Four systems I run, and the runtime underneath them. Status, schedules and last change, read from the state each system emits about itself. Open any card for the full instrument panel.
feed updated 2026-10-18newest generated_at across the four systems
Every conversation with an AI starts from nothing. On four repos that meant re-explaining the project each time, and two chats on the same day could quietly undo each other's work. CoBrain is the small runtime that fixes both. It gives every session the same short starting brief, goes and gets the rest only when the conversation calls for it, and makes two sessions take turns on any file they both want to write.
Anything that starts work enters here
Cowork sessionplanning · writing
concurrent sessiona second chat, same repo
coding sessionexclusive write access
scheduled runcron · launchd
01resident set
Every session starts from the same two pages.
A chat opens already knowing the project's rules and where everything lives. Nothing else is loaded up front, so the starting context stays small.
core.mdthe rules
manifest.mda map of every file
everything elsenot loaded
cap 15k tokens · all four repos fit under it · Portfolio also loads COWORK_CONTEXT.md
02fetch router
It fetches the rest only when you bring it up.
Certain words are wired to certain files. Say "decision" and the decision log opens. Say "ship" and the handoff does. A word list, not a search.
"decision"→decisions/INDEX.md
"ship"→HANDOFF.md
"wrap up"→MAINTENANCE.md
Portfolio 18 · Ambito 16 · career-ops 22 · OneLook 16 rows · replayed against trigger cases every run
03decision sidecar
Nothing that was decided gets quietly rewritten.
Rulings are added to the bottom of a log that is never edited in place, and mirrored to a one-line index so a later session can find one without reading the file.
DEC-229appended 09-11
DEC-228unchanged
editing in placenot permitted
Ambito 240 rows · OneLook 41 · none yet in Portfolio or career-ops
04coordination
Two chats can work on one project without clobbering each other.
Before writing to a shared file a session claims it for fifteen minutes. A second session is told who holds it and how long is left, and keeps reading until it clears.
session Aclaimsdecisions.md
session Bdenied ·read-only
session Btold:re-fetch, the file moved
900-second claims · 60 claims across four repos, all released
05lifecycle skills
Every session opens and closes the same way.
Load the context, say out loud what got loaded so a wrong start is correctable in one turn, do the work, then file everything before the chat ends.
boot→attest→work→wrap
skipping wrapflagged at the next boot
init · start · wrap · coord — four commands, Claude plugin v0.3.0
06validation gate
It checks its own wiring and says when it is broken.
Nine checks run against each repo's setup. All four pass right now.
Portfolio · career-ops · OneLookpass
Ambitopass, with warnings
every trigger case passes · one engine version across the four repos
What it saved · resident context vs. the always-on floor it replaced
residentfloor it replaced
Portfolio6.9k of 10.7k tokens · −36%
Ambito10.9k of 125k tokens · −91%
career-ops14.6k of 75k tokens · −80%
OneLook7.5k of 48k tokens · −84%
Each track is that repo's own floor, not a shared scale. Tokens are characters ÷ 4. Script-measured, loader-inclusive.
Engine
18 scripts
stdlib Python
0
dependencies
v0.3.0
Claude plugin · 4 skills
97
own decisions
2026-06-19
live
Portfolio footnote — the hand-computed, loader-excluded figure at the time was −52% (9,412 → 4,552 tokens, DEC-043). Roughly half of that came from slimming COWORK_CONTEXT.md in the same operation; the runtime-only cut before that edit was 17%.
the page you are reading is maintained by the system it describes.