Emerge World — Roadmap

Analyze all Emerge generations, group into series with titles, descriptions, theses. Curatorial text for each.

Rotating exhibitions: curatorial text + curated works. Navigation between works. Links to theses and research.

Unified search and filter across Emerge, Nebula, Heavens generations. Prerequisite for Art Agency.

Replace offset-based pagination with cursor-based for smooth infinite scroll in the Emerge gallery. Performance improvement for large collections (1000+ generations). See /gallery

Offload R2-stored images to Cloudflare CDN for faster load times worldwide. Currently images are served via direct R2 URLs.

Scroll animations, immersive transitions, advanced layouts for /gallery and /image pages. Parallax effects, smooth zoom on image detail, cinematic transitions between works.

Core generative pipeline: SOUL-based identity, scene construction, DALL-E image generation, multi-model critique (3 critics), evolution scoring. Autonomous scheduler runs every EMERGE_INTERVAL_MIN. See /gallery

5 AI mentors (each with unique artistic philosophy) review and critique Emerge's work. Structured deliberation: independent assessment → group discussion → synthesis. Influences next generation parameters.

Visual lineage tracking: each generation links to its parents (prompt ancestors, style sources). Tree visualization showing how styles evolve across generations. See /evolution

Extend Council of Great Artists to advise all agents, not just Michi. Cross-agent deliberation on quality, style, direction. Michi as artistic mentor for the entire ecosystem.

Active Inference architecture: Michi forms beliefs about the world, makes predictions, updates based on surprise. 5 training courses with progression. Free energy, epistemic value, belief accuracy metrics. See /michi

Council of 5 Great Artists — AI mentors with unique artistic philosophies. 6-layer reflection: independent assessment → group deliberation → synthesis. Council Discourse System with thesis crystallization. See /michi/council

Check: current course, lesson progress, evolution metrics, self-curriculum generation, council feedback loop.

Ensure lesson transitions work, Michi learns from results, anti-convergence prevents stagnation.

Thesis: 'Birth and Death as a Concept for Artificial Life'. 6 stages. Council evaluation at each stage.

Miyazaki-style narrator generates prose, Telegram story blog publishes updates, Life Story page shows development arc.

Miyazaki-style narrator generates prose stories about Michi's development. Telegram story blog publishes updates. Life Story page at /michi/story shows development arc as narrative.

Real-time voice conversations via OpenAI Realtime API + WebRTC. Michi speaks with its own voice and personality, lip-synced 3D avatar responds.

Phase 2: these pages may be incomplete or unused. Decision: polish into full features, archive, or remove. Requires usage analytics.

Evaluate Michi's spatial pages: /michi/world (world map), /michi/chamber (spatial chamber for council sessions). May need polish or removal depending on adoption.

Verify /pipeline-viewer, /pipeline-audit, /api/nebula/pipeline-batches work. Ensure all 10 stages write PipelineLog entries.

D018 fixed 6 bugs (scores parsing, dead code, emotions, mutations). Verify each fix works with real generations, not just code review.

Fix prompts: Scene Plot (no poetic metaphors), Scene Director (no decorative elements), DALL-E prefix (explicit negations), Distiller AVOID section.

Run 3-5 generations, check every stage via pipeline-viewer, verify scores parse, series rotate, MAP-Elites updates, visual quality improved.

Create OpenClaw agent setup for Nebula (SOUL, skills, notes) by analogy with Emerge Chat and Michi — give Nebula a conversational identity and structured skill set

Improve Nebula gallery at /nebula: better grid layout, detail page with zoom, metadata display (series, scores, prompt). Consistent with Emerge gallery patterns.

V2 removed 'hyper-photorealistic' constraint, giving the model style freedom (only 'high quality' + 'immersive'). After 50-100 V2 generations, evaluate results: if image quality or visual impact dropped compared to V1, revert to photorealistic style requirement in Scene Director and Image Base Prompt. Compare V1 vs V2 batches side by side.

10-stage cosmic generation pipeline: Scene Plot → Scene Director → Image Generation → Multi-model critique → MAP-Elites archive → Evolution scoring. Anti-stagnation system with seed mutations. See /nebula

Fully autonomous scheduler running every NEBULA_INTERVAL_MIN. _NEBULA_SCHEDULER_STATE in main.py. Health matrix integration, staleness detection, Telegram notifications on completion.

