v3.1.0 — All major LLM providers (OpenAI, Anthropic, Gemini + custom)

v3.1.0 — all major LLM providers (OpenAI, Anthropic, Gemini + custom)

Slower now works with every major LLM provider through Prism, with fewer config variables than before — provider credentials move out of Slower's config entirely.

Highlights

• Pick a provider with one variable: SLOWER_AI_SERVICE=openai|anthropic|gemini|ollama|…. Any Prism provider works; a fully custom backend registers via AiServiceManager::extend().
• Credentials delegated to Prism (config/prism.php / OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY) — Slower's own open_ai config block is gone.
• Sensible low-cost default model per provider (gpt-5.4-mini, claude-haiku-4-5, gemini-2.5-flash), overridable with SLOWER_AI_RECOMMENDATION_MODEL.

Design

• Replaced openai-php/laravel with prism-php/prism — one official package for all providers. The entire integration lives behind a single PrismDriver; the AiServiceDriver contract and AiServiceManager::extend() seam are unchanged, so Prism (and its pre-1.0 API) is isolated to one class.
• AiServiceManager maps ai_service → any Prism provider (Provider::tryFrom), with extend() taking precedence and unknown names throwing a helpful error.
• Provider/transport exceptions propagate (a null return means only "no usable text"), so failed analyses stay retryable exactly as before.

Backward compatibility

Existing OpenAI users need no changes — Prism reads OPENAI_API_KEY, and a boot-time bridge still honors a legacy slower.open_ai.api_key. Minor release.

Verification

• 126 tests / 268 assertions green on Laravel 11, 12, and 13 (PHP 8.4); PHP 8.5 covered by CI. PHPStan 2 level 5 clean, Pint clean.
• PrismDriver tested with Prism::fake() (provider/model/prompt routing, empty→null); manager resolution, custom extend (via the real singleton path), the create{Name}Driver() BC path, and config all covered.
• Architecture selected by a multi-model council (Prism won decisively; per-provider SDKs rejected) plus Fable 5's independent opinion.
• Hardened through four rounds of recursive review, which caught and fixed real bugs a single pass would have missed: a critical extend()-singleton failure (custom LLMs silently ignored), wrong per-provider model defaults (OpenAI model sent to other providers), a dropped create{Name}Driver() extension convention, and misleading error messaging. Two residual items were reasoned design choices, not defects: key validation is deferred to call-time (an eager guard would break keyless providers like Ollama), and the 30s request timeout is a sane default (raise PRISM_REQUEST_TIMEOUT if needed).

Heads-up: GitHub Actions is currently failing to provision runners account-wide (unrelated to this change), so CI hasn't validated the matrix here — it was validated locally across Laravel 11–13.