OiMy — Complete Architecture v4

# After every turn: store it (honcho builds its user model incrementally)
honcho.record_turn(user_id, user_message, assistant_response, skill_id="habits")

# Before every archetype response: get synthesized inference
insight = honcho.get_insight(user_id, query="what do we know about this user's
  habits, preferences, and what motivates them?")
# → Honcho calls its own LLM on all stored turns and returns:
# "User has mentioned wanting to cook 3× but hasn't followed through.
#  They consistently bring up their daughter as a source of motivation.
#  Prefer practical, step-by-step guidance. Respond positively to
#  being told they've done well. Likely have limited time in mornings."
# This string is injected directly into the Kimi K2.6 prompt.

# Trigger: user idle 30+ min OR explicit goodbye signal
# One GPT-5.4 Mini call processes the whole session transcript

prompt = """Extract structured facts from this conversation. Return JSON.
{
  "people": [{"name": "...", "relationship": "...", "facts": ["..."]}],
  "preferences": [{"category": "...", "value": "...", "confidence": 0.0-1.0}],
  "barriers": [{"category": "...", "specific_value": "...", "context": "..."}],
  "motivators": [{"category": "...", "specific_value": "...", "context": "..."}],
  "events": [{"description": "...", "date": "..."}],
  "skip_if_already_known": true
}
Only extract NEW information not already in the entity store."""

# Example output from session about cooking habits:
{
  "people": [{"name": "Priya", "relationship": "daughter", "facts": ["loves cooking together on Sundays", "celiac diagnosis May 2026"]}],
  "barriers": [{"category": "forgetfulness", "specific_value": "forgets to pack school lunches during busy weekday mornings", "context": "repeated across 3 turns"}],
  "motivators": [{"category": "belonging", "specific_value": "daughter lights up when they cook together on Sundays", "context": "mentioned unprompted as bright spot"}],
  "preferences": [{"category": "cooking", "value": "quick weeknight meals under 30 min preferred", "confidence": 0.85}]
}

# Barrier: category + specific value + context
{
  "user_id": "sarah-8040",
  "signal_type": "barrier",
  "category": "forgetfulness",                 ← BCT lookup dimension
  "specific_value": "forgets school lunches     ← passed to Kimi
                     during busy weekday mornings",
  "confidence": 0.82,
  "evidence_count": 4,
  "source": "entity_extraction_session_12"
}

# Motivator: category + specific value
{
  "signal_type": "motivator",
  "category": "belonging",                       ← EMA dimension
  "specific_value": "daughter lights up when       ← passed to Kimi
                     they cook together on Sundays",
  "confidence": 0.88
}

# Entity: person with facts
{
  "signal_type": "person",
  "name": "Priya",
  "relationship": "daughter",
  "facts": ["celiac diagnosis May 2026", "loves Sunday cooking together",
             "8 years old", "picky about textures"]
}

# BCT selection result + specific values assembled into prompt block
BCT_CONTEXT = """
Behavior Change context for this response:
• Technique: BCT 7.1 — Prompts/cues (anchoring NLP pattern)
• Apply: Help user identify a specific environmental cue that will trigger 
         the cooking behavior. Keep to one concrete cue.
• Barrier to address: FORGETFULNESS
  Specific: "forgets school lunches during busy weekday mornings"
  → Design the cue around this exact trigger point, not generic forgetfulness.
• Motivator to leverage: BELONGING  
  Specific: "daughter lights up when they cook together on Sundays"
  → Reference this explicitly — it's their strongest lever.
• Entity context: Priya (daughter, 8, celiac, loves Sunday cooking)
• Don't say: "you struggle with forgetfulness" — never label the user
• Do: tie the cue to Sunday cooking (already happening) as foundation
"""

# 1. Classify query domain + specific question
domain = classify_expert_domain(skill_id, user_input)
# → {skill: "occupational-therapy", area: "sensory_processing", age: "4yo"}

# 2. Retrieve from vector knowledge base (per domain)
kb = VectorKnowledgeBase(domain=skill_id)
docs = kb.retrieve(user_input, top_k=4)
# → ["OT activity: sensory bin with rice for tactile avoidance...",
#    "Research: heavy work activities reduce sensory overwhelm in 3-5yo...",
#    "Protocol: proprioceptive input before fine motor tasks..."]

# 3. Safety gate — always runs regardless of RAG result
risk = assess_risk(user_input)
gate = physical_gate.assess(skill_id, category, user_input, context)

# 4. Kimi K2.6 generates response with retrieved context
prompt = f"""
Expert domain: {skill_id} | Safety: {risk}
Retrieved knowledge:
{docs}
User question: {user_input}
User context: {profile_context}
Rules: Always include disclaimer. Never diagnose. Escalate if emergency.
"""

