Amplifier Digital Twin Universe

Using amplifier-digital-twin-universe

A Digital Twin Universe (DTU) is a complete, isolated environment stood up on demand from a declarative profile. It closes the gap between "the tests pass on my machine" and "this works where it will actually run."

The library is the tool. amplifier_digital_twin_universe holds every capability. The CLI is a thin wrapper over it and adds nothing of its own, so anything you can do from the shell you can also do from Python.

Before writing code

Confirm every capability, argument, and field name against --help or the library's own signatures before you write it. Do not fill gaps from memory.

amplifier-digital-twin-universe --help

--help is the complete, agent-facing listing: every capability, its arguments, what it returns, and which capabilities are model-backed. -h is a short summary for a person and is deliberately not the same output. If you cannot confirm something from --help or the source, say so rather than guessing.

Reference material, when --help is not enough:

docs/01-library.md   every capability as Python: signatures, parameters, returns
docs/03-cli.md       every capability as a command: flags, defaults, exit codes
docs/02-configuration.md   the environment variables this tool reads
docs/installing-incus.md   installing the runtime that launches environments

All four live at https://github.com/microsoft/amplifier-smart-tool-digital-twin-universe.

Install

As a CLI:

uv tool install git+https://github.com/microsoft/amplifier-smart-tool-digital-twin-universe

As a library:

uv add "amplifier-digital-twin-universe @ git+https://github.com/microsoft/amplifier-smart-tool-digital-twin-universe"

Prerequisites

Ask the tool rather than assuming. check measures this host and reports what is present, what is absent, and what each absent thing costs you.

amplifier-digital-twin-universe check

incus runs the environments; without it nothing launches. docker runs mock service sidecars; without it everything works except profiles that declare sidecars. git is required by the model-backed capabilities. avahi publishes .local hostnames.

Linux only. The tool does not claim platforms it has not been run on.

The shape of the surface

Deterministic, and runnable with no provider configured:

manifest           the manifest as structured data
check              what this host can do right now
validate-profile   whether a profile document is launchable
launch             stand up an environment from a profile path
exec               run one command inside and capture its output
status             one environment's state and access URLs
list               every environment on this host
check-readiness    evaluate an environment's readiness checks once
update             re-run an environment's update commands in place
file-push          copy host paths in
file-pull          copy environment paths out
destroy            tear one down

Model-backed, and failing loudly when no provider is configured:

create-profile     draft a profile from a description of what to test
install            ordered steps to make this host able to launch
doctor             diagnose a symptom against measured evidence
manage             turn a request in words into deterministic actions

A model-backed capability reasons over measured evidence and takes minutes, up to tens of minutes for a broad request. Give it a timeout sized for that, and poll rather than blocking on a short one.

Working with environments

An environment is long-lived and nothing reaps it. Destroy what you launch, and pull anything worth keeping out first, because nothing inside survives teardown.

amplifier-digital-twin-universe launch --profile ./profile.yaml
amplifier-digital-twin-universe check-readiness --id dtu-1a2b3c4d
amplifier-digital-twin-universe exec --id dtu-1a2b3c4d --command "curl -sf localhost:8000/health"
amplifier-digital-twin-universe file-pull --id dtu-1a2b3c4d --source /var/log/app.log --destination ./
amplifier-digital-twin-universe destroy --id dtu-1a2b3c4d

list is scoped to the machine, not to your session, so an environment another session launched appears there too and destroy will take it. Name the profile by path; profile names that live inside the engine's own repository do not resolve from an installed package.

manage plans without changing anything and runs only with --confirmed, which authorizes that one invocation. Read the plan before confirming it.

Reading the manifest

The manifest says what the tool is for and what it needs, so a caller can decide whether to reach for it before invoking anything. It is reachable two ways.

amplifier-digital-twin-universe manifest

from amplifier_digital_twin_universe import load_manifest

manifest = load_manifest()
manifest.name, manifest.version, manifest.platforms
manifest.to_dict()  # plain data, JSON-serializable

Read it through the library rather than by locating a file. Install layouts differ by ecosystem and no filesystem path is portable across them.

Output and failure contract

Every CLI capability writes exactly one JSON document to stdout.

A failure writes a structured envelope and exits non-zero. Never parse prose out of it, and never treat an empty result as success.

{"error": {"code": "no_provider", "message": "...", "remedy": "..."}}

0  success
2  bad invocation
3  a model-backed capability was called with no provider configured
4  a required prerequisite is missing
5  the capability ran and failed

A model-backed capability with no provider configured fails saying exactly that and names what to configure. It never falls back to a degraded deterministic answer, so a result you receive is always the result you asked for.

Choosing a surface

Import the library from Python. Shell out to the CLI from anything that cannot import Python in-process: a shell script, a CI job, or an agent that can run commands but not load a Python object. Both reach the same capabilities.