metergraph 0.6.5__tar.gz → 0.6.7__tar.gz

{metergraph-0.6.5/src/metergraph.egg-info → metergraph-0.6.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metergraph
-Version: 0.6.5
+Version: 0.6.7
 Summary: Fire-and-forget LLM spend capture for Metergraph
 Author: Pioneer Square Labs
 License-Expression: Apache-2.0
@@ -137,6 +137,14 @@ part types retain their model, usage, timing, and status metadata, but those
 parts are not replayable yet. See the runnable
 [`python-litellm-otel` example](../examples/python-litellm-otel/).
 
+On a tracer provider shared with other instrumentation,
+`MetergraphGenAIExporter(include_scopes=[...], exclude_scopes=[...])` filters on
+`span.instrumentation_scope.name` (exclude wins; with `include_scopes` set, only
+the listed scopes pass), so one call is never captured twice. The public
+`exporter.skipped` dict counts spans that produced no row, keyed by reason
+(`"scope"`, `"not-genai"`, `"no-model"`) — the first thing to check when a tee
+shows nothing.
+
 Configuration:
 
 - `METERGRAPH_APP_TOKEN` — required bearer token

{metergraph-0.6.5 → metergraph-0.6.7}/README.md

@@ -112,6 +112,14 @@ part types retain their model, usage, timing, and status metadata, but those
 parts are not replayable yet. See the runnable
 [`python-litellm-otel` example](../examples/python-litellm-otel/).
 
+On a tracer provider shared with other instrumentation,
+`MetergraphGenAIExporter(include_scopes=[...], exclude_scopes=[...])` filters on
+`span.instrumentation_scope.name` (exclude wins; with `include_scopes` set, only
+the listed scopes pass), so one call is never captured twice. The public
+`exporter.skipped` dict counts spans that produced no row, keyed by reason
+(`"scope"`, `"not-genai"`, `"no-model"`) — the first thing to check when a tee
+shows nothing.
+
 Configuration:
 
 - `METERGRAPH_APP_TOKEN` — required bearer token

{metergraph-0.6.5 → metergraph-0.6.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metergraph"
-version = "0.6.5"
+version = "0.6.7"
 description = "Fire-and-forget LLM spend capture for Metergraph"
 readme = "README.md"
 requires-python = ">=3.10"

{metergraph-0.6.5 → metergraph-0.6.7}/src/metergraph/_capture.py

@@ -20,7 +20,7 @@ from urllib.parse import urlsplit
 
 from ._context import CaptureContext, snapshot
 from ._gateway import detect_gateway, gateway_evidence, resolve_gateway
-from ._template import scrub, template_hash
+from ._template import json_value, scrub_request, template_hash
 from ._version import SDK_VERSION
 
 
@@ -330,29 +330,34 @@ def _request_id(response: Any) -> str | None:
 def _response_content(response: Any, aggregate_text: str | None = None) -> Any:
     if aggregate_text is not None:
         return aggregate_text
-    direct = _get(response, "output_text") or _get(response, "text")
-    if direct is not None:
-        return scrub(direct)
+    # Only a non-empty string is reply text. On an OpenAI Responses object
+    # `text` is the text *config*, and a function-call-only reply lives in
+    # `output`.
+    direct = _get(response, "output_text")
+    if not isinstance(direct, str) or not direct:
+        direct = _get(response, "text")
+    if isinstance(direct, str):
+        return direct
     choice = _first(_get(response, "choices"))
     message = _get(choice, "message")
     content = _get(message, "content")
     if content is not None:
-        return scrub(content)
+        return json_value(content)
     parsed = _get(message, "parsed")
     if parsed is not None:
-        return scrub(parsed)
+        return json_value(parsed)
     normalized_text = _response_text(response)
     if normalized_text is not None:
         return normalized_text
     blocks = _get(response, "content")
     if blocks is not None:
-        return scrub(blocks)
+        return json_value(blocks)
     outputs = _get(response, "output")
     if outputs is not None:
-        return scrub(outputs)
+        return json_value(outputs)
     candidates = _get(response, "candidates")
     if candidates is not None:
-        return scrub(candidates)
+        return json_value(candidates)
     return None
 
