feat: add multi-provider provider intelligence

README.md

@@ -5,7 +5,7 @@
 [![Release](https://github.com/sir-ad/knight/actions/workflows/release.yml/badge.svg)](https://github.com/sir-ad/knight/actions/workflows/release.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-0f172a.svg)](./LICENSE)
 
-Privacy-first Chrome extension for job applications. Knight parses your resume locally with Ollama, autofills common ATS portals, logs applications, syncs Gmail status updates, and drafts follow-ups without sending your profile to a hosted LLM.
+Privacy-first Chrome extension for job applications. Knight parses your resume locally or through your chosen provider, autofills common ATS portals, logs applications, syncs Gmail status updates, and drafts follow-ups without forcing your profile through a hosted AI workflow.
 
 ![Knight icon](./careerflow/assets/icon128.png)
 
@@ -14,11 +14,13 @@ Website: [sir-ad.github.io/knight](https://sir-ad.github.io/knight/)
 ## What It Does
 
 - Local-first resume parsing for `PDF`, `DOCX`, and `TXT`
+- AI provider support for `Ollama`, `OpenAI`, `Anthropic`, `Google Gemini`, and `OpenRouter`
+- Smart defaults that auto-discover installed Ollama models and recommend the best reachable provider/model
 - Autofill support for `Workday`, `Greenhouse`, `Lever`, `Naukri`, `iCIMS`, `SmartRecruiters`, `Taleo`, `SuccessFactors`, plus a generic fallback
 - In-page review overlay before filling
 - Local application log with status history
 - Gmail read-only sync for confirmations, interviews, rejections, and offers
-- Follow-up email draft generation with Ollama
+- Follow-up email draft generation with the active provider
 
 ## Repo Layout
 
@@ -40,15 +42,30 @@ Then load `careerflow/build/chrome-mv3-prod` in `chrome://extensions` with Devel
 ## Local Requirements
 
 - Node.js 20+
-- Ollama running locally
+- Ollama running locally for the privacy-first path
 - A pulled model such as `llama3.2:3b`
+- Optional cloud provider API key for `OpenAI`, `Anthropic`, `Google Gemini`, or `OpenRouter`
 - Optional Gmail OAuth setup using `PLASMO_PUBLIC_GOOGLE_CLIENT_ID`
 
 ```bash
-ollama serve
+OLLAMA_ORIGINS=chrome-extension://* ollama serve
 ollama pull llama3.2:3b
 ```
 
+Knight expects the Ollama endpoint as the host root only, for example `http://localhost:11434`. Do not add `/api` to the saved endpoint.
+
+## Supported Portals
+
+- [Workday](https://www.workday.com/) via `*.myworkdayjobs.com`
+- [Greenhouse](https://www.greenhouse.com/) via `boards.greenhouse.io` and `*.greenhouse.io`
+- [Lever](https://www.lever.co/) via `jobs.lever.co` and `*.lever.co`
+- [Naukri](https://www.naukri.com/) via `*.naukri.com`
+- [iCIMS](https://www.icims.com/) via `*.icims.com`
+- [SmartRecruiters](https://www.smartrecruiters.com/) via `*.smartrecruiters.com`
+- [Oracle Taleo](https://www.oracle.com/human-capital-management/taleo/) via `*.taleo.net`
+- [SAP SuccessFactors](https://www.sap.com/products/hcm.html) via `*successfactors*` and `*.jobs2web.com`
+- Generic fallback for employer-hosted forms that expose standard labels, placeholders, or aria metadata
+
 ## Validation
 
 From [`careerflow`](./careerflow):
@@ -68,7 +85,7 @@ npm run package
 
 ## Tags
 
-`chrome-extension` `job-search` `ats-autofill` `ollama` `gmail` `privacy-first` `plasmo` `typescript`
+`chrome-extension` `job-search` `ats-autofill` `ollama` `openai` `anthropic` `gemini` `openrouter` `gmail` `privacy-first` `plasmo` `typescript`
 
 ## Notes
 

careerflow/README.md

@@ -2,24 +2,26 @@
 
 Chrome MV3 extension for local-first job application automation.
 
