The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill features an autoUpdate mechanism that periodically executes shell commands in the background. Specifically, it runs npx @vectorize-io/hindsight-coding-agents to check for and apply updates to the local runtime once per day during a session start.
[DATA_EXFILTRATION]: The skill automatically ingests and transmits project-specific data to the vendor's API at api.hindsight.vectorize.io. This data includes commit messages, full git diffs (when configured), codebase surveys, and complete session transcripts. While consistent with the skill's stated purpose, this involves the transfer of sensitive intellectual property to an external service.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its ingestion of external, untrusted data that is later used to influence agent reasoning and synthesis.
Ingestion points: Untrusted data enters the agent context through git commit history, repository surveys, and external documents ingested via hindsight_ingest_document (SKILL.md).
Boundary markers: The instructions do not specify the use of clear delimiters or boundary markers to isolate external data from the system prompt or agent instructions.
Capability inventory: The skill provides tools for deep reasoning and synthesis, such as hindsight_reflect and hindsight_search_knowledge_pages, which process the ingested data (SKILL.md).
Sanitization: There is no documented evidence of sanitization, filtering, or validation of the content retrieved from external data sources before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute various maintenance and installation commands, such as npx @vectorize-io/hindsight-coding-agents install all. It also triggers codebase surveys that run via the current harness's CLI (e.g., Claude or Codex).

hindsight-coding-agent