The Agent Skills Directory

[COMMAND_EXECUTION]: Spawns pnpm and python subprocesses using spawnSync with shell: false. This approach ensures that arguments are not interpreted by a shell, effectively mitigating command injection vulnerabilities.
[PROMPT_INJECTION]: Features an indirect prompt injection surface due to the processing of untrusted data from the local file system. Malicious instructions embedded in files could influence the compression process or the generated memory records.
Ingestion points: main.cjs reads file paths and content from search results to build a context corpus.
Boundary markers: The skill lacks clear delimiters or instructions for the model to ignore prompts hidden within the ingested text.
Capability inventory: The tool has permissions to execute scripts and write results to the memory files used by the agent.
Sanitization: No content validation or sanitization is performed on the ingested data beyond basic ANSI code stripping.
[DATA_EXFILTRATION]: All operations are performed locally. No network-related commands or calls were found, ensuring that data exposure is confined to the project environment.

token-saver-context-compression