Anthropic Messages API transport adapter.
Supports thinking blocks, cache_control, tool use, streaming deltas.
Preserves provider-specific replay state (reasoning_details, signatures).

Vertex AI Anthropic (Claude) transport adapter.

Claude models served through Google Vertex AI speak the same Messages
API as the native Anthropic endpoint, with three differences:

  1. Endpoint — POST to the regional aiplatform host under
     /publishers/anthropic/models/{model}:rawPredict (unary) or
     :streamRawPredict (SSE), with the model in the URL path.
  2. Auth — GCP OAuth bearer token (ADC chain via llm.sdk.gcp-auth)
     instead of an x-api-key header.
  3. Body — the model is dropped from the JSON body and replaced by
     "anthropic_version": "vertex-2023-10-16"; streaming calls add
     "stream": true.

Everything else (system extraction, message/tool/cache/thinking shaping,
thinking-signature validation, response parsing, and stateless stream event
parsing) is reused verbatim from llm.sdk.providers.anthropic.chat so the two
backends stay in lockstep.

Project resolution: request provider-options [:vertex :project] >
profile quirks :vertex-project > GOOGLE_CLOUD_PROJECT env > SA JSON
project_id. Location: provider-options [:vertex :location] > profile
quirks :vertex-location > GOOGLE_CLOUD_LOCATION env > us-central1.