← The Independent 100: a study

Mario Zechner and Pi, the upstream of OMP

Deep dive, written 2026-09-24 by an AI agent session; public edition 2026-09-25 Covers Mario Zechner (#85)

Bracketed tags are evidence labels: [Verified] means we read the page, file or API response ourselves, [Reported] means a source says so and we did not check the underlying thing, and [Inference] is our own reasoning. CI is continuous integration: automated checks that run on every push. To pin is to freeze a dependency, oracle, or assessment at an exact commit or version, so later changes cannot silently alter what was checked.

Credit

The Independent 100 list was curated by dan (@irl_danB) and published on 2026-09-24 at independent.prose.md (announcement). The ranks and section names are his, and number 100 is his open wild-card slot. This study is ours and is not endorsed by him or by anyone on the list.

Summary

How Pi's small core, project-trust gate and enforced supply-chain rules compare with OMP, the fork built on Pi, and which of Pi's practices carry general lessons.

Sources

Licenses: Pi is MIT, Copyright (c) 2025 Mario Zechner [Verified: github.com/earendil-works/pi/blob/v0.87.1/LICENSE#L3]; OMP is MIT [Verified: github.com/can1357/oh-my-pi/blob/v18.3.0/LICENSE]. terminalcp and cchistory have no LICENSE file [Verified: GitHub license API, 2026-09-25, evidence/licenses.jsonl: no license file], though terminalcp's package.json declares MIT [Verified: raw.githubusercontent.com/badlogic/terminalcp/HEAD/package.json]. agent-tools is archived with no LICENSE file [Verified: GitHub license API, 2026-09-25, evidence/licenses.jsonl].

What he built

Authorship: Pi's LICENSE carries Mario Zechner's copyright line (cited above), and the Pi post on his own site describes Pi's design in the first person and links the badlogic/pi-mono repository; the MCP versus CLI post links badlogic/terminalcp the same way.

Pi is an npm workspace monorepo with packages ai, agent, tui, coding-agent, durable, telemetry, protocol, client, server, session-backends, chord and evals [Verified: github.com/earendil-works/pi/tree/v0.87.1/packages]. The CLI is @earendil-works/pi-coding-agent, binary pi. The core is deliberately small:

The Pi post argues for a minimal prompt, four tools, no approval prompts, file-based todos and plans, no MCP (Playwright MCP at 21 tools and 13.7k tokens), no background bash (use tmux), and no subagents beyond spawning pi --print for review [Reported: the Pi post].

Where OMP differs

OMP describes itself as a fork of Pi [Verified: github.com/can1357/oh-my-pi/blob/v18.3.0/README.md#L22]. Verdicts are [Inference].

AreaPi v0.87.1OMP 18.3.0Verdict
MCPnone in corebuilt in, lazy catalogOMP mostly answers the cost point
Project trustgate, default askno gatingPi
Subagentsexample onlybuilt-in taskOMP, for multi-agent work
Approvalnonetiers, default yolosame in practice
System promptabout 1k tokens14.6 KB templatePi, before any appended skills
Exact pinsenforcedcaret ranges shipPi
Release age2 (npm)3 days, repo onlyboth set in the repo
Action pins39 of 39 by SHA32 SHA, 17 tagPi

General lessons

  1. A minimum release age on the installing side. OMP's repo-level minimumReleaseAge governs its own development, not users installing OMP (see above), so a gate that protects users has to be set in their own install configuration [Inference]. Pi's .npmrc shows the pattern: min-release-age=2 with save-exact=true. Risk: urgent fixes wait unless overridden.
  2. A release surface diff (cchistory's idea). Pack the previous and new releases of a harness and diff prompt files, the tool registry and settings defaults. On OMP 18.2.11 to 18.3.0 it would show both behaviour changes above (hub, edit headers) [Inference]. cchistory has no LICENSE file, so reuse the method, not the code.
  3. Close the project-trust gap in configuration. Where the harness does not gate project config, turning off project-level MCP loading (mcp.enableProjectConfig) and keeping agent sessions out of freshly cloned third-party repositories covers part of what Pi's trust gate does [Inference].
  4. Count the whole prompt, not just the base. Mario's point about context spent before work begins applies to everything appended to the base prompt, such as a skill index; OMP's skills.includeSkills and skills.ignoredSkills settings trim that index per profile [Verified: settings schema, lines 5347-5349].
  5. Pi's dependency checks are reusable. A copy of check-pinned-deps.mjs or check-lockfile-commit.mjs keeps the MIT notice and his copyright line.

His MCP versus CLI result over 120 runs was "a wash when the tool is well designed", with the MCP arm 23% faster and 2.5% cheaper than the matching CLI [Reported: the MCP versus CLI post]. OMP loads MCP schemas lazily, which his Claude Code runs did not, so the result may not carry over to OMP [Inference].

Do not adopt

  • Pi's security scope as a policy for untrusted work. Pi's SECURITY.md documents prompt injection and untrusted repositories as out of scope for the tool, so running it on untrusted repositories needs guards beyond it [Inference].
  • Copying terminalcp or cchistory code. Neither has a LICENSE file. Reuse the method only.

Attribution we owe

What we could not verify

  • Pi's current rendered prompt size in tokens. We read the source but did not run Pi; the under-1,000-tokens figure is his, from Nov 2025.
  • Whether bun's minimumReleaseAge refuses or silently downgrades an explicit package@version install, and whether OMP's eval environment reads the global bun config.
  • Whether turning off mcp.enableProjectConfig also stops project MCP discovery through OMP's separate Claude-config loader (github.com/can1357/oh-my-pi/blob/v18.3.0/packages/coding-agent/src/discovery/claude.ts#L79-L100). Not tested.
  • Whether OMP's legacy Pi shim loads extensions written against Pi 0.87.x.
  • Pi's Terminal-Bench 2.0 leaderboard placement (known only from his post).
  • The unit of npm's min-release-age (days assumed) was not checked against npm documentation.

How to read this

A deep dive compares one person’s public work with how we work and says what we would adopt. Its proposals are our plans, not recommendations to anyone else, and its judgements of fit are one reviewer’s. Source: study/independent-100/deep/mario-zechner-pi.md.

Request a correction

Is something about this deep dive wrong or out of date, or would you rather not be covered? Open a correction (a GitHub issue form with this page’s address filled in). A request to remove a record is honoured without argument; accepted corrections are recorded in the changelog with the date.

Help fix it

Each opens a GitHub issue form with this page’s address filled in. You need a GitHub account.

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

Source on GitHub MIT licensed