GitHub

PI-Desktop: Durable, Bounded Agent Work

How sessions, host policy, and recovery make local action reviewable

6 chapters69 min total
Keywords
PI-DesktopSessionsHost-coreProjectsPermissionsPlanGoalCheckpointsProvidersPlugins
Overview

PI-Desktop frames coding-agent work as a project-bound, session-based workflow rather than a one-off chat. The series traces how host-core governs tools, paths, permissions, contracts, persistence, recovery, providers, and extensions. It presents a durable, inspectable local workbench, while retaining limits around early-preview assurance, OS isolation, remote control, and universal offline use.

Learning outcomes
  • PI-Desktop's core unit is a project-bound session, not an isolated prompt or visible folder.
  • Host-core, rather than the renderer or model, resolves identity, policy, execution, persistence, and audit.
  • Plan and Goal make intent reviewable, but approval still selects an execution posture and preserves tool and path rules.
  • Durability preserves history and bounded context, not exact process replay or complete isolation.
  • Its early-preview design suits bounded local agency, not mature release assurance, hostile-plugin isolation, or remote administration.
Keywords
PI-DesktopSessionsHost-coreProjectsPermissionsPlanGoalCheckpointsProvidersPlugins

Chapters

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans
Ch.1-3 FreeCh.4-6 Premium

FAQ

What is PI-Desktop designed to do?

PI-Desktop is a local desktop workspace for coding agents that binds projects, sessions, tools, and reviewable results. It is not presented as a full IDE replacement or hosted relay.

How does PI-Desktop keep an agent in the right project?

PI-Desktop binds tool requests to the originating session's persisted project, not the currently visible folder. Host-core then resolves and canonicalizes the path before applying policy.

What do Plan and Goal change in PI-Desktop?

Plan makes a proposed method reviewable, while Goal makes an outcome, acceptance criteria, and boundaries reviewable. Explicit approval moves the same Agent into execution with a selected permission mode.

Does PI-Desktop resume unfinished work after a restart?

PI-Desktop preserves durable records and known partial results, but it interrupts uncertain approvals and executions instead of replaying them. Model context can be rebuilt, while panel arrangements are transient.

Is PI-Desktop fully offline or completely sandboxed?

PI-Desktop is local-first, not universally offline or a complete OS sandbox. Remote providers receive request context, Bash children retain user privileges, and plugin isolation remains incomplete.