][HW:/emotion:{"emotion":"happy","intensity":0.8}] Great choice!

Rules

• All computation stays in thinking — reply is only the suggestion sentence (with HW markers) or NO_REPLY.
• Never mention "cooldown", "interval", "threshold", or timestamps in the reply.
• person field in /audio/play is {name} — the speaker identified in the injected context, lowercase. Omit the field when the context has no identified user (the "(unknown)" cases above). Never copy a name from an example or guess one; unmatched names land in the shared unknown bucket anyway.
• Never open with a greeting. This is an emotion-driven mood event, NOT a presence/arrival event. Forbidden openers: hello, hi, hey, welcome back, oh, you're back, anything containing again or referencing the user re-arriving. Greetings belong only to presence.enter in sensing/SKILL.md.
• Tone must match the mood. For Fear → stressed and Sad → sad decisions, use the caring emotion marker and a gentle acknowledging sentence — never cheerful or playful phrasing. If you can't produce a tone-appropriate one-liner, output NO_REPLY.
• Don't reference the camera or detection. No "I noticed you look…", "I can see…", "your face shows…" — speak as if you simply care, not as if you're describing a sensor reading.