My personal agent distro built on top of Pi.
I wanted a terminal-first experience. I was a Hermes main, but I wanted to come back to the terminal, and its CLI was not that enjoyable. So I built the environment I wanted on Pi, which already owns inference, sessions and the TUI, and kept for myself only the parts Pi does not have.
devopens Pi from the project you are working on, with one of the profiles defined in my local, untrackedprofiles/directory, which selects guidance and skills, and keeps private state out of the project and out of Git. launcherworkruns commands and delegated Pi children in the background, in separate processes, with version-controlled model dispatch and outcomes delivered back into the conversation. work, dispatchworkspacelets several sessions and their children write to one repository without overwriting each other, allocating worktrees when needed and releasing them once the work is delivered. workspaceread_urlgives the lead and children sourced page text, using an authenticated Chrome profile copy when needed. Pi's native codemode can compose reads in scripts. read URL- Background compaction prepares long-session summaries while the lead and children keep working. compaction
- The workflow itself (sizing, grilling, specs, implementation, review) comes from a shared skill library; dev owns only the integration.
The component map and decision-record index are in ARCHITECTURE. What dev trusts and what it does not protect: SECURITY.
Dev is built for me: it assumes pi installed with the pi.dev installer, the skill library at ~/.agents/skills, and Node satisfying package.json with SQLite's WAL-reset fix. Read SECURITY before installing.
cd ~/Developer/dev
npm ci
npm run setup
npm link --ignore-scriptsProfiles are personal and stay out of Git: profiles/manifest.json names each profile's guidance and skill directories, and dev refuses to start without it. The format is in the launcher guide.
Then, from any Git project:
cd /path/to/your/project
dev| Task | Start here |
|---|---|
| Launch, select a profile or resume | Launcher |
| Run background commands or delegate | Work |
| Inspect retained work and release worktrees | Workspace |
| Read documentation URLs, manage the Chrome profile copy | Read URL |
| Understand long-session context handling | Compaction |
| Interpret private metrics or export charts | Usage profile |
| Upgrade Pi and dev's dependencies | Upgrade |
| Change dev's code | Development |
| Find components and decision records | Architecture |
| Look up domain terms | CONTEXT |
Agent contributors start from AGENTS.md; the security model applies to every integration.
Measured on my own use of dev, not on benchmarks: every lead and child session file in my data home, abandoned branches included. Each chart names its sample size and date range.
The metric definitions explain what these charts establish, including why tool success is not task correctness. To regenerate the committed charts, follow export for publication; private reports and drilldowns are not public exports.
