The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a Python script (compact.py) to process memory entries. The script performs standard file I/O operations (reading MEMORY.md and writing to a YAML state file) and uses local command-line arguments for control flow. These operations are restricted to the skill's defined workspace and state directories.
[DATA_EXPOSURE]: The skill reads from ~/.openclaw/workspace/MEMORY.md to generate summaries. While this involves sensitive agent memory, the data is processed locally, and the resulting summaries are stored within the skill's own protected state directory (~/.openclaw/skill-state/). There is no evidence of external transmission.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (content from MEMORY.md) to generate summaries.
Ingestion points: Reads MEMORY.md in compact.py via parse_memory().
Boundary markers: None identified for the summary output, though the DAG structure acts as a logical delimiter.
Capability inventory: File read/write to state and workspace, local execution of compact.py.
Sanitization: Simple keyword extraction and length-based truncation are used for summary generation. While an attacker could theoretically embed instructions in MEMORY.md to bias the summaries, the impact is limited to the agent's internal memory representation and does not grant further system access.

memory-dag-compactor