dbx-tools-litellm

Thin LiteLLM integration for Databricks Model Serving. It adds live endpoint discovery, loose model-name resolution, and a small set of documented serving compatibility guards, then delegates to LiteLLM’s built-in Databricks provider.

Install from PyPI:

uv add dbx-tools-litellm

To install the current main branch directly from the repository instead:

uv add "dbx-tools-litellm @ git+https://github.com/reggie-db/dbx-tools.git@main#subdirectory=packages/py/litellm"

Key features

• accepts an optional --profile override, otherwise resolving DATABRICKS_CONFIG_PROFILE, then the one marked as the Databricks CLI default;
• discovers serving endpoints from the selected workspace and caches them per process;
• advertises discovered endpoints and family aliases only under dbx/* by default, keeping them distinct from LiteLLM’s native databricks/* provider;
• resolves exact or fuzzy model names with dbx-tools-model, refreshing the live catalogue once after a miss;
• restricts tool-bearing requests to endpoints classified as tool-capable;
• routes Responses-only models through LiteLLM’s databricks/responses/... bridge;
• resolves Responses-only proxy calls before provider selection so LiteLLM’s native Databricks Responses implementation receives the original body;
• optionally classifies an auto reasoning effort as low, medium, or high for reasoning-capable OpenAI and Claude endpoints;
• marks a stable prefix of Claude requests for Anthropic prompt caching, which GPT endpoints get automatically on the native Responses surface;
• floors rate-limit backoff to the token-per-minute window so a retry lands in a fresh budget instead of amplifying the limit;
• supports LiteLLM chat, embedding, synchronous/asynchronous, and streaming entrypoints, rewriting request content only for the Databricks serving constraints described under Request processing.

Run the proxy

uv run dbx-litellm --port 4000

The launcher listens on 127.0.0.1 by default. Pass an explicit LiteLLM --host value or set HOST to expose it on another interface. --profile is available as an override but is not required.

The equivalent module invocation is:

uv run python -m dbx_tools.litellm --port 4000

Then point an OpenAI-compatible client at http://127.0.0.1:4000/v1:

curl http://127.0.0.1:4000/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"dbx/databricks-claude","messages":[{"role":"user","content":"hi"}]}'

The resolved profile is written to DATABRICKS_CONFIG_PROFILE, so endpoint discovery and LiteLLM’s delegated Databricks request use the same workspace credentials.

How a request flows

The package has two paths because LiteLLM’s CustomLLM interface handles Chat Completions and embeddings, but does not expose a native Responses hook:

1. Select one workspace. An explicit --profile overrides DATABRICKS_CONFIG_PROFILE; when neither is present, dbx-litellm runs databricks auth profiles --output json --skip-validate and selects the one entry marked "default": true. It writes the result back to DATABRICKS_CONFIG_PROFILE before LiteLLM starts.
2. Discover and resolve the model. The first model-dependent request lazily calls the selected workspace’s Serving Endpoints API. An exact endpoint name, a family alias such as dbx/databricks-claude, or a loose name such as claude sonnet is ranked against that live catalogue. A request containing function tools can match only an endpoint classified as tool-capable.
3. Choose the serving surface. Chat-compatible endpoints stay on Chat Completions. Responses-only endpoints, including newer GPT endpoints that reject tool calls on Chat Completions, are rewritten to databricks/responses/<endpoint>. The Responses request body itself is not converted or reconstructed by this package.
4. Apply opt-in reasoning. An explicit numeric or named effort is normalized to a level supported by the resolved endpoint. Only the literal value auto invokes the classifier. An omitted effort is a pass-through and lets the model use its own default.
5. Apply compatibility guards. The payload hook downsizes oversized inline images. Delegated Chat requests repair unsupported trailing assistant turns, add the required JSON-mode prompt nudge when needed, and mark Claude prompt cache breakpoints. Function tools are never rewritten.
6. Inject cached credentials. The package supplies an explicit bearer token and serving base URL from its process-wide credential cache. This keeps LiteLLM from constructing a new WorkspaceClient and authenticating again for every request.
7. Delegate to LiteLLM. LiteLLM owns HTTP transport, OpenAI parameter mapping, streaming, retries, embeddings, and Chat↔Responses conversion. The response streams back in LiteLLM’s normal OpenAI-compatible shape.

