ALL uses Node.js 24.13 or later, npm and Git. macOS arm64 is the tested primary environment. A strict TypeScript build emits the CLI and hook entrypoints into dist/.
git clone https://github.com/bstack-tools/all-agent-learning-loop.git
cd all-agent-learning-loop
npm ci --ignore-scripts
npm run build
node dist/standalone.js --helpDependencies are locked. There is one npm runtime dependency, Zod; SQLite is built into Node. The project does not require a separate database, model server or paid API key.
npm run verifyThis builds the package, runs the software tests, replays fourteen deterministic scenarios, evaluates the labelled correction dataset and checks the npm package contents. Fixtures simulate approval only inside disposable tests. They do not authorise any real proposal or prove that a live host dispatched an event.
The CI workflow runs this command on macOS with Node 24.13.0. Credentialed host experiments are excluded.
npm pack
npm install --global ./bstack-all-0.2.0-rc.1.tgz
bstack-all --versionThe package contains dist/, skills, documentation and licence notices. Tests, fixtures, raw reports and private databases are excluded. Built files and tarballs are ignored by Git.
node scripts/package-smoke.mjs packs and installs into a temporary prefix, then checks the installed binary and owned configuration removal. It may need npm registry access or cached dependencies. It does not establish native host trust.
node scripts/upgrade-smoke.mjs additionally requires the retained 0.1.0 tarball in the checkout; that historical local artifact is not committed. Its recorded result is retained under reports/.
Live host scripts use existing account credentials/allowance and are deliberately outside npm run verify and CI. They create disposable projects, but still execute a real agent. Read each script before running it and obtain explicit user verification and application decisions when required. Never script trust UI responses or infer approval.
The retained live scripts reference the original disposable trial under /private/tmp/bstack-all-020-live. They are evidence harnesses, not general setup commands or a portable automated host acceptance test. Do not rerun them against the preserved fixture.
The optional Seatbelt probes require macOS and OS permission to apply the experimental policy. Browser experiments require a local Chrome installation. See evaluation, host evidence and security boundaries for what each experiment establishes.
reports/ contains recorded synthetic evaluations and controlled disposable-trial results, including unsuccessful attempts. It is development evidence, not runtime state. Reports can contain fixture paths and opaque IDs. Never add live private project history, authentication material or account configuration to it.
The release report separates software tests, replays, actual host evidence and unmeasured outcomes. Rebuilding an artifact changes its checksum; use the receipt for the specific artifact being distributed.