TaskBuddy

An adaptive goal-planning app. You state a goal, TaskBuddy breaks it into tasks, schedules everything into one timeline across all your goals, and tells you the probability you actually finish on time. When reality diverges from the plan, it proposes a recovery strategy instead of quietly going red.

Design philosophy: the LLM is used for understanding (turning messy notes into structured tasks, reading a written check-in). Every number, the priority score, the schedule, the completion odds, the recommended move, comes from deterministic code you can read. No probability is ever produced by a model.

Stack

Next.js 16 (App Router), React 19, TypeScript, Tailwind CSS v4, Amazon Aurora PostgreSQL Serverless v2, Amazon Cognito for auth, Amazon Bedrock for the LLM layer. Runs on AWS Lambda behind CloudFront, with the async half on EventBridge, SQS and Step Functions.

Getting started

pnpm install
pnpm dev

Open http://localhost:3000.

With no environment configured, the app runs in demo mode: an in-memory store seeded with sample data, and an offline heuristic parser instead of the LLM. Everything works, but data resets when the server restarts. The sidebar shows a "Demo mode" badge while keys are missing.

Going live

Copy .env.local.example to .env.local and fill it in. That file documents every variable and the reasoning behind it; the short version:

Variable	Purpose
PGHOST	Aurora writer endpoint. The app connects as taskbuddy_app with an IAM token; there is no password. Unset means demo mode.
COGNITO_USER_POOL_ID / COGNITO_CLIENT_ID	Auth. Neither is a secret, and the client secret is never stored: it is read at cold start.
SIGNUP_CODE	Required to create an account. There is no email verification.
BEDROCK_MODEL	Inference profile id. Unset falls back to a documented default.

Note what is not on that list: no database password, no API key, and no SESSION_SECRET. Every credential is either an IAM token derived from the execution role or a public identifier.

Apply the schema with aws/scripts/apply-sql.sh, which runs aws/sql/01_schema.sql through 04_cognito.sql in order and then verifies them.

Each layer degrades on its own. No database means the in-memory store; no Foundry credentials means the heuristic extractor; a failed LLM call falls back to the heuristic rather than erroring the request.

Architecture

notes or goal -> extraction -> priority scoring -> allocation -> forecast -> strategy

Layer	What it does	Where
Extraction	Notes into structured tasks, via LLM or offline heuristic	lib/extraction.ts, lib/heuristic.ts
Priority	Deterministic six-factor score	lib/priority.ts
Scheduling	Dependency ordering, day packing, within-day arrangement	lib/schedule.ts, lib/allocate.ts, lib/arrange.ts
Forecast	Monte Carlo completion probability against real availability	lib/forecast.ts
Strategy	Cross-goal recovery moves when the odds drop	lib/portfolio-strategist.ts, lib/strategist.ts
Data	Query layer over node-postgres, with an in-memory fallback	lib/store.ts, lib/db/
Auth	Cognito user pool, verified ID tokens, RLS per request	lib/auth.ts, lib/session.ts, lib/cognito.ts
LLM client	Bedrock Converse with strict JSON schemas	lib/bedrock.ts
Mutations	Server Actions	lib/actions.ts
Async jobs	Job bodies shared by actions and SQS workers	lib/job-handlers.ts, lib/jobs.ts

Priority score = Urgency*0.30 + Impact*0.25 + Dependency*0.20 + Risk*0.15 + Confidence*0.10 - Effort*0.10, mapped to Critical / High / Medium / Low / Backlog bands.

Every row is read through Postgres row-level security. The app connects as an unprivileged role and sets the current user id per transaction, so a query cannot return another account's data even if the application logic is wrong.

Cognito mints its own subject id, so it is deliberately not the app's user id. The Postgres users.id travels in the ID token as custom:app_uid, which is what lets every foreign key and every policy stay exactly as it was, and what keeps "who is this?" free of a database round trip.

Routes

Route	What it is
/	Today: agenda, time budget, completion odds, current strategy
/strategy	Cross-goal strategy, recovery moves, plan history
/board	Kanban board across every goal
/projects	Goals, each with its own forecast
/activities	Routines and recurring commitments
/create	New entry, paste notes or state a goal
/settings	Value model, availability, weekly capacity
/login, /signup	Self-hosted email and password auth

Infrastructure

The whole AWS stack is defined as CDK in aws/infra/ - six stacks: data, auth, events, web, observability, and edge. Always run cdk diff before applying.

Three properties are cost controls sitting on an exact edge, and each one reads like an ops choice:

• DB_MIN_ACU = 0 is what lets Aurora auto-pause. Setting it to 0.5 reads like "always keep a little capacity" and takes the cluster from about $10/mo to about $50/mo with no other change and no warning.
• idleTimeoutMillis: 10_000 in lib/db/pool.ts is no longer just hygiene. Aurora will not pause while any connection is open, so a pool that held connections between page views would keep capacity awake around the clock.
• No NAT gateway. One would cost more per month than everything else in the architecture combined. That is why Lambda is not VPC-attached and why the cluster is publicly routable - a trade paid for with IAM-only authentication, forced TLS against a pinned regional CA, and forced RLS.

Two cdk synth warnings (W2508, W9011) are expected on every run and are documented in aws/README.md; they cannot be suppressed, and if either ever stops appearing the network posture changed.

Long LLM work does not run in a request. Eleven call sites, seven at medium reasoning effort and one measured at 43 seconds, are moved behind an EventBridge bus into an SQS queue with a DLQ and a concurrency cap; the skill-link fan-out runs through a Step Functions Distributed Map so a throttled judgement retries instead of being silently dropped.

Everything sits in ap-southeast-1, about 30ms from Jakarta. one region covers the whole stack.

Scripts

pnpm dev      # development server
pnpm build    # production build
pnpm start    # serve the production build
pnpm lint     # ESLint

npx tsx aws/harness/offline.ts              # 26 offline assertions, no network
bash aws/scripts/preflight.sh               # is this account ready to deploy?
bash aws/scripts/build-web.sh               # assemble the Lambda bundle
TASKBUDDY_ALERT_EMAIL=... bash aws/scripts/deploy.sh