Synth Nova Manifest

Active Plan

8 min read

Active Plan

Живой tactical roadmap. Что делаем прямо сейчас. Обновляется ежедневно.

Для долгосрочной стратегии — см. MVP-Phase. Для foundational context — см. ADR-0023-Monopreneur-Principle. Для контекста текущей фазы — см. Session-Context-2026-04-18. Для открытых вопросов — см. _Open-Questions.

Monopreneur note for task owners

Все core tasks исполняются Claude Code через Founder как оператора в Termius. Это not boilerplate — это central principle per ADR-0023-Monopreneur-Principle.

External contractor engagements отмечены явно — только для GeeLark adversarial engineering где LLM не справляется без human expertise.

Статус по состоянию на 2026-04-19

Текущая фаза: Фаза 0 — Фундамент

Ключевые события последние дни:

  • Testing report 2026-04-18: three working production products confirmed
  • Стадия пересмотрена на late seed / pre-A
  • Chamber positioning: компонент agency, не standalone (Q10)
  • Chamber-first research architecture — new direction
  • Pipeline Memory (Task 5.5) — priority decision
  • Intake Agent spec готов (Python/Streamlit)
  • Production Agents Gap — Variant C (stubs) + B (3 priority agents) in backlog
  • Investment Intelligence Research Track — parallel exploratory (Q13)
  • Monopreneur Principle formally adopted via ADR-0023

В работе сейчас:

  • Parallel workers fix (Claude Code в Termius)
  • Navigator retest после Director timeout fix

Bottleneck: retest pending.

Production Agents Gap (важный контекст)

Реальный клиентский ТЗ (пример: эзотерика на 13 секций) требует:

Navigator покрывает (4 секции quality):

  • Market analysis via Researcher
  • Geo analysis (частично) via Researcher
  • Unit economics via Financial Modeler
  • Competitor analysis via Scout
  • Plus Aggregate report

Navigator не покрывает (9 секций):

  • Product matrix — Agent-ProductDesigner
  • Audience segmentation + avatars — Agent-AudienceSegmenter
  • CJM — Agent-JourneyMapper
  • DM funnels — Agent-FunnelArchitect
  • Sales scripts — Agent-Copywriter
  • Non-DM funnels — Agent-FunnelArchitect
  • Content strategy — Agent-ContentStrategist
  • Localization — Agent-Localizer
  • AI sales architecture — Agent-SalesArchitect

Стратегия: Variant C — stubs для 9 секций сейчас via Jinja + Sonnet. Priority Build 3 agents (AudienceSegmenter, FunnelArchitect, Copywriter) в backlog — all built with memory integration from start.

Активные задачи (приоритизированы)

Task 1 — Retest Navigator после Director timeout fix

Owner: Claude Code + periodic tester Status: in-progress Priority: P0 (blocks everything) Effort: 1-2 часа

3-5 runs на разных типах запросов. Замер judge score (baseline 7.25, expect ≥ 7.5). Документация results.

Blocks: Task 2, 3, 4, 5

Task 2 — P0 bug fixes из testing report

Owner: Claude Code Status: blocked by Task 1 Priority: P0 Effort: 2-3 часа

  1. Chamber confirmation dialog (HIGH) — setShowConfirm(false)
  2. LaTeX $...$ escaping (MED)
  3. GPT-4o model substitution (MED) — pin версию

LOW (backlog): Founder Action dedup, Heatmap sync.

Blocks: Task 3

Task 3 — Retest Chamber после bug fixes

Owner: Periodic tester Status: blocked by Task 2 Priority: P0 Effort: 1 час

Verify fixes + no regression.

Blocks: Task 4

Task 4 — Commit manifest updates

Owner: Claude Code Status: blocked by Task 3 Priority: P0 Effort: 30 минут

Один atomic commit с 7 файлами (обновлённый список после Monopreneur framing):

  1. 06-Business/Working-Products.md — новый
  2. 07-Roadmap/Session-Context-2026-04-18.md — internal через ignorePatterns
  3. 09-Decisions/_Open-Questions.md — санитизированный
  4. 03-Roles/Agent-Intake.md — Python/Streamlit spec
  5. 07-Roadmap/Active-Plan.md — этот файл
  6. 07-Roadmap/Research-Track-Investment-Intelligence.md — exploratory
  7. 09-Decisions/ADR-0023-Monopreneur-Principle.md — новый, foundational

Plus Manifesto update (Founder’s decision после review SUGGESTED-Manifesto-update.md).