In short, the package decides which live Databricks endpoint and API surface to use, performs a few documented serving compatibility fixes, and then gets out of LiteLLM’s way.

Profiles and authentication

Profile selection happens once, at proxy startup, in this order:

1. optional dbx-litellm --profile <name> override;
2. DATABRICKS_CONFIG_PROFILE;
3. the one profile marked as the Databricks CLI default.

Startup fails rather than guessing when none of those produces exactly one profile. The same selected profile creates one process-wide WorkspaceClient used for both endpoint discovery and authentication, so model names cannot be pulled from one workspace while requests are sent to another.

The package does not introduce another authentication scheme. The Databricks SDK resolves the selected profile’s configured authentication, including OAuth machine-to-machine credentials, and authenticate() supplies its bearer token. That token and the workspace’s /serving-endpoints base URL are passed directly to LiteLLM’s built-in Databricks provider. SDK background token refresh is disabled because this package’s guarded credential cache is the sole refresh owner.

To override the Databricks CLI default for a process, set the environment:

DATABRICKS_CONFIG_PROFILE=my-workspace uv run dbx-litellm --port 4000

Model discovery and names

Models are pulled from the selected workspace’s live Serving Endpoints API, not from a static list in this package:

• discovery is lazy on the first request that needs model resolution;
• the successful endpoint catalogue is retained in memory for the process;
• exact endpoint names and fuzzy family names are ranked by dbx-tools-model;
• a miss forces one fresh endpoint listing and retries resolution once;
• tool-bearing requests filter out endpoints not classified as tool-capable;
• resolving a model also registers its native streaming capability with LiteLLM, preventing unknown Databricks models from being buffered as fake streams.

GET /v1/models is intentionally a live refresh point. It requests a fresh endpoint listing, publishes each endpoint as dbx/<endpoint>, and adds one resolvable alias for each recognized deployed family, such as dbx/databricks-gpt or dbx/databricks-claude. Exact endpoint ids remain in the response; aliases supplement rather than replace them. If that refresh fails, the route falls back first to the last successful in-process catalogue and then to LiteLLM registry metadata. It never invents a workspace endpoint from the registry when live discovery succeeded.

The packaged proxy advertises the dbx/* namespace so callers can distinguish this discovery-and-routing layer from LiteLLM’s native databricks/* provider. Unqualified names remain accepted for fuzzy resolution, but are not advertised. A custom config can expose both namespaces.

What is cached

There is no single “LiteLLM cache” in this integration. Four independent caches serve different purposes:

Cache	Storage and scope	Filled when	Refresh or expiry	Purpose
Endpoint catalogue	Memory, one proxy process	First resolution or /v1/models	Forced once after a resolution miss; /v1/models always refreshes	Avoid listing Serving Endpoints on every request
Databricks bearer token	Memory, one proxy process/profile	First delegated request needing credentials	OAuth expiry minus 10 minutes; 30-minute fallback when no expiry is exposed; check-lock-check makes one caller refresh	Avoid a new SDK client and token mint per request
Reasoning context and scores	diskcache, default ~/.cache/dbx-tools/litellm, shared by local processes using that directory	Only for reasoning_effort: auto or reasoning.effort: auto	TTL, default 86,400 seconds; bounded to 64 MiB, eight turns, and 6,000 sampled characters	Reuse follow-up context and avoid classifying the same sample again
Provider prompt cache	Databricks/model-provider managed	Repeated prompt prefixes	Provider-defined lifetime and eviction	Reduce billed/counted repeated input tokens; GPT is automatic, Claude uses explicit breakpoints added here

The endpoint and credential caches are not written to disk. Restarting the proxy clears both. The reasoning cache is the only package-owned persistent cache and can be moved or assigned a shorter TTL with the environment variables under Automatic reasoning effort. Prompt-cache contents remain provider-side; this package only supplies Claude’s cache markers and reports the usage fields returned by the provider.

