The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands to manage the task planning lifecycle. This includes using bash to run a setup script (create-folders.sh) and invoking git for change detection (git status, git diff) and file organization (git mv). These operations are scoped to the local environment and intended for project management.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes untrusted data to influence AI behavior.
Ingestion points: Untrusted data enters the workflow via the $ARGUMENTS variable and the content of the user-provided TASK_FILE in the .specs/tasks/ directory.
Boundary markers: The skill lacks robust delimiters or "ignore instructions" warnings when interpolating untrusted file paths and task content into the prompts for sub-agents (e.g., in the prompts for sdd:researcher and sdd:tech-lead).
Capability inventory: The orchestration logic has the capability to write and move files on the local system and launch multiple foreground agents (sdd:researcher, sdd:code-explorer, sdd:software-architect, etc.) with varying model tiers.
Sanitization: No explicit sanitization, escaping, or schema validation is applied to the user-provided task specifications before they are used to generate implementation plans.

plan-task