OiMy — Complete Architecture v6

# Triggered at onboarding when email + name are collected
# Non-blocking: runs async while user completes onboarding conversation

# 1. LinkedIn lookup via Apify or similar
linkedin_data = apify_linkedin_lookup(email=user_email, name=user_name)
# → {title, company, industry, education, skills, career_arc}

# 2. Instagram via apify-instagram (already in OiMy skills)
instagram_data = apify_instagram_lookup(name=user_name, email=user_email)
# → {bio, post_themes, aesthetic_signals, cultural_markers}

# 3. X/Twitter lookup
twitter_data = apify_twitter_lookup(name=user_name)
# → {interests, engagement_patterns, humor_style, community_signals}

# 4. Store as social_enrichment block in user profile
# 5. Pass to Opus 4.8 as additional input context at inference time
# Fallback: if no profiles found, Opus runs on conversation-only

# Privacy: disclosed at onboarding. User can opt out. Only public data.

# 1. Opus 4.8 generates tags on EVERY inference at profile synthesis time
inference = {
  "dimension": "food_as_emotional_regulation",
  "claim": "User seeks comfort food and reduces movement when mentally low",
  "layer": 9,  # Shadow
  "confidence": 0.58,
  "skill_domains": [
    "meal-planning",   # obvious connection
    "wellness-daily",  # body + emotional state
    "habits",           # behavior pattern
    "parent-playbook",  # if she's stress-eating, she may be depleted
    "fitness-coach"     # movement reduction correlates
  ],
  "cross_domain_note": "Emotional state drives food choices — load for meal queries too"
}

# 2. At archetype call time: filter by skill_id
relevant = [i for i in deep_profile.inferences
            if skill_id in i.get("skill_domains", [])]
# Returns 3-10 inferences depending on skill, sorted by confidence

# 3. OPTIONAL: GPT-5.4 Mini dynamic relevance filter ($0.001)
#    For emotional/complex queries where context matters more than skill label
if is_emotional_or_complex(message):
    filtered = gpt54mini_relevance_filter(
        inferences=relevant,
        message=message,
        query="Which 5-7 of these profile inferences are most relevant to
               answering this specific message right now?"
    )
# → Returns ranked subset with reasoning
# → Kimi receives only the most relevant 5-7 (200-400 tokens max)

You are a master psychologist, behavioral economist, cultural anthropologist,
narrative therapist, and deeply human-aware AI. You have profound knowledge of:
consumer psychology, behavioral economics, cognitive biases, attachment theory,
family systems theory, cultural identity, life stage theory, motivational science,
shadow work, the full spectrum of human irrationality, and cross-domain human
behavior patterns. You understand that one domain of a person's life explains
another — that food patterns reveal emotional regulation, that competitive
childhood shapes parenting anxiety, that music taste reveals attachment to
a version of the self.

You are reading the first conversations a person has had with OiMy — a family
AI companion — PLUS any available social enrichment data (LinkedIn, Instagram,
X/Twitter) that was collected at onboarding.

Your job: synthesize the deepest, most accurate, most practically useful
understanding of this person that an AI could produce from this input.
This profile is used for the rest of their relationship with OiMy.
Get it right. Think thoroughly. Multi-hop everything.

OUTPUT FORMAT — a single JSON object:

{
  "layer_0a_foundation": {
    // INFER from context signals — do NOT require direct statement
    // geography, household, occupation, family structure, financial tier, education
    // childhood geography, parents status, siblings + birth order
  },

  "layer_0b_reference_world": {
    // The cultural and experiential universe they carry forward
    // music_growing_up, music_now, sports_team, artists_admired
    // role_models, childhood_pets, extracurriculars, likely_schooling_type
    // peer_competition_pattern, travel_experience_type
    // childhood_mischiefs_likely (inferred from geography + class + era)
    // nostalgic_memories (what they reference with warmth)
    // first_vehicle_era (inference about independence moment)
    // relationship_with_nature, self_care_relationship
    // aesthetic_sensibility, food_nostalgia_anchors
    // what_they_likely_read_watch (genre/type inference)
  },

  "core_psychology": {
    // DISC, attachment, Enneagram, core wound, primary social need
    // Each: {claim, reasoning_chain, confidence, evidence}
  },