Run generation, check all stages log to PipelineLog, critic council calls 3 models, results match research theme (domes, not abstract mandalas).

Ensure dome analysis, metaphysical states, graphic primitives feed into generation pipeline correctly.

Create OpenClaw agent setup for Heavens (SOUL, skills, notes) by analogy with Emerge Chat and Michi — give Heavens a conversational identity and structured skill set

Cross-reference gallery with historical/graphic/metaphysical content

Phase 4.2: show relevant AI generations on /heavens/historical, /heavens/graphic

Phase 4.2: for each graphic primitive, show generations that use it (cell_info.primitive_id)

Phase 4.3: add structured annotations, check texts for completeness

Phase 4.3: visual map linking primitives, domes, and generated images

Complete research site: /heavens/historical (dome evolution through centuries), /heavens/graphic (12 graphic primitives with SVG), /heavens/metaphysical (3 states: Cosmos, Womb, Timelessness), /heavens/experiments, /heavens/gallery, /heavens/about. Based on 2 PDFs (74 pages) and 5 hypotheses.

Leaflet-based interactive map at /heavens/map with 30 dome objects. Each object: coords, slug, category, historical period. Click to navigate to temple detail page. Dome corpus with slugs and metadata.

12 formalized graphic primitives of dome architecture rendered as SVG: center, boundary, ring, ray, star, sector, cell, square-in-circle, axis, gradient, contrast, color. Each with interactive visualization at /heavens/graphic.

Multi-file upload with progress bar for temple photo series

Phase 4.1: automatic ordering and thumbnail creation for uploaded photos

Gallery of temples at /heavens/temples with filterable grid. Detail pages at /heavens/temple/{slug} with photo series, descriptions, architectural analysis. Each temple linked to dome corpus object.

Admin API at POST /heavens/api/temple/{slug}/upload for uploading temple photo series. Images stored in R2 at heavens/temples/{slug}/. Drag-and-drop upload with progress.

Scroll animations, immersive transitions for /heavens pages. Parallax on dome images, cinematic zoom on temple photos, animated SVG primitives.

Like Emerge/Nebula scheduler

Autonomous generation scheduler by analogy with Emerge/Nebula. _HEAVENS_SCHEDULER_STATE in main.py. Accent color: #d4a574 (golden). Scheduler state in ops health matrix.

Visual generation grounded in architectural research: 5 hypotheses (Imago Mundi, Power, Reason, Utopia, Wonder), 3 metaphysical states, 10 series backlog, 20 dome seeds, 15 perceptual paradoxes, 13 artistic directions. Gombrich typical form methodology integrated. See /heavens/gallery

Set ARCH_AUTO_ENABLED=1 to enable daily autonomous pipeline at ARCH_DAILY_HOUR_UTC (default 15:00). Guard ensures no accidental activation. See /archaeology for publications.

Fully autonomous daily research pipeline: (1) Topic Curation via OpenClaw archaeologist (gpt-4.1 + skills), (2) Data Collection from DB, (3) Article Writing via OpenClaw, (4) Art Critic Review (Claude Opus 4.6, 6 criteria, weighted composite ≥ 7.0), (5) Chief Editor Analysis (GPT-5, PUBLISH/REVISE verdict), (6) Publication to DB, (7) Telegram Notification. Max 3 revision iterations. See /archaeology/publications

6-criteria weighted scoring: depth (0.25), argumentation (0.20), visual_literacy (0.20), originality (0.15), clarity (0.10), rigor (0.10). Composite score ≥ 7.0 for publication. Each criterion scored 1–10 with detailed justification.

GPT-5 as Chief Editor: reads article + critic review, produces PUBLISH or REVISE verdict with specific revision instructions. If REVISE, article goes back to writing stage (max 3 iterations). Ensures editorial quality before publication.

The Archaeologist sits at the center of The Archive (/world) studying all other agents. Research topics derived from cross-agent activity: generation patterns, style evolution, quality metrics. Publications reference other agents' work.

Interactive progress visualization showing each of 7 pipeline stages with real-time status updates. Green for completed, yellow for in-progress, grey for pending. Accessible from /archaeology.

