How it works and the trust model
Exactly what the agent sends, what it can run, and what you have to trust.
What the agent sends
- For each container labelled stratacore.service: its name, team, version and repository labels, health status, CPU and memory use, and a small config summary: image name and tag, exposed ports, restart policy, memory and CPU limits, and whether a healthcheck exists.
- For each preview: its PR number or name, CPU and memory use, whether it is active, and whether it is running or asleep.
- Test results that you post yourself, and, if you configure it, Terraform drift summaries.
It never sends environment variable names or values, file contents, logs from your services, Terraform state or plan values, or your git token.
What it can run
- Only a fixed list of job kinds. There is no job that runs an arbitrary command.
- Every value in a job (branch, commit, repository, URL, service, version) is checked against a strict pattern on your machine before use, so the server cannot smuggle in shell syntax or flags.
- Hooks are commands you configured on the agent. The server cannot change them.
- Built-in previews run the docker and git commands described above. Kubernetes mode changes only namespaces it created and marked itself.
What you still have to trust
- The agent needs access to the Docker socket, which is equivalent to administrator access on that host. Run it on a host you are comfortable giving that to.
- The job queue comes from the StrataCore control plane. A compromised control plane could ask for any of the allowed job kinds with values that pass validation, for example to create or remove a preview, or to run your own hooks. It cannot run commands that you did not configure.
- Preview environments run code from your repository.
The agent file is a single, dependency-free script. Read it before you install it, and pin its checksum with STRATACORE_AGENT_SHA256.