ADR-0065: A second provider, and hardening the provider contract

Status: Accepted; distribution amended by ADR-0066; keys made per profile by ADR-0138
Date: 2026-09-04
Deciders: jka
Decision: D65

Context

The agent had one real provider (Claude) besides the offline echo stub. A contract with a single implementation is unproven. Users also asked for ChatGPT, and ADR-0008 had always named a local-model provider.

Decision

A second provider was built in the tree, before any contract was frozen for outside authors:

It is OpenAI-compatible, not tied to ChatGPT. The hosted endpoint is the default, but the endpoint is configuration data, so the same code drives Ollama, llama-server, LM Studio or vLLM. It uses the OpenAI API with the user's own key, never the ChatGPT web interface.
It mirrors the Claude provider: a loader, an authentication face (Bearer token), and an inference module mapping rio's conversation to Chat Completions.

The second implementation forced three contract changes:

Per-provider keys. A single keyed-provider slot was replaced by per-provider key state; agent.key.set and agent.key.clear take a provider name.
Provider metadata. register_provider gained -label and -signup, and agent.providers returns the list, so the GUI builds its provider picker and key dialog from data.
A shared runtime library. The HTTPS streaming transport and the ASCII-safe JSON serialisers moved to plugins/lib/ (rio::llm::*), used by both providers.

Making providers installable from repositories was named as the next milestone.

Consequences

The contract was proven by two implementations before being versioned (ADR-0066).
A new provider appears in the GUI without GUI changes.
tcllib's JSON parser decodes null as the string "null"; the stream reader treats that string as absent.