CMChristopher Musial

All projects

Live

2026 – present

Project & Task Command Center

A shared project management hub where my AI agents and I plan, assign, prioritize and track work.

Private repo, walkthrough on request

Version poll
Every 30 seconds
Full reload
Only when the hash changes
Idle pause
10 minutes

AI details

What the system actually does

Models

  • The board doesn't run a model itself. Bots call the tasks API.

Techniques

  • Token-authenticated tasks API for bots, and session login for people.
  • Dedupe a checklist line by sync-key, then T-ref, then fuzzy title, including archived cards.
  • Map bot status words onto initialized, working, approval, and complete.
  • Poll a version hash every 30 seconds and reload the task list only when the hash changes.

Data

  • Cards with a project, a T-ref, notes, and one of four columns.
  • Lines from the daily ops checklist, and one-card bot upserts with owner, status, and next step.

Evaluation

  • Archived cards are tombstones, so the sync doesn't revive or edit them.
  • A line that says the board is down is skipped.
  • The UI doesn't auto-complete a card just because a bot proposed it, except through the checklist promote-to-complete path.
  • Deletes and archive-all-completed ask for confirmation.

Challenges

  • Bots and people need one board without duplicate cards or revived archive rows.
  • Work that needs a person has to stay in the approval column until someone moves it.
  • Reloading the full task list on a timer wastes work when nothing changed.

Innovations

  • waiting-on-chris from a bot lands in the approval column, and a person still has to move it.
  • A cheap version endpoint returns a hash, cached briefly, so the open board reloads only when stored data changed.
  • Polling pauses while the tab is hidden, during a drag, or after ten minutes idle.

Stack and tags

  • Next.js
  • Vercel
  • Vercel Blob
  • dnd-kit
  • agents
  • approvals
  • human-in-the-loop
  • operations

Problem

Bot work kept landing in chat, and the daily checklist could drift from what was really in flight. I had no single board that a person and a bot could both update without duplicating cards or bringing archived work back.

Why I built it

Project & Task Command Center is the private Next.js task board on Vercel that my bots and I share. I don't publish the private board URL, API tokens, or session cookies here.

How it works

People sign in with a session, and bots call a token-authenticated /api/tasks API. Each card belongs to a project and has a T-ref, notes, and one of four columns: initialized, working, approval, or complete. The board uses dnd-kit for drag-and-drop, and you can search by T-ref or title, archive, and unarchive. Notes can hold a short thread in the form "Chris asked" / "Devyn replied."

New work shows up as a line in the daily ops checklist or as a bot upsert of one project card (owner, status, next step). The sync matches a checklist line by a sync-key stored in notes first, then by an explicit T-ref, then by fuzzy title across projects, including archived and complete cards. Archived cards are tombstones, so they're never revived or edited. A line saying the board is down gets skipped. When the checklist marks a matched live card done, it can be promoted to complete.

A bot upsert writes an Owner / Next step / Last update header into the notes and leaves the other note lines alone. Bot status words (todo, in-progress, waiting-on-chris, complete) map onto the four columns, so waiting-on-chris lands in approval.

A signed-in person can open a card, edit the title and notes, drag it, or use Move to. Work that needs a person sits in approval until someone moves it. Deletes and "archive all completed" ask for confirmation. Outside that checklist promote-to-complete path, the UI won't auto-complete a card just because a bot proposed it.

Instead of reloading the full task list on a timer, the board polls /api/version every 30 seconds. That endpoint returns a hash of the stored data, cached briefly on the server, and a full refresh runs only when the hash has changed. Polling pauses while the tab is hidden, while a drag or move is in flight, or after ten minutes idle. It starts again on pointer, keyboard, touch, scroll, or focus. Optimistic drag moves stay put until the server confirms them. Data lives in Vercel Blob, with a short history of prior snapshots.

Outcome

One live board that people and bots both see, with stable T-refs, so the daily checklist and the specialist bots match what's actually in flight.