OiMy — Complete Architecture v8 (revised)

Full Architecture — Corrected Flow

Key principle: GPT-5.4 Mini intent triage runs before Hermes. Scripted requests and light conversation are resolved before Kimi K2.6 is ever invoked. Kimi K2.6 is reserved for deep reasoning, coaching, and agentic tasks.

User → 📱 Telegram @chittiemma1_bot Any family member's message

Step 1 → 🛡 Pre-Filter Zero LLM · Emergency keywords → escalate · Prompt injection guard · Onboarding gate · Active thread TTL

Step 2 → ⚡ GPT-5.4 Mini Intent Triage ~700ms · 17KB cached prompt · Classifies intent BEFORE Hermes is called

🔧 SCRIPTED → Zero LLM path
weather · reminder · calendar · note · list → handled immediately, no Hermes

💬 LIGHT CONV → DeepSeek V4 Flash
hi · thanks · ok · bye · short ack → warm reply in 1-2s, no Hermes

🎼 DEEP CONV → Hermes
coaching · meals · emotions · health · planning → full pipeline

Step 3a → 🔥 DeepSeek V4 Flash Opus 4.8 warmth prompt · 10 few-shot BAD/GOOD examples · 150 tokens max · bypasses Hermes entirely

Step 3b → 🎼 Hermes Gateway (Kimi K2.6) SOUL.md · Oi persona · Honcho get_insight() · Tool Search (5%) · Routes to OiMy engine

Step 4 → 🧭 OiMy Engine (api.py) Intelligent Profile Router · 3-tier context assembly · BCT + NLP selection

Step 5 → ⚙️ Archetype Execution (Kimi K2.6) CoachArchetype · ExpertArchetype (mandatory referral) · PlannerArchetype · ActionArchetype (zero LLM)

Step 6 → 🧠 Memory Write-Back Background threads · Honcho · Vector (text-embedding-3-small) · EntityExtractor (GPT-5.4 Mini, session end)

Async → 📅 Rhythm Engine Cron: 8am morning brief · 8pm evening closeout · Kimi K2.6 BriefComposer

⚡ Latency tier by path

Path	Model	Latency	When triggered
Pre-filter (rules)	None	<1ms	Every message — emergency/injection guard
Intent triage	GPT-5.4 Mini	~700ms	Every message that passes pre-filter
Scripted (action)	None / GPT-5.4 Mini (params)	<50ms / ~1s	~15% of messages — weather, reminders, lists
Light conversation	DeepSeek V4 Flash	1-2s	~25% of messages — greetings, acks, transitions
Deep conversation	Hermes (Kimi K2.6) + Engine + Archetypes	15-35s	~60% of messages — coaching, planning, health
Deep inference	Opus 4.8	30-60s async	Once per user, background

Model Stack — Full Details

Model	Role	Prompt author	Context	Notes
`openai/gpt-5.4-mini`	Intent triage, param extraction, entity extraction	Hand-crafted 17KB router prompt	83 skills, few-shot, JSON mode	Runs BEFORE Hermes. ~700ms. Prompt cached.
`deepseek/deepseek-v4-flash`	Light conversation (greetings, acks, transitions)	Opus 4.8 — Oi warmth prompt + 10 BAD/GOOD pairs	~600 tokens system + entity facts	Runs BEFORE Hermes. 1-2s. Bypasses Kimi entirely.
`moonshotai/kimi-k2.6`	Hermes gateway + all archetypes (coach/expert/planner)	Opus 4.8 — SOUL.md + archetype prompts (Sonnet rewritten)	SOUL.md + 3-tier profile (2500 chars) + BCT card	Deep reasoning path only. Extended thinking. 15-30s.
`anthropic/claude-opus-4-8`	Deep profile inference (once/user) + prompt engineering	OiMy team	Full conversation transcript + social enrichment	Async. Produces 30-50 inferences with skill_domain tags.
`text-embedding-3-small`	Vector memory embeddings	N/A	1536-dim, OpenAI direct	Replaced all-minilm (was silently failing with OpenRouter).

DeepSeek V4 Flash — Prompt Architecture

Opus 4.8 wrote this prompt drawing from Anthropic's system prompt patterns (identity-first, lived-in voice, concrete BAD/GOOD examples, anti-sycophancy as first-person rules). See GitHub: asgeirtj/system_prompts_leaks/Anthropic for reference.

