The Agent Skills Directory

[SAFE]: The skill performs local file operations on conversation transcripts and skill directories to generate usage analytics. No malicious behavior or data exfiltration patterns were detected. The script scripts/analyze.py handles all operations locally using the standard sqlite3 and pathlib modules.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted conversation transcripts (Category 8 surface). It identifies usage patterns by parsing transcript.jsonl files in scripts/analyze.py. The implementation includes boundary management by using regex to filter out system tags, code blocks, and non-user prose before analysis. It possesses minimal capabilities (local file reads and SQLite writes), with no ability to execute code from the transcripts or perform network operations, mitigating common indirect injection risks.
[DATA_EXFILTRATION]: While the skill accesses sensitive conversation history located in the agent's brain directory (~/.gemini/antigravity/brain), it does not contain any code to perform network requests, and it does not import network-related libraries (like requests, urllib, or socket). All extracted data remains within a local SQLite database file (skill_analytics.db).

skill-analytics