The Agent Skills Directory

[COMMAND_EXECUTION]: The included Python script executes the pgrep command to detect if the Codex application is running. This is used to implement a safety lock that prevents the skill from modifying files while the application is active, minimizing the risk of data corruption.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes data from local application files.
Ingestion points: The scripts/codex_history_repair.py script reads thread data from ~/.codex/state_5.sqlite and session metadata from rollout-*.jsonl files.
Boundary markers: Data retrieved from these files is presented to the agent in reports (e.g., in status or dry-run commands) without explicit delimiters or warnings to ignore instructions embedded within the metadata.
Capability inventory: The skill provides the agent with the ability to modify SQLite records, rewrite JSONL files, and delete/move files (via the delete-archived and sync commands in the provided Python script).
Sanitization: While the script uses standard parsing libraries, it does not perform sanitization or filtering on the metadata strings (such as cwd or provider names) before they are displayed to the agent.

codex-history-recovery