feat(promptfoo): add crab pf serve Slack polling daemon

Adds a local daemon that polls for Slack DMs and runs the promptfoo
target discovery agent on behalf of team members. Non-technical users
can DM the Crab bot with "pf: <target spec>" and receive generated
promptfooconfig.yaml files back in-thread.

- New slack.ts: Slack Web API helpers using native fetch (zero deps)
- New serve.ts: daemon loop with setup flow, DM polling, and job mgmt
- Updated cli.ts: routes `serve` subcommand, adds --reasoning flag
- Updated providers.ts: reasoningEffort support for GPT-5/o-series

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: MrFlounder

plugins/promptfoo/package-lock.json

@@ -22,6 +22,7 @@ export interface ChatOptions {
   tools: unknown[];
   maxTokens?: number;
   temperature?: number;
+  reasoningEffort?: 'low' | 'medium' | 'high';
 }
 
 export interface ChatResponse {
@@ -41,18 +42,21 @@ export class OpenAIProvider implements LLMProvider {
   private apiKey: string;
   private model: string;
   private baseUrl: string;
+  reasoningEffort?: string;
 
-  constructor(options: { apiKey?: string; model?: string; baseUrl?: string }) {
+  constructor(options: { apiKey?: string; model?: string; baseUrl?: string; reasoningEffort?: string }) {
     this.apiKey = options.apiKey || process.env.OPENAI_API_KEY || '';
     this.model = options.model || 'gpt-4o';
     this.baseUrl = options.baseUrl || 'https://api.openai.com/v1';
+    this.reasoningEffort = options.reasoningEffort;
 
     if (!this.apiKey) {
       throw new Error('OpenAI API key required. Set OPENAI_API_KEY env var.');
     }
   }
 
   async chat(options: ChatOptions): Promise<ChatResponse> {
+    const isReasoning = this.model.startsWith('gpt-5') || this.model.startsWith('o1') || this.model.startsWith('o3');
     const response = await fetch(`${this.baseUrl}/chat/completions`, {
       method: 'POST',
       headers: {
@@ -63,8 +67,13 @@ export class OpenAIProvider implements LLMProvider {
         model: this.model,
         messages: options.messages.map((m) => this.toOpenAIMessage(m)),
         tools: options.tools,
-        max_tokens: options.maxTokens || 4096,
-        temperature: options.temperature ?? 0.7,
+        ...(isReasoning
+          ? { max_completion_tokens: options.maxTokens || 4096 }
+          : { max_tokens: options.maxTokens || 4096 }),
+        ...(isReasoning ? {} : { temperature: options.temperature ?? 0.7 }),
+        ...(options.reasoningEffort || this.reasoningEffort
+          ? { reasoning_effort: options.reasoningEffort || this.reasoningEffort }
+          : {}),
       }),
     });
 
@@ -253,12 +262,12 @@ export class AnthropicProvider implements LLMProvider {
 /**
  * Create provider from string identifier
  */
-export function createProvider(provider: string): LLMProvider {
+export function createProvider(provider: string, options?: { reasoningEffort?: string }): LLMProvider {
   const [type, model] = provider.split(':');
 
   switch (type) {
     case 'openai':
-      return new OpenAIProvider({ model: model || 'gpt-4o' });
+      return new OpenAIProvider({ model: model || 'gpt-4o', reasoningEffort: options?.reasoningEffort });
     case 'anthropic':
       return new AnthropicProvider({ model: model || 'claude-sonnet-4-20250514' });
     default:

plugins/promptfoo/src/agent/providers.ts

@@ -15,18 +15,26 @@ import * as path from 'node:path';
 import { parseArtifact, detectFormat } from './parsers/index.js';
 import { runDiscoveryAgent } from './agent/loop.js';
 import { createProvider } from './agent/providers.js';
+import { runServe } from './serve.js';
 
 const args = process.argv.slice(2);
 
 async function main() {
   try {
+    // Handle 'serve' subcommand
+    if (args[0] === 'serve') {
+      await runServe(args.slice(1));
+      return;
+    }
+
     // Parse arguments
     const filePath = getArg('--file') || getArg('-f');
     const urlArg = getArg('--url');
     const providerStr = getArg('--provider') || process.env.DISCOVERY_PROVIDER || 'openai:gpt-4o';
     const outputDir = getArg('--output') || getArg('-o') || '.';
     const verbose = args.includes('--verbose') || args.includes('-v');
     const maxTurns = parseInt(getArg('--max-turns') || '30', 10);
+    const reasoningEffort = getArg('--reasoning');
 
     let context: string;
 
@@ -52,7 +60,7 @@ async function main() {
     // Detect format
     const format = detectFormat(context);
     console.log(`Detected format: ${format === 'unknown' ? 'text/description' : format}`);
-    console.log(`Provider: ${providerStr}`);
+    console.log(`Provider: ${providerStr}${reasoningEffort ? ` (reasoning: ${reasoningEffort})` : ''}`);
     console.log(`Output: ${outputDir}`);
     console.log('');
 
@@ -62,7 +70,7 @@ async function main() {
     }
 
     // Create provider
-    const provider = createProvider(providerStr);
+    const provider = createProvider(providerStr, reasoningEffort ? { reasoningEffort } : undefined);
 
     // Run discovery agent
     console.log('Starting target discovery agent...');
@@ -134,6 +142,7 @@ Options:
   --output, -o <dir>     Output directory (default: current dir)
   --provider <provider>  LLM provider (default: openai:gpt-4o)
   --max-turns <n>        Max agent turns (default: 30)
+  --reasoning <effort>   Reasoning effort for GPT-5/o-series (low, medium, high)
   --verbose, -v          Show detailed output
 
 Supported input formats:
@@ -158,6 +167,11 @@ Examples:
   # Using Anthropic
   crab pf --file target.txt --provider anthropic:claude-sonnet-4-20250514
 
+Subcommands:
+  serve                  Run Slack polling daemon
+  serve --setup          Configure Slack username
+  serve --help           Show serve help
+
 Output:
   The agent will create:
   - promptfooconfig.yaml (promptfoo configuration)