The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session data from external trajectory files to determine its sequence of actions, which could allow malicious instructions in the recorded data to influence the agent's behavior.
Ingestion points: Read .../trajectory.ndjson (SKILL.md)
Boundary markers: None present to distinguish recorded data from instructions.
Capability inventory: mcp__plugin_ruflo-core_ruflo__browser_eval, Bash, and various browser interaction tools (SKILL.md).
Sanitization: No sanitization or validation of the recorded trajectory content is mentioned.
[DYNAMIC_EXECUTION]: The skill performs dynamic execution by replaying recorded arguments through browser evaluation tools.
Evidence: Use of mcp__plugin_ruflo-core_ruflo__browser_eval with arguments loaded directly from the trajectory file.
[EXTERNAL_DOWNLOADS]: Fetches and executes CLI tools directly from the NPM registry at runtime.
Evidence: Usage of npx -y ruvector@0.2.25 and npx -y @claude-flow/cli@latest to perform status checks and memory searches.
[COMMAND_EXECUTION]: Executes shell commands to interact with session containers and search embedding namespaces.
Evidence: Use of the Bash tool to run rvf status and memory search commands.

browser-replay