Add Vertex AI provider support for Claude models

[fcorrea]

Add Vertex AI provider support for Claude models

Summary

• Add support for routing Claude requests through Google Cloud Vertex AI when ANTHROPIC_VERTEX_PROJECT_ID and CLOUD_ML_REGION (or ANTHROPIC_VERTEX_REGION) are set

• Introduce AuthMode enum (ApiKey | OAuth | Vertex { url }) in anthropic.rs to cleanly separate the three authentication paths without branching on booleans

• Implement Google Application Default Credentials (ADC) token fetching with in-memory caching, supporting:

  • GCE/Cloud Run metadata server (automatic on GCP)
  • authorized_user credentials (from gcloud auth application-default login)
  • service_account credentials (from a downloaded service account JSON key)

• Service account JWT signing uses ring (already a transitive dependency via rustls) ��� pure Rust, no subprocess or openssl dependency required

• Build the correct Vertex AI streaming endpoint: {region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/publishers/anthropic/models/{model}:streamRawPredict

• Handle Vertex-specific API differences:

  • model is stripped from the request body (it's in the URL)
  • anthropic_version: "vertex-2023-10-16" is sent in the body instead of as a header
  • No anthropic-beta header is sent (Vertex rejects unknown beta headers; prompt caching works natively via cache_control blocks in the request body)

• Add build_anthropic_vertex_route() so the model picker shows Vertex AI as a distinct provider

• Initialize AnthropicProvider when Vertex env vars are present, even without Claude OAuth credentials (startup.rs)

• Update the model picker UI to show "Vertex AI" as the provider when Vertex env vars are set (inline_interactive.rs)

Edge cases and tradeoffs

• Token refresh: Google ADC tokens are cached in-memory with a 60-second safety margin before expiry. On GCE this hits the metadata server; elsewhere it refreshes via the OAuth2 token endpoint or a signed JWT for service accounts.

• global region: When CLOUD_ML_REGION=global, the endpoint uses https://aiplatform.googleapis.com (no subdomain), matching Anthropic's SDK behavior. Multi-region identifiers us and eu are also accepted by Vertex.

• Prompt caching: Fully supported on Vertex — cache_control blocks flow through in the request body unchanged. The anthropic-beta: prompt-caching-2024-07-31 header is intentionally omitted because Vertex rejects unrecognized beta headers.

• Service account key format: ring accepts PKCS#8 DER (the format Google generates for service account JSON keys) with a PKCS#1 DER fallback. The PEM headers are stripped and the body base64-decoded before passing to ring.

• Code size: anthropic.rs grew significantly. The ADC/JWT machinery could be extracted to a separate module in a follow-up.

Validation

• Tested locally with ANTHROPIC_VERTEX_PROJECT_ID, CLOUD_ML_REGION, and authorized_user ADC credentials. Full multi-turn conversations with tool use and prompt caching worked correctly through Vertex — this PR was written while jcode was actively running through Vertex AI.

• cargo fmt --all -- --check: passes

• cargo check -p jcode: passes

• cargo clippy on changed files: no new warnings

• Panic budget: no new unwrap/expect in our changes

• Pre-existing budget failures (jcode-import-core, jcode-tui-mermaid, unrelated files) are present on master before this change

Environment variables

Variable	Required	Notes
ANTHROPIC_VERTEX_PROJECT_ID	Yes	GCP project ID
CLOUD_ML_REGION	Yes (or ANTHROPIC_VERTEX_REGION)	e.g. us-east5, global, us, eu
GOOGLE_APPLICATION_CREDENTIALS	No	Path to service account JSON; falls back to ~/.config/gcloud/application_default_credentials.json

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews