The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local system commands (like git and date) and internal vendor-specific scripts (namespaced plan-marshall:*) to automate plan directory creation and configuration. These operations are limited to the project's own workspace and specialized tooling.
[DATA_EXFILTRATION]: The skill communicates with GitHub to fetch issue details. As GitHub is a well-known service and the data is retrieved as part of the primary task of issue-to-plan conversion, this represents intended functionality without risk of unauthorized exfiltration.
[PROMPT_INJECTION]: The skill handles untrusted data from descriptions and issues. It utilizes a two-step writing pattern for request.md to avoid shell-level command injection and stores user content within designated markdown sections to maintain clear boundaries between system instructions and user data.
Ingestion points: Task descriptions, lesson records, and GitHub issue bodies retrieved in SKILL.md Step 4.
Boundary markers: Uses a structured request.md template with a ## Original Input header to isolate untrusted content.
Capability inventory: Includes file system writes, plan directory deletion, and execution of internal management scripts.
Sanitization: Employs a specific path-allocation flow for writing large blocks of text to files, preventing content from being interpreted as shell arguments.

phase-1-init