Pi — a minimal, extensible terminal coding harness

What it is

A terminal coding agent that is deliberately small at the core. Its own framing: "a minimal terminal coding harness… designed to stay small at the core while being extended", and "Adapt Pi to your workflows, not the other way around." MIT licensed, per the project site.

The extension surface is the product:

• Extensions — TypeScript modules adding "tools, commands, events, and custom UI".
• Skills — "Agent Skills for reusable on-demand capabilities."
• Prompt templates — "Reusable prompts that expand from slash commands."
• Sessions — JSONL session files with "session management, branching, and tree navigation", plus export.
• Context engineering — customisable system prompts, "context compaction and branch summarisation", auto-compaction.

Operation modes: interactive terminal UI, print/JSON, RPC, and an SDK. The site claims 15+ model providers (Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq and others), "hundreds of models", and 50+ extension examples.

What it deliberately leaves out

The overview names the omissions as a design position, not a roadmap gap: sub-agents, MCP, plan mode, and permission popups are absent by choice, to be built as extensions instead. Read the exclusions as the actual thesis — the claim is that a harness core should be small enough to reason about, with capability added deliberately rather than shipped by default.

Caveat on that list: it appears on the marketing overview. The docs index read the same day describes only "a small core, extended through modular components" and does not restate the exclusions. Treat the specific four as the site's claim, not as verified behaviour.

Why this rung — and the honest problem with it

The Binding Constraint Thesis measures harness-induced variance at 7.80× model-induced variance (18.48 vs 2.37 pp²), with ranking reversals in 6 of 9 model pairs. The rung exists because the harness decides how many tokens a finished task costs.

Pi does not move a number on that rung. It publishes no benchmark, no success rate, no token accounting. By the strict reading of the source bar — every source must move a number on one of the five rungs — it does not clear it, and this file says so rather than dressing a tool up as evidence.

What it is instead is an instrument. The thesis says the harness dominates; Pi is a harness whose Execution, Tool, Context and Observability layers are configurable and inspectable in the open. That makes it the cheapest available way to stop reading about the effect and measure it:

• It is a concrete artefact to map against the paper's proposed ETCSOVG Harness Card (Execution, Tool, Context, Scheduling, Observability, Verification, Governance) — a disclosure standard with, so far, few worked examples.
• Its JSONL session files with branching are the raw material for trajectory comparison; the factorial's whole method is holding one axis fixed while varying the other.
• Its compaction and system-prompt controls are exactly the Context layer the paper says goes undisclosed.
• Minimal-vs-extended is a Minimal/Full harness contrast you can actually configure, against the same model.

That is the seat argument. Everything in this rung is currently read it standing. Running a harness Connor controls is the only route to adjacent or ran it — and it is the same axis /tools/harness-cost asks the reader to supply from their own measurement.

Limitations

• No measurement of any kind. No benchmark, no success rates, no cost or latency figures. Nothing here may be cited as a number.
• Vendor self-description, read from the site and docs index only. The source was not read, and no claim has been checked against it.
• datePublished and repo are null deliberately — neither was established from a primary source on 2026-08-06. The MIT licence is the site's claim; it has not been read from a LICENSE file.
• The exclusion list is single-sourced to the overview page (see caveat above).
• Nobody has run it. Until Pi is actually installed and exercised, its presence here changes the seat on nothing. A tool in the KB that is never run is the same failure as a dataset nothing feeds.

Provenance

tool, primary, vendor-authored, unmeasured. Registered as an instrument for the harness thesis at Connor's direction, not as evidence for it. The distinction is the point of this file.