$ cat ~/map/zenpi.md

ZenPi

Minimal self-improving harness around pi that turns recurring friction into evidence-backed, human-approved single improvements.

tools✓ Oct 16

Facts

repository: github.com/TannerMidd/ZenPi ↗
GitHub Stars: ★ 15
Forks: 0
license: MIT
npm / mo: —
last push: Oct 3, 2026

install

git clone --branch v0.8.1 --depth 1 https://github.com/TannerMidd/ZenPi.git && cd ZenPi && ./zenpi install

## TL;DR

Minimal, self-improving harness around the pi coding agent — it records privacy-minimized local friction signals, qualifies them, and turns each into one human-approved, verified, and reversible improvement via the /harness-improvement loop.

## Key points

$Grows by evidence, not accumulation: when a real, reusable capability gap gets in the way, ZenPi records a sanitized, bounded, deduplicated local signal, measures recurrence and reach, and gives the human one clear next improvement — collection stays off until explicitly enabled and reports are never uploaded.
$/harness-improvement opens a menu of qualified and review-needed items; choosing one explicitly authorizes that exact smallest-sufficient improvement, and the zenpi-improve skill inspects the evidence, implements the narrowest intervention, and runs direct acceptance checks.
$The completion gate verifies registry integration, runs npm run check and supported closed validators, then retires an item only when everything passes — every retirement keeps its evidence, gates, and changed files; failed checks leave it selected, and later friction returns it to the review menu.
$Adds /zen (focused execution mode), /files (browse files, view diffs, leave review comments), /zen-subagents (remember subagent setups), and the private /wishlist with status, history [gap-id], and loop-health reporting (retirements, reopen rate, open reviews, time-to-retire).
$Command guard analyzes every model-initiated command before execution in Guard, Strict, or Off modes — Guard denies confirmed host-wide catastrophe outright and asks before Git destroys work; approvals are exact-call and session-scoped, a critical attempt locks the session until reviewed, and /guard changes the mode.
$Needs Node.js 22.19+, npm, and Git; the installer sets up pi (0.84.4+) if it is missing. ./zenpi plan previews without changing anything, ./zenpi install asks for confirmation, ./zenpi doctor re-verifies managed state and capability validators; everything is backed up and reversible.

## FAQ

Q Does it send any data anywhere?

No. Wishlist reports are sanitized, bounded, deduplicated by task, stored locally only, and never uploaded; ZenPi never persists command text and does not read pi credentials, sessions, or history. Skim reports before sharing them yourself.

Q Is the command guard a sandbox?

No — it is defense in depth over its documented bash, powershell, read, write, edit, and native-subagent seams. Direct human shell access, unclassified custom tools, TOCTOU changes, and external processes remain outside its hard boundary; use OS permissions, a least-privilege account, a container, or a VM when code or data is hostile.

## References

## Related entries

Last updated: Aug 31, 2026