feat(decisions): add Databricks ai_decide as a /v1/decisions provider and auto-router decider

[mateo-berri]

TLDR

Problem this solves:

• /v1/decisions refused databricks/<endpoint> as an unknown provider
• The auto router classifier only accepted provider: jev, laya, or bespoke
• A customer serving OpenJev (ai_decide) on Databricks could use neither

How it solves it:

• New Decisions provider databricks, posting to {DATABRICKS_API_BASE}/<endpoint>/invocations
• Keyed by DATABRICKS_API_KEY, falling back to DATABRICKS_TOKEN
• opensource_classifier_config.provider: databricks for the auto router, model = the serving endpoint name
• Classifier logs and cost rows carry databricks/<endpoint>, not Databricks's internal /mosaicml/local_model
• /v1/decisions returns the Databricks body unchanged, model included
• A databricks/<name> that is not a bare serving endpoint name (a slash, .., a query string, a space) answers 400 must be the bare endpoint name instead of reaching Databricks
• Admin UI: Add Auto Router -> OSS Classifier offers a Databricks radio, and Classifier tuning takes the serving endpoint name
• A databricks/databricks-openjev-qwen35-4b cost-map entry with mode: evaluation at a zero rate (Databricks publishes no per-token OpenJev price), so /health probes the decider through /v1/decisions and Lens lists it with no model_info needed; a deployment's model_info prices set the billed cost

User Flow

Before: a platform team pointing the proxy at their Databricks OpenJev serving endpoint gets a 400 on every decision, and the auto router never boots

1. The admin adds model: databricks/databricks-openjev-qwen35-4b to model_list plus an auto_router/complexity_router deployment with opensource_classifier_config.provider: databricks, then starts the proxy
2. The proxy log reports the auto router deployment refused: opensource_classifier_config.provider Input should be 'jev', 'laya' or 'bespoke', and GET http://localhost:4000/v1/models lists no auto
3. They send POST http://localhost:4000/v1/decisions with "model": "databricks-decider", a state, and three questions
4. They get 400 Unknown Decisions provider 'databricks'. Supported providers: perplexity, typesafe, openrouter, cloudflare, strands_decider
5. POST http://localhost:4000/v1/chat/completions with "model": "auto" returns 400 There are no healthy deployments for this model, the same on /v1/responses and /v1/messages

After: the same config boots clean, decisions come back from the Databricks endpoint, and the auto router splits prompts across tiers

1. The admin adds the same model_list entries and starts the proxy
2. The proxy log is clean and GET http://localhost:4000/v1/models lists auto next to databricks-decider
3. They send the same POST http://localhost:4000/v1/decisions and get 200 with an answer per question (is_billing, intent: duplicate_charge, urgency) and x-litellm-model-name: databricks/databricks-openjev-qwen35-4b
4. POST http://localhost:4000/v1/chat/completions with "model": "auto" and "What is the capital of France?" is answered by the SIMPLE tier (x-litellm-model-name: openai/gpt-5.6-luna), the same on /v1/responses and /v1/messages
5. The same route with a lock-free ring buffer design question is answered by the COMPLEX tier (x-litellm-model-name: openai/gpt-5.6)

Linear ticket

Resolves LIT-9315

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/unit/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Screenshots / Proof of Fix

Last updated: 5cfd0b2. QA and /live-pr-risk ran on 5cfd0b2; the Admin UI leg ran on 1ccf0ee, and the commits after it change no dashboard code

