Canonical case study

Verify an AI-assisted repository change

A disposable local example for a developer who wants to add a health-check endpoint, tests, and documentation without losing the plan, verification result, recovery boundary, or evidence record.

Scope: This is a reproducible local procedure. It does not claim that a model changed a production repository, guarantee correctness, or substitute for project review.

1. Create a disposable repository

npm install -g @flotic/minitok
minitok doctor
output="$(minitok promote demo)"
printf '%s\n' "$output"
fixture="$(printf '%s\n' "$output" | sed -n 's/^Demo fixture created: //p')"
minitok migrate "$fixture"
cd "$fixture"

The demo fixture is local-only and does not call a model or apply a change.

2. Preview the repository task

minitok run --dry-run "Add a health-check endpoint, tests, and documentation. Preserve existing APIs."
minitok status
minitok run list
minitok run show <run-id>

A model-backed dry-run requires the provider configured by the user and the applicable entitlement. The fixture itself requires neither.

3. Inspect the result

  • The plan uses the real fixture files and preserves the existing API constraint.
  • Changed files match the requested endpoint, tests, and documentation.
  • The project's own verification command passes or reports a clear failure.
  • Failed checks stop or guide bounded repair rather than being hidden.
  • The run record identifies checks, retries, stop reasons, and remaining human review.

4. Preserve evidence and recover safely

  • .minitok/last-run.json contains the redacted latest run summary.
  • .minitok/evidence/runs/ contains structured evidence when enabled.
  • .minitok/last-run.patch may preserve an unapplied patch for manual review.

A stopped run is not an applied change or deployment result. Review a preserved patch before applying it manually.

5. Confirm the local first-success boundary

Only confirm this after minitok migrate and the repository's own checks passed. This records an anonymous aggregate signal only; it does not send source code, prompts, run logs, provider data, or an identity.

Limits

For a tiny edit with no meaningful repository checks, a simple editor may be faster. minitok is not an LLM provider and does not guarantee correctness, security, deployment success, or production readiness.

Related: reproducible demo, verification guide, recovery guide, and evidence guide.