The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill contains instructions to read from ~/.claude/.credentials.json using grep to check for the presence of authentication tokens. Accessing platform-specific credential files is a high-risk operation that can lead to token exposure.
[COMMAND_EXECUTION]: The skill requires the execution of numerous shell commands and custom binaries that are not permitted by the allowed-tools configuration (which is restricted to Read, Glob, and Grep). Proscribed actions include:
Modifying project files and performing git operations (git add, git commit, git rm).
Executing setup and configuration binaries from hidden directories (~/.claude/skills/gstack/bin/gstack-config, gstack-team-init).
Running the primary benchmark binary (gstack-model-benchmark).
Manipulating the file system using touch and find.
[COMMAND_EXECUTION]: The skill uses eval to execute the output of a local binary (gstack-slug), which is a dangerous dynamic execution pattern as it executes arbitrary code returned by the subprocess.
[DATA_EXFILTRATION]: By reading sensitive configuration files and checking for environment variables like ANTHROPIC_API_KEY, the skill accesses private credentials that could be exposed or mishandled during the benchmarking process.
[COMMAND_EXECUTION]: The skill provides instructions to modify the project's CLAUDE.md and .gitignore files, which could be used to hide malicious activity or modify the agent's core routing behavior in other sessions.

benchmark-models