The dependency doctrine: adopt, wrap, or write

Every dependency in Aegis carries one of three designations, recorded in the project spec. Nothing enters the tree without one.

Adopt

Use the library's public interface directly as our contract. Justified only when the interface is itself a standard whose ecosystem we want plugin authors to inherit: LangGraph's graph and checkpointer, the official MCP SDK, OpenTelemetry, FastAPI, pydantic. Adopting means programming against that ecosystem on purpose — a feature, not a leak.

Wrap

Use it behind an Aegis Protocol, imported in exactly one module, version pinned. For anything that moves fast or that we may swap: LiteLLM (the default ModelProvider backend), LLM Guard (the default guardrail adapter), Presidio (the PII pack). A wrapped dependency breaking is a one-file fix; nothing else in the codebase knows it exists.

Write

Only where no good wheel exists and the surface is small: the entry-point registry glue, the graph assembler, the verdict spine, secret:// resolution, key hashing. If a "write" item grows past thin glue, that is a design smell to revisit.

Why this is enforced, not aspirational

The import-linter contract in CI fails the build if a wrapped library is imported outside its adapter module, or if a policy pack imports anything but public aegis APIs. Doctrine that is not executable is opinion.

One consequence worth naming: tiered extras. The slim install (pip install aegis-gateway) carries the kernel, server, and CLI only. Heavy dependencies (spaCy models, transformers) arrive only with the pack that needs them ([pii], [llm-guard], [rag]). A framework whose install pulls torch before the first request would contradict everything above.