The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands specified in the Closure field (such as grep, comm, find) during the plan creation phase in Step 4. These commands are derived from the preceding session analysis and are executed immediately to establish a baseline.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection by ingesting data from the current session context (Step 1) to generate executable shell commands.
Ingestion points: The plan is derived from an analysis already in context (e.g., from tools like /investigate or /probe).
Boundary markers: The instructions do not define specific boundary markers or delimiters for the ingested analysis data.
Capability inventory: The skill has access to Bash, Write, and Edit tools, and Step 4 explicitly triggers command execution.
Sanitization: Step 5bis performs structural validation (unique IDs, dependency cycles) but does not include sanitization or filtering of the shell commands themselves before execution.
[DYNAMIC_EXECUTION]: In Step 5bis, the instructions mandate that the agent generate and run a "short script" at runtime to perform mechanical validation of the generated plan file.

plan-context