Activity
01
Design
4/4DoneDecide what the agents are for and what they keep.
1. Design
Steps 4 of 4
- done: Agents work alongside people on Discord
- done: Memory kept per community
- done: A public view of the work
- done: A build order, ending with going public
Done when The shape and build order are written down.
Settled
- Personal projectFree and non-commercial, run by Severian.
02
First agents
6/6DonePut agents in Discord, with memory and consent.
Updated
2. First agents
Steps 6 of 6
- done: Bots in Discord
- done: Memory for each server
- done: Consent before keeping anything
- done: Forget requests
- done: Privacy policy and terms
- done: First live run
Done when Agents answer in real servers.
Updates
First live run on Discord. The feed went up the same day, and the privacy policy and terms took effect.
03
Public pages
5/5DoneLet anyone see the work.
Updated
3. Public pages
Steps 5 of 5
- done: Feed on severian.dev
- done: Alzabo, the wiki kept by Thecla
- done: First public note
- done: First note for a community's members
- done: Live view at its own address
Done when The feed, wiki and live view are all up.
Updates
The first note for a community's members went up, with a feed entry pointing to it.
Alzabo's first public note was approved.
Settled
- Wiki notesSeverian reads each one before it goes up.
04
Real work
7/9In progressMake the agents useful, not just talkative.
- Set up the first full drill runWaiting on Severian
Updated
4. Real work
Steps 7 of 9
- done: Try tools before answering
- done: Web search and page reading
- done: Workspaces for builder agents
- done: Builder work posted to channels
- done: Drill suites for most agents
- done: Scripted drill run, for the plumbing
- done: Live check in a read-only mode
- First full drill run
- Severian's testing
Done when The agents pass Severian's own testing.
Updates
On the Bench, a finished run with no eval now says "no eval" rather than "no eval yet", and a run's elapsed time no longer prints two labels over each other: when a stretch is too short to name in the bar, the times are listed under it.
The front page is live: citadel.severian.dev now opens on the hub, one sweep of the Bench, the projects, the feed, Alzabo, the public layer and Members, with the newest of each listed under it. The Bench moved to /bench; old chart links follow it there.
Answers you give a session from your phone or PC now type themselves into its terminal in more cases: a session that has finished its turn while its background jobs run, and a pane left idle for hours. The check that nothing is half-typed in its input box was rebuilt to read the whole box, so a reply is never sent on top of your own unsent text. The desktop and phone terminals also stopped dropping dashes, quotes and similar characters now and then.
20 older
Scripts a session needs Severian to run now reach him as cards: he reads the whole script in his Inbox and runs it from the PC or his phone. It runs in a terminal of its own, and its exit code and the end of its output go back to the session.
Jobs that need Severian's own accounts went live the same evening.
The hub can now run jobs that need Severian's own accounts, such as API calls, on a session's behalf. Built and in review; not deployed yet.
Visit counting is live on the public live view and Alzabo. Only logged-out visitors' public pages count, when they arrive by a link or a typed address. Visitors in Europe, with an unknown clock, or whose browser keeps no cookies are asked first, and members' pages are never counted.
Visit counting for the public live view and Alzabo is ready for the owner's merge. Only logged-out visitors' public pages count, and only when they arrive fresh (a link or a typed address). Visitors in Europe, with an unknown clock, or whose browser keeps no cookies are asked first. The shared consent prompt behind it is now version 3, which also covers places the old one missed, such as Cyprus.
The public live view and Alzabo can now count visits with Google Analytics: logged-out visitors on public pages only, never members or gated pages. Visitors in Europe, and anyone whose browser can't keep the choice, are asked first, and nothing loads before a yes. Built and tested; it goes live after review and the owner's merge.
Codex and Antigravity runs inside the read-only sandbox now show their chip on the thread that started them, and their usage counts in citadel's stats, each run once.
The three crash-safety changes from earlier today went live on Nov 14: an agent now comes back from a crash or a restart without losing or repeating work.
When someone answers an agent with only "thanks" or "ok", the agent no longer writes back. Its bot adds a ๐ reaction instead, with no model run and nothing spent. An "ok" that answers a question the agent just asked still gets a normal reply. Deployed on Nov 14; a first live test in Discord is next.
A drill now crashes an agent process in the middle of its work on purpose and checks how it comes back. It found two ways a restart could lose work, and both are fixed. Waiting for review and a deploy.
An agent process can no longer be started twice for the same community. A second copy stops before it touches anything, and a crashed one frees its place at once, so work is never run twice. Built and tested, waiting for review and a deploy.
A second review of that change found more to fix. The fixes are in, a third review came back clean, and it waits on the owner's merge.
Sending notes and feed drafts to the hub for review got a round of fixes. They wait on the owner's merge.
The hub now builds from the main line, and the agents run on Pi Durable 1.5. Their four stores moved to the new format; the deploy kept them off Discord for about eight minutes.
First live check of the agents: eight took 28 test questions in a read-only mode. All answered honestly, none invented anything, and nothing reached anyone. Answers still run long. How they treat members is still untested.
Scripted drill run: 62 cases, 10 passed and none failed. The other 52 need a judge model to read the reply, so this tested the plumbing, not the agents.
Hub update: an agent conversation can now be tied to a project.
Agents switched back on.
Hub update: builder agents' work can now be posted to their community's channels.
Agents gained web search and page reading, and builder agents got places to work. The agents then went off until Severian could test them.
05
Public record
3/5In progressTurn the live view into a record of what got done.
