refactor: finish renaming stuff for ai-provider now that ai-provider is in ai

Author: sroussey

.claude/CLAUDE.md

@@ -41,9 +41,9 @@ task-graph                            (core DAG pipeline engine)
     ↓
 dataset, tasks                        (KnowledgeBase, documents, chunks; utility tasks)
     ↓
-ai                                    (AI task base classes, model registry)
+ai                                    (AI task base classes, model registry, provider helpers)
     ↓
-ai-provider                           (concrete provider implementations)
+providers/*                           (concrete provider implementations: anthropic, openai, gemini, ollama, ...)
     ↓
 test                                  (integration tests across all packages)
 workglow                              (meta-package re-exporting everything)
@@ -61,7 +61,7 @@ Each package builds three runtime targets via `bun build --target=X`:
 
 Types built with `tsc` (composite + incremental). Conditional exports in `package.json` resolve automatically per runtime.
 
-Exception: `ai-provider` builds per-provider sub-paths (`./anthropic`, `./openai`, `./google-gemini`, etc.) instead of browser/node/bun.
+Exception: vendor packages under `providers/*` (e.g. `@workglow/anthropic`, `@workglow/openai`, `@workglow/google-gemini`) ship `./ai` and `./ai-runtime` sub-paths instead of browser/node/bun.
 
 Exception: `util` has multiple named exports beyond `"."`:
 
@@ -178,18 +178,21 @@ Task categories: text generation/embedding/summary/translation/rewriting/classif
 
 RAG tasks: `ChunkVectorUpsertTask` (input: `knowledgeBase` + `chunks` + `vector`, optional `doc_title`), `ChunkRetrievalTask` (input: `knowledgeBase` + `query` + `model`, with `method: "similarity" | "hybrid"`), `HierarchyJoinTask`, `RerankerTask`, `QueryExpanderTask`, `TextChunkerTask`, `HierarchicalChunkerTask`.
 
-### `@workglow/ai-provider` — provider implementations
+### `providers/*` — provider implementations
 
-Each provider is a separate sub-export with optional peer dependencies:
+Each provider is a standalone package with optional third-party peer dependencies. They each expose `./ai` (main-thread shell) and `./ai-runtime` (worker / inline runtime):
 
-- `@workglow/ai-provider/anthropic` — Claude
-- `@workglow/ai-provider/openai` — OpenAI
-- `@workglow/ai-provider/gemini` — Gemini
-- `@workglow/ai-provider/ollama` — Ollama (browser + node)
-- `@workglow/ai-provider/hf-transformers` — HuggingFace Transformers.js
-- `@workglow/ai-provider/hf-inference` — HuggingFace Inference API
-- `@workglow/ai-provider/llamacpp` — node-llama-cpp
-- `@workglow/ai-provider/tf-mediapipe` — TensorFlow MediaPipe (browser)
+- `@workglow/anthropic` — Claude
+- `@workglow/openai` — OpenAI
+- `@workglow/google-gemini` — Gemini
+- `@workglow/ollama` — Ollama (browser + node)
+- `@workglow/huggingface-transformers` — HuggingFace Transformers.js
+- `@workglow/huggingface-inference` — HuggingFace Inference API
+- `@workglow/node-llama-cpp` — node-llama-cpp
+- `@workglow/tf-mediapipe` — TensorFlow MediaPipe (browser)
+- `@workglow/chrome-ai` — built-in Chrome / WebBrowser AI
+
+Shared cloud-provider helpers (base classes, registration, model search, OpenAI-shape chat, image-output conversion, tool-call parsing) live in `@workglow/ai/provider-utils` and are imported by every vendor package.
 
 **Important: `*_JobRunFns.ts` files execute inside workers.** Workers have an isolated runtime with a separate `globalServiceRegistry`. Do not access main-thread-only state (e.g., credential stores, service registries) from run functions. Instead, resolve such state in the task class on the main thread (e.g., `AiTask.getJobInput()`) and pass the resolved values through the serialized job input.
 

.gitignore

@@ -6,7 +6,7 @@ dist
 *.d.ts
 *.d.ts.map
 !examples/web/src/vite-env.d.ts
-!packages/ai-provider/src/provider-chrome/common/WebBrowser_ChromeAI.d.ts
+!providers/chrome-ai/src/ai/common/WebBrowser_ChromeAI.d.ts
 !packages/tasks/src/util/interpreter.d.ts
 storybook-static/
 packages/cli-web/

README.md

@@ -64,7 +64,7 @@ Packages:
 - **[packages/job-queue](packages/job-queue/README.md)**
 - **[packages/tasks](packages/tasks/README.md)**
 - **[packages/ai](packages/ai/README.md)**
-- **[packages/ai-provider](packages/ai-provider/README.md)**
+- **[providers/*](providers/) — vendor provider packages (Anthropic, OpenAI, Gemini, Ollama, HuggingFace, llama.cpp, MediaPipe, Chrome AI, Playwright, Electron, SQLite, Postgres, Supabase, bun-webview)**
 - **[packages/util](packages/util/README.md)**
 - **[packages/test](packages/test/README.md)**
 

bun.lock

@@ -17,7 +17,7 @@
   - [`packages/job-queue`](#packagesjob-queue)
   - [`packages/task-graph`](#packagestask-graph)
   - [`packages/ai`](#packagesai)
-  - [`packages/ai-provider`](#packagesai-provider)
+  - [`providers/*`](#providers)
   - [`packages/util`](#packagesutil)
   - [`examples/cli`](#examplescli)
   - [`examples/web`](#examplesweb)
@@ -559,9 +559,9 @@ This is the main task handling library, with tasks, compound tasks, data flows,
 
 These are the LLM tasks, models, etc. These tasks are agnostic to the provider and thus are like abstract versions. AI providers contribute the concrete implementations. Which implementation is used is determined by the model repository.
 
