n1r.core — what the runtime owns
n1r.core — what the runtime owns
Section titled “n1r.core — what the runtime owns”
n1r.coreis the N1-owned runtime surface. It hides substrate details and owns the cross-cutting mechanics every Nexus product needs.
n1r.core is the foundation that takes an authored pack and turns it into a bounded, observable, scoped runtime execution. It exists so product packs do not need to own auth, tokens, model quirks, telemetry policy, or trusted-fact generation.
What n1r.core owns
Section titled “What n1r.core owns”| Responsibility | What that means |
|---|---|
| Pack loading and validation | Reads the authored files, checks them, and refuses invalid packs before runtime |
| OAuth2 token lifecycle | Acquires the token for the run or session, refreshes it when needed, and hands it to Pinax |
| Model resolution | Resolves model ids and chooses provider-native routing where required |
| LLM attribution and observability | Emits PHI-safe model telemetry with run, pack, subject, and model attribution |
| Hard run bounds | Enforces a real wall-clock timeout and cancellation path |
| Artifact emission | Returns structured clinical reasoning artifacts and lineage |
| Clinician-authored instrumentation | Applies the same runtime instrumentation to database-backed clinician definitions |
The important auth boundary
Section titled “The important auth boundary”n1r.core owns token acquisition and refresh. Pinax receives the token and runtime context it is handed. Pinax then uses that token when calling api-client. api-backend remains the policy authority.
That boundary keeps token lifecycle in one runtime layer and keeps capability code separate from identity logic.
The important model boundary
Section titled “The important model boundary”n1r.core resolves models by registry id, not by raw provider URL. It can use provider-native routes where needed. Gemini reasoning paths are the clearest case: the proven pattern is native Gemini through the LiteLLM gateway root when required, not flattening everything through one OpenAI-compatible path.
The important runtime boundary
Section titled “The important runtime boundary”Tool-call limits are not enough. n1r.core must enforce hard wall-clock bounds around async runs. A cooperative stop signal is not a real production limit.
The important authoring boundary
Section titled “The important authoring boundary”n1r.core must treat file-authored packs and active database-backed definitions as one runtime class after normalization. Clinician-authored agents do not get a weaker runtime path. They receive the same token, observability, timeout, artifact, and lineage behavior automatically.