feat(billing): add billing source classification for sessions and update cost handling

backend/app/api/claude_code.py

@@ -66,6 +66,7 @@ def _session_to_read(s: reader.ClaudeSession) -> ClaudeSessionRead:
             total=s.tokens.total,
         ),
         cost_usd=s.cost_usd,
+        billing_source=s.billing_source,
         message_count=s.message_count,
         first_message_at=s.first_message_at,
         last_message_at=s.last_message_at,

backend/app/schemas/claude_code.py

@@ -93,7 +93,8 @@ class ClaudeSessionRead(SQLModel):
     title: str | None = None
     models: list[str]
     tokens: SessionTokensRead
-    cost_usd: float
+    cost_usd: float           # always $0 for subscription sessions
+    billing_source: str       # "subscription" | "api"
     message_count: int
     first_message_at: datetime | None = None
     last_message_at: datetime | None = None

backend/app/services/claude_code_reader.py

@@ -23,6 +23,32 @@ logger = get_logger(__name__)
 
 ACTIVE_WINDOW_MINUTES = 30
 
+# ---------------------------------------------------------------------------
+# Billing classification
+# ---------------------------------------------------------------------------
+
+# Entrypoints that run against the user's Claude subscription (flat-rate, not
+# per-token API billing).  Sessions from these clients cost $0 in real money
+# even though they consume tokens.
+_SUBSCRIPTION_ENTRYPOINTS: frozenset[str] = frozenset({
+    "claude-vscode",
+    "claude-jetbrains",
+    "claude-web",
+})
+
+
+def _billing_source(entrypoints: set[str]) -> str:
+    """Return 'subscription' or 'api' based on the session's entrypoints.
+
+    Any session that touched a subscription client is treated as subscription
+    (cost = $0).  Sessions with no entrypoint metadata default to subscription
+    because JSONL files are written by local Claude Code clients, not the API.
+    """
+    if not entrypoints or entrypoints & _SUBSCRIPTION_ENTRYPOINTS:
+        return "subscription"
+    return "api"
+
+
 # ---------------------------------------------------------------------------
 # Pricing (USD per million tokens) — mirrors runtime_usage.DEFAULT_MODEL_PRICING
 # ---------------------------------------------------------------------------
@@ -86,7 +112,8 @@ class ClaudeSession:
     title: str | None
     models: list[str]
     tokens: SessionTokens
-    cost_usd: float
+    cost_usd: float           # $0 for subscription sessions, real cost for api sessions
+    billing_source: str       # "subscription" | "api"
     message_count: int        # assistant turns
     first_message_at: datetime | None
     last_message_at: datetime | None
@@ -221,11 +248,15 @@ def _parse_session_file(path: Path) -> ClaudeSession | None:
         return None
 
     model_list = sorted(models)
-    primary_model = model_list[0] if model_list else ""
-    cost = _price(primary_model, tokens.input, tokens.output, tokens.cache_read, tokens.cache_write)
-    for m in model_list[1:]:
-        # Additional models — approximate with same token split (rare)
-        cost += _price(m, 0, 0, 0, 0)
+    billing = _billing_source(entrypoints)
+
+    if billing == "subscription":
+        cost = 0.0
+    else:
+        primary_model = model_list[0] if model_list else ""
+        cost = _price(primary_model, tokens.input, tokens.output, tokens.cache_read, tokens.cache_write)
+        for m in model_list[1:]:
+            cost += _price(m, 0, 0, 0, 0)
 
     now = datetime.utcnow()
     is_active = bool(last_ts and (now - last_ts) < timedelta(minutes=ACTIVE_WINDOW_MINUTES))
@@ -238,6 +269,7 @@ def _parse_session_file(path: Path) -> ClaudeSession | None:
         models=model_list,
         tokens=tokens,
         cost_usd=round(cost, 6),
+        billing_source=billing,
         message_count=message_count,
         first_message_at=first_ts,
         last_message_at=last_ts,