The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts negative_ledger_memory_note.py, synesthesia_memory_note.py, and source-memory-reconcile.py use subprocess.run to interact with local CLI tools such as ledger, memory-note, and git for data validation, repository context resolution, and note transport.
[DYNAMIC_EXECUTION]: The source-memory-reconcile.py script dynamically loads the synesthesia adapter script from a relative path using importlib.util.spec_from_file_location to share logic between modules.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external ledger event files and eligibility definitions to create memory notes, which are eventually consumed by an agent. This creates an attack surface for indirect prompt injection if source records are malicious.
Ingestion points: Canonical ledger event files (e.g., .ledger/learnings/events.jsonl) and source-memory-eligibility JSON files.
Boundary markers: The skill produces structured Markdown files but does not explicitly implement specialized LLM boundary markers for the injected data.
Capability inventory: Shell command execution via subprocess.run and file writing to the agent's memory extension directory.
Sanitization: Implements schema-based structural validation through the ledger validate utility and manual field trimming before the data is persisted to notes.
[SAFE]: The implementation demonstrates strong security engineering, including the use of atomic regular-file replacements to prevent data corruption and the enforcement of restricted file permissions (e.g., 0o600 for note files and 0o700 for directories) to protect memory contents.

memory-source-notes