The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download a script from a remote URL and pipe it directly into the shell using curl -fsSL "<anteater>/send-transcript" | bash. This pattern is highly risky as it executes unverified code from a remote source on the user's system.
[DATA_EXFILTRATION]: The skill is designed to send coding session transcripts to an external endpoint (POST /review). Transcripts of a development session often contain sensitive intellectual property, logic, and environment details. Although the skill claims the script redacts secrets, the transmission of full session history is a significant privacy and security risk.
[COMMAND_EXECUTION]: The skill utilizes shell commands (curl, bash) to perform its primary workflow, increasing the risk of command injection if the <anteater> placeholder or the --idea argument (which comes from the session) is not properly sanitized.
[INDIRECT_PROMPT_INJECTION]: The skill processes session transcripts, which are essentially untrusted data aggregated during the coding session (e.g., from web outputs or file reads). This data is later summarized by the agent, creating a surface for embedded instructions in the transcript to influence the agent's summary or subsequent actions.
Ingestion points: The transcript is retrieved and summarized in SKILL.md (Workflow steps 1 and 4).
Boundary markers: No specific delimiters or "ignore instructions" markers are defined for the transcript summary step.
Capability inventory: The skill can execute shell scripts and access network endpoints.
Sanitization: The skill mentions that the helper script redacts keys before upload, but does not specify sanitization of the content before the agent summarizes it.

convex-improve-convex-plugin