Portfolio PM · Reference

The Portfolio Card Pipeline

What actually happens between "you type an ask on a card" and "the work is live and the card is done."

Reference page · four machines cooperate on every card · three of the stops are AI judgment, the rest is plumbing

YOUR BROWSER PM BRAIN · CLOUD VM WORKSHOP · YOUR PC GITHUB + PRODUCTION Your ask Your review PM writes ticket ◆ PM judges it ◆ Close gate Dispatcher Coder builds ◆ PR gate + CI Receipt + door check writes the card "looks good, close it" claims the job (only while PC is awake) judged against YOUR words ◆ AI judgment call (3) — everything else is deterministic plumbing
One card, four machines. Blue-outlined stops are AI judgment; every other hop is mechanical and verifiable.

The stops, in plain English

  1. 1Filed.

    You type the ask on a card (or create a new project). Your exact words are saved — they outrank every translation made later.

  2. 2Routed.

    The PM reads it, writes a work ticket, and queues it for the coding bench. Your verbatim words ride along with the ticket and are marked authoritative.

  3. 3Building.

    The workshop on your PC hands the job to a coding agent (the bench rotates between three). It builds in isolation and ships through the same pull-request gate all code uses.

  4. 4Checked.

    The platform independently verifies the change is merged and actually running in production (the "door check" — not the coder's word, the platform's). Then the PM judges the result against your original words.

  5. 5Your review.

    The card lands in "Ready for your eyes" — or "Missed the mark" if the finished work does not match what you asked, in the PM's honest judgment.

  6. 6Done.

    You say "close it" and the card closes on your authority. The PM records a closing memo for the archive; imperfect paperwork is noted, never a reason to keep the card open.

Where it can stall — and what happens now

A build run dies. The first technical failure requeues itself automatically — same instructions, next coding agent. Only a second identical failure comes to you, because two identical failures usually mean the ticket needs a change.
Your PC is asleep. The workshop lives on your machine, so queued work waits until it wakes. Nothing is lost; the board should read as weather, not neglect.
The work does not match the ask. The card is flagged "Missed the mark" instead of quietly turning green — that is the read-back doing its job. Your words drive the redo.

dvo88.com reference · the live board is at /portfolio · this page describes the pipeline as designed and shipped through 2026-08-21