The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection vulnerability surface. It is designed to ingest and render reasoning traces that contain untrusted data generated by AI agents.
Ingestion points: The AgentTrace class in SKILL.md and TraceTreeBuilder in references/reasoning-trace-visualization.md accept raw string content for reasoning steps.
Boundary markers: Delimiters or explicit warnings to ignore instructions within the traces are absent in the rendered output.
Capability inventory: The skill has the capability to send data to external observability backends via OpenTelemetry gRPC and HTTP exporters, and write traces to local files.
Sanitization: No sanitization is performed on the ingested content before it is printed to the terminal in references/reasoning-trace-visualization.md or displayed in the UI components in references/reasoning-trace-visualization.md (e.g., in the ascii_renderer.py and trace-viewer.ts snippets).
[CREDENTIALS_UNSAFE]: Detection of a hardcoded credential within the documentation.
Verbatim: GF_SECURITY_ADMIN_PASSWORD: "agent-observability" found in references/opentelemetry-agent-integration.md.
Evidence: The password is used in a docker-compose.observability.yaml example provided for setting up a local Grafana instance.
[OBFUSCATION]: Several reference files (e.g., architecture-patterns.md, code-organization.md, deployment-pipelines.md, error-handling.md, performance-optimization.md, security-best-practices.md, state-management.md, testing-strategies.md) contain 150 repetitive sections of text and code. While no specific malicious payload was identified within the repetition, this structure of extreme information density and redundancy is characteristic of adversarial attempts to overwhelm security reviewers and automated analysis tools.

agent-observability