Shipping to Production

Previous Next

Shipping to production

tuvl ship turns a project into a container image and a Helm chart. It is the same gate as CI: nothing ships that does not validate.

tuvl keys generate --write          # a persistent signing key (TUVL_SIGNING_KEY)
tuvl ship -t registry.example.com/triage:1.0.0 --push

What it does

  1. Gate. Runs the validation (--strict also fails on warnings). A workflow with a pending agent (V010) is an error at ship time, and the project's tuvl.lock must be current.
  2. Deploy artifacts. Writes deploy/Dockerfile (pinned to the locked tuvl version), .dockerignore and a chart in deploy/chart/<name>/. Existing files are kept unless you pass --force.
  3. Image. Runs docker build, and docker push with --push.

The .dockerignore is an allowlist: only models, workflows, agents, artifacts, datasources, llms, federation, config, the lock and .tuvl/system.yaml / telemetry.yaml enter the image. Specs, task lists, tests, agent rules and skills, compose files and .env stay out; a development file in the image is refused (V027).

The container runs in production mode: no dev routes, no Insight UI, JSON logs, telemetry on.

Options

Option Meaning
-t, --tag Image reference. Defaults to <project>:<version>
--no-build Write the Dockerfile and chart, skip docker build
--push Push the image after a successful build
--force Overwrite existing generated files
--strict Treat validation warnings as errors
--with-laya / --without-laya Include the Laya decision server in the chart. Default: on when an AgentModel uses provider: laya

Deploy

kubectl create secret generic triage-env --from-env-file=.env
helm install triage deploy/chart/triage --set image.repository=registry.example.com/triage

The secret must hold TUVL_SIGNING_KEY, POSTGRES_PASSWORD and the variables of every model the workflows use. A production engine refuses to start while one is missing.

For separate API and worker pods:

helm install triage deploy/chart/triage --set split=true --set worker.replicaCount=3

Laya in the chart

When a model uses provider: laya (or with --with-laya), the chart gains a Laya Deployment and Service, and the engine's LAYA_API_BASE points at it. laya.enabled in values.yaml switches it.

model: english auto / multilingual
Image docker.io/tuvl/laya-serve:<version>-en docker.io/tuvl/laya-serve:<version>
Memory request / limit 2 Gi / 3 Gi 4 Gi / 5 Gi

tuvl ship chooses the image and memory from your models. Set LAYA_API_KEY in the secret to require a bearer key on the server. Leave Laya out with --without-laya (or --set laya.enabled=false) when you use a hosted model instead.

Redis

The chart does not include Redis. Run more than one worker (split=true, or several replicas) and you need one: point a type: redis datasource at it with REDIS_HOST, REDIS_PORT and REDIS_PASSWORD. See Redis configuration.

See also