[reubenmiller]

Summary

Adds support for the thin-edge.io service commands. Container and container-group services can now be started, stopped and restarted from the Commands tab of the service in Cumulocity, using the workflow that tedge-agent ships.

This requires thin-edge.io > 2.0.1, which is currently only the main channel. Older versions behave exactly as before.

Changes

• Service plugins: tedge-container service-plugins container|container-group <action> <name>, which implements the Service Plugin API.
  • Installed as /usr/share/tedge/service-plugins/{container,container-group}, root:root, mode 0755.
  • Actions: start, stop, restart, pause, unpause.
  • Exit codes: 2 for an unsupported action, 1 for a failure, with the reason on stderr.
  • Arguments are never parsed as flags.
• Action declaration: the daemon publishes retained te/<service>/cmd/<action> capabilities for start, stop and restart on each container service.
  • Published once per process, only after the service registered successfully, and again after a re-registration.
  • Controlled by a new [service_commands] section: enabled = "auto" | true | false, and actions. In "auto", actions are declared only when tedge config get service.plugin_paths succeeds and the plugin is installed.
  • pause and unpause are not declared, as tedge-agent ships no workflow for them. The README shows how to add the workflow and declare them manually.
• Self-protection: no actions are declared on the container that hosts thin-edge.io (e.g. tedge-container-bundle). The plugins also refuse to act on it, because stopping it would disconnect the device and restarting it would interrupt the running workflow.
• Container-group lookup: the new LookupGroupService matches the exact compose project and service labels, and includes stopped containers. The previous lookup used a container-name substring match and only covered running containers.
  • The container-group log plugin uses it too, so it now finds the logs of stopped containers.
  • The service plugins reject a compose service scaled to more than one container.
• Docs: a README section on service commands. The design, specs and tasks are in the archived OpenSpec change openspec/changes/archive/2026-10-20-add-service-plugin-actions/.

Testing

• Unit tests for the plugin dispatch and exit codes, plugin path parsing, action filtering, the declaration bookkeeping (declare once, re-declare after deregistration, retry after a failed publish, skip the hosting container), and invalid container-group names. go test ./..., go vet and golangci-lint all pass.
• New system test suite tests/service-commands.robot. Its tests are skipped when the device's thin-edge.io lacks service plugin support (<= 2.0.1).
  • Ran locally against debian-14-docker-28 (tedge 2.0.2~425) with a Cumulocity tenant. All 6 tests passed, though not all in one run. In the final full run, the container-group test failed during its nginx install because the local Docker VM was out of memory and the kernel killed tedge-agent. That happened before any service command ran, and the test passed when rerun.
  • Coverage: c8y_SupportedServiceCommands; RESTART/STOP/START via c8y_ServiceCommand for a container and for a container-group service; plugin exit codes; pause/unpause via tedge service; capabilities cleared when a container is removed; enabled=false.
• goreleaser snapshot build: confirmed the plugin files and their modes in the deb, rpm and apk packages.

Known limitations

• The test devices run their own container engine (docker-in-docker), so the protection of the hosting container is only covered by unit tests.
• Detecting the hosting container reuses Self(), which can match the wrong container when containers use network_mode: host.
• Whether podman treats starting a running container, or stopping a stopped one, as success has not been checked yet. The podman matrix jobs will show it.

🤖 Generated with Claude Code

[github-actions]

Robot Results

✅ Passed	❌ Failed	⏭️ Skipped	Total	Pass %	⏱️ Duration
62	0	0	62	100	22m4.539568s