The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests historical AI conversation logs from $HOME/.codex/sessions/ to identify cognitive blind spots. This creates an attack surface where instructions embedded in previous conversations could influence the agent's current analysis.
Ingestion points: JSONL session files containing previous user and assistant messages.
Boundary markers: No specific delimiters or "ignore previous instructions" markers are used when interpolating log data into the prompt.
Capability inventory: The skill can write files to the ~/Context/ directory and perform network requests via curl.
Sanitization: The skill uses jq to filter out some system-level tags (e.g., <environment_context>) but does not sanitize the user-provided text content for instructions.
[COMMAND_EXECUTION]: The skill relies on shell commands (find, jq, curl, date) to locate session files, extract message content, and interact with the WeRead API.
[EXTERNAL_DOWNLOADS]: Makes network requests to i.weread.qq.com (WeRead) to search for books and retrieve chapter information. This is a well-known service and the operations are consistent with the skill's stated purpose.
[DYNAMIC_EXECUTION]: Executes a local TypeScript tool (Tools/WeReadWebUrl.ts) via the bun runtime. This script performs algorithmic encoding of IDs to generate web links and does not perform additional network or file operations.

ljg-blind