deploy(S4): Blender headless pipeline + WebAR client + backend fixes

backend/services/token_manager.py

@@ -26,15 +26,24 @@ from backend.services.usb_vault import KeyDomain, resolve_vault_key
 def get_runtime_location() -> str:
     return "cloud" if os.environ.get("SPACE_ID") else "pc"
 
-def _persona_system_prompt(persona: str) -> str:
+# Only the conversational path should be shaped in the assistant's voice.
+# Structured task types (blueprint JSON, code implementation, research digests,
+# captcha/osint payloads) must stay clean — a "You are JARVIS…" preamble there
+# corrupts JSON/code output.
+_PERSONA_TASK_TYPES = {"general", "gaming"}
+
+def _persona_system_prompt(persona: str, task_type: str = "general") -> str:
     """The identity prompt for the EXPLICITLY requested persona.
 
     The mobile app (and other callers) pass the persona they want per-turn, so
     we select on that string directly rather than the persisted global mode.
-    Without this, LLM replies carry no JARVIS/FRIDAY character at all — the
-    identity file that is supposed to be the source of truth was being ignored
-    on the primary /api/chat path.
+    Without this, conversational LLM replies carry no JARVIS/FRIDAY character —
+    the identity file that is supposed to be the source of truth was being
+    ignored on the primary /api/chat path. Returns "" for non-conversational
+    task types so structured output is never polluted.
     """
+    if task_type not in _PERSONA_TASK_TYPES:
+        return ""
     try:
         from modules.assistant_identity import (
             JARVIS_PERSONALITY_PROMPT, FRIDAY_PERSONALITY_PROMPT,
@@ -245,7 +254,7 @@ def _nvidia_fallback_call_sync(prompt: str, task_id: str, task_type: str, partia
     if not partial_so_far:
         # NVIDIA models take no separate system channel here, so the persona
         # prompt is prepended (same place the hive-mind context goes).
-        _sys = _persona_system_prompt(persona)
+        _sys = _persona_system_prompt(persona, task_type)
         actual_prompt = (_sys + "\n\n" if _sys else "") + omni_context + actual_prompt
 
     # Delegate to the NVIDIA Vault API directly
@@ -276,7 +285,7 @@ def _gemini_call_sync(prompt: str, task_id: str, task_type: str,
 
     # Shape the reply in the active assistant's voice (JARVIS/FRIDAY). Skip on a
     # resume continuation so we don't restate the persona mid-sentence.
-    _sys = _persona_system_prompt(persona) if not partial_so_far else None
+    _sys = _persona_system_prompt(persona, task_type) if not partial_so_far else None
     model = genai.GenerativeModel(MODEL, system_instruction=_sys) if _sys else genai.GenerativeModel(MODEL)
     actual_prompt = _build_resume_prompt(prompt, partial_so_far, _extract_last_word(partial_so_far))
     if not partial_so_far: