- Fix USD price display: all prices now show in KRW (₩)
- Add Korea region auto-detection: extracts region preference from user messages
- Fix low-spec recommendation for high-performance requirements:
- Add extractTechStack() to detect PostgreSQL, Redis, MongoDB keywords
- Enhance inferExpectedUsers() to consider tech stack complexity
- SaaS/B2B services now recommend 4GB+ RAM servers
- Fix __DIRECT__ tag appearing in output:
- Reorder message concatenation in server-agent.ts
- Add stripping logic in conversation-service.ts and api.ts
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When user requests server recommendation while in an existing session
(e.g., selecting state), reset the session and start fresh instead of
continuing the old session context.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add response_format: { type: 'json_object' } for review mode to force JSON response
- Convert expectedDau and expectedConcurrent from string to number before API call
- Add enhanced KV session debugging with key names in logs
Fixes:
- AI returning plain text instead of JSON in review mode
- 400 error from recommend API due to string values in expected_users
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously cancellation only worked in 'selecting' or 'ordering' states.
Now users can cancel server consultation at any stage using keywords:
취소, 다시, 처음, 리셋, 초기화, 다시 시작, 처음부터
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add expectedDau and expectedConcurrent fields to ServerSession
- Update system prompts to explain DAU vs concurrent users concept
- AI now asks for clarification when users mention visitor counts
- Use concurrent users (5-10% of DAU) for server recommendations
- Update inference rules: personal=10, business=50 concurrent users
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Server Expert AI now reviews recommendation results before showing to user
- Changed flow: get recommendations first → AI reviews → show with comments
- AI provides specific advice based on actual recommended specs
- Reviews include: spec adequacy, bandwidth warnings, CDN suggestions
Before: AI gave generic advice without seeing recommendations
After: AI reviews actual results and gives contextual feedback
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add cancel logic for selecting/ordering states
- Keywords: 취소, 다시, 처음
- Delete session and return to normal conversation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add 'selecting' and 'ordering' status to ServerSession
- Add lastRecommendation field to store recommendation results
- Keep session alive after recommendation (don't delete immediately)
- Add selection pattern matching (1번, 첫번째, 1번 선택 등)
- Add order confirmation message with inline buttons
- Add server_order/server_cancel callback handlers
- Add ServerOrderKeyboardData type for button data
Flow: recommend → select number → confirm with buttons → order/cancel
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add server-agent.ts with 30-year senior architect persona
- Implement KV-based session management for multi-turn conversations
- Add search_trends (Brave Search) and lookup_framework_docs (Context7) tools
- Function Calling support with max 3 tool calls per request
- Auto-infer tech stack and expected users from use case/scale
- Prohibit competitor provider mentions (AWS, GCP, Azure, etc.)
- Simplify main AI system prompt, delegate complex logic to expert AI
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