 
@@ -402,7 +407,7 @@ def _tool_argument(value: Any) -> Any:
             return json.loads(value)
         except json.JSONDecodeError:
             return value
-    return scrub(value)
+    return json_value(value)
 
 
 def _tool_policies(request: Mapping[str, Any]) -> dict[str, str]:
@@ -591,7 +596,7 @@ def _tool_events(
             block = _get(chunk, "content_block")
             if _get(block, "type") == "tool_use":
                 key = str(_get(chunk, "index", len(anthropic_deltas)))
-                initial_input = scrub(_get(block, "input", {}))
+                initial_input = json_value(_get(block, "input", {}))
                 anthropic_deltas[key] = {
                     "id": str(_get(block, "id") or key),
                     "name": str(_get(block, "name") or ""),
@@ -769,7 +774,7 @@ class CallState:
             if self.context.capture_text is not None
             else self.runtime.options.capture_text
         )
-        request_clean = scrub(self.request)
+        request_clean = scrub_request(self.request)
         request_json, request_truncated = self.runtime._text(
             json.dumps(request_clean, separators=(",", ":"), default=repr),
             "request",

{metergraph-0.6.5 → metergraph-0.6.7}/src/metergraph/_gateway.py

@@ -43,6 +43,15 @@ OPENROUTER = GatewayContract(
 # Registry of supported gateways, keyed by canonical name.
 GATEWAYS: dict[str, GatewayContract] = {OPENROUTER.name: OPENROUTER}
 
+# Evidence-only cost provenance: unlike a GatewayContract these assert no
+# gateway identity, qualify every endpoint, and are kept out of GATEWAYS so
+# detect_gateway() and resolve_gateway() never accept them -- an explicit
+# wrap(..., gateway=...) override names a gateway contract only.
+LANGFUSE_SPAN_COST = "langfuse.observation.cost_details.total"
+OPENINFERENCE_SPAN_COST = "openinference.llm.cost.total"
+
+EVIDENCE_COST_SOURCES = frozenset({LANGFUSE_SPAN_COST, OPENINFERENCE_SPAN_COST})
+
 # Model/provider identifiers are bounded to the same practical limit the rest of
 # the SDK applies to those strings.
 _MAX_IDENTIFIER_LEN = 512
@@ -115,6 +124,22 @@ def _billing_amount(value: Any) -> float | None:
     return number
 
 
+def evidence_for_mapped_cost(cost: Any, cost_source: Any) -> dict[str, Any]:
+    """Build capture-row cost evidence from a mapped vendor-reported cost.
+
+    ``cost_source`` must be one of the SDK-fixed provenance strings in
+    EVIDENCE_COST_SOURCES (e.g. a ``MappedCall.cost_source``); anything else
+    emits nothing, so provenance stays SDK-controlled. Malformed amounts are
+    omitted. The result never contains a ``gateway`` key.
+    """
+    if cost_source not in EVIDENCE_COST_SOURCES:
+        return {}
+    amount = _billing_amount(cost)
+    if amount is None:
+        return {}
+    return {"reported_cost_usd": amount, "reported_cost_source": cost_source}
+
+
 def gateway_evidence(
     gateway: str | None, endpoint: str, response: Any
 ) -> dict[str, Any]:
@@ -125,12 +150,17 @@ def gateway_evidence(
     emitted only on the gateway's qualified endpoints. A provenance source
     appears only alongside the value it describes. The served provider is never
     inferred.
+
+    ``gateway`` may also name an evidence-only contract by its SDK-fixed
+    provenance string (see EVIDENCE_COST_SOURCES).
     """
     if gateway is None:
         return {}
     contract = GATEWAYS.get(gateway)
     if contract is None:
-        return {}
+        # An evidence-only provenance string: the observed scalar is read
+        # from the response's "cost" field.
+        return evidence_for_mapped_cost(_get(response, "cost"), gateway)
 
     evidence: dict[str, Any] = {"gateway": contract.name}