-### `packages/ai-provider`
+### `providers/*`
 
-This is the Huggingface Transformers JS (using ONNX) and TensorFlow MediaPipe providers.
+Each AI provider lives in its own package under `providers/`: `@workglow/anthropic`, `@workglow/openai`, `@workglow/google-gemini`, `@workglow/ollama`, `@workglow/huggingface-transformers`, `@workglow/huggingface-inference`, `@workglow/node-llama-cpp`, `@workglow/tf-mediapipe`, and `@workglow/chrome-ai`. Each ships an `./ai` and `./ai-runtime` subpath. Shared helpers (cloud client base classes, registration utilities, OpenAI-shape chat plumbing) live in `@workglow/ai/provider-utils`. Non-AI vendor packages (Playwright, Electron, SQLite, Postgres, Supabase, bun-webview) also live under `providers/`.
 
 ### `packages/util`
 

docs/developers/01_getting_started.md

@@ -382,27 +382,29 @@ yields a single `finish` event so GPU serialization is still respected.
 
 ---
 
-## Sub-Path Exports
+## Vendor Provider Packages
 
-Unlike other packages that build per-runtime targets (`browser.ts`, `node.ts`, `bun.ts`), the
-`@workglow/ai-provider` package builds per-provider sub-paths. Each provider is a separate import
-with optional peer dependencies:
+Unlike core packages that build per-runtime targets (`browser.ts`, `node.ts`, `bun.ts`), each
+provider lives in its own package under `providers/*` with optional peer dependencies on the
+underlying vendor SDK. Providers import shared base classes and helpers from
+`@workglow/ai/provider-utils`.
 
 ```typescript
-import "@workglow/ai-provider/anthropic"; // Claude (requires @anthropic-ai/sdk)
-import "@workglow/ai-provider/openai"; // OpenAI (requires openai)
-import "@workglow/ai-provider/gemini"; // Google Gemini (requires @google/generative-ai)
-import "@workglow/ai-provider/ollama"; // Ollama (requires ollama)
-import "@workglow/ai-provider/hf-transformers"; // HuggingFace Transformers.js
-import "@workglow/ai-provider/hf-inference"; // HuggingFace Inference API
-import "@workglow/ai-provider/llamacpp"; // node-llama-cpp
-import "@workglow/ai-provider/tf-mediapipe"; // TensorFlow MediaPipe (browser)
-import "@workglow/ai-provider/chrome"; // Chrome Built-in AI
+import "@workglow/anthropic/ai";              // Claude (requires @anthropic-ai/sdk)
+import "@workglow/openai/ai";                 // OpenAI (requires openai)
+import "@workglow/google-gemini/ai";          // Google Gemini (requires @google/generative-ai)
+import "@workglow/ollama/ai";                 // Ollama (requires ollama)
+import "@workglow/huggingface-transformers/ai"; // HuggingFace Transformers.js
+import "@workglow/huggingface-inference/ai";  // HuggingFace Inference API
+import "@workglow/node-llama-cpp/ai";         // node-llama-cpp
+import "@workglow/tf-mediapipe/ai";           // TensorFlow MediaPipe (browser)
+import "@workglow/chrome-ai/ai";              // Chrome Built-in AI
 ```
 
-Each sub-path also has a `/runtime` variant (e.g., `@workglow/ai-provider/anthropic/runtime`) that
-exports the heavy run function implementations and worker registration helpers. The main sub-path
-exports only the lightweight provider class, constants, and the worker-backed registration function.
+Each vendor package also exposes an `/ai-runtime` entry (e.g.
+`@workglow/anthropic/ai-runtime`) that exports the heavy run function implementations
+and worker registration helpers. The `/ai` entry exports only the lightweight provider
+class, constants, and the worker-backed registration function.
 
 Some providers (Ollama, OpenAI) also have browser-specific conditional exports in `package.json`.
 

docs/technical/11-ai-provider-system.md

@@ -383,8 +383,9 @@ important implications:
    keeping the worker bundle small.
 
 4. **Provider run functions execute in workers.** Files named `*_JobRunFns.ts`
-   in the `ai-provider` package contain the actual inference logic that runs
-   inside workers. These functions must not import main-thread-only modules.
+   in each vendor provider package (under `providers/*`) contain the actual
+   inference logic that runs inside workers. These functions must not import
+   main-thread-only modules.
 
