Implement hybrid top-N analyze_papes

Author: lhy0718

.env.example

@@ -1,2 +1,4 @@
 # Optional: AutoResearch setup can write this for you on first run.
 SEMANTIC_SCHOLAR_API_KEY=
+# Required if the primary provider is OpenAI API or PDF analysis mode is set to Responses API.
+OPENAI_API_KEY=

README.ko.md

@@ -13,11 +13,13 @@ npm link
 autoresearch
 ```
 
-선택 사항: Semantic Scholar API key를 `.env`에 넣을 수 있습니다.
+필수: Semantic Scholar API key를 `.env`에 넣어야 합니다. (또는 첫 실행 setup wizard에서 입력)
+선택: 기본 provider를 `OpenAI API`로 선택하거나 PDF 분석을 `Responses API`로 설정할 경우 `OPENAI_API_KEY`도 필요합니다.
 
 ```bash
 cp .env.example .env
 echo 'SEMANTIC_SCHOLAR_API_KEY=your_key_here' >> .env
+echo 'OPENAI_API_KEY=your_openai_key_here' >> .env
 ```
 
 개발 모드:
@@ -33,8 +35,21 @@ npm run dev
 1. 빈 프로젝트에서 `autoresearch`를 실행합니다.
 2. `.autoresearch/config.yaml`이 없으면 setup wizard가 자동 시작됩니다.
 3. wizard가 설정/스캐폴드를 만든 뒤 대시보드로 진입합니다.
-4. setup wizard는 선택적으로 Semantic Scholar API key를 묻고, 입력한 값은 `.env`에 기록합니다.
-5. 실행 시 AutoResearch는 `process.env` 또는 `.env`의 `SEMANTIC_SCHOLAR_API_KEY`를 읽습니다.
+4. setup wizard는 Semantic Scholar API key를 필수로 묻고, 입력한 값은 `.env`에 기록합니다.
+5. setup wizard에서 기본 LLM provider를 선택합니다.
+   - `codex`: 메인 워크플로를 Codex ChatGPT 로그인으로 실행
+   - `api`: 메인 워크플로를 OpenAI API 모델로 실행 (`OPENAI_API_KEY` 필요)
+6. setup wizard에서 PDF 분석 모드도 선택합니다.
+   - `codex`: PDF를 로컬에서 텍스트 추출 후 Codex로 분석
+   - `api`: PDF를 Responses API로 직접 전달해 분석 (`OPENAI_API_KEY` 필요)
+7. 기본 provider로 `api`를 선택하면 setup wizard와 `/settings`에서 OpenAI API 모델도 선택할 수 있습니다.
+   - 현재 카탈로그: `gpt-5.4`, `gpt-5`, `gpt-5-mini`, `gpt-4.1`, `gpt-4o`, `gpt-4o-mini`
+8. PDF 분석에서 `api`를 선택하면 setup wizard와 `/settings`에서 Responses API PDF 모델도 선택할 수 있습니다.
+   - 현재 카탈로그: `gpt-5.4`, `gpt-5`, `gpt-5-mini`, `gpt-4.1`, `gpt-4o`, `gpt-4o-mini`
+9. `/model`은 현재 기본 provider를 따릅니다.
+   - Codex provider: Codex 모델 선택기
+   - OpenAI API provider: OpenAI API 모델 선택기
+10. 실행 시 AutoResearch는 `process.env` 또는 `.env`의 `SEMANTIC_SCHOLAR_API_KEY`, `OPENAI_API_KEY`를 읽습니다.
 
 ## CLI 정책
 

README.md

@@ -13,11 +13,13 @@ npm link
 autoresearch
 ```
 
-Optional: set your Semantic Scholar API key in `.env`.
+Required: set your Semantic Scholar API key in `.env` (or enter it during first-run setup).
+Optional: if you choose `OpenAI API` as the primary provider or `Responses API PDF analysis`, also set `OPENAI_API_KEY`.
 
 ```bash
 cp .env.example .env
 echo 'SEMANTIC_SCHOLAR_API_KEY=your_key_here' >> .env
+echo 'OPENAI_API_KEY=your_openai_key_here' >> .env
 ```
 
 Development mode:
@@ -33,8 +35,21 @@ Without `npm link`, you can still run `node dist/cli/main.js`.
 1. Run `autoresearch` in an empty project.
 2. If `.autoresearch/config.yaml` is missing, setup wizard starts automatically.
 3. Wizard creates scaffold/config and opens the dashboard.
-4. Setup wizard asks for an optional Semantic Scholar API key and writes it to `.env`.
-5. At runtime, AutoResearch reads `SEMANTIC_SCHOLAR_API_KEY` from `process.env` or `.env`.
+4. Setup wizard requires a Semantic Scholar API key and writes it to `.env`.
+5. Setup wizard lets you choose the primary LLM provider:
+   - `codex`: use Codex ChatGPT login for the main workflow
+   - `api`: use OpenAI API models for the main workflow (`OPENAI_API_KEY` required)
+6. Setup wizard also lets you choose PDF analysis mode:
+   - `codex`: download/extract PDF text locally, then analyze with Codex
+   - `api`: send the PDF directly to the Responses API (requires `OPENAI_API_KEY`)
+7. If you choose `api` as the primary provider, setup wizard and `/settings` also let you choose the OpenAI API model.
+   - Current catalog: `gpt-5.4`, `gpt-5`, `gpt-5-mini`, `gpt-4.1`, `gpt-4o`, `gpt-4o-mini`
+8. If you choose `api` for PDF analysis, setup wizard and `/settings` also let you choose the Responses API PDF model.
+   - Current catalog: `gpt-5.4`, `gpt-5`, `gpt-5-mini`, `gpt-4.1`, `gpt-4o`, `gpt-4o-mini`
+9. `/model` follows the active primary provider:
+   - Codex provider: Codex model selector
+   - OpenAI API provider: OpenAI API model selector
+10. At runtime, AutoResearch reads `SEMANTIC_SCHOLAR_API_KEY` and `OPENAI_API_KEY` from `process.env` or `.env`.
 