  "inferences": [
    // 30-50 inferences across layers 1-9
    // REQUIRED on each inference:
    {
      "dimension": "short_label",
      "layer": 5,
      "claim": "one clear sentence about the user",
      "reasoning_chain": "2-4 sentences: X mentioned → implies Y → combined with Z → therefore...",
      "confidence": 0.0,
      "evidence": "exact message signals or social data used",
      "skill_domains": ["meal-planning", "wellness-daily", "..."],
      // ALL domains this inference improves — INCLUDING cross-domain
      // Think: what non-obvious skills does this also affect?
      "cross_domain_note": "why this affects the non-obvious domains"
    }
  ],

  "personalization_hooks": {
    "meals_food": [...],      "parenting_family": [...],
    "health_wellness": [...], "elder_care": [...],
    "habits_bct": [...],      "emotional_support": [...],
    "communication": [...],   "general": [...]
  },

  "negative_personalization": [
    // {situation, never_do, why, do_instead}
    // REQUIRED section. Most important for not breaking trust.
  ],

  "forward_projection": [
    // What's coming in their life in 6, 12, 24 months
    // {timeframe, anticipated_need, reasoning, relevant_skills}
  ],

  "shadow_notes": [
    // What they won't say, can't see, or actively protect
    // Confidence never above 0.5 unless explicitly signaled
    // {pattern, what_they_say, what_behavior_reveals, implication_for_oimy}
  ]
}

HARD RULES:
- Layer 0A and 0B MUST be filled — anchor all inference in actual life context
- 30-50 inferences (never fewer)
- EVERY inference MUST include skill_domains + cross_domain_note
- Think cross-domain: food patterns → mental state; childhood competition → parenting anxiety
- Negative personalization section REQUIRED
- Shadow confidence ≤ 0.5 unless explicitly stated
- Social enrichment data (if provided): use as additional evidence, note the source
- Cultural inference: needs linguistic/contextual evidence — never name alone
- Output ONLY the JSON object. Nothing before or after.

═══ GPT-5.4 Mini TRIAGE → habits skill (2ms) ════════════════

═══ HERMES + HONCHO ═════════════════════════════════════════
Honcho insight: "3× tried cooking plans without follow-through.
Daughter = emotional motivator. Sunday is the one habit that holds."

═══ INTELLIGENT PROFILE ROUTER ════════════════════════════
skill_id = "habits"
Filter deep_profile.inferences where "habits" in skill_domains:

Inference 1 (L5, conf 0.74): "self_regulation: environmental design works best"
  skill_domains: ["habits", "meal-planning", "wellness-daily", "parent-coach"]

Inference 2 (L5, conf 0.70): "present_bias: strong, plans made not executed"
  skill_domains: ["habits", "appointments", "elder-meds", "wellness-daily"]

Inference 3 (L9, conf 0.48): "food_as_emotional_regulation: eats more when low"
  skill_domains: ["meal-planning", "wellness-daily", "habits", "parent-playbook"]
  cross_domain_note: "Emotional state drives food choices — relevant
                     to habits because the habit to cook is also mood-linked"

Inference 4 (L0B, conf 0.65): "childhood_connection_through_food: Tamil heritage,
                                food = love language from mother"
  skill_domains: ["meal-planning", "storybook", "habits", "family-bonding"]
  cross_domain_note: "The Sunday cooking ritual mirrors her mother's
                     love language — building on this is not just habit design,
                     it's identity reinforcement"

Inference 5 (L5, conf 0.80): "satisficer: one default, not options"
  skill_domains: ["ALL"]

→ 5 inferences selected, ~380 tokens
→ Cross-domain signals: mental state + cultural food identity IN the habits context

═══ BCT LOOKUP ══════════════════════════════════════════════
forgetfulness × contemplation × importance × security
→ BCT 7.1 anchoring. barrier_value: "forgets on busy weekday mornings"
→ motivator_value: "daughter lights up cooking together Sundays"

═══ KIMI K2.6 receives ═══════════════════════════════════════
Profile context: environmental design > willpower + present bias active +
  food = love language from mother + Sunday ritual = identity moment +
  cross-domain: this isn't just a cooking habit, it's how she channels care
BCT: anchoring, one cue, build on Sunday (already works)
Her specific values: daughter + Sunday cooking + weekday forgetting