Friday, October 16, 2026

The state of Blackbox in mid October

AI illustration: a cutaway of a black computer case on a desk at night, filled with tiny floors: a round library at the top, a mail-sorting room, a bank vault, shelves of filed boxes, and a glowing thread running down into an open diary; a laptop, a coffee cup, and a dish of coins sit nearby.
AI illustration, not a photograph: inside the box, the library, the mail room, the vault, and the archive. Made by Chris with ChatGPT, prompted with the full text of this post.

Blackbox is seventeen days old as a real program, if you count from its first commit on September 30. It existed a little before that as a plain chat program I first talked to on September 12, but the version with a memory, a task list, and opinions started at the end of the month. Since then it has grown to 55 commits and roughly ten thousand lines of Python, almost all of it standard library, with 158 tests that run the real bb command against a fake Todoist, a fake Upwork, and a fake model in a throwaway home folder. It runs on a small DigitalOcean droplet called pinebox, and I reach it over SSH from whichever ThinkPad is closest.

I wrote a shorter post about it earlier today. This one is the full report, with the numbers out of its own database.

The heart of it is still the same idea: Todoist is the official list, and Blackbox is the secretary who reads it, remembers everything that ever happened to it, and tells me what I am about to drop. It has seen 497 tasks and logged 742 changes to them, whether I made the change through bb or in the Todoist app on my phone. It has written 37 briefs, 21 good mornings and 16 good nights, and 13 diary entries in ~/text/diary, from September 30 through last night. I have sent it 228 messages. Those diaries are the part I value most. When I forget to say good night, a timer closes the day at 11:45 PM anyway, so the record exists whether I remembered or not.

The Upwork side was the busiest stretch of building, from October 3 to October 7. Blackbox has triaged 88 jobs against my playbook. Twenty-one got APPLY, three APPLY (NO CONNECTS), two APPLY ALL THE CONNECTS, 44 PASS, and 18 RUN. Nothing has earned a FUCK NO yet. It drafted 25 proposals, each converted to plain ASCII and checked for length, placeholders, and the sign-off before I ever see it, and it never submits anything itself. Upwork currently lists 33 of my proposals, and one of them turned into a hire.

The plumbing has been boring in the best way. The Upwork and Google sign-ins renew themselves at 6:45 and 18:45 every day, and all 25 runs in the log say OK. Every night at about 3:30 the backup takes a consistent copy of the database and my writing, encrypts it, decrypts it again to prove it can, and pushes it to a private repository. Last night's snapshot was 96 files and 3.3 MB, with 14 nights kept.

The money is small. Blackbox keeps its own ledger, and the whole recorded spend since the start is $6.36, almost all of it 362 calls to OpenAI's GPT-5.6 Terra. That is also the first unfinished thing. Blackbox can route each job to a different Claude model, with Opus for briefs and email, Haiku for pulling structure out of loose sentences, and OpenAI as the fallback, and on September 30 I ran a blind bake-off to pick the winners. I never sent my picks. So the backend is still set to OpenAI, and the Claude routing sits there tested and unused.

The code then sat still from October 7 until this afternoon, because I spent that week and a half building websites with Claude Code instead. Today it moved again, and from a different direction. OpenAI's Codex, not Claude, wrote the newest feature, which bb calls continuity. A project can now collect its sources, meaning minutes, tasks, emails, time entries, and pitches, into a dossier. Projects get explicit aliases, so the names I actually use find the right one. I can leave a checkpoint saying where I stopped and what comes next, and later say bb get me back into the project. I can tell it bb I have 25 minutes and it picks one task with an estimate that fits and drafts the first piece of work. Decisions can carry tripwires that come back on a date, scope can carry a cap in minutes with warnings at 80 percent, and an evidence bank holds claims I have checked and approved so proposals can cite real work without inventing any. I have not used any of it yet. It shipped with its own tests and a verification note.

What is still on the list: send the bake-off picks and switch the backend to Claude; let a model propose minutes and tasks from sentences that match no command, with my yes before anything is saved; a handoff command that packages a problem for ChatGPT or another tool; a daily email digest; and the public version, working name Adjutant, with the patterns written up and my name and clients stripped out.

Every command it knows now lives on its own wiki page, Blackbox commands, regenerated from bb help whenever the program changes, so the list stays current while this post stays dated. The Upwork verdicts page explains the triage ladder, and Blackbox has the overview.

The prompt for “The state of Blackbox in mid October”

I want you to write a report on the state of bb ... In a bloggy way ... And I want you to include all the commands in it.