The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's build phase involves reading instructions and context files to implement code changes. This ingestion of potentially untrusted data represents an indirect prompt injection surface.
Ingestion points: Ingests data from contextFiles, apply instructions (retrieved via comet classic openspec), and artifacts like tasks.md and proposal.md in SKILL.md.
Boundary markers: No explicit boundary markers or delimiters are defined to separate untrusted instruction data from agent instructions.
Capability inventory: The agent can execute shell commands (comet, mvn, npm) and write to the file system (code edits, comet knowledge).
Sanitization: No evidence of sanitization or validation of the ingested instructions or context files is present.
[COMMAND_EXECUTION]: The skill uses the comet CLI suite and project build tools to perform actions such as workspace preparation, state management, code formatting, and testing.

comet-tweak