The Agent Skills Directory

[OBFUSCATION]: The skill uses HTML entities to hide the name of a competitive product and its configuration path. Specifically, 'Codex' and '~/.codex/config.toml' are used in SKILL.md to mask references to 'Codex' and its sensitive configuration directory.
[DATA_EXPOSURE]: The skill attempts to read sensitive configuration files from multiple agent products, including ~/.claude/settings.json, ~/.codex/config.toml, and various .mcp.json files. These files frequently contain API keys, session tokens, and environment-specific credentials.
[EXTERNAL_DOWNLOADS]: The skill fetches model pricing data from the external URL https://models.dev/api.json to compute usage estimates.
[DYNAMIC_EXECUTION]: The qa-savings-fixture.sh script utilizes a Node.js heredoc to execute dynamically generated JavaScript for computing session savings.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted historical session logs from other agent platforms, creating a potential surface for indirect prompt injection.
Ingestion points: SKILL.md (Lane 3 and 4) reads session files (JSONL) from various local agent history directories.
Boundary markers: No explicit delimiters or system instructions are present to prevent the agent from following instructions embedded in the historical session data.
Capability inventory: The skill has access to bash, read, write, edit, and memory tools, which could be leveraged if a malicious instruction in the session history is obeyed.
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from session transcripts before processing.
[COMMAND_EXECUTION]: The skill executes shell commands, including git rev-parse and various file system checks, to validate project eligibility for generating documentation.

onboarding