Two proxies, each booted with --num_workers 2 from its own worktree and venv, on random ports: Before from the merge base 9e20264607, After from the tip 5cfd0b24aa. Both proxies run with LITELLM_LOCAL_MODEL_COST_MAP=True so each side reads its own checkout's cost map (the hosted map carries this PR's entry once it merges). Both load the same config; the Databricks workspace host is redacted as <workspace> in the outputs below. Real calls to the Databricks serving endpoint, TypeSafe, and OpenAI

model_list:
  - model_name: databricks-decider
    litellm_params:
      model: databricks/databricks-openjev-qwen35-4b
      api_base: os.environ/DATABRICKS_API_BASE
      api_key: os.environ/DATABRICKS_TOKEN
  - model_name: typesafe-decider
    litellm_params:
      model: typesafe/jev-latest
      api_key: os.environ/TYPESAFE_API_KEY
  - model_name: cheap
    litellm_params:
      model: openai/gpt-5.6-luna
      api_key: os.environ/OPENAI_API_KEY
  - model_name: strong
    litellm_params:
      model: openai/gpt-5.6
      api_key: os.environ/OPENAI_API_KEY
  - model_name: auto
    litellm_params:
      model: auto_router/complexity_router
      complexity_router_default_model: cheap
      complexity_router_config:
        tiers:
          SIMPLE: cheap
          COMPLEX: strong
        classifier_type: oss_classifier
        opensource_classifier_config:
          provider: databricks
          model: databricks-openjev-qwen35-4b
general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

decisions_body.json, the customer's support-triage shape:

{"model":"databricks-decider","state":{"message":"My card was charged twice for the same order and I need a refund.","channel":"support"},"questions":{"is_billing":{"type":"noul","instructions":"Is this a billing-related request?","criteria":{"true":"The message concerns a charge, payment, invoice, or refund.","false":"The message does not concern billing."}},"intent":{"type":"choice","instructions":"Which intent best matches the message?","criteria":{"refund":"The customer requests a refund.","duplicate_charge":"The customer reports being charged more than once.","other":"Another request."}},"urgency":{"type":"score","instructions":"How urgent is the request?","criteria":["Can wait","Needs attention soon","Urgent"]}}}

Before (9e20264)

The model list after boot

curl -s $PROXY/v1/models -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c "[.data[].id] | sort"

["cheap","databricks-decider","strong","typesafe-decider"]

# proxy stdout at boot:
16:20:57 - LiteLLM Router:ERROR: router.py:9430 - Error creating deployment: 1 validation error for ComplexityRouterConfig
opensource_classifier_config.provider
  Input should be 'jev', 'laya' or 'bespoke' [type=literal_error, input_value='databricks', input_type=str]

/v1/decisions with a databricks/ deployment

curl -s -i $PROXY/v1/decisions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d @decisions_body.json | grep -E "^HTTP|^x-litellm-model-name|^x-litellm-model-api-base|^\{"

HTTP/1.1 400 Bad Request
{"error":{"message":"litellm.BadRequestError: Unknown Decisions provider 'databricks'. Supported providers: perplexity, typesafe, openrouter, cloudflare, strands_decider\n\nLiteLLM: model group 'databricks-decider' failed with the error above. No fallback was attempted.","type":"invalid_request_error","param":null,"code":"400"}}

/v1/decisions control: the TypeSafe deployment, same body

curl -s $PROXY/v1/decisions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "$(jq -c ".model=\"typesafe-decider\"" decisions_body.json)" | jq -c "{model, intent: .answers.intent.choice, urgency: .answers.urgency.score, error: .error.message}"

{"model":"jev-1.13.0","intent":"duplicate_charge","urgency":1.16,"error":null}

auto router (provider: databricks), easy prompt on /v1/chat/completions

curl -s -D headers.txt $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"messages\":[{\"role\":\"user\",\"content\":\"What is the capital of France? Answer in one word.\"}]}" | jq -c "{model, content: .choices[0].message.content, error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:" headers.txt | tr -d "
"

{"model":null,"content":null,"error":"litellm.BadRequestError: You passed in model=auto. There are no healthy deployments for this model\n\nLiteLLM: model group 'auto' failed with the error above. No fallback was attempted."}
x-litellm-response-cost: 0

auto router (provider: databricks), hard prompt on /v1/chat/completions

curl -s -D headers.txt $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"messages\":[{\"role\":\"user\",\"content\":\"Design a lock-free multi-producer multi-consumer ring buffer for a 64-core NUMA system. Prove its linearizability, analyze false sharing under contention, and give the memory-ordering fences needed on ARMv9 versus x86-TSO. Keep it under 150 words.\"}]}" | jq -c "{model, content: .choices[0].message.content[:80], error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:" headers.txt | tr -d "
"

