IntelliOps — Roadmap

Each sprint is scoped to one week.

Planos instrucionais detalhados (código, padrões, dicas) estão em docs/sprints/:

Planos de implementação detalhados (task-by-task com código) estão em docs/superpowers/plans/:

Sprint 1 — Calibration

Replace httpx full-page fetch with Firecrawl — handles Cloudflare, JS-rendered pages, and blocked sites; falls back to Tavily snippet on failure
Validate each mode prompt produces the correct deliverables
Run all 4 modes with tests/run_mode_{mode}.py and evaluate report quality and search coverage
Adjust search cap per mode if comparative modes (Competitor Intel, Vendor Evaluation) show thin findings
Re-run after adjustments and compare token usage and output quality against baseline

Sprint 2 — New Intelligence Modes

Each mode = new file in backend/prompts/modes/ + structured form inputs + entry in MODE_FILES.

Market Mapping — map players, segments, and positioning across a sector
Leadership Intel — executive background, track record, and professional connections
Funding & Deal Intelligence — investment rounds, M&A activity, and capital movements
Risk Assessment — multi-dimensional scorecard: reputational, financial, regulatory, geopolitical
Regulatory Watch — regulatory changes by sector and jurisdiction
Talent Signal — hiring patterns as a proxy for undisclosed strategic direction
Partnership & Ecosystem Mapping — alliances, integrations, and partner ecosystem

Sprint 3 — Evals

LLM-as-judge eval script — runs a fixed set of queries per mode and scores response quality
Eval criteria per mode (e.g. coverage of required sections, citation density, factual specificity)
Eval results saved to tests/evals/ for tracking quality over time
Baseline established for all modes before any further prompt changes

Sprint 4 — Production Readiness

Dockerfile for FastAPI + LangGraph stack
docker-compose.yml — brings up backend + LangGraph server together
CI pipeline — runs integration tests and evals on push
.env.example review — ensure all required and optional vars are documented

Sprint 5 — Report History & Search

Motivação: relatórios de inteligência são gerados ao vivo via web search — o valor está na frescura dos dados, não no histórico. RAG (busca vetorial em relatórios antigos) introduz complexidade de infraestrutura (pgvector, embedding model, chunking strategy) por benefício marginal: o usuário que quer informação atualizada vai re-rodar o agente, não perguntar para um relatório de 3 meses atrás. Busca por metadados é suficiente para o caso de uso.

Armazenar relatórios no Supabase após stream (researches + reports tables do Sprint 6)
GET /reports — lista relatórios com filtros: modo, empresa, data
Full-text search nos relatórios via PostgreSQL tsvector (sem embedding, sem pgvector)
Dashboard de histórico no frontend: card grid com filtro por modo, empresa, data
/report/:id — visualização do relatório salvo com export (PDF, Obsidian, Slack)

Sprint 6 — SaaS Frontend (React)

Stack: Vite + React SPA · Supabase (Auth + PostgreSQL + Storage)

Auth

Email/password login (Supabase Auth)
Google OAuth login (Supabase Auth)
AuthGuard — protected route wrapper
User profile (/settings)

Dashboard

Card grid per research (company, mode, date, status)
Filter sidebar: mode, date presets (today / week / month), company search
Card click opens /report/:id

Research

Mode-specific form (ResearchForm) replicating current build_query() logic
Real-time SSE streaming (StreamViewer) — Activity + Report panels
Auto-save report on done event

Saved Report

Rendered markdown view at /report/:id
Export: PDF, Obsidian, Slack

Backend (FastAPI)

POST /reports/save — persists research + report to Supabase after stream ends

Database (Supabase)

profiles table (id = auth.uid, full_name, avatar, org_id)
researches table (id, user_id, mode, company, query, status, token_count, created_at)
reports table (id, research_id, markdown_content, pdf_url, updated_at)
Row Level Security by user_id on all tables
Supabase Storage — PDFs per research

Sprint 7 — Intelligence Terminal (Base)

Spec: docs/superpowers/specs/2026-04-08-intelligence-terminal-design.md

Tabelas Supabase: entities, relationships, dossiers, entity_drafts
Extrator de entidades pós-run (Haiku-4.5) — processa relatório e gera rascunho JSON
Endpoint POST /entities/review — HITL gate antes de persistir no grafo
Endpoint GET /entities/graph — carrega nós e arestas para o frontend
Painel HITL de curadoria no frontend (aprovar / descartar / editar entidades)
Grafo D3.js: force-directed, iniciais dentro do nó, label abaixo, cor por tipo, tamanho por weight
Hover sobre nó → tooltip glass flutuante próximo ao nó
Click no nó → split view (grafo dimmed + dossier com abas: Perfil / Conexões / Histórico / Jobs)
Filtros por tipo de entidade (company, person, org, event)

Sprint 8 — Monitoring Jobs Engine