The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing multiple gh (GitHub CLI) commands (edit, comment, view, list) to synchronize state with remote repositories. While it correctly uses --body-file to mitigate some command-line injection risks, it still involves executing shell commands based on internal agent state.
[DATA_EXFILTRATION]: The skill is designed to automatically transmit sensitive local data to GitHub, including verbatim conversation history (Last Turn), full response snapshots (Last Reply), and local markdown log contents (Plan Snapshot). If the targeted repository is public, this leads to significant information disclosure of the agent's internal workings and user interactions.
[PROMPT_INJECTION]: The resume functionality introduces an indirect prompt injection surface.
Ingestion points: Data is ingested via gh issue view --comments (referenced in SKILL.md).
Boundary markers: The instructions do not specify any delimiters or safety filters to isolate the recovered state from the rest of the conversation.
Capability inventory: The agent's subsequent actions are directly influenced by the Next, Goal, and Status fields recovered from these external comments.
Sanitization: There is no evidence of validation or sanitization of the comment content before it is used to restore the agent's operational context.

checkpoint-persistence