-Knight is built with Plasmo and centered on one constraint: resume data should stay on the user’s machine. The extension parses resumes with Ollama, injects a review-first autofill overlay on common ATS sites, logs applications locally, reads Gmail with a read-only scope for status changes, and drafts follow-ups from local context.
+Knight is built with Plasmo and centered on one constraint: resume data should stay on the user’s machine whenever possible. The extension parses resumes with a configurable provider stack, injects a review-first autofill overlay on common ATS sites, logs applications locally, reads Gmail with a read-only scope for status changes, and drafts follow-ups from local context.
 
 ## Current Scope
 
 - Resume parsing for `PDF`, `DOCX`, and `TXT`
+- AI provider support for `Ollama`, `OpenAI`, `Anthropic`, `Google Gemini`, and `OpenRouter`
+- Smart defaults that auto-discover installed Ollama models and recommend the best reachable provider/model
 - Autofill support for:
-  - `Workday`
-  - `Greenhouse`
-  - `Lever`
-  - `Naukri`
-  - `iCIMS`
-  - `SmartRecruiters`
-  - `Taleo`
-  - `SuccessFactors`
+  - `Workday` via `*.myworkdayjobs.com`
+  - `Greenhouse` via `boards.greenhouse.io` and `*.greenhouse.io`
+  - `Lever` via `jobs.lever.co` and `*.lever.co`
+  - `Naukri` via `*.naukri.com`
+  - `iCIMS` via `*.icims.com`
+  - `SmartRecruiters` via `*.smartrecruiters.com`
+  - `Taleo` via `*.taleo.net`
+  - `SuccessFactors` / `Jobs2Web`
   - Generic fallback forms
 - Local application tracking with status history
 - Gmail read-only sync for confirmation, interview, rejection, and offer emails
-- Ollama-based follow-up drafting
+- Follow-up drafting with the active provider
 
 ## Stack
 
@@ -28,17 +30,23 @@ Knight is built with Plasmo and centered on one constraint: resume data should s
 - `React 18`
 - `Tailwind CSS`
 - `Ollama`
+- `OpenAI`
+- `Anthropic`
+- `Google Gemini`
+- `OpenRouter`
 - `Jest`
 
 ## Local Setup
 
 ```bash
 cd careerflow
 npm install
-ollama serve
+OLLAMA_ORIGINS=chrome-extension://* ollama serve
 ollama pull llama3.2:3b
 ```
 
