Run /find-inspiration against the first-tier candidate repos

[nonreagent]

Kind: task
Sources: webpro/dotfiles, paulirish/dotfiles, jessfraz/dotfiles
Effort: S per run (three separate runs, not one merged pass)

Problem

Two find-inspiration runs have happened: runs/2026-07-04-priors.json and runs/2026-08-19-samandmoore.json. Between them, decisions.jsonl now holds 18 decided ids, so the obvious veins (deploy engine, version manager, secrets, agent hooks) are drained. Without new sources the skill has nothing undecided left to show.

Candidates and expected yield

Ranked by undecided ideas each is likely to surface, not by general repo quality. Each was checked against the GitHub API for activity and tree shape on 2026-08-20.

webpro/dotfiles — 1.2k stars, pushed 2026-07-22, bash

Hits the two areas this repo has already invested in, with a different design.

• bin/is-macos, is-ubuntu, is-arch, is-arm64, is-supported, is-executable — runtime predicate helpers where we use os=Darwin manifest conditions and .Darwin/.Linux filename suffixes.
• install/ holds one manifest per ecosystem: Brewfile, Caskfile, Codefile (VS Code extensions), npmfile, Rustfile, pacmanfile, duti (file associations). We have Brewfile alone.
• test/bin.bats uses bats where test/test_deploy.sh and test/test_shell.sh are bespoke. CI runs a real installation, not a smoke test.

paulirish/dotfiles — 4.4k stars, pushed 2026-08-19

agents/ is a pnpm workspace: 12 skills, TypeScript, node --test suites with checked-in fixtures, typecheck, and a preflight script.

It answers a question we have no answer to: how do you test a skill? make check-skill-frontmatter and make check-skills verify shape and symlink integrity. Nothing runs a skill and asserts on the result. Also the most direct read on the "artifacts outlive the box" rule in home/.agents/rules/workflow.md.

jessfraz/dotfiles — 3.6k stars, pushed 2026-08-19, bash

Closest philosophical match: bash-first, Linux and macOS, Makefile-driven.

• test.sh shellchecks every file file(1) identifies as shell — broader coverage than a per-file test suite, and it would catch scripts our current suite never looks at.
• .codex/AGENTS.md plus .codex/notify.py is cross-harness agent config from a non-Claude direction. We already maintain home/.copilot/instructions/ and home/.gemini/ mirrors, so this is a live problem here.

Definition of done

• Three runs, one per repo, each producing a runs/YYYY-MM-DD-<owner>.json.
• decisions.jsonl and the new runs/ files committed — the log is the skill's memory and has to survive a fresh clone.
• Issues filed for adopt and spike items via bin/triage-issues.py, dry-run first.
• A README attribution row per repo, added only once an adopted item actually lands.

Notes

Run them separately. The decided-id filter dedups across sources, so overlap costs nothing, but merging three profiles into one comparison loses the per-repo framing that makes a difference legible.

Second-tier candidates, held back deliberately: citypaul/.dotfiles (agent config as a released product), nicknisi/dotfiles (same home/ layout, but overlaps heavily with already-rejected ids), alrra/dotfiles. The one unexplored foundation question is nix-darwin plus home-manager (dustinlyons/nixos-config, mitchellh/nixos-config) — worth a run only if replacing Homebrew, asdf and deploy.sh together is genuinely on the table.