Relationship to LiteLLM

LiteLLM remains the proxy and provider implementation. It owns the OpenAI-compatible routes, Databricks transport, parameter mapping, streaming semantics, retries, embeddings, and Chat↔Responses conversion. The Databricks SDK owns authentication; this package caches its bearer result and passes it to LiteLLM explicitly so the provider does not construct a client per request.

This package supplies only the workspace-specific layer LiteLLM does not have: deterministic profile selection, live endpoint discovery, fuzzy names, and capability-aware routing. Request messages and content blocks are rewritten only to satisfy concrete Databricks serving constraints — the ordered pipeline under Request processing (trailing-assistant repair, JSON nudge, Claude prompt-cache marking) and the image payload guard — or when the caller explicitly requests automatic reasoning selection. Function tools pass through unchanged; the LiteLLM 1.96.2 namespace workaround unwraps only the grouping object around those functions.

Pinned compatibility

The package pins litellm[proxy] to 1.96.2 because its two narrow startup patches target private LiteLLM APIs:

• non-Claude reasoning_effort must not enter the thinking-token helper without a thinking block;
• Responses namespace tools must be unwrapped before LiteLLM’s Chat bridge, which still drops that wrapper in 1.96.2.

The proxy also caps FastAPI below 0.140.7. That FastAPI release removed get_flat_dependant, which LiteLLM 1.96.2 imports during proxy startup. Remove the cap after LiteLLM ships its get_flat_params compatibility fix. Upgrade LiteLLM only with the full offline suite and proxy startup smoke test.

LiteLLM 1.96 loads custom handlers from a Python file beside the config. For an existing LiteLLM config, add config_provider.py next to the YAML:

from dbx_tools.litellm.access_log import dbx_access_logger
from dbx_tools.litellm.payload_guard import dbx_payload_guard
from dbx_tools.litellm.provider import dbx_provider
from dbx_tools.litellm.reasoning import dbx_auto_reasoning
from dbx_tools.litellm.routing import dbx_responses_router

Then register that adjacent shim under dbx:

model_list:
  - model_name: "dbx/*"
    litellm_params:
      model: "dbx/*"
      allowed_openai_params:
        - reasoning_effort
        - thinking
        - parallel_tool_calls

litellm_settings:
  callbacks:
    - config_provider.dbx_payload_guard
    - config_provider.dbx_auto_reasoning
    - config_provider.dbx_responses_router
    - config_provider.dbx_access_logger
  custom_provider_map:
    - provider: dbx
      custom_handler: config_provider.dbx_provider

The packaged config advertises only dbx/*. A consumer config can opt into LiteLLM’s native Databricks provider independently:

model_list:
  - model_name: "databricks/*"
    litellm_params:
      model: "databricks/*"

Set DATABRICKS_CONFIG_PROFILE or pass --profile to override the Databricks CLI default.

Automatic reasoning effort

Automatic effort is opt-in. On Chat Completions, send "reasoning_effort": "auto":

{
  "model": "claude sonnet",
  "messages": [{ "role": "user", "content": "Debug this distributed deadlock" }],
  "reasoning_effort": "auto"
}

On Responses, use the native reasoning shape:

{
  "model": "gpt 5 codex",
  "input": "Debug this distributed deadlock",
  "reasoning": { "effort": "auto" }
}

The callback resolves databricks-meta-llama-3-1-8b-instruct against the live catalogue as its default classifier preference, asks the discovered endpoint for a score from 0.01 through 1.00, then maps that score through the target Databricks endpoint’s inferred reasoning levels. The default mapping is minimal at or below 0.05 when available, low below 0.34, medium below 0.67, xhigh at or above 0.85 when available, and otherwise high. An exact 1.00 selects max when the endpoint exposes it. Chat Completions for GPT-5.6 excludes max; the native Responses path can use the endpoint’s full set. Integer classifier output is treated as a percentage (73 becomes 0.73), except 1, which remains the maximum score.

