The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to discover the environment and verify model availability, including gh auth status, bun --version, orca orchestration run-create, and various agent-specific CLI tools like claude, grok, and codex.
[DATA_EXFILTRATION]: The skill accesses sensitive local files including agent configuration files (~/.claude/settings.json, ~/.codex/config.toml, ~/.grok/config.toml) and detailed session history logs (.jsonl files in Codex and Grok session directories). While the data is used to generate a local skill, reading these private paths constitutes significant data exposure.
[DYNAMIC_EXECUTION]: The core functionality is the automated generation of a new skill (select-model) based on environment discovery. This involves creating new markdown instructions that are interpreted as logic by the agent in future tasks, effectively performing dynamic instruction generation.
[PERSISTENCE]: The skill establishes persistence by installing the generated select-model skill into ~/.agents/skills/ and creating relative symlinks in product-specific directories like ~/.claude/skills/ and ~/.grok/skills/, ensuring the generated logic is active across future agent sessions.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect injection by ingesting data from external model catalogs and local session logs (e.g., in SKILL.md Step 2 and references/grok.md). This untrusted content is interpolated into the generated select-model skill without explicit sanitization or safety boundary markers, which could allow malicious data in a log or catalog to influence future agent decisions.

create-model-selection-skill