OiMy Engine / Architecture

Message Processing Flow
with Human 3.0 Integration

v8.4 post-H3.0
Updated Jun 21, 2026
Hetzner 116.203.107.13
Skills 89 active
NPS baseline 8.4 / 10
Example trace
"My husband and I keep fighting about screen time every night. I told myself I'd set proper limits last week but I never actually did it."
user_id: family_sarah_4721  /  channel: telegram  /  session message: #8
Request flow
01 Inbound
Telegram webhook receives the message. JWT validated against the LINQ signing secret. User ID resolved. Request enters the OiMy chat handler.
Resolved
user_id: "family_sarah_4721" channel: "telegram" auth: "bearer JWT - valid" msg_count: 7 // prior messages; entering message #8
02 Intent Router gpt-5.4-mini
Classifies message intent and selects the scaffold. Routes to chat_light for fast companion responses, or coach / expert / planner for deeper workflows. Low-latency by design - GPT-5.4-mini keeps this under 200ms.
Classification
skill_id: "screen-time" workflow: "chat_light" confidence: 0.91 latency_ms: 178
03 Signal Collector
Runs in parallel with routing. Extracts latent signals: stress markers, detected barriers, urgency cues, relationship flags. Output feeds the deep profile pipeline and the BCT engine downstream.
Signals extracted
relationship_stress: true barrier: "inconsistency" frustration_level: "moderate" topics: ["screen_time", "marriage", "parenting"]
04 H3.0 Engine New
Human 3.0 framework by Dan Koe. Every message calls tick(): increments message count, extracts commitment language via regex, and checks for false transformation signals by comparing new messages against unresolved prior commitments. Grok 4.3 inference fires every 10 messages to classify life quadrant and coaching engagement level from conversation context.
tick() - msg #8
// commitment detected in message commitment: "set proper screen time limits" topic: "screen_time" session_ts: "2026-06-21T17:23Z" // prior commitment on same topic found (6 days ago) signal_logged: true signal_count: 2 // threshold: 2 - probe now eligible // cadence: Grok 4.3 inference fires at msg #10 msg_count: 8 next_infer: 10 quadrant: "spirit" // from prior inference at msg #0 level: "individualist"
05 Deep Profile Pipeline
Infers psychological profile from message history. Maslow level, decision motivator, social needs classifier. Budget-gated: runs on cadence rather than every message. Results are persisted and merged into subsequent context builds.
Inferred profile
maslow_level: "love_and_belonging" decision_motivator: "avoid_conflict" social_need: "partnership" bct_tags: ["planning_gap", "self_efficacy_low"]
06 Profile Context Builder Modified
Assembles the tiered system prompt context. H3.0 now injects a third tier after [Psych], surfacing the user's coaching engagement level, primary life domain, and a BCT hint sentence. The tier is omitted entirely when H3.0 profile fields are still null, so no regression for new users.
Context injected into system prompt
// [Psych] tier - existing [Psych] Maslow: love/belonging Motivation: avoid_conflict // [H3.0] tier - new [H3.0] Engagement: individualist Domain: spirit BCT hint: Offer frameworks, not directives. Avoid "you should / must". Use identity- based framing. Tone: collaborative.
07 False Transform Probe New
Checks whether the user has accumulated 2+ false transformation signals on the current topic within a 14-day cooldown window. A false transformation signal is logged when a prior commitment is re-raised without evidence of progress. When the threshold is met and cooldown is clear, a gentle curiosity probe is prepended to the system prompt. Never lectures. Maximum one probe per topic per 14 days.
Probe decision
topic: "screen_time" signal_count: 2 // threshold: 2 last_probed: null // never probed before cooldown_clear: true probe_fires: true // text injected at top of system prompt: probe_text: "We've talked about screen time a few times - I'm curious how it's actually been going day to day, not just intentions?"
08 BCT Engine Modified
Behavior Change Technique selection is now gated by h30_level. When level is set, the scoring layer boosts techniques appropriate to that engagement style and suppresses mismatched ones. Null level is a clean fall-through to prior selection logic with zero behavioral change.
BCT selection - individualist path
level: "individualist" // boosted BCTs boosted: ["values_clarification", "identity_reframing", "autonomous_goal_setting"] // suppressed BCTs suppressed: ["prescriptive_action_plan", "authority_directive"] selected: "identity_reframing"
09 Quadrant Tiebreaker Modified
When the intent router returns ambiguous results (confidence below 0.70), the H3.0 quadrant is used as a tiebreaker to prefer scaffolds in the user's primary life domain. Additive only: unambiguous routing is never overridden.
Tiebreaker check
router_confidence: 0.91 threshold: 0.70 // confidence above threshold - tiebreaker not invoked tiebreaker_used: false quadrant: "spirit" // would prefer parenting/relationship scaffolds
10 Companion Model grok-4.3
Full system prompt assembled: scaffold instructions + [Psych] tier + [H3.0] tier + BCT hint + false transform probe. Routed through _call_openrouter_companion. Hard fallback to claude-haiku-4-5 if circuit breaker opens.
Model call
model: "x-ai/grok-4.3" via: "openrouter" context: "scaffold + [Psych] + [H3.0] + probe" bct: "identity_reframing" fallback: "anthropic/claude-haiku-4-5"
11 Response Delivered
Governance checks applied. Response normalized and returned to Telegram. The false transform probe is woven in naturally as an opening curiosity question. BCT framing avoids "you should" language. Tone is collaborative, identity-based as appropriate for an individualist user profile.
Final response
We've talked about screen time a few times - I'm curious how it's actually been going day to day, not just the intentions. The nightly fight usually isn't about screens. It's about you and your husband not holding the same line. What does a rule that actually works for both of you look like - not the ideal version, the one you'd both actually follow?
Model stack
Companion + Coach
Grok 4.3
x-ai/grok-4.3 / OpenRouter
Companion Fallback
Claude Haiku 4.5
anthropic/claude-haiku-4-5
Intent Router
GPT-5.4-mini
openai/gpt-5.4-mini / OpenRouter
H3.0 Inference
Grok 4.3
x-ai/grok-4.3 / OpenRouter
Deep Inference
Claude Opus 4.8
anthropic/claude-opus-4-7 / OpenRouter
Tall Model
Kimi K2.6
moonshotai/kimi-k2.6 / OpenRouter
Direct API
DeepSeek V3
deepseek-chat / api.deepseek.com
NPS Eval - running now
Opus 4.8 + Grok 4.3 + GLM 5.2
30 users / 10 each / GPT-5.5 QA