The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a suite of verification commands that allow an agent to query and control the host environment. This includes tools for Kubernetes management (kubectl), macOS service orchestration (launchctl), VM management (colima), and network probes (curl).
[DATA_EXFILTRATION]: The documentation maps the locations of sensitive system resources, creating a discovery risk. It explicitly identifies paths for environment configurations (~/.config/system-bus.env), gateway settings (~/.joelclaw/gateway/.pi/settings.json), and Kubernetes secrets (pi-auth, agent-identity).
[PROMPT_INJECTION]: The skill identifies an indirect prompt injection surface by grounding its reasoning in external, potentially untrusted data sources.
Ingestion points: Reads the last 50 lines of ~/Vault/system/system-log.jsonl and agent configurations in ~/.joelclaw/gateway/AGENTS.md.
Boundary markers: No delimiters or instructions to ignore embedded commands within the log data are present.
Capability inventory: High-privilege access to the system via kubectl, launchctl, and curl for infrastructure modification.
Sanitization: No validation or sanitization logic is described for the ingested logs or markdown files.
[COMMAND_EXECUTION]: The verification section contains an embedded Python script (`python
<<'PY') used to read and parse local source files (packages/system-bus/src/inngest/functions/index.host.ts`) to count registered functions.

system-architecture