pi — the harness by subtraction

A system prompt under a thousand tokens, stable across releases, plus hierarchical AGENTS.md from global to project scope. Nothing is injected behind the operator's back — what the model was told is exactly what you can read.

The agent finds its own material the way a person would: grep, find, read. No retrieval index, no embedding store, no codebase map — the position is that a model trained to navigate repositories should navigate, and that a visible search trail beats an invisible ranking.

Compaction triggers when context crosses the window minus a reserve (16,384 tokens by default): the newest 20,000 tokens survive untouched and everything older is summarized, with file lists accumulated across successive compactions. Manual /compact accepts steering instructions, and leaving a session-tree branch can write a summary of the abandoned work into the trunk.

Read, write, edit, bash — plus grep, find, and ls — defined with TypeBox schemas, streaming progress, and results split into a model-facing part and a UI-facing part. MCP is refused outright: manifests are too expensive and their results are not composable. The replacement is a small CLI tool and a short README; the agent shells out, and output can land on disk and be piped onward.

The deepest mechanism on the sheet. Sessions are JSONL trees — every entry carries a parent id, and the live position is a leaf. Pick any earlier message and continue from it: a new branch in the same file. /tree navigates, folds, and labels the whole structure; /fork and /clone split sessions; plans and todos live as PLAN.md and TODO.md in the repo, versioned, instead of as modes in the harness.

One event-driven loop, embeddable four ways — interactive TUI, one-shot JSON, an RPC subprocess speaking JSONL over stdio, and an in-process SDK. Long-running work is deliberately pushed out to tmux, where the process is observable and a human can attach mid-debug. The harness runs the turn; the operator runs the day.

Three registers, all plain files: Agent-Skills-standard SKILL.md directories with progressive disclosure (only descriptions ride in the prompt; the body loads on demand), markdown prompt templates invoked as slash commands, and TypeScript extensions that register tools, commands, and shortcuts — shareable as packages on pi.dev.

Observability is the design's default state rather than a feature: every file the agent read, every command it ran, every sub-agent transcript is on screen, and tmux panes make long processes watchable. Verification machinery, though, is yours to add — an extension can intercept every tool result, and a test run is a bash call away, but nothing demands receipts unprompted.

The core is deliberately frozen — a stable prompt and a fixed toolset are the point — so improvement accretes around it: skills, templates, and extensions in the global and project directories, shared as installable packages. The harness does not compound; your kit does.