The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md and README.zh-CN.md files promote an installation method that pipes a remote script directly to sh (curl -fsSL https://raw.githubusercontent.com/levi-qiao/longgraph-skill/main/install.sh | sh). This pattern allows for the execution of arbitrary, unverified code from an external GitHub repository.
[COMMAND_EXECUTION]: The executor.md and supervisor.md templates instruct the agents to execute shell commands provided in the ops.md file, such as git and pytest. This creates a direct path for arbitrary command execution if the shared configuration files are compromised.
[INDIRECT_PROMPT_INJECTION]: The skill relies on agents reading and obeying instructions from shared Markdown files (ledger.md, directives.md, findings.md) that are updated during runtime. Ingestion points: Agent nodes ingest instructions and findings from other nodes and the workspace. Boundary markers: The provided templates lack delimiters or instructions to prevent the agent from processing malicious prompts embedded in the ingested data. Capability inventory: The agents have the ability to run shell tools and git commands. Sanitization: There is no evidence of input validation or sanitization of the data before it is interpreted by the LLM.
[EXTERNAL_DOWNLOADS]: The skill documentation directs users to download and execute code from the author's GitHub repository (levi-qiao), which is not a pre-verified trusted organization, introducing risk from unvetted third-party infrastructure.

loop-graph