Blocks: Task 5, 5.5, 5.6, 5.9

Task 5.5 — Pipeline Memory (RAG для learning loop)

Owner: Claude Code Status: blocked by Task 4 Priority: P0 (biggest judge score gain) Effort: 2-3 часа

Два шага:

  1. Ingest reports в RAG — new pipeline_memory collection (не manifest_chunks). Bulk ingest 25+ runs (judgement, log, status) + automated incremental via file watcher.

  2. Pre-stage RAG lookup hook — каждый агент перед запуском делает RAG query по имени + task context. Top 5 learnings prepended к system prompt. Latency < 500ms.

Guardrails: top-5 filter, no negative overfitting from edge cases, periodic audit через 100 runs.

Acceptance:

  • 25+ existing runs ingested
  • Judge score на следующих 3-5 runs > 7.25 baseline
  • Automated ingest new runs

Impact: +0.5-1.0 judge score на ВСЕ существующие агенты. Exponentially compounds. Все будущие агенты строятся с memory integration сразу.

Blocks: Task 5, 5.6, 5.9 (они benefit from memory)

Task 5 — Navigator Extended mode (stubs для 9 секций)

Owner: Claude Code Status: blocked by Task 5.5 Priority: P0 Effort: 2-3 дня

Add stage после Aggregate — Extended Sections Generator. Optional по flag в spec от Intake Agent.

9 Jinja templates в src/synth_brain/extended/templates/ (product- matrix, audience-avatars, CJM, DM funnels, sales scripts, non-DM funnels, content strategy, localization, AI sales architecture).

Claude Sonnet (не Haiku — quality важна). Parallel execution. Каждая секция помечена [Preliminary — dedicated agent pending].

Self-check guardrails: если stub generic → flag INSUFFICIENT_RESEARCH.

Cost +$0.50-1.00, duration +3-5 мин.

Blocks: Task 6

Task 5.6 — Quick/Full Analysis Modes

Owner: Claude Code (parallel с Task 5.9) Status: blocked by Task 4 Priority: P0 (demo-ready value) Effort: 1 день

Quick mode (5 мин, $0.50): Scout + Researcher + Aggregate только. Без FM, Rating, Product Designer, Audience, Funnel, Content, Copywriter.

Full mode (20-30 мин, $3-5): все agents + Extended stubs если требуется.

Mode selection:

  • Intake Agent (when built) определяет по запросу
  • Либо explicit user choice с cost/duration preview

Impact: demo-value для инвестора. Quick mode = wow-effect за 5 минут. Full mode остаётся для serious depth.

Task 5.9 — Chamber-first Research Architecture

Owner: Claude Code (parallel с Task 5.6) Status: blocked by Task 4 Priority: P0 (positioning moat) Effort: 1-2 дня

Scout и Researcher используют Chamber в начале как primary knowledge source (not only final arbiter).

def scout_competitors(query):
    # Step 1: Chamber first (2 мин, $0.19)
    chamber_result = chamber.query(
        prompt=f"List competitors for {query} with confidence 0-10",
        models=['claude', 'gpt-4', 'gemini', 'grok', 'deepseek']
    )
 
    # Step 2: Consensus + hallucination check
    consensus = measure_cross_model_agreement(chamber_result)
    hallucination_signals = check_specific_patterns(chamber_result)
 
    # Step 3: Branch
    if consensus >= 0.8 and hallucination_signals.low:
        return format(chamber_result)  # 80% queries, fast path
    else:
        verified = web_search_verify(chamber_list)  # 20% queries
        return merge(chamber_result, verified)

Impact:

  • Scout: 8-12 мин → 2-3 мин (80% queries)
  • Researcher: 10-15 мин → 3-5 мин (80% queries)
  • Cost примерно тот же (Chamber $0.19 ≈ web search API costs)
  • Quality: 5-LLM consensus starting point

Positioning value: превращает Chamber из “multi-LLM deliberation” в primary knowledge infrastructure. Moat для investor story.

Task 6 — Intake Agent Phase 1 (MVP-lite)

Owner: Claude Code + Co-Developer (planning) Status: blocked by Task 5, 5.5 Priority: P1 Effort: 4-5 дней

См. Agent-Intake Phase 1.