{"model":null,"content":null,"error":"litellm.BadRequestError: You passed in model=auto. There are no healthy deployments for this model\n\nLiteLLM: model group 'auto' failed with the error above. No fallback was attempted."}
x-litellm-response-cost: 0

auto router, easy prompt on /v1/responses

curl -s -D headers.txt $PROXY/v1/responses -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"input\":\"What is 2+2? Answer with just the number.\"}" | jq -c "{model, output_text: [.output[]?.content[]?.text] | join(\"\"), error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:" headers.txt | tr -d "
"

{"model":null,"output_text":"","error":"litellm.BadRequestError: You passed in model=auto. There are no healthy deployments for this model\n\nLiteLLM: model group 'auto' failed with the error above. No fallback was attempted."}
x-litellm-response-cost: 0

auto router, easy prompt on /v1/messages

curl -s -D headers.txt $PROXY/v1/messages -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"max_tokens\":20,\"messages\":[{\"role\":\"user\",\"content\":\"Name one primary color. One word.\"}]}" | jq -c "{model, text: [.content[]?.text] | join(\"\"), error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:" headers.txt | tr -d "
"

{"model":null,"text":"","error":"litellm.BadRequestError: You passed in model=auto. There are no healthy deployments for this model\n\nLiteLLM: model group 'auto' failed with the error above. No fallback was attempted."}
x-litellm-response-cost: 0

/model/info for the decider deployment, no model_info in the config

curl -s $PROXY/model/info -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c ".data[] | select(.model_name==\"databricks-decider\") | {model_name, model: .litellm_params.model, mode: .model_info.mode, input_cost_per_token: .model_info.input_cost_per_token, output_cost_per_token: .model_info.output_cost_per_token}"

{"model_name":"databricks-decider","model":"databricks/databricks-openjev-qwen35-4b","mode":null,"input_cost_per_token":null,"output_cost_per_token":null}

/health for the decider deployment

curl -s $PROXY/health -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c "{healthy: [.healthy_endpoints[] | select(.model | startswith(\"databricks\")) | {model, api_base}], unhealthy: [.unhealthy_endpoints[] | select(.model | startswith(\"databricks\")) | {model, api_base, exception_status, error: .error[:160]}]}"

{"healthy":[],"unhealthy":[{"model":"databricks/databricks-openjev-qwen35-4b","api_base":"https://<workspace>.azuredatabricks.net/serving-endpoints","exception_status":400,"error":"litellm.BadRequestError: DatabricksException - {\"error_code\":\"BAD_REQUEST\",\"message\":\"BAD_REQUEST: Invalid endpoint type: chat/completions is not supported by q"}]}

After (5cfd0b2)

The model list after boot

curl -s $PROXY/v1/models -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c "[.data[].id] | sort"

["auto","cheap","databricks-decider","strong","typesafe-decider"]

/v1/decisions with a databricks/ deployment

curl -s -i $PROXY/v1/decisions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d @decisions_body.json | grep -E "^HTTP|^x-litellm-model-name|^x-litellm-model-api-base|^\{" | tr -d "\r"

HTTP/1.1 200 OK
x-litellm-model-name: databricks/databricks-openjev-qwen35-4b
x-litellm-model-api-base: https://<workspace>.azuredatabricks.net/serving-endpoints
{"model":"/mosaicml/local_model","answers":{"is_billing":{"type":"noul","noul":0.9796676459153025},"intent":{"type":"choice","choice":"duplicate_charge","confidence":0.9652768882846836,"probabilities":{"refund":0.022973337505349804,"duplicate_charge":0.9768512588564555,"other":0.00017540363819449625}},"urgency":{"type":"score","score":1.9016174776422046,"confidence":0.8524262164633067,"legend":{"0":"Can wait","1":"Needs attention soon","2":"Urgent"},"probabilities":{"0":0.006936897169280873,"1":0.08450872801923372,"2":0.9085543748114854}}},"usage":{"input_tokens":238,"output_tokens":0}}

