The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various local shell commands to manage project state and run tests. Evidence includes calls to git (status, branch, stash), test runners (pytest, pnpm test), and a specialized CLI tool called herdr used for log monitoring and session management.
[PROMPT_INJECTION]: The skill is exposed to indirect prompt injection risks because it processes untrusted data from the local repository, including source code, documentation, and application logs. These inputs influence the creation of 'user stories' and the 'Scope Contract,' which could theoretically be manipulated by an attacker who has modified the project's files.
[COMMAND_EXECUTION]: The agent uses the computer-use MCP and browser-based automation to perform runtime validation of UI goals. This involves high-privilege interaction with a browser environment, though the skill provides hygiene guidelines (session recording, auth-state management) to mitigate accidental exposure.
[SAFE_PRACTICE]: The skill includes explicit security guards, such as a prohibition on using sudo, instructions to halt if production credentials are required but not provided, and a 'never auto-merge' policy for generated pull requests. It also implements a peer-review protocol where one agent validates the work of the other before proceeding.

goal-loop