The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill documents a workaround for secret resolution issues by suggesting that users inline authorization headers (e.g., Bearer tokens or Basic auth) directly into the workflow definitions (Section 5, Gotcha 3). This practice is highly discouraged as workflow definitions are often shared, logged, or stored in version control systems, potentially exposing sensitive credentials.
[PROMPT_INJECTION]: The skill describes a durable workflow pattern (Section 6) that fetches data from external, untrusted sources (CoinGecko and Luno APIs) and interpolates the raw responses directly into an LLM prompt.
Ingestion points: Data is ingested via HTTP GET tasks (fetch_coingecko and fetch_luno) in the workflow definition.
Boundary markers: The provided prompt template lacks boundary markers, delimiters, or explicit instructions to the agent to ignore any embedded commands within the fetched data.
Capability inventory: The workflow engine has the capability to perform network operations, execute HTTP tasks, and trigger further agent runs.
Sanitization: There is no mention of sanitizing, escaping, or validating the external content before it is processed by the LLM, creating a surface for indirect prompt injection.
[COMMAND_EXECUTION]: The skill requires the execution of local shell commands for server setup and management, including starting the AgentSpan server process (agentspan server start) and performing health checks using curl.

rterm-agentspan