/v1/decisions control: the TypeSafe deployment, same body

curl -s $PROXY/v1/decisions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "$(jq -c ".model=\"typesafe-decider\"" decisions_body.json)" | jq -c "{model, intent: .answers.intent.choice, urgency: .answers.urgency.score, error: .error.message}"

{"model":"jev-1.13.0","intent":"duplicate_charge","urgency":1.16,"error":null}

auto router (provider: databricks), easy prompt on /v1/chat/completions

curl -s -D headers.txt $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"messages\":[{\"role\":\"user\",\"content\":\"What is the capital of France? Answer in one word.\"}]}" | jq -c "{model, content: .choices[0].message.content, error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:|^x-litellm-classifier-cost" headers.txt | tr -d "\r"

{"model":"auto","content":"Paris","error":null}
x-litellm-model-name: openai/gpt-5.6-luna
x-litellm-response-cost: 8.4e-06
x-litellm-classifier-cost: 0.0

auto router (provider: databricks), hard prompt on /v1/chat/completions

curl -s -D headers.txt $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"messages\":[{\"role\":\"user\",\"content\":\"Design a lock-free multi-producer multi-consumer ring buffer for a 64-core NUMA system. Prove its linearizability, analyze false sharing under contention, and give the memory-ordering fences needed on ARMv9 versus x86-TSO. Keep it under 150 words.\"}]}" | jq -c "{model, content: .choices[0].message.content[:80], error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:|^x-litellm-classifier-cost" headers.txt | tr -d "\r"

{"model":"auto","content":"Use LCRQ: linked, fixed-size power-of-two rings. Each cache-line-aligned slot is","error":null}
x-litellm-model-name: openai/gpt-5.6
x-litellm-response-cost: 0.053304000000000004
x-litellm-classifier-cost: 0.0

auto router, easy prompt on /v1/responses

curl -s -D headers.txt $PROXY/v1/responses -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"input\":\"What is 2+2? Answer with just the number.\"}" | jq -c "{model, output_text: [.output[]?.content[]?.text] | join(\"\"), error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:|^x-litellm-classifier-cost" headers.txt | tr -d "\r"

{"model":"auto","output_text":"4","error":null}
x-litellm-model-name: openai/gpt-5.6-luna
x-litellm-response-cost: 9.800000000000001e-06
x-litellm-classifier-cost: 0.0

auto router, easy prompt on /v1/messages

curl -s -D headers.txt $PROXY/v1/messages -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d "{\"model\":\"auto\",\"max_tokens\":20,\"messages\":[{\"role\":\"user\",\"content\":\"Name one primary color. One word.\"}]}" | jq -c "{model, text: [.content[]?.text] | join(\"\"), error: .error.message}"; grep -i -E "^x-litellm-model-name|^x-litellm-response-cost:|^x-litellm-classifier-cost" headers.txt | tr -d "\r"

{"model":"auto","text":"Blue","error":null}
x-litellm-model-name: openai/gpt-5.6-luna
x-litellm-response-cost: 8.8e-06
x-litellm-classifier-cost: 0.0

/model/info for the decider deployment, no model_info in the config

curl -s $PROXY/model/info -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c ".data[] | select(.model_name==\"databricks-decider\") | {model_name, model: .litellm_params.model, mode: .model_info.mode, input_cost_per_token: .model_info.input_cost_per_token, output_cost_per_token: .model_info.output_cost_per_token}"

{"model_name":"databricks-decider","model":"databricks/databricks-openjev-qwen35-4b","mode":"evaluation","input_cost_per_token":0.0,"output_cost_per_token":0.0}