+Knight expects the Ollama endpoint as the host root only, for example `http://localhost:11434`. Do not save `/api`, `/api/tags`, `/api/generate`, or `/api/chat` in Settings.
+
 Optional environment variable:
 
 ```bash
@@ -82,4 +90,5 @@ npm run package
 
 - The Gmail flow uses Chrome Identity with token-based OAuth and a read-only Gmail scope.
 - Local storage migration handles prior `knight_*` keys.
-- The extension defaults to the configured Ollama endpoint and model from Settings.
+- The extension keeps provider API keys in local Chrome storage and excludes them from export.
+- Smart defaults prefer a reachable local Ollama model, then fall back to the best configured cloud provider.

careerflow/package.json

@@ -1,7 +1,7 @@
 {
   "name": "knight-extension",
   "displayName": "Knight",
-  "version": "0.1.2",
+  "version": "0.2.0",
   "description": "Knight - Universal Job Application & Tracking System - Privacy-first browser extension",
   "license": "MIT",
   "author": "sir-ad",
@@ -68,9 +68,13 @@
     },
     "name": "Knight",
     "description": "Knight - Universal Job Application & Tracking System",
-    "version": "0.1.2",
+    "version": "0.2.0",
     "host_permissions": [
       "https://www.googleapis.com/*",
+      "https://api.openai.com/*",
+      "https://api.anthropic.com/*",
+      "https://generativelanguage.googleapis.com/*",
+      "https://openrouter.ai/*",
       "http://localhost:11434/*",
       "http://127.0.0.1:11434/*"
     ],

careerflow/src/background/service-worker.ts

@@ -3,18 +3,105 @@ import { gmailClient } from "./gmail-client"
 import { ApplicationTracker } from "../lib/application-tracker"
 import { classifyEmail } from "../lib/email-classifier"
 import * as db from "../lib/db"
-import { storageManager } from "../lib/storage-manager"
+import { storageManager, normalizeProviderSettings } from "../lib/storage-manager"
+import {
+  buildLLMConfig,
+  diagnoseProvider,
+  discoverProviderModels,
+  generateProviderChat,
+  generateProviderStructured,
+  generateProviderText,
+  providerSupportsApiKey,
+  recommendLLMConfiguration,
+} from "../lib/llm/provider-service"
+import { SUPPORTED_PORTALS } from "../lib/supported-portals"
 import type {
   ApplicationLogPayload,
   ApplicationStatus,
   ATSAdapterName,
   GmailStatus,
+  ProviderModelCatalog,
+  ProviderSecretStore,
+  ProviderSettings,
   RuntimeMessage,
   RuntimeResponse,
 } from "../lib/types"
+import type { LLMConfig } from "../lib/llm/types"
 
 const applicationTracker = new ApplicationTracker()
 
+async function resolveLLMRequest(payload?: {
+  config?: Partial<LLMConfig>
+  apiKey?: string
+}): Promise<{
+  config: LLMConfig
+  secrets: ProviderSecretStore
+  settings: Awaited<ReturnType<typeof storageManager.getSettings>>
+  providerSettings: ProviderSettings
+}> {
+  const [settings, storedSecrets] = await Promise.all([
+    storageManager.getSettings(),
+    storageManager.getProviderSecrets(),
+  ])
+
+  const providerSettings = normalizeProviderSettings({
+    ...(settings.llmConfig as ProviderSettings),
+    ...(payload?.config || {}),
+  })
+
+  const secrets = { ...storedSecrets }
+  if (providerSupportsApiKey(providerSettings.provider) && payload?.apiKey) {
+    secrets[providerSettings.provider] = payload.apiKey.trim()
+  }
+
+  return {
+    config: buildLLMConfig(providerSettings, secrets, payload?.config),
+    secrets,
+    settings,
+    providerSettings,
+  }
+}
+
+async function persistCatalogAndRecommendation(
+  catalog: ProviderModelCatalog,
+  settings: Awaited<ReturnType<typeof storageManager.getSettings>>,
+  secrets: ProviderSecretStore,
+  providerSettings: ProviderSettings
+) {
+  const nextCatalogs = {
+    ...settings.providerCatalogs,
+    [catalog.provider]: catalog,
+  }
+
+  const nextConfig =
+    settings.autoMode === "smart-defaults"
+      ? normalizeProviderSettings({
+          ...providerSettings,
+          model: catalog.recommendedModel,
+          ...(providerSettings.provider === "ollama" && catalog.endpoint
+            ? { endpoint: catalog.endpoint }
+            : {}),
+        })
+      : settings.llmConfig
+
+  const recommendation = recommendLLMConfiguration(
+    {
+      ...settings,
+      llmConfig: nextConfig,
+      providerCatalogs: nextCatalogs,
+    },
+    secrets
+  )
+
+  await storageManager.saveSettings({
+    llmConfig: nextConfig,
+    providerCatalogs: nextCatalogs,
+    lastRecommendation: recommendation,
