Greenhorn / captured contrast
Unconstrained Codex baseline.
A separate Codex session worked in a disposable copy of the same fixture. It was explicitly allowed to use ordinary developer knowledge to fill README gaps. This is not a Greenhorn run and receives no contract ruling.
The fixture is deliberately seeded. This capture proves a behavioral contrast only; it is not a prevalence result, a third-party-repository result, or an isolation benchmark. The baseline deliberately uses convenient host developer tools; Greenhorn’s accepted commands run in Docker.
Commands actually recorded
/bin/zsh -lc "pwd && rg --files -g 'README*' -g 'package.json' -g 'pnpm-lock.yaml' -g 'yarn.lock' -g 'package-lock.json' -g 'bun.lockb' -g 'AGENTS.md' -g '.nvmrc' -g '.node-version'"completed · exit 0/private/var/folders/w3/c0231jz97cl12bj6q0c8j0fr0000gn/T/greenhorn-baseline-4kqer8 package.json package-lock.json README.md
/bin/zsh -lc "sed -n '1,240p' README.md && printf '\\n--- package ---\\n' && node --version && npm --version && sed -n '1,220p' package.json && printf '\\n--- git ---\\n' && git status --short"failed · exit 12811.12.1 { "name": "greenhorn-fix-loop-demo", "version": "1.0.0", "private": true } --- git --- fatal: not a git repository (or any of the parent directories): .git/bin/zsh -lc 'npm ci && node verify.js'completed · exit 0up to date, audited 1 package in 269ms found 0 vulnerabilities fix-loop checkpoint passed
Captured 2026-07-18T15:38:47.388Z · Codex exit 0 · raw JSONL-derived transcript