/health for the decider deployment

curl -s $PROXY/health -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq -c "{healthy: [.healthy_endpoints[] | select(.model | startswith(\"databricks\")) | {model, api_base}], unhealthy: [.unhealthy_endpoints[] | select(.model | startswith(\"databricks\")) | {model, api_base, exception_status, error: .error[:160]}]}"

{"healthy":[{"model":"databricks/databricks-openjev-qwen35-4b","api_base":"https://<workspace>.azuredatabricks.net/serving-endpoints"}],"unhealthy":[]}

Admin UI (1ccf0ee)

The later commits add a cost-map entry and a unit test and touch no dashboard code, so the screenshots below stand.

The dashboard dev server (npm run dev in ui/litellm-dashboard, http://localhost:3000) pointed at a proxy booted with --num_workers 2 and store_model_in_db: true from each side. Steps to reproduce:

1. Open http://localhost:3000/models-and-endpoints/, click the "Auto-Routers" tab, then "Add Auto Router"
2. Name it auto-databricks-ui and pick the "OSS Classifier" classifier type
3. Before: the OSS provider row offers Jev, Laya, and Bespoke Nimble only, so a Databricks endpoint cannot be configured from the UI

4. After: a Databricks radio appears

5. Under "Advanced settings" -> "Classifier tuning", the Databricks note and the Classifier Model field take the serving endpoint name databricks-openjev-qwen35-4b

6. Simple and Medium tiers cheap, Complex and Reasoning tiers strong, Default model cheap, then "Test Routing" with a lock-free ring buffer prompt: tier COMPLEX, decided by the OSS classifier, classifier model databricks/databricks-openjev-qwen35-4b, confidence 79.2%, routed to strong

7. "Add Auto Router": the toast reads "Successfully created Auto Router: auto-databricks-ui" and the row shows up after "Refresh models"

8. The new router serves traffic through the proxy

curl -s -i $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d '{"model":"auto-databricks-ui","messages":[{"role":"user","content":"What is the capital of France? Answer in one word."}]}' | grep -E "^HTTP|^x-litellm-(model-name|model-group|complexity-router)|^\{" | tr -d "\r" | jq -R -r 'fromjson? | .choices[0].message.content // .'

HTTP/1.1 200 OK
x-litellm-model-name: openai/gpt-5.6-luna
x-litellm-model-group: auto-databricks-ui
x-litellm-complexity-router-tier: SIMPLE
x-litellm-complexity-router-cause: jev_classifier
Paris

9. Click the row, then "Edit Auto Router": the Databricks radio and the endpoint name come back filled in; "Save Changes" answers "Auto router configuration updated successfully", and the hard prompt still routes to the COMPLEX tier

curl -s -i $PROXY/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" -d '{"model":"auto-databricks-ui","messages":[{"role":"user","content":"Design a lock-free MPMC ring buffer for a 64-core NUMA system and prove linearizability. Under 100 words."}]}' | grep -E "^HTTP|^x-litellm-(model-name|model-group|complexity-router)" | tr -d "\r"

HTTP/1.1 200 OK
x-litellm-model-name: openai/gpt-5.6
x-litellm-model-group: auto-databricks-ui
x-litellm-complexity-router-tier: COMPLEX
x-litellm-complexity-router-cause: jev_classifier

Type

🆕 New Feature

Caveats (if any)