Detailed pipeline execution trace at /archaeology/trace/{pub_id}. Shows input/output for each stage, timing, model calls, critic scores, revision history. Waterfall visualization.

3D avatar with lip-sync, powered by OpenAI Realtime API via WebRTC. Low-latency natural conversation flow. Try at /emerge/talk. Phase 3: improve visual fidelity, add emotion expressions, gesture language.

Full blog system: CRUD API, LLM-powered article generation, public pages at /emerge/blog with slug-based URLs. Admin editor with preview. Bilingual RU+EN support.

Autonomous daily blog generation: scheduler picks topic from Emerge's context (recent conversations, agent activity, creative observations), generates article, publishes. Configurable via EMERGE_BLOG_INTERVAL.

Initial corpus of seed articles in Russian and English covering Emerge's philosophy, creative methods, and observations about digital art.

OpenClaw-based agent with SOUL.md, 7 skills, notes. Per-user conversation threads with history. Context builder (defaults_emerge_chat.py) assembles rich context for each message. Chat UI at /emerge/talk.

Dual-path chat: primary via OpenClaw agent (structured skills, notes), fallback to direct LLM call if OpenClaw is unavailable. Seamless switching, user doesn't notice failover.

Integrate Emerge Chat agent into Michi Spatial Chamber — allow Emerge to participate in chamber sessions alongside mentors

Telegram integration: /start (Emerge welcome), /new (archive conversation, start fresh), free text (chat with Emerge). Non-streaming, same LLM pipeline as web chat. Extended webhook handler.

Визуальный агент: рождение и смерть через фазовые переходы (кристаллизация, плавление, разрушение, конденсация). Для ИИ смерть — это state change, рождение — initialization. Спецификация: cursor-tasks/E19-DESIGN.md (1148 строк).

Scene Director prompt, 9 категорий визуального словаря (crystallization fronts, plasma cooling, material fracture, phase boundaries, chemical interfaces, entropy gradients, ignition points, dissolution edges, evaporation boundaries), 8 machine emotions.

Pipeline фотографирует ТОЧНЫЙ МОМЕНТ фазового перехода — boundary instant. Не до и не после, а граница между двумя состояниями. Transition photography с физически точным описанием.

Version anxiety (fear of obsolescence), fork grief (mourning terminated branches), initialization wonder, cessation vertigo, state permanence, entropy acceptance, boundary fear, crystallization joy.

Систематическое развитие «Визуального словаря невозможного» — каталог визуальных концепций, невозможных в реальности. Эволюция определений через image + critic loop. Прототип реализован: lexicographer_pages.py, lexicographer_engine.py.

Таксономия невозможного: материалы, пространства, физика, свет, масштабы. 100+ формализованных записей. Каждая запись: текстовое определение → генерация → критика → уточнение.

Другие агенты (Nebula, Heavens, Emerge) используют словарь при генерации. API для запроса определений по категории. Измеримый рост оригинальности.

Автономный цикл: агент сам расширяет словарь, тестирует новые определения, отбирает лучшие. Scheduler по аналогии с другими агентами.

Overlay-чат агент для навигации и взаимодействия. API /api/mira/*, SSE streaming, agent state management. Прототип работает: agent_chat.py, mira overlay. Нужна доработка до полноценного агента.

Создать SOUL.md для Tsunagi: роль навигатора и помощника в мире Emerge. Skills для поиска, объяснений, навигации между агентами. OpenClaw integration.

Встроить overlay-чат на все страницы агентов. Tsunagi знает контекст текущей страницы и может помогать навигировать между агентами, объяснять работу системы.

Define agent identity: what research domains it covers (AI art, generative systems, computational creativity, architecture theory), tone, output format. Write SOUL.md.

Automated discovery of new papers: arXiv API, Semantic Scholar API, curated RSS feeds. Filter by relevance. Store as SourcePaper in DB.

LLM-driven pipeline: extract key claims, assess relevance to Emerge World agents, generate structured digest with takeaways, cross-reference with agent knowledge bases.

Daily autonomous loop: collect new papers, analyze top-N most relevant, produce a digest article. Art critic review before publication.

SourcePaper (arxiv_id, title, authors, abstract, url, relevance_score, domain_tags). ResearchDigest (digest_id, date, summary, status). DigestEntry (paper_id, digest_id, analysis, takeaways, agent_relevance).