{
  "barrier": "forgetfulness",
  "stage": "contemplation",
  "hughes_need": "importance",
  "decision_style": "security",
  "primary_bct": "7.1 Prompts/cues",
  "secondary_bct": "8.1 Behavioral practice/rehearsal",
  "nlp_pattern": "anchoring",
  "opener": "When you set up that visual reminder in the kitchen,
             you'll notice how naturally the habit starts to stick.",
  "rationale": "Anchoring creates a reliable trigger (security need),
               validated by their importance as family caregiver."
}
// Then technique file loaded: data/technique_directory/bct_7_1.json
// Contains: gemma4_prompt_text (now kimi_prompt_text), dos/donts,
//            example_exchange, best_paired_with, family_context_notes
// And NLP file: data/nlp_directory/nlp_anchoring.json
// Contains: origin, when_to_use, gemma4_prompt_text, example_openers

# Barrier detection: keyword first, GPT-5.4 Mini fallback
barrier_category = detect_barrier_category(message)
# → "forgetfulness" (from keywords: "keep forgetting", "always forget")
# Specific value extracted at session end by entity extraction

# Motivator EMA update: Hughes decision style
motivator_ema = update_hughes_ema(user_id, message)
# security: 0.72↑  belonging: 0.61→  autonomy: 0.48→
# (EMA smoothed — doesn't flip on a single message)

# Goals: regex + periodic LLM extraction
goals = update_goals_inferred(user_id, message)
# family_cooking: {detail: "cook at home more", conf: 0.75, evidence: 3}

# Stage of change: classifier runs every 15 msgs
stage = classify_stage(user_id, recent_messages)
# contemplation → preparation → action → maintenance

# 07:00 local: morning brief · 19:30 local: evening closeout
SignalCollector.collect()   # calendar · meds · school · meals · habits · open loops
DailyState.compile(signals) # 12 fact types · VisibleScope · NudgeBudget
OpportunityDetector.score() # 5 positive × 3 penalty · budget-gated · suppression
BriefComposer.compose()     # structure = deterministic · prose = Kimi K2.6
# → delivered via Hermes → Telegram

══ PRE-FILTER ════════════════════════════════════════════════════════
No emergency · No injection · No active thread → pass through. Cost: $0. Time: <1ms.

══ GPT-5.4 Mini TRIAGE (2ms, cached) ════════════════════════════════
17KB prompt already cached. Message contains behavioral change intent + emotional hook.
Path → CONVERSATIONAL (not scripted: not a simple weather/reminder/list request)
Skill → {"skill": "habits", "confidence": "high"}

══ HERMES (Kimi K2.6) ═══════════════════════════════════════════════
SOUL.md loaded. Tool Search: oimy-engine skill selected (habits query).
Honcho get_insight() called:
  "User has tried cooking plans 3× in prior sessions, never followed through.
   Consistently references daughter as emotional anchor. Messages are terse —
   prefer numbered steps. Likely time-constrained on weekdays. Responds well
   to plans that build on existing habits (Sundays already happen)."
Context passed to engine.

══ CONTEXT PREPARATION ═══════════════════════════════════════════════
DEEP_PROFILE: high-C · anxious attachment · needs validation before solution
Entity store:
  - daughter: Priya, 8, celiac, loves Sunday cooking (extracted session 12)
  - barrier: forgetfulness → "forgets school lunches during busy mornings"
  - motivator: belonging → "daughter lights up when cooking together Sundays"
goals_inferred: family_cooking [0.75, 3× evidence]
stage_of_change: contemplation (repeated intent, no sustained action)

══ BARRIER + SIGNAL DETECTION ════════════════════════════════════════
Keywords: "keep meaning to", "always forget" → barrier_category: "forgetfulness"
Specific value noted in message: "when we cook together on Sundays" → motivator lever
EMA check: belonging = 0.61 (rising), security = 0.72 (dominant)

══ BCT MATRIX LOOKUP (0ms, dict) ════════════════════════════════════
key = (forgetfulness, contemplation, importance, security)
→ primary_bct:   "7.1 Prompts/cues"
→ secondary_bct: "8.1 Behavioral practice/rehearsal"
→ nlp_pattern:   "anchoring"
→ opener:        "When you set up that visual cue, you'll notice how
                  naturally the habit starts to stick."

Technique file: bct_7_1.json
→ prompt_text: "Guide user to identify ONE specific environmental cue that
                will trigger the behavior. Ask: what's already in their
                environment at the right time? Keep to one concrete cue."

══ KIMI K2.6 PROMPT (assembled) ════════════════════════════════════
System includes:
  - SOUL.md: OiMy persona
  - BCT guidance: BCT 7.1 anchoring, with exact specific values:
    · barrier specific: "forgets school lunches on busy mornings"
    · motivator specific: "daughter lights up cooking together Sundays"
    · entity: Priya (daughter, celiac, loves Sunday cooking)
  - Honcho insight: "3× tried cooking plans, Sundays already work for her"
  - DEEP_PROFILE: high-C, validate first, numbered steps

══ RESPONSE (Kimi K2.6) ════════════════════════════════════════════

# GPT-5.4 Mini processes this session's transcript
New facts extracted:
  - motivator confirmed: "daughter lights up when cooking together on Sundays" → stored
  - barrier confirmed: "forgets things on busy weekday mornings" → evidence_count +1
  - preference inferred: "prefers Sunday as prep day" [confidence: 0.78]
  - Priya: new fact added: "helps with Sunday cooking"
  
Honcho record_turn() called for each exchange in session.
→ Honcho's user model updated. Next get_insight() will reflect these patterns.