Provider selection and data policy

Model and provider choices

Policy	Permitted behavior
no-model	Direct MCP tools and deterministic report inventory
metadata-only	Report model receives structural counts, tool/outcome metadata, IDs and hashes; candidate vulnerabilities are rejected
local-agent	Local/container/SSH CLI profile or loopback openai-compatible profile; hosted providers are rejected before context is sent
approved-provider	Configured provider or CLI profile, including OpenRouter, receives the bounded packet; external-client work also needs the recorded provider approval reference

Assessment profiles are local Codex, local Claude, container Codex, guarded OpenAI-compatible, or guarded OpenRouter. Daybreak Blue/Red are optional access programs and do not change scope, tool policy or evidence standards. API keys are required for OpenRouter and remote OpenAI-compatible endpoints; a loopback compatible endpoint may run without a key. Local Codex and Claude use their own configured credentials. A missing Daybreak grant cannot block an ordinary ready CLI or API profile. See guarded local and OpenRouter models.

Plan policy

A managed run requires:

• active scope and testing window;
• nonempty objectives, authorized methods, stop conditions, and exact allowedToolNames;
• known work ownership;
• modelDataPolicy: "local-agent" for self-hosted CLI processing or a loopback OpenAI-compatible endpoint, or "approved-provider" for an approved provider route such as OpenRouter;
• a client-work provider reference only when external client data uses approved-provider.

local-agent rejects hosted OpenAI, Anthropic, OpenRouter, and non-loopback compatible profiles before context is sent. no-model permits deterministic report generation. metadata-only permits a report model to see structural metadata and cannot start a tool-using assessment.