The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands to manage the git environment, analyze PRs, and run tests.
It resolves the repository's trunk branch and executes git diff to identify changes.
It utilizes the GitHub CLI (gh) to view PR information and check out specific branches or PRs.
It invokes browser automation via the agent-browser CLI to interact with the application.
[COMMAND_EXECUTION]: The skill automatically initiates project-specific development processes.
It identifies and starts the development server (e.g., bin/dev, npm run dev) without human intervention to facilitate testing.
It includes an autonomous 'Fix Loop' that can modify application code and commit changes back to the repository using git commit via the ce-commit skill.
[DATA_EXFILTRATION]: The skill reads high-value project documentation to establish testing context.
It accesses files such as STRATEGY.md, VISION.md, and PERSONAS.md to align test cases with product goals.
It retrieves PR metadata and diffs from GitHub, which may contain sensitive intellectual property.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data as part of its core logic.
Ingestion points: Processes PR diffs, branch content, and live web page content from the development server. These sources are identified in the file SKILL.md under Phase 1 and Phase 4.
Boundary markers: Lacks explicit delimiters or instructions to ignore embedded commands within the analyzed code or web content.
Capability inventory: Possesses significant capabilities including shell execution, file system writes, and the ability to commit code changes.
Sanitization: No evidence of sanitization or filtering of the content retrieved from diffs or browser interactions before it is processed by the model.

ce-dogfood