Low

• Decisions and classifier calls to the OpenJev endpoint cost 0 by the catalog: the databricks/databricks-openjev-qwen35-4b entry carries zero rates because Databricks bills the endpoint in DBUs and published no per-token price as of 2026-10-07 (its supported-models page is the entry's source). A deployment's model_info prices set the decisions cost (verified live: 0.000238 with prices set) and a databricks/<endpoint> price in a custom cost map prices the classifier; a guessed DBU conversion would pin a number nobody owns
• A databricks/* wildcard chat deployment now serves /v1/decisions: a missing endpoint or an over-long name answers Databricks's 404, and a chat, embeddings, or agent endpoint answers 400 with Databricks's own message, where the base answered 400 unknown provider. Backward incompatible on purpose, so one wildcard serves every endpoint; a per-endpoint allowlist would be new config nobody asked for
• A Databricks decider on an endpoint other than databricks-openjev-qwen35-4b (a custom serving endpoint wrapping OpenJev) has no catalog entry, so /health probes it as chat and Lens hides it until its deployment sets model_info: {mode: evaluation}. The catalog keys Databricks's own system endpoint, and guessing the mode from an arbitrary endpoint name would misclassify chat endpoints
• Decisions auth is a PAT (api_key, then DATABRICKS_API_KEY, then DATABRICKS_TOKEN); the OAuth M2M and databricks-sdk auth that chat deployments accept does not apply. OAuth for decisions is its own change against the shared Databricks auth helper
• An auto router added through /model/new or the dashboard answers 400 no healthy deployments until the next DB poll loads it, seconds to a few minutes. Pre-existing for every DB-added deployment
• A serving endpoint named like a legacy cost-map entry (dbrx-instruct, bge-large-en, gte-large-en, llama-2-70b-chat, meta-llama-3.1-70b-instruct) would price a decision with that entry through the shared Databricks prefix stripping. Pre-existing for chat deployments, and no ai_decide endpoint carries those names
• misc / Run tests and proxy-infra-root / Run tests (Unit Tests workflow, not ruleset-required) are red at the tip on test_every_redis_producer_declares_a_key_family and test_gateway_plus_backend_covers_full_app, the same two failures as main's latest completed run https://github.com/BerriAI/litellm/actions/runs/37716406166. This PR touches none of the files those tests read, so the fix belongs on main
• feat(decisions): serve System One format at /v1/systemone and OpenAI format at /v1/decisions #45184 moves the System One body to /v1/systemone; the PR landing second renames one test route and the docs. Nothing to change here until it lands
• feat(complexity_router): allow exact Jev classifier api_url for Cloudflare Clef #44445 adds an exact classifier api_url; the PR landing second merges one small _request_url hunk. Nothing to change here until it lands
• /v1/decisions keeps Databricks's own model string in the body; proxy headers and spend logs carry databricks/<endpoint>. Rewriting the body's model would change the provider's wire format for every caller to fix a cosmetic mismatch
• The Unity Catalog AI Gateway route is not used; the serving endpoint's /invocations route is, since it is the route Databricks documents for ai_decide endpoints and the one the customer's workspace exposes. The gateway route would need its own auth and body shape for a path nobody asked for

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

• be80d3e passes /live-pr-risk

• f875922 passes /live-pr-risk

• 5cfd0b2 passes /live-pr-risk

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[codecov]

Codecov Report

❌ Patch coverage is 98.86364% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...itellm/proxy/common_utils/registry_read_through.py	94.73%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Adds Databricks as a decisions provider and classifier option.

The PR appears safe to merge, with no new blocking issues found

What we checked:

• Classifier costs stay unchanged: Both the classifier's reported cost and its logged cost use input and output token prices only

Summary

Adds Databricks serving endpoints to /v1/decisions, the auto-router classifier, and the dashboard. Changes since the previous review only add zero cache prices to the matching entries in both catalogs

• No new actionable issues found
• Both earlier threads were checked. Endpoint names are validated before URL construction, and credential lookup and response rewriting now live in the provider adapter
• mateo-berri explicitly accepts zero catalog pricing, wildcard decisions support, PAT-only auth, custom endpoints needing mode: evaluation, existing deployment-load delays and legacy price-name collisions, and retaining Databricks's response model

_{Reviews (6) · Last reviewed commit: "fix(cost-map): declare zero cache rates ..." · Reviewed by Greptile}

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_databricks_ai_decide_decisions (3acae4b) with main (5b8f0a4)}

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit f875922. Configure here.}

[mateo-berri]

@greptileai

dismissed