The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external data sources which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The skill ingests 'concrete anchors' such as logs, error messages, artifacts, and test documents as specified in SKILL.md.
Boundary markers: There are no instructions to use delimiters or ignore embedded instructions when processing these external data sources.
Capability inventory: The agent has access to shell command execution via host tools, test runners (Playwright, Cypress), and mobile automation adapters (argent, agent-device) as detailed in references/app-device.md and references/browser-react.md.
Sanitization: No sanitization or filtering of external content is mentioned before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill utilizes a template-based approach to generate and execute shell scripts for human-in-the-loop reproductions.
Evidence: SKILL.md instructs the agent to use scripts/hitl-loop.template.sh by copying it, editing the steps, and running it to reproduce issues.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the installation of several external third-party tools and packages.
Evidence: references/app-device.md suggests installing @swmansion/argent and agent-device via npm install -g. references/browser-react.md suggests installing playwright (Microsoft), cypress, and vitest.
Mitigation: The instructions explicitly state: 'Do not silently install tools. Suggest installation only when the missing adapter blocks the required evidence loop,' which preserves user oversight.
[COMMAND_EXECUTION]: The skill workflow relies heavily on executing shell commands to detect environment state, run tests, and capture diagnostic data.
Evidence: Usage of command -v, npm view, argent --version, and various automation commands like agent-device open are documented throughout the reference files.

diagnose