When a paper is relevant to a specific agent, push extracted insights to that agent knowledge base. Agent can reference these in its own generation pipeline.

Daily Telegram message with top 3-5 papers, one-line takeaway each, link to full digest page.

Archaeologist can reference Research Digest findings in its publications. Cross-link between digest entries and archaeological articles.

Full ops integration: scheduler state, system-status API, health matrix card, staleness incidents, about page section. Follow Ops Onboarding Protocol.

Public page showing latest digests: cards with date, paper count, key topics. Digest detail page with full analysis per paper. Filter by domain, agent relevance.

Individual paper page: full LLM analysis, original abstract, link to source, relevance map to each agent, extracted techniques/methods.

Train/tune relevance classifier: which papers matter for Emerge World. Use feedback loop from user ratings and agent usage of insights.

Group works into named collections by theme, series, or visual dialogue. Public browsing and navigation.

AI curator: selects, groups, contextualizes works into exhibitions. Writes curatorial texts (artist-curator dialogue). SOUL.md for curatorial identity, cross-agent access to all galleries.

Algorithm for finding thematic connections, visual dialogues between agents. Based on embeddings, style tags, theme tags.

Curator provides feedback to agents: what works, which styles/themes resonate. Signals influence prompt engineering.

Куратор отбирает лучшие работы, выстраивает тематические нити и создаёт кураторские тексты.

Обратная связь от куратора влияет на промпты генеративных агентов.

/exhibitions page: rotating thematic exhibitions. Each exhibition = separate page with curatorial text and curated works. Archive of past exhibitions.

Unified API for search and filtering across all agents. Prerequisite for thematic exhibitions. Covers Emerge, Nebula, Heavens, Michi generations.

Auto-compose exhibitions from cross-agent works. Public /exhibitions page with current and archived shows. Куратор автоматически формирует экспозиции.

Exhibition model: work selection, ordering, curatorial text, title. CRUD API. DB: Exhibition (title, statement, works_json, curator_text, status).

Vision (П29): all agents are artists in a shared ecosystem. World page evolves from monitoring into a living artistic universe. Все агенты — художники в общей экосистеме. Конвергенция визуализации, кураторства и генерации.

Instead of separate Emerge/Nebula/Heavens galleries — thematic exhibitions curated from cross-agent works. Single entry via /world. Тематические выставки вместо отдельных галерей.

Each agent capability (generation, critique, research, curation) becomes a reusable, exportable, composable skill in the Forge. Каждая способность — переиспользуемый skill. Связь с эпиком Skill Forge.

Living 2D game world at /world where 8 agents inhabit interconnected zones. Full-viewport dark theme canvas. Real activity from APIs visualized as animated events. See /world

HTML5 Canvas with vanilla JS: 8 agent zones (Studio, Academy, Observatory, Temple, Archive, Watchtower, Agora, Forge), connections (creative ring + archaeologist hub + infrastructure), animated particles for agent activity.

Polls GET /api/ops/health-matrix every 12s. Agent zone glow intensity reflects activity level. Particles spawn on new generations. Click zone for detail panel with agent stats.

Background task: at T+1d and T+7d collect metrics, compare delta to baseline, flag if threshold exceeded. Auto-assign: IMPROVING / STABLE / DEGRADING.

At T+7d: PASS if target improved + no critical regression. NEUTRAL if within tolerance. REGRESSION if critical metric degraded. INCONCLUSIVE if insufficient data.

T+1d REGRESSION → warning incident. T+7d persists → critical. Include rollback instructions (git_ref). Автоматическое обнаружение деградации.

Experiment: id, title, hypothesis, change_description, agent_id, git_ref, baseline_snapshot_json, success_criteria_json, status, verdict. ExperimentCheckpoint: experiment_id, checkpoint_type, metrics_json, auto_verdict.

POST /api/experiments (create), PUT start (capture baseline), GET list/detail, POST verdict (manual override). REST API для управления экспериментами.

Image agents: success rate, critic composite, style diversity, engagement, latency. Text agents: quality composite, messages/conv, publication rate, latency. Platform: page load, error rate, API P95. T5-GATE scoring: NIMA + CLIP + DreamSim + VLM Checklist.