Intake Agent подключается к Pipeline Memory с самого начала (separate intake_memory namespace). Знает про Extended mode и Quick/Full modes. При spec требующем 13 секций — honestly коммуницирует scope split (“4 full research + 9 preliminary + Quick/Full choice”).

Blocks: Intake Phase 2

Upcoming (P2 — следующие 2-3 недели)

Intake Agent Phase 2 (1 неделя)

  • Chamber integration for brainstorm gaps
  • Voice interface (Whisper + OpenAI TTS)

Production Agents Priority Build (3-4 недели, sequential)

Решение 2026-04-18: all 3 in priority, built sequentially, all with memory integration from start.

Agent 1 — Agent-AudienceSegmenter (неделя)

  • Avatars с depth: fears, desires, triggers, objections
  • Использует research + pipeline memory
  • Заменяет stub audience-avatars.md.jinja

Agent 2 — Agent-FunnelArchitect (неделя)

  • DM funnels + non-DM funnels (landing, webinar, quiz)
  • Platform-specific
  • Заменяет stubs dm-funnels и non-dm-funnels

Agent 3 — Agent-Copywriter (неделя)

  • 5 sales scripts по price matrix
  • С учётом avatars from Agent-AudienceSegmenter
  • Заменяет stub sales-scripts.md.jinja

После: coverage grows 4/13 quality → 7/13 quality + 6/13 stubs.

Web search caching (demoted)

После Chamber-first research — caching нужен только для 20% fallback cases. Lower priority чем раньше.

Haiku для simple Researcher queries

Малый effort. Интегрируется в Pipeline Memory rules (“for simple queries use Haiku”).

GeeLark operational debt (1-2 дня)

Owner: Разработчик-1 (external contractor engagement per ADR-0023)

  • Rotate Gemini key
  • Delete duplicate flows
  • Fill descriptions

Production Agents Backlog (P3 — после priority build)

6 agents в очереди, не все требуют dedicated builds — некоторые могут оставаться stubs:

  • Agent-ProductDesigner (high priority)
  • Agent-JourneyMapper (medium)
  • Agent-ContentStrategist (medium, planned)
  • Agent-Localizer (lower)
  • Agent-SalesArchitect (lower)

Criteria для stub → agent promotion:

  1. Stub регулярно выдаёт INSUFFICIENT_RESEARCH flag
  2. Client feedback указывает quality gap
  3. Revenue-correlated section

Research Tracks (parallel P4)

Backlog (P4 — без конкретных дат)

  • Constitution translation на русский (Q9)
  • Media-weight terminology (Q1)
  • Product framing finalization (Q2) — мostly resolved через Chamber positioning
  • Fundraise preparation session (Q4) — с Monopreneur hook
  • 4 scenarios spec (Q8)
  • Big filename unification migration

Blockers

Текущие

Ни одного open blocker.

Watch list

  • LLM budget при scale — Intake Phase 2 + Extended stubs + Chamber- first = additional $0.70-1.20 per task. Monitor в Reports UI.
  • Researcher quality — без 8.0 quality секции страдают, stubs тем более.
  • Stub quality drift — periodic review каждых 100 runs.
  • Rate limiting conflicts — unified budget pool нужен.
  • Monopreneur purity drift — quarterly audit: какой % работы requires external contractors? Target: core 95%+ LLM.

Daily ritual

Утром:

  1. Review Active Plan — queued / blocked / in-progress
  2. Разблокированные → in-progress
  3. Новые задачи добавить

Вечером:

  1. Обновить статусы
  2. Записать findings в session notes
  3. Update blockers

Еженедельно:

  1. Archive completed tasks в Sprint-Week-N.md
  2. Re-prioritize backlog
  3. Update MVP-Phase если milestone
  4. Monopreneur purity check — не дрейфуем ли к обычной команде?

Update log

ДатаChangeАвтор
2026-04-18Initial creation — 5-step stabilization + Intake roadmapFounder + Strategy Claude
2026-04-18Added Production Agents Gap — Variant C + B in backlogFounder + Strategy Claude
2026-04-19Added Task 5.5 Pipeline Memory (priority decision)Founder + Strategy Claude
2026-04-19Added Task 5.6 Quick/Full modes, Task 5.9 Chamber-first researchFounder + Strategy Claude
2026-04-19Added Research Tracks section (Investment Intelligence)Founder + Strategy Claude
2026-04-19Monopreneur framing: corrected task owners, added ADR-0023 referenceFounder + Strategy Claude

Связанные документы

Graph View

Backlinks