Skip to content

Proposals and approval

The model proposes, a person approves. Two seylo-mcp tools change nothing by themselves. They create something for a person to decide on:

Tool Creates Who decides Where
dashboard_propose A dashboard draft A releaser publishes or discards it Observability → Dashboards
scrape_target_propose A manifest-edit request in the approval queue An approver or admin approves or rejects it Operate → Approvals

A coding agent can’t decide on its own proposal, or on anyone else’s. seylo-mcp has no tool that approves, publishes or discards anything.

When an agent calls dashboard_propose, Seylo first checks the spec with the same validator the Dashboards page uses. Then it saves the dashboard as a draft. Nothing is published, and nobody’s dashboard changes.

On Observability → Dashboards, drafts appear in a separate card, Proposed by coding agents. The draft:

  • isn’t in the ordinary dashboard list until a releaser publishes it;
  • shows Proposed by with the token’s label and the time;
  • belongs to the person who issued the token;
  • can carry a short note from the agent explaining what it is for. Seylo redacts the note before storing it.

Open a draft to see it drawn with live data before you decide. Its queries run with your access, as any dashboard’s do.

A releaser presses Publish or Discard. Publishing makes it an ordinary dashboard. A discarded draft is never published. Each decision is recorded in the audit log.

A token can have 10 drafts waiting at once. The eleventh is refused until someone publishes or discards one.

scrape_target_propose asks for a Deployment’s metrics to be collected by the annotated-pods job of the observability chart. It adds three annotations to the Deployment’s pod template, which is where the scrape job reads them. On the Deployment’s own metadata they would do nothing:

spec:
template:
metadata:
annotations:
prometheus.io/scrape: "true"
prometheus.io/port: "9090"
prometheus.io/path: "/metrics"

Before filing, Seylo starts from the manifest it tracks for the workload, or from the live object if it doesn’t track one. It then checks who owns the Deployment and runs a server-side dry run of the change. Only then does it file the request.

The request always goes into the approval queue, even if your organisation has approvals turned off.

On Operate → Approvals the request is an ordinary manifest edit, with a summary in this form:

manifest-edit(shop-dev): checkout [deployment] - add prometheus.io scrape annotations (port 9090, path /metrics) (coding agent: Priya's Claude Code)

Its notes start with Proposed by coding agent "Priya's Claude Code" (MCP scrape_target_propose). They also include the agent’s reason, if it gave one. The ownership report travels with the request.

The request is filed in the name of the person who issued the token.

Seylo checks ownership again, then applies the change and reads it back. Because the pod template changes, the Deployment’s pods roll. About one scrape interval (30 seconds) after the new pods are ready, the agent can confirm the target with metrics_series and the selector up{job="annotated-pods"}.

If the request is rejected, nothing is applied.

What is refused before it reaches a person

Section titled “What is refused before it reaches a person”

These checks run when the agent calls the tool. A refusal is returned to the agent with the reason, and no draft or request is created.

Situation Seylo’s answer begins
Argo CD, Flux or Helm manages the Deployment Not proposed: Deployment checkout is managed by and names the controller. A coding-agent token can’t acknowledge another owner, so the annotations belong in that owner’s source, such as the Git repository or Helm values
The cluster’s dry run rejects the change Not proposed: the cluster's dry run rejected the change.
The workload doesn’t exist No Deployment "checkout" in shop-dev.
The workload is a StatefulSet or DaemonSet "checkout" is not a Deployment. Ask a person to add the annotations
The annotations are already there Not a refusal. The status is unchanged and nothing is filed
The dashboard spec is invalid The dashboard spec is not valid, so nothing was saved., followed by each problem and where it is in the spec
The token already has 10 drafts waiting This token already has 10 dashboard drafts waiting for a person, the most allowed.
The capability wasn’t ticked on the token The tool isn’t listed for the token at all

Alert-rule and silence proposals aren’t available.

request_secret records that the application needs a Secret: its name, its key names and why. It never carries a value. Seylo writes the request to the audit log and replies with where a person must create it: in the vault, if the secrets-store CSI driver syncs that Secret, or as an ordinary Kubernetes Secret if it doesn’t. Seylo doesn’t put a task in a queue or send a notification. The agent tells its user what to create.