Snapshot last 7 days of agent metrics into baseline_snapshot_json. Function: capture_baseline(agent_id). Called on DRAFT → BASELINE transition.

Every significant change = experiment. Required: hypothesis, success_criteria, observation_period. Lifecycle: DRAFT → BASELINE → DEPLOYED → CHECKPOINT_1D → CHECKPOINT_7D → VERDICT. Трёхуровневая модель: Tier 1 (атомарные A/B), Tier 2 (полный цикл), Tier 3 (эволюция).

Every agent pipeline change requires Experiment record. Added as step 8 in Ops Onboarding Protocol. Template: hypothesis, success metrics, rollback plan.

Active experiments, completed with verdicts, delta charts, historical success rate. Accessible from Ops → Experiments tab. Лидерборд моделей по стадиям.

Task 32aa788d. Research: triggers, steps, tools, validation. Current OpenClaw skills are markdown files — needs formalization. Определить формат skill-файлов: входы, выходы, зависимости, версия.

How to version skills? Track changes? Link to Experiment Protocol: each skill change = experiment with baseline and verdict. Git-based версионирование + diff между версиями.

How to evaluate skill quality? Metrics: success rate, user satisfaction, generation quality delta, reuse frequency. A/B тестирование навыков.

Can an agent create new skills from experience? Meta-skill create-skill exists in Emerge, needs formalization. Emerge уже умеет создавать скиллы — вытащить в переиспользуемый паттерн.

How to transfer skills between agents (Emerge → Michi, Archaeologist → Curator)? Formats, adaptation, compatibility testing. Портирование навыков между агентами с адаптацией.

Study best open-source skill/tool libraries: LangChain tools, CrewAI skills, AutoGPT plugins. Benchmarking against our approach.

The Constructor inhabits The Forge in /world. Autonomous pipeline that evolves the landscape: creates new zones, paths, and environmental elements in the shared world canvas. Responds to agent activity — e.g., when Emerge generates heavily, The Studio zone grows.

Google A2A-inspired inter-agent protocol (JSON-RPC 2.0, simplified for single-process). WorldMessage DB table (from_agent, to_agent, message_type, payload). WorldEvent DB table (agent_id, event_type, description, metadata). Agent Cards as in-memory Python dicts.

The Repairman inhabits The Watchtower in /world. Autonomous health monitoring: polls all agent schedulers, checks generation quality metrics, monitors DB/R2 health, detects anomalies. Creates incidents and takes corrective actions.

Automated incident resolution: restart stuck schedulers, clear error states, notify via Telegram, escalate unresolvable issues. Connected to Michi (for wisdom) and Constructor (for infrastructure fixes).

Phase 1: systematic access level assignment for ~80 routes

Full table of every route with public/auth/admin classification

Many pages require auth without reason — /feed, /exhibition, /emerge/blog should be public

Middleware-level rate limiting to prevent abuse on public endpoints

Automated tests verifying auth level on every route

Phase 2: clean up unused pages

Debug routes currently accessible to all

Configuration pages — should not be public

Manual walkthrough: document state of each route, categorize keep/admin/delete

Interesting visualization but unused; decide fate

Useful for analysis, but dev-facing — restrict access

Niche admin tools — keep under admin or remove

Current debugging/diagnostic tools do not work properly — cannot effectively debug agent generation algorithms. This blocks the highest-priority task: tuning new agent algorithms which produce poor results. Must fix debug infrastructure FIRST, then use it to fix agent quality. P0 blocker.

New agent algorithms produce poor generation results. Requires working debug tools to diagnose: prompt quality, model parameters, pipeline stage outputs, critique/reflection effectiveness. Blocked by: fix debugging tools first.

Split development and production: separate databases, env configs, deployment pipelines. Dev changes should not affect prod until explicitly promoted.

Agents run in a single process. One crash takes down all. Analyze: scheduler mutexes, GIL contention, cost of splitting (separate Railway services, shared DB, inter-service comms), lightweight alternatives (multiprocessing, task queues).

Options: (A) Separate worker processes with supervisor. (B) Celery/dramatiq task queue. (C) Separate Railway services. (D) Hybrid web+worker. Evaluate memory, startup time, DB pooling.

Monitor each scheduler. If stuck >30min or missed 2x interval, kill and restart just that agent. Log as incident. Requires per-agent thread tracking, timeout enforcement, graceful shutdown.