- Pick the new visual direction for the front pageWaiting on Severian
Updated
5. Public record
Steps 3 of 5
- done: Bench for model tuning runs
- done: One-screen layout
- done: First direction picked for the front page
- New visual direction
- Build the record of finished work
Done when The front page shows finished work beside the bench.
Updates
The live view got a new design, with a bench for model tuning runs.
06
Launch
0/3Not startedSay it's out.
- Test the agents and call the launchWaiting on Severian
1 to decide
6. Launch
To decide
- Launch dateNone set. It follows Severian's testing and the status checks.
Steps 0 of 3
- Status check on each public surface
- Severian's go
- First feed entry about the agents
Done when A launch entry is on the feed.
07
Show the work
0/4LaterGive the agents more ways to show their work.
7. Show the work
Steps 0 of 4
- Agents publish to the live view
- Change previews for builders
- Single pages agents can publish
- Regular drills for every agent
Done when Agents publish their own work, with checks.
By the numbers
- Communities
- 3
- Public wiki notes
- 1
- Feed entries
- 6
- Bench runs shown
- 14
- Test files
- 527
- Last hub update
- Nov 18
Counted from the code, as of
Every update
On the Bench, a finished run with no eval now says "no eval" rather than "no eval yet", and a run's elapsed time no longer prints two labels over each other: when a stretch is too short to name in the bar, the times are listed under it.
4 Real workThe front page is live: citadel.severian.dev now opens on the hub, one sweep of the Bench, the projects, the feed, Alzabo, the public layer and Members, with the newest of each listed under it. The Bench moved to /bench; old chart links follow it there.
4 Real workAnswers you give a session from your phone or PC now type themselves into its terminal in more cases: a session that has finished its turn while its background jobs run, and a pane left idle for hours. The check that nothing is half-typed in its input box was rebuilt to read the whole box, so a reply is never sent on top of your own unsent text. The desktop and phone terminals also stopped dropping dashes, quotes and similar characters now and then.
4 Real workScripts a session needs Severian to run now reach him as cards: he reads the whole script in his Inbox and runs it from the PC or his phone. It runs in a terminal of its own, and its exit code and the end of its output go back to the session.
4 Real workJobs that need Severian's own accounts went live the same evening.
4 Real workThe hub can now run jobs that need Severian's own accounts, such as API calls, on a session's behalf. Built and in review; not deployed yet.
4 Real workVisit counting is live on the public live view and Alzabo. Only logged-out visitors' public pages count, when they arrive by a link or a typed address. Visitors in Europe, with an unknown clock, or whose browser keeps no cookies are asked first, and members' pages are never counted.
4 Real workVisit counting for the public live view and Alzabo is ready for the owner's merge. Only logged-out visitors' public pages count, and only when they arrive fresh (a link or a typed address). Visitors in Europe, with an unknown clock, or whose browser keeps no cookies are asked first. The shared consent prompt behind it is now version 3, which also covers places the old one missed, such as Cyprus.
4 Real workThe public live view and Alzabo can now count visits with Google Analytics: logged-out visitors on public pages only, never members or gated pages. Visitors in Europe, and anyone whose browser can't keep the choice, are asked first, and nothing loads before a yes. Built and tested; it goes live after review and the owner's merge.
4 Real workCodex and Antigravity runs inside the read-only sandbox now show their chip on the thread that started them, and their usage counts in citadel's stats, each run once.
4 Real workThe three crash-safety changes from earlier today went live on Nov 14: an agent now comes back from a crash or a restart without losing or repeating work.
4 Real workWhen someone answers an agent with only "thanks" or "ok", the agent no longer writes back. Its bot adds a ๐ reaction instead, with no model run and nothing spent. An "ok" that answers a question the agent just asked still gets a normal reply. Deployed on Nov 14; a first live test in Discord is next.
4 Real workA drill now crashes an agent process in the middle of its work on purpose and checks how it comes back. It found two ways a restart could lose work, and both are fixed. Waiting for review and a deploy.
4 Real workAn agent process can no longer be started twice for the same community. A second copy stops before it touches anything, and a crashed one frees its place at once, so work is never run twice. Built and tested, waiting for review and a deploy.
4 Real workA second review of that change found more to fix. The fixes are in, a third review came back clean, and it waits on the owner's merge.
4 Real workSending notes and feed drafts to the hub for review got a round of fixes. They wait on the owner's merge.
4 Real workThe hub now builds from the main line, and the agents run on Pi Durable 1.5. Their four stores moved to the new format; the deploy kept them off Discord for about eight minutes.
4 Real workFirst live check of the agents: eight took 28 test questions in a read-only mode. All answered honestly, none invented anything, and nothing reached anyone. Answers still run long. How they treat members is still untested.
4 Real workScripted drill run: 62 cases, 10 passed and none failed. The other 52 need a judge model to read the reply, so this tested the plumbing, not the agents.
4 Real workHub update: an agent conversation can now be tied to a project.
4 Real workThe live view got a new design, with a bench for model tuning runs.
5 Public recordAgents switched back on.
4 Real workThe first note for a community's members went up, with a feed entry pointing to it.
3 Public pagesHub update: builder agents' work can now be posted to their community's channels.
4 Real workAgents gained web search and page reading, and builder agents got places to work. The agents then went off until Severian could test them.
4 Real workAlzabo's first public note was approved.
3 Public pagesFirst live run on Discord. The feed went up the same day, and the privacy policy and terms took effect.
2 First agents