Skip to content

vertexai.md

Latest commit

 

History

History
57 lines (43 loc) · 2.41 KB

vertexai.md

File metadata and controls

57 lines (43 loc) · 2.41 KB
summary read_when
Vertex AI provider data sources: gcloud ADC credentials and Cloud Monitoring quota usage.
Debugging Vertex AI auth or quota fetch
Updating Vertex AI usage mapping or login flow

Vertex AI provider

Data sources + fallback order

  1. OAuth via gcloud ADC (only path used in fetch()):
    • Reads application_default_credentials.json from the gcloud config directory.
    • Uses Cloud Monitoring time-series metrics to compute quota usage.

OAuth credentials

  • Authenticate: gcloud auth application-default login.
  • Project: gcloud config set project PROJECT_ID.
  • Fallback project env vars: GOOGLE_CLOUD_PROJECT, GCLOUD_PROJECT, CLOUDSDK_CORE_PROJECT.

API endpoints

  • Cloud Monitoring timeSeries:
    • Usage: serviceruntime.googleapis.com/quota/allocation/usage
    • Limit: serviceruntime.googleapis.com/quota/limit
    • Resource: consumer_quota with service="aiplatform.googleapis.com".

Mapping

  • Matches usage + limit series by quota metric + limit name + location.
  • Reports the highest usage percent across matched series.
  • Displayed as "Quota usage" with period "Current quota".

Token Cost Tracking

Vertex AI Claude usage is logged to the same local files as direct Anthropic API usage (~/.claude/projects/). CodexBar identifies Vertex AI entries using two methods:

Detection Methods

  1. Model name format (primary): Vertex AI uses @ as version separator

    • Vertex AI: claude-opus-4-5@20251101
    • Anthropic API: claude-opus-4-5-20251101

  2. Metadata fields (fallback): Entries with provider metadata

    • metadata.provider: "vertexai"
    • Keys containing vertex or gcp

Requirements

To see Vertex AI token costs:

  1. Enable the Vertex AI provider in Settings → Providers
  2. Enable "Show cost summary" in Settings → General
  3. Use Claude Code with Vertex AI (your cv alias sets ANTHROPIC_MODEL=claude-opus-4-5@20251101)

Note: The model name must include the @ format for detection to work. If Claude Code normalizes model names to - format when logging, the entries won't be distinguishable from direct Anthropic API usage.

Troubleshooting

  • No quota data: Ensure Cloud Monitoring API access in the selected project.
  • No cost data: Check that ~/.claude/projects/ exists and contains .jsonl files from Claude Code usage with Vertex AI metadata.
  • Auth issues: Re-run gcloud auth application-default login.