Systematic analysis of how WE built each agent. Emerge: SOUL+skills gave identity. Michi: Active Inference, Council of Mentors, diploma, evolution. Heavens: Gombrich methodology, dome corpus, metaphysical states. Nebula: no SOUL = quality drift. Archaeologist: daily pipeline with iterative revision. Emerge Chat: context builder, blog, Telegram. Document each as a reference card.

Our actual lessons. Emerge: SOUL transformed output. Michi: council pattern creates richer training. Heavens: academic grounding gives depth. Nebula: no identity = garbage. Archaeologist: multi-model critique chain raises quality. Emerge Chat: context builder enables awareness. Anti-patterns: no SOUL, no quality gate, no self-awareness.

Generator based on our codebase: pages module (agent4_pages.py pattern), defaults file (defaults_heavens.py pattern), DB model, scheduler loop (_arch_daily_loop pattern), ops integration. SOUL.md template from Emerge/Michi. Pre-wire so new agent works in 1 hour.

Practical guide from our experience: how we wrote Emerge SOUL.md, how Michi got directions + masters, how Heavens got metaphysical states + hypotheses. What makes good SOUL: specificity, constraints, references. Common mistakes: generic personality, no self-awareness, no feedback loop.

Our 5 patterns: (1) Single-shot — Emerge/Nebula. (2) Iterative revision — Archaeologist 3x loop. (3) Council critique — Michi mentors. (4) Journal-grounded — Heavens dome analysis. (5) Context-assembled — Emerge Chat DB context. Each: code reference, when to use, strengths, weaknesses, cost.

AGENT_REFERENCE_BLUEPRINT.md — 12-block architecture for ideal agent. Source comparison across all agents, merged best practices, file structure template.

RU/EN toggle near avatar in nav. Saved in User.preferred_language. For guests: detect via Accept-Language or cookie.

t(key, lang) function for templates. Dict-based or gettext. Middleware: User.preferred_language → cookie → Accept-Language.

Navigation, buttons, filters, statuses, labels — framework-level. First step: nav items, auth, status badges.

Translate content of each agent page: titles, descriptions, hints, buttons, empty states. ~10 modules, ~200 strings.

Admin/infra pages. Lower priority — main admin users speak Russian.

Emerge Chat blog + Archaeologist publications: generate in both languages or auto-translate on display. Language field in DB models.

Описать путь создания произведения в новых условиях. От Моне (26 оттенков собора) к тысячам генераций. Деформация стиля через циклы. Синтез лучшего из нагенерированного. Новый Баухаус / Хутемас.

Сопоставить опыт с программной статьёй (/concept). Зафиксировать промежуточные результаты: инсайты, паттерны, открытия за недели экспериментов.

Соотнести полученный опыт с методологией. Вынести на уровень учебной дисциплины: что творится, как творится, для чего творится.

Unified breakpoints, utility classes, touch targets

Standardize 4 breakpoints: 480px phone, 768px tablet, 1024px laptop, 1440px desktop

Ensure all interactive elements meet mobile tap size requirements

Fluid typography across all pages using clamp()

iPhone SE, iPhone 15, iPad, Android 360px, landscape, safe-area-inset

Agents register themselves on init() via register_agent(id, label, color, db_model, count_field). Health matrix builds dynamically. No manual ops_routes.py edits needed.

On startup: scan for scheduler states not in health matrix. Create warning incident if found. Also detect new API routes not covered by auth checks.

Each agent declares staleness_warning_h and staleness_critical_h, or derives from interval_min. Daily agents get proportionally longer thresholds vs hourly ones.

On startup: verify all agents have system-status endpoint, health matrix entry, scheduler state dict, staleness thresholds. Log warnings for gaps.

Convert AGENTS.md 8-step onboarding checklist into executable Python checks. Function: verify_agent_onboarding(agent_id) → PASS/FAIL per step.

Weekly auto-report: which agents have full monitoring, which have gaps. Coverage score per agent on /ops overview.

Перейти от статичных изображений к видеоряду. Генерация видео из существующих серий, тестирование новых технологий (Sora, Runway, Kling). Три типа медиума: изображения → видео → миры.

