The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external command-line utilities to verify credentials and model availability.\n
Evidence: Calls claude auth status --json, codex login status, and grok models to verify model lanes.\n
Evidence: Utilizes codex exec and Grok CLI to perform connectivity probes.\n- [PERSISTENCE]: The skill modifies system-level instruction files that govern agent behavior across multiple sessions.\n
Evidence: Appends an include directive (@~/.claude/pstack-models.md) to ~/.claude/CLAUDE.md, which is a configuration file loaded during agent initialization.\n
Evidence: Inserts a configuration block between markers in ~/.codex/AGENTS.md to persist model assignments.\n- [DYNAMIC_EXECUTION]: The skill spawns child agents and manages background processes during its verification phase.\n
Evidence: Launches "Claude-native agents" (e.g., pstack-fable-high) as one-turn probes.\n
Evidence: Executes spawn_agent in the Codex environment to test reasoning efforts.\n
Evidence: Launches external processes in the background with retained handles to verify multi-lane connectivity.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests external data to generate instruction sets, creating a potential vector for indirect influence.\n
Ingestion points: Reads the model matrix and descriptor grammar from provider-dispatch.md and existing sheet files in home directories.\n
Boundary markers: Uses XML-style comments () in Codex configuration files, but these serve as delimiters rather than security sanitization.\n
Capability inventory: Can write to persistent instruction files and execute CLI commands or child agents based on the parsed configuration.\n
Sanitization: Mentions normalization of specific rolling-alias model names, but lacks generic sanitization or escaping of the parsed content before it is written into system-level instruction files.

setup-pstack