The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md and README.zh-CN.md files provide an installation command using curl -fsSL ... | sh. This command fetches a script from the official repository of the skill author (github.com/levi-qiao/longgraph-skill) to initialize the local environment and create symlinks for the /longgraph command. As the script originates from the vendor's own infrastructure for set-up purposes, it is documented as a core function without malicious intent.
[COMMAND_EXECUTION]: The executor node is designed to run arbitrary shell commands defined as 'gates' (e.g., pytest, make test) within the project workspace to verify task completion. These commands are executed within the local development context as part of the agent's primary work cycle.
[INDIRECT_PROMPT_INJECTION]: The skill architecture involves agents reading and writing shared state files like ledger.md and directives.md, creating a potential ingestion point for instructions embedded in the repository. To mitigate this, the prompts in executor.md and supervisor.md include explicit boundary markers and instructions to ignore previously loaded skills and stick to the provided Markdown contract. Capability inventory includes subprocess calls for gates and file-writing to state artifacts. Sanitization is achieved through structured Markdown schemas and the use of watermarks to fold and archive directive history, reducing the surface for persistent injection.

loop-graph