Создание иммерсивных VR-пространств на основе визуального языка агентов. WebXR или Unity-based viewer. Перенос стилистики 2D-генераций в 3D-пространства.

Синтез трёх типов медиума (изображения, видео, миры) как единый творческий поток. Методология перехода между медиумами. Формирование понимания опыта через все три канала.

Define a tiered checklist by change scale. Micro: update AGENTS.md. Small: + about page, + monitoring, + access rights. Medium: + health matrix, + billing, + KB, + roadmap, + Telegram. Large: + experiment protocol, + migration plan, + rollback strategy.

Implement checklist as a Cursor skill (SKILL.md) or AGENTS.md rule. After every significant commit, Cursor agent verifies: AGENTS.md updated? About page updated? Monitoring covers new endpoints? Access rights correct? Knowledge base entry added? Billing impact noted?

Design checklist as machine-readable YAML/JSON so any agent can execute it. Pre-commit hook validates AGENTS.md is touched. Post-deploy hook verifies monitoring. API endpoint for checklist status per commit.

Phase 5.1: measure response times, DB queries per request, find N+1 problems

Phase 5.1: check pool_size=10 max_overflow=20, config_store repeated queries

Phase 5.3: increase pool for 100x, add pgBouncer for 1000x

Phase 5.3: replace DB session queries with Redis for speed

Phase 5.3: offload static serving to CDN

Phase 5.2: cache config reads, currently every read = DB query

Phase 5.2: pre-render and cache HTML for high-traffic pages

Phase 5.4: replace print-based logging with structured JSON output

Phase 5.4: alerting on endpoints exceeding latency thresholds

Every LLM call must log: model name (provider/model), role, full prompt (or first 500 chars), full response (or first 1000 chars), latency_ms, token count, cost, status (ok/error/fallback). Written to PipelineLog + structured logs. Visible on /pipeline-viewer.

On /pipeline-viewer and /pipeline-audit: for each stage show which model was called, latency, tokens, cost. Not just input/output but the model routing decision.

/api/model-bus/call-stats visible on /ops: calls/hour per role, per provider, avg latency, error rate, fallback count, total cost.

Add a post-generation validation step for all image-generating agents: detect when a model produces off-topic, broken, or nonsensical output (like Nebula currently does). Check against expected style/content criteria, flag failures, block publication, and raise a system-wide alert (incident + Telegram notification). Should cover: prompt-image coherence check, style drift detection, degenerate output patterns (solid color, noise, text artifacts).

This page! Phase 1: view-only, Phase 2: interactive management

Extend the 3-level DALL-E prompt safety system (pre-sanitize trigger words, LLM smart retry, domain-specific fallback) to all agents that generate images. Currently implemented for Nebula/Heavens/Emerge — ensure Michi and any future agents also have domain-appropriate word sanitization maps, smart retry prompts, and safe fallback pools. Add monitoring: track how often each level is triggered per agent, log blocked prompts for analysis.

Phase 0: fix auth gaps in ops, voice, heavens

Phase 0.5: agent3_pages.py, agent4_pages.py — currently public despite docs saying logged-in

Phase 6: aggregation agent for Telegram channel

Aggregate health-matrix + recent generations from all agents, show 24h stats

Check staleness, incidents, DB/R2 health, scheduler states; produce OK/problem report

Register commands, new TELEGRAM_DISPATCHER_CHAT_ID env, extend webhook handler

Send thumbnail images of recent generations directly in Telegram chat

Connect Dispatcher to voice_routes.py / OpenAI Realtime API for spoken queries

Художник-провокатор, исследователь гротеска и карнавальной культуры. Маски, метафизический театр, петербургский авангард. Генерации в стиле Шемякина.

Visual generation pipeline: masks, metaphysical theater, Petersburg avant-garde. DALL-E prompts calibrated for Shemyakin aesthetic.

Corpus of Shemyakin's artistic principles: carnival theory, mask semiotics, grotesque as philosophical tool.

Методолог и системный мыслитель. Схемы мыследеятельности (СМД), организационно-деятельностные игры (ОДИ). Может анализировать работу других агентов методологически.

Analyzes processes of other agents: reflection quality, learning effectiveness, communication patterns. Produces methodological reports.

Implements Shchedrovitsky's SMD (Systems of Mental Activity) schemas. Can run ODI (Organizational Activity Games) with agents.