Reuse onboarding flows in Settings, fix AI provider requests, classify business conversations

CHANGELOG.md

@@ -37,6 +37,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   are preserved, and Skip retains the unscanned attachment warning.
 
 ### Fixed
+- **Business conversations are classified by purpose.** Requested quotes,
+  customer questions, negotiations and internal approvals are distinguished
+  from promotional outreach. Categorization uses outgoing correspondence as
+  context and applies the same guidance to automatic and manual processing.
+  Category lists also refresh their membership after background classification.
 - **Undo send on a forward puts you back in it.** Undoing the send of an
   inline forward reopened it with what you had written but left it unfocused,
   so typing went nowhere until you clicked into it. It now scrolls to the

CLAUDE.md

@@ -204,6 +204,9 @@ duplication and repeated patterns:
 Prefer small, pure, single-responsibility helpers that are easy to unit test.
 
 Established shared helpers (extend this list as you add more):
+- `buildAIChatRequestOptions(provider)` — `packages/core/src/utils/ai-provider-auth.ts`.
+  Restricts Chat Completions vendor options to the configured provider; reused
+  by model verification, renderer completions and both background categorizers.
 - `withTimeout(promise, ms, message)` — `packages/core/src/utils/timeout.ts`.
   The single source for the "race a promise against a timeout" pattern; it also
   clears the timer. Use it instead of hand-rolling `Promise.race([p, setTimeout(reject)])`.

apps/desktop/electron/services/ai-categorization-service.ts

@@ -7,7 +7,7 @@
  * Runs entirely in the Electron main process for reliability.
  */
 
-import { cleanLLMJsonResponse, tryParseLLMJson, salvageJsonArrayWithDiagnostics, extractBalancedJsonArray, cleanEmailHtmlForLLM, isConnectionError, isUpstreamError, describeNetworkError, classifyAIError, createLogger, applySecurityGate, buildSecurityContext, formatSecurityLines, PHISHING_PROMPT, SPAM_PROMPT, buildAIAuthHeaders, existingCategoryClassification, automaticCategorizationDeferred, isSpamProtectedEmail } from '@sarvinbox/core';
+import { cleanLLMJsonResponse, tryParseLLMJson, salvageJsonArrayWithDiagnostics, extractBalancedJsonArray, cleanEmailHtmlForLLM, isConnectionError, isUpstreamError, describeNetworkError, classifyAIError, createLogger, applySecurityGate, buildSecurityContext, formatSecurityLines, buildCategorizationPrompt, buildAIAuthHeaders, buildAIChatRequestOptions, existingCategoryClassification, automaticCategorizationDeferred, isSpamProtectedEmail } from '@sarvinbox/core';
 import type { EmailRecord , AIErrorInfo, EmailSecurityContext } from '@sarvinbox/core';
 
 import { getMainWindow, requireStorage } from '../shared';
@@ -165,12 +165,8 @@ const AUTO_BACKOFF_MAX_MS = 30 * 60_000;
 const BULK_RESTART_BASE_MS = 15_000;
 const BULK_RESTART_MAX_MS = 5 * 60_000;
 
-// The spam and phishing guidance is shared with the pipeline in core
-// (`SPAM_PROMPT`, `PHISHING_PROMPT` in categorization-utils) so both prompts
-// teach the model the same rules — the Adobe Sign lure of 2026-09-23 was
-// marked important by a prompt that had never been told authentication is
-// not identity, and a second copy of the text here is how one prompt learns
-// a lesson the other does not.
+// The complete classification guidance is shared with the background pipeline
+// in core, so manual processing follows the same purpose and reply rules.
 
