Adagio CLI

adagio-cli is the Python command-line interface for Adagio pipeline execution.

For user-facing documentation and product guides, please reference the docs:

• Adagio Docs

The adagio frontend is used to build pipelines that can be run with this package on the command line It can be found here:

• Adagio

Development

Set up the project and run the test suite with:

uv sync --group dev
uv run pytest

Runtime environments

Every plugin action requires an explicit execution environment. A runtime config passed with --config can define one default and override it per plugin or task; the CLI never guesses an image from a plugin name.

Conda environments are supported with kind = "conda":

version = 1

[defaults]
kind = "conda"
prefix = "/opt/conda/envs/qiime2-2026.1"

[plugins]
dada2 = { kind = "conda", prefix = "/opt/conda/envs/q2-dada2" }

The environment must already exist and contain QIIME 2 plus the plugins needed by the pipeline. Adagio enters it with conda run; it does not create or manage the environment.

Task resource requests

A runtime config can also declare the requested shape of each task execution:

[resources.tasks."denoise-node"]
cpus = 4
memory = "8 GiB"

cpus is a positive whole number and memory is a positive, unit-bearing quantity. Each entry describes one independently schedulable execution of that task, not an aggregate pipeline allocation. If a task is later expanded into parallel subtasks, each subtask requests this shape and the scheduler determines the total concurrent allocation.

The current serial executor parses and validates these fields but intentionally does not apply them yet. Omitting either field leaves that resource at the executor's default.

Plugin submission defaults

QAPI submissions can persist the environment that Adagio should use for the submitted plugins:

adagio qapi build --plugin my-plugin \
  --default-conda-prefix /opt/conda/envs/my-plugin

adagio qapi build --plugin my-plugin \
  --default-docker-image registry.example.org/my-plugin:2026.1

These options are explicit and mutually exclusive. Omitting both leaves the plugin without a default; the Adagio app will flag that plugin until an environment is selected for a run.

Remote artifact inputs

An artifact input accepts a local path or an HTTPS URL, both on the command line and in an arguments JSON file:

adagio run pipeline.adg --config runtime.toml --cache-dir ./cache \
  --input-seqs 'https://example.org/data/seqs.qza' --log-dir ./logs

{"version": 1, "inputs": {"seqs": "https://example.org/data/seqs.qza"}}

Pass that file with --arguments arguments.json. The flag name (seqs here) comes from the pipeline's input name.

The consuming task downloads the artifact inside its execution environment; the orchestration process does not fetch it. The original archive is loaded without re-importing, preserving its UUID and QIIME provenance. The downloaded type must match the declared input type and consuming action. Collections may contain both local paths and URLs; artifact URLs can also supply metadata when the artifact supports a metadata view. Metadata inputs also accept HTTPS URLs to TSV files, which QIIME loads directly without importing them as artifacts.

Downloads stream to disk with bounded retries and share a locked staging directory for the run. Every new run fetches again before QIIME result-cache lookup, including runs with reuse enabled. To verify a specific version, append #sha256=<64 hexadecimal digits> to the URL. A mismatch fails the input. Transfer resumption and persistent download caching are not implemented yet.

The staging directory lives in the run's temporary work directory, which is under the system temporary directory (TMPDIR on macOS and Linux). Point TMPDIR at a larger volume for large inputs. A download whose reported size exceeds the free space there fails before transfer; disk errors are not retried. Download time appears as input_download_seconds in the task timings.

Raw data imports, remote manifests, visualizations, login flows, and embedded URL credentials are not supported. The URL need not end in .qza; its contents are validated. Workers need network access to the source. No wget or curl installation is needed.

Task logs contain transfer progress. --log-dir also saves per-task *_inputs.json receipts with the source, final URL, byte count, SHA-256, type, and original UUID for artifacts (metadata TSVs have no artifact UUID). Connected runs include these receipts in their task-finished events. URL query strings are omitted from receipts and download logs; the full source is represented by a hash and remains in the supplied run arguments.

Catalog pipelines

Run a pipeline from the Adagio pipeline catalog:

adagio pipeline show @adagio/microbial-diversity
adagio run @adagio/microbial-diversity --cache-dir /path/to/cache --arguments run-arguments.json

@adagio/<slug> first resolves against a nearby local adagio-pipelines checkout when one is available. If no local catalog is found, Adagio fetches pipeline.adg from cymis/adagio-pipelines on GitHub, checking official before community.

During adagio run, remote catalog pipelines are downloaded under the selected --cache-dir and reused by source name and slug on later runs. adagio pipeline show uses a temporary download when it fetches from GitHub because it does not take a cache directory.

Private GitHub access is explicit: set GITHUB_TOKEN or GH_TOKEN to a token that can read cymis/adagio-pipelines; with a token, the CLI fetches through the GitHub contents API. The CLI does not read browser, git, or gh credentials automatically.

Cache progress and diagnostics

A task reports Preparing task during host-side setup, including input hashing, then the environment's download/start phases. Inside the environment, it reports Preparing cache lookup while resolving and loading inputs, then Checking cache. A hit reports Using cached result while restoring and saving that result, and finishes as Reused cached result. Only a miss reports Running action. Tasks with reuse disabled report Preparing task then Running action. Environment startup remains visible separately.

Each successful task records monotonic timings in the task_finished event's timings metadata and in its node log (persisted by --log-dir). The app's node log viewer exposes the same breakdown. Useful fields, in seconds:

Totals overlap their component stages: do not add every field together. Missing fields mean the stage was not measured (including older worker manifests). Tasks with reuse disabled omit the cache-stage timing fields. These are timings of successful tasks, not a profiler for failed actions. No cache identity, validation, or reuse policy is changed.

To profile a representative pipeline, run it twice against a new isolated cache with identical inputs, parameters and environment. Preserve both log directories:

profile_dir=$(mktemp -d)
adagio runtime --spec pipeline.adg --arguments arguments.json --config runtime.toml \
  --cache-dir "$profile_dir/cache" --output-dir "$profile_dir/first/outputs" \
  --log-dir "$profile_dir/first/logs"
adagio runtime --spec pipeline.adg --arguments arguments.json --config runtime.toml \
  --cache-dir "$profile_dir/cache" --output-dir "$profile_dir/repeat/outputs" \
  --log-dir "$profile_dir/repeat/logs"

Confirm Reused cached result: true before interpreting the repeat as a cache hit; mutable plugins, changing inputs, or random defaults can still require an action. Compare repeated runs with the same cache size and input sizes as the reported slow workload before deciding which stage to optimize.

An opt-in Docker integration test exercises a two-step pipeline with synthetic feature tables, three cache-hit repeats, changed parameters, changed inputs, and reuse disabled. Use an installed QIIME image containing feature-table:

ADAGIO_DOCKER_TEST_IMAGE=ghcr.io/cymis/qiime2-plugin-feature-table:2026.1 \
  PYTHONPATH=src python -m pytest -q -s tests/test_docker_cache_integration.py

The test uses an isolated temporary cache and prints the path to cache-profile.json, containing the image ID, phases, timings, and output artifact IDs for each run. Cache hits must preserve artifact IDs and never report running; misses must execute the actions.

Connected deployments require the backend and runtime server with adapter-schemas 0.5.0. Deploy backend support for the new phases before the runtime update. Progress travels through a mounted JSON-lines file; workers do not receive runtime credentials or need network access to report phases.