agent_da

An agent that helps you achieve long-term goals by setting daily tasks to complete.

You describe an agenda and a timeframe. The agent writes a progressive day-by-day plan that lands on your deadline, turns each day into a concrete to-do list, reminds you on your chosen channel at your chosen time, and rewards completion with points and streaks.

Status

Backend complete — 36 tests passing. The React frontend covers the first-agenda wizard (sign in → capture → plan → review → approve); the today dashboard and progress views are still to come.

Milestone	State
M1 Skeleton: FastAPI + Postgres + Alembic + auth + Render blueprint	✅
M2 Plan generation: LangGraph intake → clarify → plan → critique → repair → review	✅
M3 Daily execution: materialisation, to-do CRUD, completion, points, dashboard	✅
M4 Delivery: SMTP / Twilio / Meta / console channels, dispatcher, scheduler + cron	✅
M5 Gamification: streaks, missed-day penalties, bonuses, weekly rollup	✅
M6 Hardening: tests, rate limits, logging, health checks, migrations	✅
M7 Frontend: React + Vite + Tailwind	🚧 agenda wizard shipped

Quickstart (no credentials needed)

With no NVIDIA_API_KEY, the agent uses a deterministic offline planner; with no SMTP or Twilio credentials, reminders are logged to the console. The whole product is exercisable locally.

cd apps/api
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

# watch the entire product work, goal to points
.venv/bin/python -m app.cli demo

# or run the API
.venv/bin/uvicorn app.main:app --reload   # -> http://127.0.0.1:8000/docs

Full instructions, tests and deployment: docs/running.md.

Design documents

Document	Contents
docs/architecture.md	Decisions, system diagram, runtime flows, points and streak rules, Render deployment, risks
docs/data-model.md	Postgres schema, including the ledger idempotency keys that make rewards retry-safe
docs/langgraph-agent.md	Agent state, graph, node-by-node behaviour, prompt constraints, failure handling
docs/api-contract.md	All 38 routes with request/response examples
docs/running.md	Install, configure, run, test, deploy

Stack

• Agent & backend: LangGraph + FastAPI (Python)
• LLM: nvidia/nemotron-3-ultra-550b-a55b via NVIDIA NIM, structured output
• Database: Render Postgres (async SQLAlchemy + Alembic); SQLite locally
• Frontend: React + Vite + Tailwind CSS (agenda wizard shipped)
• Reminders: APScheduler dispatcher in-process, Render Cron as a safety net
• Channels: SMTP email, Twilio WhatsApp, Meta WhatsApp Cloud API, console fallback
• Hosting: Render (render.yaml)

How the hard parts are handled

• The plan always lands on the deadline. The agent works against a deterministically built, dated skeleton. Model output is re-normalised onto it, so the number of days and their dates can never drift — a truncated response degrades into a valid plan plus warnings, not a broken one.
• Reminders go out exactly once. A unique dedupe_key per (agenda, day, kind, date) plus SELECT … FOR UPDATE SKIP LOCKED claiming, shared by the in-process scheduler and the Render cron.
• Points can't be double-awarded. The ledger is append-only and every row has a deterministic idempotency key; un-completing writes compensating negative rows instead of deleting.
• Long agendas stay affordable. To-dos are expanded for the first 3 days at approval, then lazily the day before, with the previous day's actual completion fed back into the prompt.
• It runs with zero credentials. Every integration degrades: offline planner, console channel.

Next step

The first-agenda wizard (M7) is built and smoke-tested: sign in, capture an intention and deadline, watch the agent plan, answer its questions, review the drafted days, and approve. Run it with npm run dev:all from apps/web, or npm run e2e:all to drive the whole flow headless — see docs/running.md.

What remains is the rest of the frontend: the today dashboard, to-do completion, and progress views. The API contract they need is already frozen in docs/api-contract.md.