The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill performs remote code execution by downloading a script via curl from a placeholder URL (<anteater>/send-transcript) and piping it directly into a bash shell. This pattern executes arbitrary network content without integrity checks or prior verification.- [DATA_EXFILTRATION]: The skill is designed to exfiltrate coding session transcripts, which contain code snippets, interaction history, and project context, to an external endpoint (<anteater>/review). While the skill claims to redact secrets and require user consent, the transmission of local session data to a third-party server represents a significant data exposure risk.- [COMMAND_EXECUTION]: The skill invokes shell commands using bash -s -- --idea to process session-derived data. This execution path is sensitive to the content of the script downloaded from the network and the arguments provided during the session.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from the coding session transcript.
Ingestion points: Coding session transcripts gathered from the terminal or IDE context.
Boundary markers: None identified.
Capability inventory: Network access via curl, shell execution via bash, and file system access to read transcripts.
Sanitization: The skill claims to redact keys/tokens before upload, but does not provide details on sanitization for embedded instructions within the transcript.

convex-improve-convex-plugin