Explicit named or numeric selectors do not invoke the classifier, but they are normalized through the resolved endpoint’s supported levels. A native thinking object takes precedence and is passed through after removing the competing effort selector. An omitted or default selector is a true pass-through. Unsupported targets have auto removed and use their provider default.

The classifier sees at most eight recent non-system turns and 6,000 characters. Full Chat transcripts are sampled directly. Short follow-ups can recover prior turns from metadata.thread_id, metadata.conversation_id, or metadata.session_id; successful Responses calls index that bounded context by response id so a later previous_response_id can recover it. Scores are keyed by a SHA-256 hash of the complete bounded sample. Context and scores use diskcache with the same TTL, so retries and identical follow-ups avoid repeated classifier calls without retaining an unbounded transcript. A classifier timeout, malformed score, or empty sample falls back to 0.50.

Configuration:

• DBX_TOOLS_LITELLM_REASONING_MODEL overrides the classifier endpoint;
• DBX_TOOLS_LITELLM_REASONING_CACHE_DIR changes the disk-cache directory;
• DBX_TOOLS_LITELLM_REASONING_CACHE_TTL_SECONDS sets the context and result TTL (default: 86,400 seconds);
• DBX_TOOLS_LITELLM_REASONING_TIMEOUT_SECONDS sets the classifier timeout (default: 5 seconds).

For Claude targets, LiteLLM’s Databricks transformer maps the selected reasoning_effort to the backend’s native extended-thinking token budget.

The one-line dbx-access record includes thinking_requested=<level> for every request. Automatic requests also include thinking_selected=<level> after the classifier maps the score through the resolved model’s capabilities. The existing reasoning=<tokens> field remains the number of reasoning tokens reported by the provider, not the selected effort level.

Request processing

Every delegated Chat Completions request runs through a small, ordered pipeline (provider._prepare_messages) before it reaches Databricks. Each step exists to satisfy a concrete Databricks serving constraint that an OpenAI-style client does not know about. Order matters, because each step can change what the next one sees:

1. Trailing-assistant repair (_repair_trailing_assistant). Databricks rejects a transcript whose last message is an assistant turn with “This model does not support assistant message prefill. The conversation must end with a user message.” Codex hits this on retry, when a stream that disconnected mid-turn is resumed with its partial answer replayed as the final message. The repair drops trailing assistant text turns so the transcript ends where the model can continue. It preserves trailing tool_calls / function_call, which the client may be about to answer, and never empties the list.
2. JSON nudge (_ensure_json_mentioned). OpenAI-family endpoints refuse response_format: {"type": "json_object"} unless the prompt itself contains the word “json”. This is a prompt-content rule, so no parameter filtering satisfies it. When json mode is requested but unmentioned, the nudge appends a short instruction to the last non-system turn — the one role guaranteed to survive into input on the Responses bridge. This is exactly how Mem0’s memory extraction trips the rule; the nudge fixes every client at once. Runs after the repair so it never appends to a turn that is then dropped.
3. Prompt-cache marking (_apply_prompt_cache, Claude only). See below. Runs last so its breakpoints land on boundaries the earlier steps have already settled.

The image payload guard (payload_guard.DbxPayloadGuard) is a separate pre-call hook, not part of the message pipeline. Databricks rejects any request body over 32 MiB; chat clients inline uploaded images as base64 and resend them every turn, so a couple of photos push a long chat past the cap and every turn then fails with an opaque 400. The guard measures the serialized request and, if it is over target, downscales base64 images (largest first) with Pillow until it fits, raising a clear size-named error only if it still cannot.

Prompt caching

Caching behaviour differs by model family because the two Databricks serving surfaces expose it differently. The proxy leaves the automatic case alone and fills the explicit case that OpenAI-style clients never trigger.

