The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes local Codex session transcripts and logs. While the instructions focus on metadata (size, age) and file management (archival), the content of these files is untrusted and could contain instructions targeting the agent.
Ingestion points: $CODEX_HOME/sessions (JSONL transcripts) and Codex CLI logs.
Boundary markers: The instructions specify a read-only "Audit" mode by default and require explicit "Apply" mode for modifications.
Capability inventory: File moving/copying (archival/backup), configuration modification (TOML), and scheduled task creation.
Sanitization: No specific content sanitization is mentioned; behavior is restricted by the metadata-focused nature of the audit.
[DATA_EXPOSURE]: The skill targets sensitive local data including session transcripts, configuration files, and databases located in ~/.codex and ~/Library/Logs/com.openai.codex. The instructions include explicit safeguards against exfiltration, prohibiting the placement of this data in shared folders or cloud-synced destinations and specifically forbidding the copying of authentication credentials.
[PERSISTENCE_MECHANISMS]: The skill includes an "Automate" mode that allows the agent to create recurring scheduled tasks for auditing and reporting. This is a legitimate maintenance feature triggered only upon explicit user request and is constrained to read-only operations.
[COMMAND_EXECUTION]: The skill references an external command or tool $jww-handoff to handle task transitions during archival. This tool is treated as a host-supported operation.

jww-codex-maintenance