@pi-unipi/long-horizon

Long-horizon execution for Pi — mode-gated /goal /ralph /swarm /graph orchestration with a TypeSafe prompt judge

Packages

Package details

extension

Install @pi-unipi/long-horizon from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:@pi-unipi/long-horizon
Package
@pi-unipi/long-horizon
Version
3.0.0-alpha.0
Published
Sep 25, 2026
Downloads
5,909/mo · 1,706/wk
Author
neuron-mr-white
License
MIT
Types
extension
Size
184.3 KB
Dependencies
1 dependency · 3 peers
Pi manifest JSON
{
  "skills": [],
  "themes": [],
  "prompts": [],
  "extensions": [
    "./index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

@pi-unipi/long-horizon

Long-horizon execution for Pi/UniPi. A small jev judge (TypeSafe System One) picks the cheapest mode that can finish a task — direct, plan, or goal — and the engine keeps goal-mode runs honest across many turns.

  • Mode gating: ambiguous work is judged before the runner starts, so simple edits stay direct and real multi-step work gets a plan or a goal.
  • Goal mode: an owned goal-state file, continuation prompts, and a loop that stops when the goal is met — no drifting forever.
  • Orchestration: /ralph loops, /swarm parallel lanes, and /graph dependency graphs over tasks.
  • Judge: provider-agnostic decision calls (OpenRouter by default) with the key resolved from settings or the environment.

See skills/ for the agent-facing command docs.