 ## CLI Policy
 

src/app.ts

@@ -3,6 +3,7 @@ import {
   ensureScaffold,
   hasSemanticScholarApiKey,
   loadConfig,
+  resolveOpenAiApiKey,
   resolveAppPaths,
   resolveSemanticScholarApiKey,
   runSetupWizard,
@@ -14,19 +15,19 @@ import { CodexCliClient } from "./integrations/codex/codexCliClient.js";
 import { AgentOrchestrator } from "./core/agents/agentOrchestrator.js";
 import { launchTerminalApp } from "./tui/TerminalApp.js";
 import { InMemoryEventStream } from "./core/events.js";
-import { CodexLLMClient } from "./core/llm/client.js";
+import { CodexLLMClient, OpenAiResponsesLLMClient, RoutedLLMClient } from "./core/llm/client.js";
 import { LocalAciAdapter } from "./tools/aciLocalAdapter.js";
 import { SemanticScholarClient } from "./tools/semanticScholar.js";
 import { DefaultNodeRegistry } from "./core/stateGraph/nodeRegistry.js";
 import { CheckpointStore } from "./core/stateGraph/checkpointStore.js";
 import { StateGraphRuntime } from "./core/stateGraph/runtime.js";
 import { askLine } from "./utils/prompt.js";
 import { AppConfig } from "./types.js";
+import { ResponsesPdfAnalysisClient } from "./integrations/openai/responsesPdfAnalysisClient.js";
+import { OpenAiResponsesTextClient } from "./integrations/openai/responsesTextClient.js";
 
 export async function runAutoresearchApp(): Promise<void> {
   const paths = resolveAppPaths(process.cwd());
-  await ensureScaffold(paths);
-
   const firstRunSetup = !(await configExists(paths));
   let config;
   if (firstRunSetup) {
@@ -36,14 +37,21 @@ export async function runAutoresearchApp(): Promise<void> {
   } else {
     config = await loadConfig(paths);
   }
+  await ensureScaffold(paths);
 
   const runStore = new RunStore(paths);
   const codex = new CodexCliClient(paths.cwd, {
     model: config.providers.codex.model || "gpt-5.3-codex",
     reasoningEffort: config.providers.codex.reasoning_effort || "xhigh",
     fastMode: config.providers.codex.fast_mode === true
   });
-  const titleGenerator = new TitleGenerator(codex);
+  const openAiText = new OpenAiResponsesTextClient(() => resolveOpenAiApiKey(paths.cwd), {
+    model: config.providers.openai.model,
+    reasoningEffort: config.providers.openai.reasoning_effort
+  });
+  const titleGenerator = new TitleGenerator(() =>
+    config.providers.llm_mode === "openai_api" ? openAiText : codex
+  );
   const initialRunId = await maybeCreateInitialRun({
     firstRunSetup,
     runStore,
@@ -52,14 +60,19 @@ export async function runAutoresearchApp(): Promise<void> {
   });
 
   const eventStream = new InMemoryEventStream();
-  const llm = new CodexLLMClient(codex);
+  const codexLlm = new CodexLLMClient(codex);
+  const openAiLlm = new OpenAiResponsesLLMClient(openAiText);
+  const llm = new RoutedLLMClient(() =>
+    config.providers.llm_mode === "openai_api" ? openAiLlm : codexLlm
+  );
   const aci = new LocalAciAdapter();
   const semanticScholarApiKey = await resolveSemanticScholarApiKey(paths.cwd);
   const semanticScholar = new SemanticScholarClient({
     apiKey: semanticScholarApiKey,
     perSecondLimit: config.papers.per_second_limit,
     maxRetries: 3
   });
+  const responsesPdfAnalysis = new ResponsesPdfAnalysisClient(() => resolveOpenAiApiKey(paths.cwd));
 
   const nodeRegistry = new DefaultNodeRegistry({
     config,
@@ -68,7 +81,8 @@ export async function runAutoresearchApp(): Promise<void> {
     llm,
     codex,
     aci,
-    semanticScholar
+    semanticScholar,
+    responsesPdfAnalysis
   });
 
   const checkpointStore = new CheckpointStore(paths);
@@ -80,6 +94,7 @@ export async function runAutoresearchApp(): Promise<void> {
     runStore,
     titleGenerator,
     codex,
+    openAiTextClient: openAiText,
     eventStream,
     orchestrator,
     initialRunId,