The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by aggregating data from external or untrusted sources to brief subagents.
Ingestion points: The skill reads data from $ARGUMENTS, tasks.md, and project-specific context files such as AGENTS.md, CLAUDE.md, and CONTEXT.md (SKILL.md, Phase 0 and Phase 5).
Boundary markers: There are no explicit delimiters or 'ignore embedded instructions' warnings used when passing these ingested contents to the clean-context subagents.
Capability inventory: Dispatched subagents have the ability to execute project code (tests) and commit changes to the repository via the Agent tool and speckit-checkpoint-commit skill.
Sanitization: No sanitization, escaping, or structural validation is performed on the repository data before it is interpolated into the subagent prompts.
[COMMAND_EXECUTION]: The skill instructs subagents to dynamically discover and execute project test runners (e.g., pytest, npm test, cargo test, make test) based on repository configuration files like package.json or Makefile. While this is intended functionality for implementation verification, it represents the execution of arbitrary code found within the project environment.

speckit-opsmill-implement