The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the openspec command-line tool (openspec list --json) to retrieve structured data about active change proposals and the current state of the workspace context.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local codebase and existing specification files to provide context for discussions, creating a potential injection surface.
Ingestion points: Reads source code files, design documents, and change artifacts located in openspec/changes/ and the broader codebase.
Boundary markers: None explicitly implemented; the agent is instructed to "reference them naturally" in the conversation.
Capability inventory: Shell command execution for project listing and write access to markdown documentation files (spec.md, design.md, tasks.md, proposal.md).
Sanitization: No specific content filtering or sanitization of ingested codebase data is described.
[SAFE]: The skill features robust internal instructions that explicitly prevent the agent from writing executable code or implementing features. Its operational scope is strictly confined to "capturing thinking" via markdown artifacts and investigative reads, which significantly reduces the potential impact of any malicious input encountered during discovery.

openspec-explore