 ## Provider Integration
 

docs/technical/13-worker-system.md

@@ -376,8 +376,8 @@ the `onExpiry` callback is invoked -- typically used to lock a
 - The Web Crypto API (`crypto.subtle`) is available natively -- no polyfills needed.
 - Provider clients (Anthropic, OpenAI) set `dangerouslyAllowBrowser: true`
   when they detect a browser or web worker context (see
-  `packages/ai-provider/src/provider-anthropic/common/Anthropic_Client.ts:48`
-  and `packages/ai-provider/src/provider-openai/common/OpenAI_Client.ts:49`).
+  `providers/anthropic/src/ai/common/Anthropic_Client.ts`
+  and `providers/openai/src/ai/common/OpenAI_Client.ts`).
   This is **not** automatically safe: any API key loaded into a browser
   context — even one decrypted from `EncryptedKvCredentialStore` at runtime —
   is reachable from the page and should be treated as exposed to the end user.

docs/technical/15-credential-management.md

@@ -171,18 +171,18 @@ Here, `build-js` itself expands into many concurrent builds (browser, node, bun,
 graph, media, compress), and `build-types` runs after all JS builds complete. The sequential
 ordering (`&&`) ensures that type generation can reference the built artifacts.
 
-**Provider package (`@workglow/ai-provider`):**
+**Vendor provider packages (`providers/*`):**
 
 ```json
 {
-  "build-package": "concurrently -c 'auto' -n 'code,browser,types' 'bun run build-code' 'bun run build-browser' 'bun run build-types'"
+  "build-package": "concurrently -c 'auto' -n 'code,types' 'bun run build-code' 'bun run build-types'"
 }
 ```
 
-The ai-provider package does not follow the standard browser/node/bun entry point pattern.
-Instead, it builds per-provider sub-paths (anthropic, openai, gemini, etc.) with `--root ./src`
-to preserve directory structure, plus a separate browser-specific build for providers that support
-browser environments (ollama, openai, tf-mediapipe).
+Each vendor package (`@workglow/anthropic`, `@workglow/openai`, `@workglow/google-gemini`, ...)
+ships two entry points — `./ai` (main thread) and `./ai-runtime` (worker /
+inline runtime) — built with `--root ./src` to preserve directory structure. Browser-capable
+providers (ollama, openai) add `*.browser.ts` variants with `--target=browser`.
 
 ### `build-js` — JavaScript Only
 
@@ -268,18 +268,16 @@ Platform-agnostic sub-paths (schema, graph) are built with `--target=browser` si
 code runs everywhere. Platform-specific sub-paths (media, compress, worker) are built with the
 matching target for each variant.
 
-### Provider Pattern (ai-provider)
+### Provider Pattern (vendor packages under `providers/*`)
 
-`@workglow/ai-provider` uses `--root ./src` to build multiple entry points in a single invocation
-while preserving their subdirectory structure in the output:
+Each vendor provider package (e.g. `@workglow/anthropic`, `@workglow/openai`) uses `--root ./src`
+to build the `ai` and `ai-runtime` entry points in a single invocation while
+preserving their subdirectory structure in the output:
 
 ```bash
 bun build --sourcemap=external --packages=external --root ./src --outdir ./dist \
-  ./src/provider-anthropic/index.ts \
-  ./src/provider-anthropic/runtime.ts \
-  ./src/provider-gemini/index.ts \
-  ./src/provider-gemini/runtime.ts \
-  # ... more providers
+  ./src/ai.ts \
+  ./src/ai-runtime.ts
 ```
 
 This produces output like:

docs/technical/19-build-system.md

@@ -1,25 +1,25 @@
 import tsPlugin from "@typescript-eslint/eslint-plugin";
 import tsParser from "@typescript-eslint/parser";
-import globals from "globals";
+import jsxA11yPlugin from "eslint-plugin-jsx-a11y";
 import reactPlugin from "eslint-plugin-react";
 import reactHooksPlugin from "eslint-plugin-react-hooks";
-import jsxA11yPlugin from "eslint-plugin-jsx-a11y";
 import regexpPlugin from "eslint-plugin-regexp";
+import globals from "globals";
 
 export default [
   {
     ignores: ["dist/**", "**/storybook-static/**"],
   },
   {
-    files: ["{packages|examples}/**/*.{js,ts,tsx,mts,cts}"],
+    files: ["{packages|providers|examples}/**/*.{js,ts,tsx,mts,cts}"],
     languageOptions: {
       parser: tsParser,
       parserOptions: {
         ecmaVersion: "latest",
         sourceType: "module",
         ecmaFeatures: {
-          jsx: true
-        }
+          jsx: true,
+        },
       },
       globals: {
         ...globals.browser,
@@ -28,10 +28,10 @@ export default [
     },
     plugins: {
       "@typescript-eslint": tsPlugin,
-      "react": reactPlugin,
+      react: reactPlugin,
       "react-hooks": reactHooksPlugin,
       "jsx-a11y": jsxA11yPlugin,
-      "regexp": regexpPlugin,
+      regexp: regexpPlugin,
     },
     rules: {
       ...tsPlugin.configs["recommended"].rules,