Skip to content

Deploy from the coding agent

When the hosted coding agent has written your change and the Kubernetes manifests to run it, you can deploy it from the session. Seylo deploys the manifests in the session’s working tree to a cluster you have connected, through the same checks as any other change: the policy rules, a dry run on the cluster, approval by a person (or a deploy allowance), and a read-back of every object.

  • The cluster is connected and its agent is online.
  • The session’s sandbox is running. If it is asleep, wake it: Deploy reads the manifests from the session’s files.
  • The image comes from your CI. Building the image inside Seylo is not live yet: the image comes from your CI. Open the pull request, let your CI build and push the image, and use its tag or digest. The Deploy dialog shows Build it in Seylo as unavailable, with the reason.
  1. In Build → Coding agent, open the session and choose Deploy.

  2. Choose the cluster, then an existing namespace or a new one. Seylo tells you what will happen there:

    • Deploying grants the coding agent access to it for this session only - you may release in that namespace, so you can go ahead.
    • The coding agent may deploy this session here - access was already granted.
    • You don’t have access to deploy in this cluster - request access - see request access.
  3. Under Image, keep From my CI.

  4. Choose Review what will be deployed. Seylo lists the manifest files it found, the objects that will exist in the namespace, anything it will not deploy and why, the images each container will run, and who decides each object.

  5. Fix what stands in the way, then choose Check again:

    • Untick files you do not want deployed.
    • Replace any image without a fixed tag (or using :latest) with the tag or digest your CI built, for example registry.example.com/acme/launchpad:1.4.2.
    • If an object is managed by Argo CD, Flux or Helm, tick the acknowledgement for that tool by name - or, better, make the change in the Git repository or Helm values that tool deploys from.
    • Type the values of any Secret into its write-only fields. See Secrets.
  6. Choose Send for approval. The deploy goes to Approvals; nothing changes on the cluster until someone who may approve it does. If a deploy allowance covers the whole deploy, the button reads Deploy instead, and the deploy goes straight through, recorded as yours.

What the deploy contains Approved by
Workloads (Deployment, StatefulSet, DaemonSet, Job, CronJob), Services, Ingresses, NetworkPolicies, ConfigMaps, Secrets (keys only), HPAs, PodDisruptionBudgets, ServiceAccounts, PersistentVolumeClaims, instances of custom resources the cluster serves, a new namespace Someone with the approver role in the namespace
Roles and RoleBindings, ClusterRoles and their bindings, CustomResourceDefinitions, privileged, hostPath or host-network workloads, a namespace outside the cluster’s allowed namespaces An admin
A role that grants cluster-admin or wildcards, or a blocking policy finding Always an admin, whatever an allowance names

The Deploy dialog shows who decides each object, and why. One approval is enough, and you may approve your own request if you hold the role. It is not two-person control.

Each of these is listed in the plan with what to do instead.

Not deployed Why, and what to do
A Secret, or an environment variable, that carries a value A value the agent has seen is already in its conversation. Rotate it, and let the agent write the Secret with keys only (below)
A bare Pod or ReplicaSet Its controller owns it. Deploy the Deployment, Job or StatefulSet that runs it
A ResourceQuota or LimitRange They are the namespace owner’s limits; an admin sets them
A PersistentVolume or StorageClass, or another kind that reconfigures the cluster for every tenant A person applies it under Resources. Deploy a PersistentVolumeClaim instead of a volume
A cluster-scoped custom resource It has no namespace boundary, so a person applies it under Resources
An object that names a different namespace Remove metadata.namespace, or deploy to that namespace
A template ({{ }}, for example a Helm chart) Render it first (helm template) and deploy the output
An image with no tag, or :latest Give the tag or digest your CI built

A Namespace object in your repository is skipped: choose New namespace instead, and Seylo creates it as part of the deploy.

The coding agent writes a Secret with its keys only, every value empty. Seylo refuses a manifest, or a tool call, that carries a secret value: a value the agent has seen is already in its conversation, so rotate it.

When you deploy, type the values into the write-only fields in the Deploy dialog. They are sent once, encrypted, written to the cluster, and removed from Seylo afterwards. No page, log, audit entry or export shows them, not even to you.

  • An existing Secret is kept as it is if you leave all its fields empty. To replace it, fill every field: a partial rewrite would remove the keys you left empty, so it is refused.
  • Secrets kept in a vault work too: deploy an ExternalSecret or a SecretProviderClass, which refers to the vault and carries no value.

If you may not deploy to a namespace, Deploy says You don’t have access to deploy in this cluster - request access. Add a short note for the approver (no secrets) and choose Request access. The request appears in Approvals; when someone who may approve in that namespace approves it, the coding agent may deploy your session there, and you deploy as above.

A namespace outside the cluster’s allowed namespaces can only be added by an admin, so an admin approves that request.

By default every deploy is approved by a person. An admin can give a named person a standing deploy allowance, so that their Deploy goes straight through, recorded as theirs. The person still reads the plan and presses Deploy; the allowance removes the second person, never the first.

An admin grants one under Admin → Access → Deploy allowances. An allowance:

  • is for one person, on one cluster, for named namespaces (at most 20, never a wildcard: a namespace created later is not covered);
  • names the kinds it covers, and any admin-level changes it covers, such as Roles and RoleBindings. Anything it does not name goes to approval;
  • has a daily limit of deploys (10 unless the admin sets another, at most 50). Past it, deploys go to approval until 00:00 UTC;
  • always ends, at most 90 days after it is granted. The Until picker offers no later day, and a later date typed in shows the latest one allowed under the field and keeps Grant off;
  • applies by default only when the person presses Deploy in Seylo. An admin can also let it apply to deploys proposed through that person’s coding-agent tokens. That is off by default, because there the agent presses the button.

An allowance never covers a deploy the cluster’s own policy refuses, a role that grants cluster-admin or wildcards, or a blocking policy finding.

Revoking an allowance takes effect at once: the person’s next deploy goes to approval.

The session’s Deploys tab shows:

  • Access - each namespace the coding agent may deploy this session to, who granted it and when, and Revoke. Revoking also stops any deploy of the session still waiting for approval.
  • Deploys - each deploy’s approval, its rollout as the cluster reports it, and which objects applied.
  • Undo - rolls a Deployment back to its previous revision, image and configuration, through the normal rollback (and Approvals, when approval is on). A first deploy has nothing to roll back to; remove it under Resources if it should not run.

A deploy token with the propose_deploy capability lets your own coding agent file plain manifests for the token’s namespace as one deploy for a person to approve. They are judged by the same rules as a deploy from Seylo’s coding agent and filed in Approvals. A token:

  • never creates anything cluster-wide;
  • never sends a secret value. A Secret must already exist, and the agent sends its keys with empty values only;
  • never overrides an Argo CD, Flux or Helm owner: such an object is refused, naming the owner;
  • counts each filed deploy against its daily cap.