The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill extracts and processes historical conversation data to generate summaries and progress updates. This creates an attack surface where malicious instructions embedded in past conversations could attempt to influence the agent's analysis or the resulting progress records.
Ingestion points: Historical conversations and prompts are read from the sdk_sessions and user_prompts tables in the local claude-mem.db database via the scripts/request-groups.mjs script.
Boundary markers: The skill instructions specify that the agent and subagents should use verbatim quotes and specific JSON formats for data exchange.
Capability inventory: The subagent used for reviewing requests has access to memory search tools (mem_search, timeline, get_observations) and a CLI tool for recording progress.
Sanitization: The requestText function in the provided script includes logic to strip system-level tags (e.g., <instructions>, <system-reminder>) and environment context from ingested prompts before they are used for analysis.

request-progress