@adia-ai/llm 0.8.4 → 0.8.5

package/CHANGELOG.md

@@ -1,9 +1,17 @@
 # Changelog
 
-## [Unreleased]
+## [0.8.5] — 2026-07-17
+
+### Added
+- **Gemini production passthrough** (`server.js`, operator ruling 2026-07-17): exe.dev's LLM gateway carries anthropic/openai/fireworks only, so an explicitly selected gemini model on `ui-kit.exe.xyz` had no server path. `server.js` gains a narrow gemini-only passthrough at `/api/llm/gemini/*` — forwards verbatim to `generativelanguage.googleapis.com` with the server-side `GEMINI_API_KEY` injected (the browser's sentinel header is discarded; path allowlisted, not an open proxy; SSE relayed with the same reader pattern as `handleChat`). `deploy/Caddyfile` + `infra/caddy/llm-gateway.snippet` route `/api/llm/gemini/*` to `server.js` ordered before the gateway handle.
+
+### Fixed
+- **Gemini passthrough hardening per review**: per-IP sliding-window rate limit (30 req / 5 min, unbounded-map guarded) bounds worst-case quota burn under the same no-auth public-demo trust model as the existing `/api/chat` smart proxy; 1 MB body cap returns 413 while accumulating, not after; path allowlist tightened to real Gemini API prefixes (`/v1`, `/v1beta` — `/^\/v\d/` previously accepted arbitrary versions); client disconnect now aborts the upstream call via `AbortController` hooked to `req` close before the body read, combined with the existing 120s timeout via `AbortSignal.any` (previously attached post-fetch, SSE branch only, so a vanished client kept work alive until timeout).
+
+## [0.8.4] — 2026-07-16
 
 ### Fixed
-- **The model id decides the provider** (`createAdapter({ model })`): provider resolution was env-key detection only (or hardcoded `anthropic` on production browser hosts), so every non-default model was POSTed to the wrong provider's API and 404'd — only Claude models could generate in a browser. Precedence is now `opts.provider` → the model id (new export `detectProviderFromModel` in `adapters/index`) → `LLM_PROVIDER` env → env-key detection.
+- **The model id decides the provider** (`src/llm-bridge.ts`, `src/adapters/index.ts` + built `llm-bridge.js`/`adapters/`): provider resolution was env-key detection only (or hardcoded `anthropic` on production browser hosts), so every non-default model was POSTed to the wrong provider's API and 404'd — only Claude models could generate in a browser. Precedence is now `opts.provider` → the model id (new export `detectProviderFromModel` in `adapters/index`) → `LLM_PROVIDER` env → env-key detection.
 - **Passthrough proxying preserves the upstream URL path + query**: `passthroughRequest` overwrote the whole URL with the static `proxyUrl`, dropping gemini's `models/<model>:<action>?alt=sse` routing — every gemini passthrough call hit the proxy root. The proxy prefix now joins `buildRequest()`'s own path/query; legacy full-path proxyUrls (`/api/llm/anthropic/v1/messages`) reduce to the identical final URL. `resolveBaseUrl` entries are provider prefixes (`gemini: /api/llm/gemini`; Vite gains the route, `google` kept as a legacy alias).
 - **`google` → `gemini` provider aliasing** (gh#268 L-1): `provider: 'google'` and `GOOGLE_API_KEY` auto-detection produced an always-throwing bridge (`Unknown provider "google"`); both canonicalize to the registry key `gemini` now. `GEMINI_API_KEY` is canonical.
 - **Gemini `stopReason` raw pass-through** (gh#268 L-2, invariant 1): `parseResponse`/`parseStream` forwarded a hardcoded `'end'`, hiding truncation from the downstream detector; `finishReason` (`MAX_TOKENS` / `SAFETY` / …) passes through raw, `'end'` only when absent.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@adia-ai/llm",
-  "version": "0.8.4",
+  "version": "0.8.5",
   "description": "Provider-agnostic LLM client \u2014 anthropic / openai / gemini adapters with a unified chat() + streamChat() facade. Used by AdiaUI's chat-shell and the A2UI generation pipeline; works in browser (with proxyUrl) and Node.",
   "type": "module",
   "exports": {