The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes 'reusable knowledge' claims submitted by agents that are later consolidated into a shared memory store.
Ingestion points: Files written to ~/.agents/shared-memory-inbox/pending/ are the primary entry point for untrusted data.
Boundary markers: YAML frontmatter is used to separate metadata from the natural language content of the submission.
Capability inventory: The skill has capabilities to write to memory files, execute Python scripts, and perform network operations via scp.
Sanitization: A validation script (validate_submission.py) performs schema checks and identifies common secret patterns (API keys, private keys), but does not sanitize the claim body for prompt injection payloads.
[COMMAND_EXECUTION]: The skill directs the agent to use various shell commands for its core logic.
Evidence: Uses mkdir and rmdir to implement a file-based lock mechanism (.curator.lock).
Evidence: Executes the scripts/validate_submission.py script using python3.
Evidence: Commands the use of scp to transfer consolidated notes to other systems.
[DATA_EXFILTRATION]: The skill facilitates the movement of processed data to external hosts.
Evidence: The curation process includes a step to scp memory notes to remote hosts specified in the evidence_host metadata field, which could be used to transmit internal memory contents to arbitrary network locations.

shared-memory-intake