# LLM provider layer built on rig-core instead of hand-written clients

QuiCommit's LLM layer (commit/tag/changelog generation) moved from six hand-written HTTP clients to `rig-core` 0.41 (locked, `reqwest` + `rustls` features only, umbrella `rig` crate not used). Rationale: the hand-written layer accumulated duplicated SSE parsing, string-matching retry logic, inconsistent per-provider timeouts, and provider-specific bugs (Anthropic silently ignoring configured base_url, Ollama constructor panics); rig-core 0.41 natively supports all six providers and models reasoning/thinking as first-class typed content (streaming `Reasoning`/`ReasoningDelta` events), which matches QuiCommit's thinking-mode display feature. In exchange we accept rig's 0.x breaking-change velocity (upgrades require manual provider-diff review), no application-level retry on non-streaming requests, and the Ollama endpoint moving to `/api/chat`.

**Considered Options**: keep hand-writing (rejected — every new provider meant a full HTTP/SSE/retry suite; six near-duplicate implementations already); per-provider SDKs like async-openai plus hand-written rest (rejected — multiple API styles, no unified interface); rig-core (accepted — one dependency covers every current provider and normalizes reasoning content across them).

**Consequences**: user-visible behavior changes recorded in CHANGELOG for 0.6.0 (Ollama `/api/chat`, no non-streaming retry, credential-verify endpoint changes, Anthropic base_url now honored); rig-core upgrades must re-verify provider API changes before bumping the locked version.