> hey, #opus 5.5, use #fframes and create a magnificent 2-minute #demoscene piece themed around the #backrooms, take a less conventional approach — skip the greetings and other classic formalities
Real abstractions change the space of possible moves.
Some distinctions become visible. Others practically vanish.
Bad abstractions lock in the wrong distinctions and make thinking rigid.
Good ones expand the search space while cutting noise.
Abstractions don’t just package information.
They shape thinking — and its efficiency.
On a raw stream of events, thought stays surface-level: too many details, weak hierarchy, high inertia from earlier moves.
Seems like I finally got the perfect screenshot 🔥
The pi-telegram extension for remote control of my agents just got a big update!
github.com/llblab/pi-telegra…
1/4
Built Pi State Flow, an extension for the #Pi coding agent inspired by `SKILL.state`.
Instead of carrying the full history of completed runs forward, it materializes the information needed to continue into an explicit state.
github.com/llblab/pi-state-f…
3/4
Unlike #SKILLstate, state isn’t patched after every tool call.
Pi’s native tool loop remains intact, with one validated handoff committed when the complete run finishes. The full session remains inspectable outside the model context.
6/7
The runtime is intentionally thin.
The depth lives in the protocol: `grow-loop` decides whether another iteration should run; `while-true` executes one bounded worker slice.
Thin runtime, thick protocol.
7/7
It feels less like “run forever” and more like growing a project fractally: each iteration reveals and narrows the next useful edge.
Pi agent folks, would love feedback:
pi install npm:@llblab/pi-grow-loop
github.com/llblab/pi-grow-lo…
