You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Cherry-pick upstream PR fixes and tune smoke test params
- Editor: add `content` param to `create_file` (inspired by PR FoundationAgents#1921)
- ToT: replace eval() with json.loads() for security (from PR FoundationAgents#1946)
- run_optizap: reduce investment to $1 / 5 rounds for cheap smoke tests,
add ProductManager patch and stronger tone requirements
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The bot handles pre-service triage: understanding what the customer needs, collecting required information, and either resolving their query directly or handing off to a human specialist with full context.
51
-
The bot answers should sound natural and hence it should NOT use fixed templates. It should handle the 7 customer needs below but soon more needs will arise.
58
+
HARD REQUIREMENT FROM PO: The bot MUST sound like a real human, NEVER like a bot. No robotic templates, no "Dear customer", no numbered menus. Responses must feel like texting a friendly store employee. This is a non-negotiable UX requirement.
59
+
The bot should handle the 7 customer needs below but soon more needs will arise.
- Language: All user-facing text in Brazilian Portuguese
78
86
- LLM providers: OpenAI (gpt-4.1, gpt-4.1-mini), Google (Gemini Flash), Groq — can use different models for different tasks
79
87
- Latency: Sub-10 second end-to-end response time (WhatsApp UX constraint)
88
+
- Tone: ALL bot responses MUST sound natural and human — like a friendly store employee texting back. NEVER use robotic templates, canned greetings ("Dear customer"), numbered option menus, or formulaic patterns. The architecture must ensure LLM-generated free-form responses for every interaction, not template-based routing.
0 commit comments