The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates shell scripts (eval/eval.sh, prepare.sh) and then executes them at runtime during the 'Test & Iterate' phase (Phase 5). This constitutes script generation followed by execution, which is a potential risk if the generated content is influenced by malicious input.
Evidence: bash eval/eval.sh and test -f prepare.sh && bash prepare.sh in Phase 5.
[COMMAND_EXECUTION]: The skill uses shell commands to interact with the local file system and external services, including repository initialization and task management tools.
Evidence: Usage of git, gh (GitHub CLI), and hive CLI commands in Phase 6.
[CREDENTIALS_UNSAFE]: The skill prompts the user to provide a sensitive admin key and reads from an environment variable (HIVE_ADMIN_KEY) to perform administrative actions on an external server.
Evidence: Instructions in Phase 6.3b to read HIVE_ADMIN_KEY or prompt the user for an admin key to be used in the hive task create command.
[INDIRECT_PROMPT_INJECTION]: The skill creates a instruction set (program.md) that will be interpreted by other agents. This creates a vulnerability surface where malicious user input provided during the design phase can be embedded as instructions for subsequent agents.
Ingestion points: Phase 1 (Problem definition, artifact description, task data).
Boundary markers: Absent. The program.md template does not include specific safety boundaries for the generated content.
Capability inventory: The skill has full file-write capabilities and shell execution via bash.
Sanitization: Absent. The skill interpolates user-provided strings directly into the markdown and script templates.

hive-create-task