TaskStax — Agent-Driven Task Planner
A personal project in active development: a task manager where an LLM agent turns goals into milestones and tasks and can change the app through tools, while priority scoring, dependency resolution and daily scheduling stay deterministic code. Not public yet; the repository opens once it is finished.
The Design Constraint
The agent does the reasoning that needs language: decomposing a vague goal into milestones, splitting a task, writing a weekly reflection. Anything that has a right answer is plain code: the priority score, dependency order and the daily plan. That split keeps the numbers repeatable and testable, and it limits what a model can get wrong.
The agent is bounded planning inside a graph I designed, not an open-ended autonomous loop. A supervisor routes between planner, reflection and tool-execution nodes, and it stops for approval before it applies a batch of changes.
What It Does
- Goal decomposition with review — The planner returns milestone trees and tasks as structured, schema-validated output. A human-in-the-loop interrupt lets the user edit or reject the plan before anything is written to the backlog.
- Deterministic priority score — A weighted score over importance, urgency, deadline proximity and effort, computed in code rather than asked of the model.
- Dependency resolution — Task blocking is resolved with a topological sort (Kahn’s algorithm). Circular dependencies are rejected with the cycle reported.
- Daily timebox scheduler — Unblocked tasks are packed greedily into the minutes available that day, so a plan never exceeds the capacity the user set.
- Permissioned agent tools — Stax, the in-app assistant, changes the app through tool calls. Each capability sits behind a per-user AI permission; for example, theme changes are refused unless the user enabled them in settings. The UI is kept in sync over server-sent events.
- Memory and reflection — User context lives in a hybrid memory store on Postgres with pgvector, with upsert and deduplication. Weekly reflections are generated from recorded activity events rather than from the model’s guess.
Frontend
The web app is a Next.js 16 dashboard over the API. It shows tasks and the daily plan, lets the user review, edit or approve the agent’s proposed plan before anything is written to the backlog, and updates live over server-sent events when the agent or API changes data.
Isolation and Quality
Every query, memory search and agent tool call is scoped to the authenticated user. Passwords use PBKDF2-HMAC-SHA256 and the API issues signed JWTs. The repository has unit tests for the priority, graph, scheduler, auth and access-control logic, and integration tests that check multi-tenant isolation through the API.
Status
The core backend, agent graph and dashboard are built and I am finishing the remaining features. I will publish the repository when it is complete.
Technology Stack
| Backend | Python 3.13, FastAPI, SQLAlchemy 2.0, Alembic, uv monorepo |
|---|---|
| Agents | LangGraph supervisor with planner, reflection and tool nodes, Gemini with structured output |
| Data | PostgreSQL on Supabase with pgvector for the hybrid memory store |
| Frontend | Next.js 16 dashboard, server-sent events for live sync |
| Delivery | Docker multi-stage build, Docker Compose, pytest, Ruff |