The routing prompt is the architectural linchpin. It decides 60-70% goes to DeepSeek vs 30-40% to Kimi. Must be precise about the boundary.
| Category | Examples | Why DeepSeek |
|---|---|---|
| Conversational / emotional | "I'm exhausted," "ugh mornings," "feeling overwhelmed," "good morning," "thanks" | No specialized expertise needed. Warmth + context = best response. |
| Simple food/activity suggestions | "what should we have for dinner," "any activity ideas for the weekend," "quick lunch for Lily" | With context of allergies/preferences, DeepSeek can handle this well. |
| Opinions and parenting philosophy | "what do you think about gentle parenting," "is screen time really that bad," "should I let him quit soccer" | DeepSeek has opinions and can engage at depth. No BCT needed. |
| Follow-up and continuity | "did that work btw," "Lily loved the pasta," "Ethan is still refusing" | Conversational context is the main ingredient. DeepSeek excels here. |
| Simple health / reassurance | "Lily has a small cold," "Ethan seems tired today" | Context-aware reassurance + one referral sentence = done. |
| Light planning / logistics | "help me think through this week," "what should I prioritize," "remind me about X" | Conversational planning doesn't need full BCT pipeline. |
| Personal connection | "you've been helpful lately," "I feel like I keep failing at this," "just needed to vent" | This IS the relationship. DeepSeek must own it. |
| Category | Examples | Why Kimi |
|---|---|---|
| BCT behavior change | "I want to build a habit of X," "help me stick to Y," "why do I keep failing at Z" | Requires BCT technique selection, stage classification, identity framing |
| Complex medical | "Lily has fever 102 for 3 days + ear pain + not eating," "Ethan had a seizure" | Requires expert archetype, mandatory referral, clinical pattern recognition |
| Multi-constraint planning | "week meal plan: peanut allergy + picky + low-carb + 5 days + 30min max," "monthly schedule with school + sports + appointments" | Requires PlannerArchetype with constraint satisfaction |
| Expert domains | IEP questions, occupational therapy, speech therapy, elder medication management | Requires specialized skill archetypes with domain knowledge |
| Structured goal-setting | "I want to create a family fitness plan," "help me set up a reading routine" | Benefits from Journey archetype, milestone tracking, BCT |
What DeepSeek receives at runtime for every response it handles. Total: ~800-1200 tokens well within DeepSeek's context window.
[SYSTEM PROMPT — Opus 4.8 Master Prompt] Oi's identity, personality, quirks, backstory (~300 words) How to use injected context (~150 words) Tone calibration rules (~100 words) Anti-sycophancy rules (~50 words) 10 BAD/GOOD examples (~200 words) [FAMILY CONTEXT — injected runtime, T1] Child: Ethan (11) — peanut allergy, soccer Tuesdays Child: Lily (7) — picky eater (beige foods), ear infection Pet: Biscuit (labradoodle) David: low-carb, works late 3 days/week [PROFILE — injected runtime, T2+T3] Communication: S-type · Social need: belonging · Attachment: anxious Key inferences: achievement anxiety (projects onto Ethan), fear-based procrastination, strongest when given one concrete next step [HONCHO SYNTHESIS — injected runtime] "User engages best in evenings. Soccer + school are her main stress points this week. Responded well to the chicken idea last session. Tends to spiral when given too many options." [INSTAGRAM SIGNALS — injected runtime, when available] Follows: meal prep accounts, family lifestyle, home organization Aesthetic: warm/neutral tones, practical vs aspirational Recent interest: quick weeknight dinners, kids' sports [CONVERSATION HISTORY — last 8 turns as messages[]] User: ugh mornings are a disaster again Oi: Yeah, Mondays do that. What's blowing up? User: ethan forgot his bag and lily wont eat ... [USER MESSAGE] "what do you think about the whole gentle parenting thing"
| Model | % of messages | Latency | What it handles | Prompt author |
|---|---|---|---|---|
openai/gpt-5.4-mini | 100% (classifier) | 700ms | Intent + depth classification — runs on every message | Hand-crafted 17KB prompt with expanded DeepSeek routing |
deepseek/deepseek-v4-flash | 60-70% | 2-4s | All conversational turns — with full personality, context, memory, Instagram signals | Opus 4.8 — master companion prompt |
moonshotai/kimi-k2.6 | 30-40% | 15-30s | BCT coaching · medical · complex planning · expert archetypes | Opus 4.8 SOUL.md + Sonnet 4.6 archetype prompts |
anthropic/claude-opus-4-8 | Once/user async | 30-60s | Deep profile inference (30-50 inferences, all layers) + prompt engineering | — |
text-embedding-3-small | Each turn bg | 200ms | Vector memory | — |
| Feature | Status | Notes |
|---|---|---|
| Architecture v8.2 routing design | 🔄 In Progress | Opus 4.8 writing master DeepSeek prompt + expanded routing logic now |
| DeepSeek full context injection (history + Honcho + Instagram) | 🔄 In Progress | Will replace current entity-facts-only injection once Opus completes |
| GPT-5.4 Mini routing prompt expansion | 🔄 In Progress | DeepSeek domain: light greetings → full conversational companion |
| DeepSeek personality/backstory/quirks/opinions | 🔄 In Progress | Maya-inspired + Claude-style. Rendering now by Opus 4.8. |
| Instagram enrichment (FullContact + Apify) | ⚠️ Needs key | FullContact silent lookup wired. Add FULLCONTACT_API_KEY to env. |
| Honcho workspace | ⚠️ Needs setup | Create workspace at app.honcho.dev → activates dialectic memory |
| NPS 7-8 consistently | ⚠️ In progress | Peak 7 achieved. v8.2 prompt changes are the key unlock. |
| Hermes SOUL.md v2 | ✅ Live | Oi backstory, 8 anti-sycophancy rules, 7 warmth rules, 4 relationship modes |
| 3-Tier Profile Router | ✅ Live | T1+T2+T3 context injection into every Kimi call |
| deep-inference-v6.txt (Opus 4.8) | ✅ Live | 467 lines, 11 layers, 9 frameworks |
| EntityExtractor | ✅ Live | GPT-5.4 Mini batch, session end |
| Rhythm Engine | ✅ Live | 8am + 8pm Hetzner crontab |