The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes complex data from external sources and infrastructure environments to make operational decisions, creating a surface for injection.
Ingestion points: The agent is instructed to read and analyze outputs from kubectl, terraform show -json, and hcloud server list (in SKILL.md and references/immutable-node-replacement.md).
Boundary markers: The instructions lack explicit delimitation or warnings to treat outputs from these external tools as untrusted data.
Capability inventory: The skill grants the agent powerful capabilities, including shell execution, SSH root access, Terraform plan application, and firewall rule modification.
Sanitization: No sanitization or validation of the ingested infrastructure metadata is specified before it is interpolated into subsequent management commands.
[COMMAND_EXECUTION]: The skill utilizes a broad suite of system administration tools including ssh, terraform, kubectl, and hcloud. While necessary for cluster management, these provide the agent with full control over the target infrastructure and local environment.
[DATA_EXFILTRATION]: The skill requires access to highly sensitive configuration and credential files. While it explicitly forbids committing these to version control, their availability in the environment represents a risk of exposure.
Evidence: Commands access and manipulate kubeconfig files, .tfstate files, and SSH private keys (referenced in SKILL.md under 'SSH Access Sanity' and state pull sections).
[EXTERNAL_DOWNLOADS]: The skill performs remote operations to fetch code and configuration, which is standard for the workflow but involves external dependencies.
Evidence: Uses git pull to update the local repository and terraform init -upgrade to download provider binaries and infrastructure modules.

upgrade-cluster