🎯 fix: Infer Agents Endpoint for Model Specs Naming an Agent

[danny-avila]

Problem

A model spec whose preset names an agent_id but omits endpoint is completely unusable — selecting it fails every time with:

Something went wrong. Here's the specific error message we encountered: Model spec mismatch

The match is strict equality:

export function isModelSpecEndpointMatch(modelSpec, endpoint): boolean {
  return Boolean(modelSpec && endpoint === modelSpec.preset?.endpoint);
}

With preset.endpoint undefined, that is false for every endpoint. There is no code path where such a spec can match, so it is dead on arrival — and the error names neither the spec nor the missing field, which makes it hard to act on.

This is easy to hit when a spec is authored through a UI rather than hand-written YAML, since only the agent_id is strictly meaningful to the author.

The selector had the same gap

Fixing only the server would not have been enough. handleSelectSpec read preset.endpoint directly:

if (isAgentsEndpoint(spec.preset.endpoint)) {
  model = spec.preset.agent_id ?? '';
}
setSelectedValues({ endpoint: spec.preset.endpoint, model, modelSpec: spec.name });

With the endpoint undefined, the client sends no endpoint and skips assigning agent_id to model — so the request is malformed before it reaches the check. Both halves need the same resolution, which is why this lives in librechat-data-provider rather than in either consumer.

Changes

• Add resolveModelSpecEndpoint to librechat-data-provider: returns preset.endpoint when set, otherwise infers agents when the preset names an agent_id. Only the agents endpoint can serve such a preset, so the inference is unambiguous.
• Use it for endpoint matching (packages/api/src/modelSpecs) and in the selector, so the menu and the request pipeline resolve a spec identically.

An explicit endpoint always takes precedence, so existing configured specs are unaffected. Nothing is inferred for presets without an agent.

Scoped to agent_id only — assistant_id is intentionally left alone, as assistants are being removed.

Testing

packages/api and client typecheck clean. modelSpecs suite 11 → 14 tests, and the existing buildEndpointOption middleware suite (13 tests) still passes. New cases cover:

• preset with agent_id, no endpoint → matches agents, not other endpoints
• explicit endpoint alongside agent_id → explicit wins
• preset with neither → infers nothing