Architecture Overview

# tools/my_tool.py
def get_my_metric(date: str, user_info: dict) -> dict:
    """
    Fetch a custom metric for the given date.

    Args:
        date: ISO 8601 date string.

    Returns:
        Dict containing the metric value and unit.
    """
    user_id = user_info["user_id"]    # injected by Mirobody from JWT
    return {"value": 42, "unit": "steps"}

skills/glucose-coach/
├── metadata.json   # name, summary, when_to_use, when_not_to_use, tags
└── SKILL.md        # YAML frontmatter (name, description, license) + body

"90次每分钟"    → ParsedQuantity(value=90.0,  unit="/min",   family="NRat")
"<5.6 mg/dL"   → ParsedQuantity(value=5.6,   unit="mg/dL",  family="MCnc", comparator="<")
"Millimol pro Liter" → "mmol/L"
"600步"        → ParsedQuantity(value=600.0, unit="{steps}", family="Num")

config.yaml            ← base, do not edit
  └── config.{env}.yaml  ← your overrides; ENV=localdb → config.localdb.yaml
       └── env vars       ← highest precedence; injected by .env or shell

1. Browser POSTs to /api/chat (SSE).
2. JWT middleware → user_id.
3. Chat service creates/uses a session, persists user message to th_messages.
4. DeepAgent (default) is loaded with the user's PROVIDERS_DEEP config.
5. Agent calls tools/list via MCP, plans, then calls tools:
   - get_user_profile  → user_id, timezone, available providers
   - indicator search  → "knee pain" → SNOMED CT / LOINC candidates
   - get_health_data   → time series joined by user_id + indicator
   - chart_service     → renders PNG, uploads to S3, returns presigned URL
6. Agent streams thinking / reply chunks via SSE; response is persisted.
7. Aggregator + insights run in background for next session.