Support request labels on Vertex calls for GCP cost attribution #142

Environment

• google-antigravity SDK v0.1.7
• Python 3.11+
• Linux (GitHub Actions) and macOS

Description

It would be great to attach GCP request labels to Vertex calls when running with vertex=True. Vertex AI accepts a labels map on GenerateContentRequest (docs), and those labels flow into Cloud Billing.

Today there's no way to set them, so multiple SDK-based workloads in one GCP project collapse into a single Vertex AI line item.

Evidence

The Vertex request type compiled into the harness already carries labels:

$ strings google/antigravity/bin/localharness | grep -i 'GenerateContentRequest.*Label'
...google.cloud.aiplatform.master.GenerateContentRequest.LabelsEntry...
...prediction_service_go_proto.(*GenerateContentRequest).GetLabels

But no message in the harness config proto (the Python → Go contract) exposes a labels field:

>>> import google.antigravity.connections.local.localharness_pb2 as pb
>>> from google.protobuf.message import Message
>>> [f.name for f in pb.VertexEndpoint.DESCRIPTOR.fields]
['base_url', 'http_headers', 'project', 'location', 'options']
>>> [m for m in dir(pb)
...  if isinstance(getattr(pb, m), type) and issubclass(getattr(pb, m), Message)
...  and any('label' in f.name for f in getattr(pb, m).DESCRIPTOR.fields)]
[]

And it isn't on the Python config model either:

>>> from google.antigravity.connections.local.local_connection_config import LocalAgentConfig
>>> [f for f in LocalAgentConfig.model_fields if 'label' in f]
[]

Proposed Fix

The Vertex request already accepts labels (see above); the gap is that neither the config proto nor the Python model carries them. A fix would:

1. Add a labels map field to the harness config proto, e.g. on VertexEndpoint.
2. Set it on GenerateContentRequest.labels for each Vertex call in the harness.
3. Expose a labels: dict[str, str] option on the Vertex config, as a LocalAgentConfig shorthand alongside project / location.
4. Pass it through the config-to-proto conversion so the value reaches the harness.

Workarounds

1. Dedicated GCP project per workload — project-level attribution, but heavyweight and doesn't scale to workloads that share infrastructure.
2. Client-side token accounting from UsageMetadata, priced offline — parallel bookkeeping that can't reconcile against the actual invoice (and metadata was historically dropped on finish/tool turns, Finish/tool turns drop SDK usage metadata while equivalent plain text turns do report it #24).
3. Calling google-genai directly with GenerateContentConfig(labels=...) — loses the harness (skills, sandbox, policies).