The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill explicitly references and requires the use of a specific private SSH key file located at /Users/karim/.ssh/id_ed25519 for all remote operations.
[COMMAND_EXECUTION]: The skill contains numerous instructions to execute powerful shell commands, including:
Terminating processes using pkill based on command patterns.
Deleting cloud infrastructure resources via hcloud server delete and terraform destroy -auto-approve.
Performing system-level changes such as transactional-update and modifying systemd services via SSH.
[DATA_EXFILTRATION]: The skill instructions involve reading sensitive files (private SSH keys) and executing network-capable tools like hcloud, packer, and git, which could be used to transmit local data to external endpoints.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from system logs, journalctl output, and cloud provider API responses to drive subsequent agent actions, including automated code fixes and resource cleanup.
Ingestion points: Reads journalctl logs, hcloud server lists, and scenario summary files (summary-*.md).
Boundary markers: None present; the agent is instructed to treat these external logs as the "source of truth."
Capability inventory: Includes full shell execution, ssh access to remote nodes, file writing/modification, and cloud API management.
Sanitization: No evidence of sanitization or validation of the ingested log content before it influences agent decisions.
[DYNAMIC_EXECUTION]: The skill employs several dynamic execution techniques:
Creating temporary directories using mktemp -d and synchronizing code for runtime validation.
Dynamically constructing SSH commands with flags that bypass security checks (StrictHostKeyChecking=no).
Building system images using packer which dynamically fetches and executes plugins.

running-stabilization-loop