Local-first, single-user research agent. Producer-consumer pipeline: LangGraph orchestrates DeepAgents-style stage agents (framing → review → research → synthesis → audit) over your own documents (Docling-parsed) plus pluggable web search (Tavily / Exa / Perplexity). Output is a Markdown report with [^src_id] citations backed by a per-run Evidence table.

• Backend: Python 3.12, FastAPI, SQLAlchemy 2 async, Alembic, Postgres + pgvector.
• Agents: LangGraph state machines + langchain_core chat models. No LiteLLM.
• RAG: Docling for parsing, pgvector for retrieval. No external vector DB.
• Frontend: Next.js 16 (App Router) + React 19 + Tailwind v4 + shadcn-style UI.
• Streaming: Server-Sent Events with replay (?last_event_id= resume).
• Storage of secrets: Fernet-encrypted at rest in Postgres.

backend/    FastAPI app, agents, workers, alembic migrations
frontend/   Next.js 16 app (chat, run workspace, settings)
infra/      docker-compose for Postgres + pgvector
docs/       Plans, ADRs
scripts/    check.sh — run all gates locally

• Python 3.12 + uv
• Node 20+ + pnpm
• Docker (for Postgres)

docker compose -f infra/docker-compose.yml up -d

cp .env.example .env
# generate FERNET_KEY (see comment in .env.example) and paste it in
cp backend/.env.example backend/.env       # if you prefer per-package envs
cp frontend/.env.example frontend/.env.local

cd backend
uv sync
uv run alembic upgrade head
uv run uvicorn app.main:app --reload --port 8000

cd frontend
pnpm install
pnpm dev

Open http://localhost:3000.

Visit /settings:

• Add at least one LLM provider key (Anthropic / OpenAI / Google / Ollama).
• Pick a search provider (Tavily / Exa / Perplexity) and add its key.
• Optionally override the model used per role (framing / review / research / synthesis / audit).

Keys are encrypted with FERNET_KEY before being written to Postgres.

1. Pick a task on the home page (V1 ships Market Entry; M&A skeleton present, full graph pending in M8).
2. Enter a goal.
3. The framing agent produces a clarifying questionnaire — fill it in.
4. The pipeline streams agent messages, gate verdicts, artifact updates, and live token/cost usage to the run workspace.
5. The final report renders in the center pane with [^src_id] chips. Click a chip to scroll its source card into view in the right pane.
6. Cancel anytime via the "Cancel run" button in the Usage panel.

Four panes:

• Left: chat stream + agent trace (messages, gate verdicts, artifact writes).
• Center: streaming Markdown report with clickable citation chips.
• Right top: usage panel — running token totals + USD cost + cancel.
• Right bottom: sources sidebar — one card per cited source.

bash scripts/check.sh        # backend + frontend in one shot

or individually:

# backend
cd backend && uv run pytest && uv run ruff check . && uv run mypy app

# frontend
cd frontend && pnpm typecheck && pnpm lint && pnpm build

• docs/superpowers/plans/2026-04-25-consulting-research-agent-v1.md — V1 milestone plan.
• backend/README.md — backend module map and conventions.
• frontend/README.md — frontend module map.

V1 milestones M0–M7 complete: M0 infra → M1 settings/providers → M2 SSE plumbing → M3 chat shell → M4 ingestion + RAG → M5 run lifecycle + framing schema → M6 market-entry agent pipeline → M7 polish (token/cost tracking, citation chips, evidence sidebar, 4-pane workspace).

M8 (M&A consultant) tracked separately.