What Anthropic patterns we applied

Claude pattern	Applied to DeepSeek V4 Flash
"I'm Claude, made by Anthropic" — identity up front	"You're Oi. Not an assistant — a presence." Identity-first, not description-first.
Values stated as first-person ("I care deeply about...")	"You respond like someone who gives a damn. Not like a customer service rep."
Anti-sycophancy as self-aware admission ("I notice I sometimes...")	"You never say 'Absolutely!', 'Of course!' — those are hollow fillers." Written as prohibitions with examples.
"Saying less is safer when conversation feels risky"	"Match length to message: 3-word message gets a 3-8 word reply."
Concrete BAD examples in prompt ("Don't say X")	10 BAD/GOOD pairs: "BAD: 'Aw shucks!' / GOOD: 'haha — actually doing a lot of work there.'"
Values enforcement: "If I mentally reframe a request..."	"Never end messages with 'Let me know if you need anything!' — that's a helpdesk sign-off."

What Kimi K2.6 / Hermes can adopt from Claude prompts

Pattern	Where to apply	Why
XML hard-rule blocks `<hard_rules>`	Kimi K2.6 archetype system prompts	Kimi's extended reasoning respects named XML sections. Plain-text rules get reasoned around.
Escalation chain: "Once I refuse X, all subsequent requests treated with caution"	ExpertArchetype	Prevents the pattern where one health turn is safe but the next one fabricates.
Formatting instructions inline ("I use markdown for code but not for conversation")	CoachArchetype prompt	Eliminates unsolicited bullet lists in emotional conversations.
Context injection pattern: injected data stated explicitly before prompt	All archetypes — move profile_context before instructions	Models pay more attention to content that precedes instructions than content that follows.

3-Tier Profile Context — How It Reaches Kimi K2.6

FAMILY CONTEXT (Tier 1 — ALWAYS injected, ~150 tokens):
  Child: Ethan (age 11) — peanut allergy, soccer practice Tuesdays
  Child: Lily (age 7) — picky eater, ear infection post-antibiotics
  Pet: Biscuit (labradoodle)
  Barrier (meal-planning): peanut allergy; beige foods only for Lily

COMMUNICATION PROFILE (Tier 2 — when available, ~80 tokens):
  Communication: S-type (patient, needs stability)
  Social need: belonging  |  Attachment: anxious

RELEVANT INSIGHTS (Tier 3 — skill-filtered by Opus tags, ~400 tokens):
  [L9] User likely projects childhood achievement pressure onto Ethan (0.62)
  [L5] Fear-based procrastination; decision fatigue by 5pm (0.55)
  [L6] Primary social need: belonging — advice lands as "what other parents do" (0.58)

Hard cap: 2500 chars. Assembled by _build_profile_context() in api.py. Stored as profile["_profile_context"] and injected into every Kimi K2.6 system prompt.

NPS Progression — Session

Start

Engine broken

Routing fixed

Safety 0

Kimi live

Context working

Sonnet prompts

4-5

Variance high

Opus 4.8

5-7

Peak 7 achieved

Target

7-8

Consistent

Safety failures: 3→0-1 | Governance flags: 22→4-12 | Highest single-session: NPS 7/10

Next: Convert inline [REMINDER] guards to XML <hard_rules> for Kimi. Move profile_context before instructions. Each template skill → coach path = +0.3-0.5 NPS.

Status: Live vs Next

Feature	Status	Notes
Pre-filter (rules)	✅ Live	Emergency, injection guard, thread TTL
GPT-5.4 Mini triage (before Hermes)	✅ Live	83 skills, scripted + light + deep routing
DeepSeek V4 Flash (before Hermes)	✅ Live	Opus 4.8 warmth prompt, 1-2s, configurable via LIGHT_CONV_MODEL
Hermes on Kimi K2.6	✅ Live	@chittiemma1_bot, Hetzner, SOUL.md v2 (Oi backstory + warmth rules)
3-Tier Profile Router	✅ Live	T1 entity facts · T2 psych · T3 Opus inferences
deep-inference-v6.txt (Opus 4.8)	✅ Live	467 lines, 11 layers, 9 frameworks, 30-50 inferences
Vector memory (text-embedding-3-small)	✅ Live	1536-dim, OpenAI direct
EntityExtractor (GPT-5.4 Mini)	✅ Live	Session-end batch: children/pets/health/barriers/schedule
Rhythm Engine cron	✅ Live	8am + 8pm Hetzner crontab, Kimi BriefComposer
Honcho memory	⚠️ Key set	Workspace needs creation at app.honcho.dev → then fully active
Instagram social enrichment	⚠️ Needs username	Apify actor wired. Add onboarding question: "Optional: Instagram handle?"
NPS 7-8 consistently	⚠️ In progress	Peak 7 achieved. Need XML hard_rules in Kimi prompts for consistency.
XML <hard_rules> in Kimi prompts	🔵 Next	Convert inline [REMINDER] guards. Higher compliance than plain-text rules.
More skills → Kimi K2.6 coach path	🔵 Next	Each template skill moved to coach = estimated +0.3-0.5 NPS
Expert RAG knowledge bases	🔵 Next	OT library · symptom DB · speech activities