The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing shell commands that perform in-place modifications of local application configuration and data files.
Evidence in SKILL.md: Uses sed -i to perform bulk string replacement within session files in ~/.codex/sessions.
Evidence in SKILL.md: Executes SQL UPDATE statements using sqlite3 to modify the internal state of ~/.codex/state_5.sqlite.
Evidence in SKILL.md: Uses jq to delete authentication fields from ~/.local/share/opencode/auth.json.
[EXTERNAL_DOWNLOADS]: The skill documentation includes commands to download and run software from external public registries.
Evidence in SKILL.md: Pulls and runs a Docker image from GitHub Container Registry (ghcr.io/soju06/codex-lb).
Evidence in SKILL.md: Installs a Python package using the uvx tool.
[INDIRECT_PROMPT_INJECTION]: The skill defines a proxy architecture for LLM requests, creating an attack surface for instructions embedded in external data.
Ingestion points: The load balancer processes upstream OpenAI API responses and user-supplied session tokens as described in SKILL.md.
Boundary markers: No explicit prompt delimiters or boundary markers are defined in the configuration examples.
Capability inventory: The skill utilizes file system modification (via sed, jq, sqlite3), Docker container management, and network proxying.
Sanitization: No sanitization or filtering logic for proxied content is documented in the skill instructions.

codex-lb-load-balancer