The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill design incorporates a pattern where output from the openspec instructions command is treated as "prompt-level behavior contracts" and "operational guidance," which could allow external data to influence the agent's logic.
Ingestion points: The agent executes openspec instructions archive --change "<name>" --json and openspec instructions specs --change "<name>" --json, then parses the JSON response.
Boundary markers: The skill body contains instructions to keep external fields separate from built-in steps and to report conflicts. However, it explicitly directs the agent to "read and consider" the context and "follow entries" that are applicable, which creates an opening for instructions hidden in the data.
Capability inventory: The skill possesses capabilities to execute shell commands (mkdir, mv), read local files (tasks.md), and write to the filesystem via the openspec-sync-specs workflow.
Sanitization: The instructions include a check to ensure guidance is "applicable and compatible" and prohibit copying the guidance text verbatim into outputs, though these rely on the agent's interpretation rather than hard code-based sanitization.
[COMMAND_EXECUTION]: The skill uses the Bash tool to perform administrative file operations within the workspace.
Evidence: The skill executes mkdir -p to create archive directories and mv to transfer directories to the archive location. These commands utilize variables like planningHome.changesDir and changeRoot derived from tool outputs.

openspec-archive-change