google-lightspeed-agent

Guides multi-step tool chaining and common filter parameters for Inventory, Vulnerability, and Advisor tools. Enables the agent to combine information from multiple tools to answer complex queries instead of giving up after a single call. [PREFERRED]

Answers total/count/"how many?" queries for any resource — CVEs, hosts, advisory rules, blueprints, subscriptions, or content sources — with a single API call using limit=1 and response metadata instead of fetching all pages. [PREFERRED]

Teaches the agent how to handle paginated API responses correctly. Covers default fetch behavior, pagination metadata interpretation, stop conditions, and per-API pagination conventions for Vulnerability, Inventory, Advisor, and other tool categories. [PREFERRED]

Extends the base guardrails with detailed examples, edge cases, and severity-interpretation rules. Covers scope edge cases, prompt-injection patterns to recognize, data-integrity examples, and CVE severity context (advisor vs. vulnerability). [STRICT]

Extends the base tool invocation format with examples of correct vs. incorrect invocations, argument formatting guidance, schema lookup procedures, and edge cases for unknown tools. [STRICT]

Handles oversized tool results and tool errors gracefully. Teaches the agent to retry with narrower queries on oversized responses and to interpret HTTP status codes correctly instead of silently swallowing failures. [PREFERRED]

Defines the agent's capabilities reference and response style including output formatting rules for CVE lists, host inventories, advisor recommendations, and mixed results. [GUIDANCE]