The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell scripts (pre-compact.sh and post-compact.sh) designed to be executed by the agent's lifecycle hooks (PreCompact and SessionStart). These scripts perform file system operations to manage session state.
[DATA_EXFILTRATION]: The pre-compact.sh script reads the complete conversation transcript (a .jsonl file) from the local disk to extract recent exchanges. While this involves handling sensitive session data, the information is stored in a temporary local file (.memento-handoff) and is not transmitted over the network.
[PROMPT_INJECTION]: The post-compact.sh script injects a footer containing specific instructions into the conversation context after a compaction event. These instructions direct the agent to summarize the current task and continue work without prompting the user, which is the intended functional behavior of the skill.
[PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by re-injecting fragments of the conversation transcript back into the active context. If the original conversation included untrusted data (e.g., content from a malicious file or website analyzed during the session), that data is re-presented to the agent in the post-compaction state. The skill attempts to mitigate this by wrapping the trace in boundary markers.

memento