The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the tracing of LLM inputs and outputs, which inherently involves handling untrusted data from model responses.
Ingestion points: The skill intercepts OpenRouter request bodies and completion responses to extract metadata like session_id and cost (SKILL.md Step 2, 3).
Boundary markers: No explicit prompt boundary markers are provided in the code snippets.
Capability inventory: The skill modifies application code to inject headers/body fields and transmits trace data to external endpoints (ingest.maple.dev).
Sanitization: The instructions recommend enabling 'Privacy Mode' on the destination to prevent prompts and completions from leaving the user's controlled environment (SKILL.md Step 4).
[DYNAMIC_EXECUTION]: The skill provides logic to wrap the global fetch function or modify request bodies at runtime to inject tracing metadata.
The TypeScript openRouterFetch implementation dynamically parses and re-serializes JSON request bodies during the API call lifecycle (SKILL.md Step 3).
The Python openrouter_extra_body function dynamically constructs dictionary objects based on the active OpenTelemetry span context to parent traces correctly (SKILL.md Step 3).
[EXTERNAL_DOWNLOADS]: The skill references standard, well-known libraries and official service endpoints for tracing and LLM communication.
References official packages including @opentelemetry/api, openai, @openrouter/sdk, and the openrouter PyPI package.
Connects to official ingestion endpoints for the Maple service at ingest.maple.dev and ingest.eu.maple.dev.

maple-agent-tracing-openrouter