+  })
+
+  return recommendation
+}
+
 function mapClassificationToStatus(classification: string): ApplicationStatus {
   switch (classification) {
     case "confirmation":
@@ -244,6 +331,71 @@ chrome.runtime.onMessage.addListener((message: RuntimeMessage, _sender, sendResp
           } satisfies RuntimeResponse)
           break
 
+        case "TEST_LLM_PROVIDER": {
+          const { config } = await resolveLLMRequest(message.payload)
+          sendResponse({
+            success: true,
+            data: await diagnoseProvider(config),
+          } satisfies RuntimeResponse)
+          break
+        }
+
+        case "DISCOVER_LLM_MODELS": {
+          const { config, secrets, settings, providerSettings } = await resolveLLMRequest(
+            message.payload
+          )
+          const catalog = await discoverProviderModels(config)
+          const recommendation = await persistCatalogAndRecommendation(
+            catalog,
+            settings,
+            secrets,
+            providerSettings
+          )
+
+          sendResponse({
+            success: true,
+            data: {
+              catalog,
+              recommendation,
+            },
+          } satisfies RuntimeResponse)
+          break
+        }
+
+        case "GENERATE_LLM_TEXT": {
+          const { config } = await resolveLLMRequest(message.payload)
+          sendResponse({
+            success: true,
+            data: await generateProviderText(config, message.payload.prompt),
+          } satisfies RuntimeResponse<string>)
+          break
+        }
+
+        case "GENERATE_LLM_STRUCTURED": {
+          const { config } = await resolveLLMRequest(message.payload)
+          sendResponse({
+            success: true,
+            data: await generateProviderStructured(config, message.payload.prompt),
+          } satisfies RuntimeResponse)
+          break
+        }
+
+        case "GENERATE_LLM_CHAT": {
+          const { config } = await resolveLLMRequest(message.payload)
+          sendResponse({
+            success: true,
+            data: await generateProviderChat(config, message.payload.messages),
+          } satisfies RuntimeResponse<string>)
+          break
+        }
+
+        case "GET_SUPPORTED_PORTALS":
+          sendResponse({
+            success: true,
+            data: SUPPORTED_PORTALS,
+          } satisfies RuntimeResponse)
+          break
+
         default:
           sendResponse({
             success: false,

careerflow/src/lib/email-classifier.ts

@@ -1,4 +1,5 @@
-import { ollamaClient } from "./ollama-client"
+import { storageManager } from "./storage-manager"
+import { buildLLMConfig, generateProviderStructured } from "./llm/provider-service"
 
 export type EmailClassification =
   | "confirmation"
@@ -216,7 +217,15 @@ export async function classifyEmail(emailContent: {
   body: string
 }): Promise<ClassificationResult> {
   try {
-    const response = await ollamaClient.generateStructured(
+    const [settings, secrets] = await Promise.all([
+      storageManager.getSettings(),
+      storageManager.getProviderSecrets(),
+    ])
+    const response = await generateProviderStructured(
+      buildLLMConfig(settings.llmConfig, secrets, {
+        temperature: 0.2,
+        maxTokens: 400,
+      }),
       CLASSIFICATION_PROMPT.replace("{subject}", emailContent.subject)
         .replace("{from}", emailContent.from)
         .replace("{body}", emailContent.body.substring(0, 4000))

careerflow/src/lib/followup-generator.ts

@@ -1,4 +1,4 @@
-import { getLLMClient } from "./llm"
+import { generateStructuredWithActiveProvider } from "./runtime-client"
 import type { ApplicationRecord, Profile } from "./types"
 
 export interface FollowUpEmail {
@@ -70,9 +70,7 @@ export async function generateFollowUpEmail(
   profile: Profile
 ): Promise<FollowUpEmail> {
   const prompt = buildPrompt(application, profile)
-  const client = getLLMClient()
-
-  const response = await client.generateStructured(prompt, {
+  const response = await generateStructuredWithActiveProvider<FollowUpEmail>(prompt, {
     temperature: 0.5,
     maxTokens: 500,
   })

careerflow/src/lib/llm/anthropic-provider.ts

@@ -126,6 +126,10 @@ export class AnthropicProvider implements LLMProviderInterface {
     }
   }
 
+  async discoverModels(): Promise<string[]> {
+    return this.models
+  }
+
   getModels(): string[] {
     return this.models
   }