The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The README.md and installation instructions for Codex, Cursor, and Grok Build utilize a remote script for setup: curl -fsSL https://raw.githubusercontent.com/levi-qiao/longgraph-skill/main/install.sh | sh. This script, hosted on the author's GitHub repository, automates the process of cloning the repository and symlinking skill directories. As this is a vendor-owned resource used for standard installation, it does not represent a malicious threat.
[COMMAND_EXECUTION]: The skill templates (executor.md, supervisor.md) and host-specific reference files (e.g., claude-code.md, grok.md) provide instructions for agents to execute shell commands. These commands are used for build processes, running test suites (like pytest), and environment verification. This is standard functionality for an agent designed to perform software engineering tasks.
[INDIRECT_PROMPT_INJECTION]: The core architecture relies on agents reading from and writing to shared Markdown files (ledger.md, directives.md). This ingestion of state files constitutes an attack surface for indirect prompt injection. If an agent records untrusted data from an external source (like a webpage or a malicious file) into the ledger.md without sanitization, it could potentially influence the behavior of the supervisor or the next executor iteration. However, the skill includes explicit 'red lines' and 'method guards' to minimize such risks.
[DYNAMIC_EXECUTION]: The skill facilitates the dynamic generation of prompts and the execution of project-specific commands defined in the ops.md file. The supervisor agent is also designed to independently verify work by running dynamically determined test commands. This behavior is documented and central to the skill's purpose of autonomous task verification.

longgraph