The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes npx @claude-flow/cli@latest to install the required CLI components. This package is hosted on a public registry and does not originate from a recognized trusted organization or well-known service.
[REMOTE_CODE_EXECUTION]: The use of npx results in the download and immediate execution of remote JavaScript code in the local environment. This execution pattern lacks specific version pinning or integrity verification.
[COMMAND_EXECUTION]: The skill executes multiple shell commands, including /gaia for benchmark operations and npx for metadata storage and task tracking. These commands interact with the local file system and network.
[DATA_EXPOSURE]: The environment validation phase checks for the presence of ANTHROPIC_API_KEY and HF_TOKEN. While it only echoes the prefix of these keys for verification, it highlights the skill's reliance on and potential exposure of highly sensitive credentials.
[INDIRECT_PROMPT_INJECTION]: The skill processes benchmark trajectories and results which contain untrusted agent-generated data.
Ingestion points: Reads result files from ~/.cache/ruflo/gaia/results-latest.json and processes trajectories.jsonl during packaging and comparison phases.
Boundary markers: No specific boundary markers or instructions to ignore embedded commands are present when the agent processes the benchmark content.
Capability inventory: The skill has access to shell execution, file system writes, and persistent memory storage through MCP tools.
Sanitization: There is no evidence of sanitization, validation, or escaping of the ingested JSON data before it is parsed or used to generate reports.

gaia-submission