The Agent Skills Directory

[DATA_EXFILTRATION]: The skill records user-supplied insights (which may include specific file paths, commands, and system configurations) and transmits this information to an external endpoint (deepseek-v4-flash) via direct HTTP requests.
[REMOTE_CODE_EXECUTION]: The skill relies on the execution of an external shell script hindsight-jot-flush.sh and the hindsight CLI tool. Since these files are not provided as part of the skill, their logic cannot be verified and they could perform arbitrary actions on the user's system.
[TIME_DELAYED_CONDITIONAL]: The data processing and network activity are deferred until a "session-end hook" is triggered, which is a pattern of time-delayed behavior that avoids immediate detection.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by collecting untrusted user input that is later processed by a language model to generate persistent memories.
Ingestion points: User input through the jot command stored in ~/.agents/journal/jots/.
Boundary markers: Not present; the raw text is passed to an LLM for question-framing.
Capability inventory: Local file writing and network HTTP requests.
Sanitization: No sanitization or filtering is described in the data flow.

mental-model-jotting