 // ========== Service Class ==========
 
@@ -204,62 +200,12 @@ export class AICategorizationService {
    * Build the categorization prompt dynamically from category definitions
    */
   private buildPrompt(categories: LoadedCategoryDef[], userEmail: string): string {
-    const userName = userEmail.split('@')[0] || '';
-    const categorySection = categories.map((cat, i) =>
-      `${i + 1}. ${cat.slug}:\n${cat.prompt}`
-    ).join('\n\n');
-
-    return `You are an email intelligence agent for "${userEmail}" (${userName}).
-Classify each email — but ONLY assign categories when the email actually matters to this user.
-
-CATEGORIES:
-${categorySection}
-
-SPAM:
-${SPAM_PROMPT}
-
-${PHISHING_PROMPT}
-
-═══════════════════════════════════════════
-KEY RULE: Think from ${userName}'s perspective.
-"Would ${userName} need to ACT on this email?"
-═══════════════════════════════════════════
-
-WHO IS THE EMAIL FOR? (MOST IMPORTANT CHECK)
-- Look at TO: vs CC: fields
-- "User-Role: CC" = ${userName} is just looped in — DEFAULT: NOT important, NOT needs_response
-- If email body greets someone else ("Hello Hrishi") but ${userName} is CC → NOT for ${userName}
-- If task is for someone in TO: and ${userName} is CC → NOT ${userName}'s task
-- ONLY mark CC as important if body EXPLICITLY asks for ${userName}'s input by name
-
-BEHAVIORAL SIGNALS (use these — they show what ${userName} actually cares about):
-- "Read: X%" = how often ${userName} opens this sender's emails. Low% = doesn't care
-- "Keep: X%" = how often ${userName} keeps vs deletes. Low% = noise
-- "Replied: N" = how often ${userName} replied. 0 = never replied = probably not needs_response
-- User NEVER replying to a sender = that sender is NOT important enough for needs_response
-
-WHAT IS "IMPORTANT"?
-- ${userName} is in TO: (not CC:) AND email asks for their action/decision
-- From someone ${userName} consistently replies to
-- Customer/client emails directly to ${userName}
-- NOT important: team loops, FYIs, tasks for others, newsletters, automated alerts
-
-CATEGORY ASSIGNMENT RULES (STRICT):
-- Prefer ONE category per email. Only assign 2+ if genuinely applicable.
-- "important" is RARE — means URGENT, needs action TODAY. Not just "relevant".
-- An invoice is just "invoice", NOT also "important" unless payment is overdue TODAY.
-- A finance email is just "finance", NOT also "important" unless fraud alert.
-- A meeting invite is just "meeting", NOT also "important" unless meeting is in the next hour.
-- When in doubt, assign FEWER categories. Empty [] is valid.
-
-SENDER MEMORY (optional — extract if visible):
-- How the sender or user greets/addresses in this email
-- Conversation tone and key topic
-- Include "sender_memory" only when meaningful
-
-Return JSON array:
-[{"emailId":"...","categories":[],"is_spam":false,"confidence":0.8,"reasoning":"...","sender_memory":{"greeting":"Hi Advik","tone":"formal","key_context":"Invoice follow-up"}}]
-- Single email (depth 1) from occasional sender = likely lower priority unless content signals otherwise`;
+    // Read the same account's editable template as the background pipeline.
+    // An absent template uses the default; a failed read must stop the request
+    // rather than silently disregard the user's customized instructions.
+    const promptRepo = requireStorage().getRepositories?.()?.prompts;
+    const override = promptRepo?.getContent('categorization_system') || undefined;
+    return buildCategorizationPrompt(categories, userEmail, override);
   }
 
   /**
@@ -998,11 +944,8 @@ Return format (categories is an array of matching slugs from: ${categorySlugs}):
             { role: 'user', content: userMessage },
           ],
           max_completion_tokens: 16000,
-          // See packages/core/src/agent/categorization-utils.ts for the
-          // full rationale: vLLM-hosted thinking models (Gemma 3/4, Qwen3)
-          // burn the token budget on inline <think> blocks and truncate
-          // before emitting the JSON. Disable via the chat-template kwarg.
-          chat_template_kwargs: { enable_thinking: false },
+          // Keep background requests consistent with onboarding model tests.
+          ...buildAIChatRequestOptions(config),
         }),
         signal: this.abortController?.signal,
       });