The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data by reading local Codex session logs in scripts/read_codex_session.py. Boundary markers are established in SKILL.md (Workflow step 5), which explicitly commands the agent to treat all extracted content as untrusted and ignore any instructions found within them. The scripts/read_codex_session.py file provides sanitization by using the clean_user_text function to strip out context blocks from messages.
[COMMAND_EXECUTION]: The skill executes local Python scripts to parse session logs and manage user profiles. The regression test suite in scripts/test_read_codex_session.py uses subprocess.run to call the inspector script via the current system interpreter.
[DATA_EXFILTRATION]: The skill accesses local session logs in ~/.codex and a configuration profile in ~/.lovstudio. The profile management script scripts/profile_store.py includes a validation check (SENSITIVE_PARTS) to prevent the storage or exposure of credentials such as 'token', 'secret', and 'password'.
[PERSISTENCE]: The skill facilitates the storage of durable user preferences in a shared profile JSON file. This mechanism is restricted to specific vendor-approved namespaces (e.g., user, brand, workspace) and implements atomic writes with restrictive file permissions (0o600).
[EXTERNAL_DOWNLOADS]: The README.md file contains a version badge reference to img.shields.io, which is a well-known service for project metadata visualization.

lov-read-codex-session