The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled script, scripts/mo-knowledge-layer.mjs, via Node.js to evaluate the state of the target project. This script uses spawnSync to call local git commands, including ls-tree, cat-file, and log, to read metadata and file content from a specified commit SHA. The agent is also tasked with running E2E scenarios through its available command-line tools.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external Git repositories, which serves as a surface for indirect prompt injection.
Ingestion points: The knowledge layer script reads files such as AGENTS.md, Makefile, and package.json directly from the Git repository's commit history using the git cat-file command.
Boundary markers: The skill relies on specific patterns and markdown parsing, using a bundled version of the micromark library, to locate status markers rather than executing the raw content of the ingested files.
Capability inventory: The skill has the capability to execute its internal scripts and run E2E scenarios through agent tools. It also references cleanup operations performed via the orca command-line utility.
Sanitization: The logic is constrained to extracting project signals through regular expressions and parsing, which helps mitigate the risk of executing malicious payloads found within repository data.
[DATA_EXFILTRATION]: The skill's instructions identify sensitive local configuration paths, including ~/.claude.json and ~/.codex/config.toml, as potential sinks for test fixtures. The skill demonstrates a security best practice by explicitly forbidding the agent from modifying these shared configuration files directly, instead instructing it to provide the user with the necessary commands for manual cleanup.

mo-e2e