Module discovery 

Optional model discovery (capability-negotiated). RFC 0018 §5.4.

agentd may learn what an endpoint serves via a tiny handshake — off unless an endpoint looks discovery-capable, silent on failure, never fatal, never on the hot path, never at startup before a side effect (RFC 0018 decision 5).

The probe is one hand-rolled HTTP GET /v1/models over the EXISTING intel transport (super::endpoints::Endpoint::discover_models) — no new client, no streaming, zero new deps. For an OpenAI-compatible endpoint it parses { "data": [ { "id": "…" } ] }; the anthropic dialect has no list endpoint → it contributes nothing (the configured model is dialed regardless).

The SURFACE that consumes this — agentd://intelligence + the capabilities manifest intelligence.models — is supervisor-side (RFC 0018 §4.4 / RFC 0015 §capabilities). The architecture adaptation of the RFC’s “lazy on first complete_once” (a child-side hook) is to probe supervisor-side, lazily + cached, on read of the served surface: those reads are infrequent and operator-driven, so a cached probe there is the right seam, and it keeps the additive discovery field off the control protocol. The cache + TTL live in [crate::mcp::server::ServeCtx]; this module is the pure probe.

Structs

DiscoveryResult

The discovery outcome for the served surface (RFC 0018 §5.4):

Constants

DEFAULT_TIMEOUT

The default per-probe timeout (RFC 0018 §5.4 — “a SHORT timeout, it’s best-effort”). Discovery is off the hot path; a slow/wedged endpoint must not stall the operator-driven read that triggered the probe.

Functions

discover

Probe every OpenAI-compatible endpoint in list for its served models and fold the result into a DiscoveryResult (RFC 0018 §5.4). model is the configured model id, always unioned in (it is usable regardless of whether any endpoint answered discovery). Best-effort + silent: a per-endpoint probe that fails (404 / connection / non-JSON) simply contributes no models and does not flip discovery — it is NEVER fatal, NEVER a failover-class error (§5.4).