fix: route Databricks models through canonical protocols

[alongubkin]

Summary

• route each Databricks catalog model through its canonical upstream protocol and endpoint instead of mirroring the caller's public API
• normalize request/response differences needed by real providers, including Databricks token limits, reasoning blocks, and named tool choice
• add ignored live suites covering Chat Completions, Responses, Messages, tools/tool results, and streaming across direct providers and clouds
• wire provider credentials into the credentialed test environment and exclude paid live binaries from ordinary CI

Real-provider verification

• OpenAI: Chat, Responses, Messages, forced tool call + result, streaming
• Anthropic: Messages, Chat, Responses, forced tool call + result, streaming
• Databricks GPT-OSS: Chat, Responses with reasoning, Messages, forced tool call, streaming
• Databricks Gemma model service: Chat, Responses, Messages
• AWS Bedrock GPT-OSS: Chat, Responses, Messages with real SigV4
• GCP Vertex Gemini: Chat, Responses, Messages
• Azure Foundry Claude: Messages, Chat, Responses, streaming (quota-aware live harness)

Claude-specific Bedrock and Vertex tests remain ignored pending model access grants in the test accounts; the available model paths above were exercised successfully through the real Alien router.

Validation

• cargo test -p alien-ai-gateway
• cargo test -p alien-core (549 unit tests; doc tests green)
• cargo clippy -p alien-ai-gateway -p alien-core --all-targets (green with existing warnings)
• git diff --check
• bash -n scripts/gen-env-test.sh
• all live suites listed above

pnpm format-and-lint remains blocked by a pre-existing formatting error in examples/nextjs-app/next-env.d.ts; this PR does not modify that file.

Linear: ALIEN-642

[greptile-apps]

Greptile Summary

The PR routes Databricks catalog models through their canonical upstream protocols and adds the request, response, tool-choice, reasoning, and streaming translations needed across supported providers.

• Updates model catalog metadata and gateway route selection.
• Extends cross-protocol normalization and provider-specific token handling.
• Adds ignored live-provider coverage and excludes paid suites from ordinary CI.
• Makes live-provider credentials optional during local environment generation while supplying them in credentialed workflows.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
crates/alien-ai-gateway/src/protocol.rs	Extends cross-protocol request, response, tool-choice, reasoning-block, and streaming conversions.
crates/alien-ai-gateway/src/router/mod.rs	Routes catalog models through their canonical upstream protocols and applies provider-specific request normalization.
crates/alien-core/src/ai_catalog.rs	Updates canonical protocol and endpoint metadata for Databricks models.
scripts/gen-env-test.sh	Emits optional live-provider credentials without blocking ordinary local environment generation.
.github/workflows/ci-fast.yml	Excludes the new paid live suites from the ordinary hermetic test run.
.github/workflows/cloud-tests.yml	Supplies live-provider credentials to the credentialed environment-generation job.
.github/workflows/e2e-cloud.yml	Supplies live-provider credentials when generating cloud E2E environment artifacts.

Sequence Diagram

sequenceDiagram
    participant Client
    participant Gateway
    participant Catalog
    participant Provider
    Client->>Gateway: Request via Chat, Responses, or Messages
    Gateway->>Catalog: Resolve public model and canonical protocol
    Catalog-->>Gateway: Upstream model, endpoint, and protocol
    Gateway->>Gateway: Translate request when protocols differ
    Gateway->>Provider: Canonical provider request
    Provider-->>Gateway: Native response or stream
    Gateway->>Gateway: Translate to caller protocol
    Gateway-->>Client: Compatible response or stream

Loading

_{Reviews (2): Last reviewed commit: "fix: keep live provider credentials opti..." | Re-trigger Greptile}