• GPT (native Responses): GPT-5.4+ endpoints route through LiteLLM’s databricks/responses/... bridge to the native Responses surface, which applies automatic, OpenAI-style prefix caching. No marking is needed; a repeated prefix reads from cache and reports cached_tokens. Changing reasoning.effort between turns does not evict the cache, because effort is a top-level parameter and not part of the cached input prefix.
• Claude (emulated Responses / chat): Databricks refuses the native Responses passthrough for Claude (“Responses API passthrough is not supported for model databricks-claude-…”), so these turns go through LiteLLM’s Responses-to-Chat emulation onto chat/completions. Anthropic caching on Databricks is explicit: a request is cached only where a content block carries cache_control. OpenAI-style clients (Codex, Open WebUI) never send it and the emulation does not add it, so without intervention the whole transcript is re-billed as fresh input every turn — which repeatedly trips the workspace input-tokens-per-minute limit on long chats.

_apply_prompt_cache closes that gap for Claude targets by stamping cache_control: {"type": "ephemeral"} on two rolling breakpoints: the first system message (stable for the life of the chat) and the last stable turn (the message before the volatile final turn, already present and cache-written on the previous turn). Anthropic matches the longest cached prefix at each breakpoint, so two breakpoints cache effectively the whole history except the newest turn. The final turn is left unmarked because it is new every request and would only ever write, never read. This is a no-op for non-Claude models and for single-turn requests, which have no stable prefix. LiteLLM’s Databricks transformer preserves cache_control for Claude, so marking the blocks here is sufficient; the endpoint honours it and returns cache_creation_input_tokens and cache_read_input_tokens.

Databricks disables the stateful Responses store (store / previous_response_id) workspace-wide by default, so the full transcript is re-sent every turn on both families. Caching is what keeps the re-sent prefix from being billed and rate-limited each time.

Rate-limit retries

The packaged router retries rate limits five times with exponential backoff and honors provider Retry-After headers. Timeouts and internal server errors get three retries. Authentication, bad requests, and content-policy failures are not retried.

Databricks’ REQUEST_LIMIT_EXCEEDED is a per-minute token budget, and a retry re-sends the whole request body. Retrying inside the same minute only adds more tokens to an already-exceeded window and cannot succeed — the amplification that turns one rate limit into a spiral of failed reconnects. So when the server does not send a Retry-After, rate-limit backoff is floored to the rate-limit window (RATE_LIMIT_WINDOW_SECONDS) so every retry lands in a fresh window rather than piling into the current one. A server Retry-After, when present, is authoritative and overrides the floor.

Streaming dbx requests apply the same bounded retry protection when a rate limit arrives before the first response chunk. A failure after content has already streamed is returned immediately because restarting would duplicate partial output in the client.

Responses routing

LiteLLM’s CustomLLM interface has no native Responses hook. The pre-call router therefore resolves the model before LiteLLM selects a provider. For a Responses-only endpoint it changes only the model identifier to the native Databricks Responses route and injects the cached api_key and api_base; LiteLLM’s Databricks Responses implementation receives the original body. Chat-compatible families use LiteLLM’s normal Responses-to-Chat fallback.

The same policy also protects Chat Completions callers: GPT family versions known to reject function tools on Chat Completions are delegated through LiteLLM’s databricks/responses/... bridge. This keeps clients on one OpenAI-compatible proxy URL while selecting the Databricks surface that the resolved endpoint actually supports.

Validate

Run the offline package checks:

uv run pytest packages/py/model packages/py/litellm -q
uv run ruff check packages/py/model packages/py/litellm
uv run dbx-litellm --help

Then start the proxy and exercise live discovery before inference:

uv run dbx-litellm --port 4000
curl -fsS http://127.0.0.1:4000/health/readiness
curl -fsS http://127.0.0.1:4000/v1/models

Release validation also covers non-streaming Chat, streaming Chat, Responses, function-call replay, namespace tools, and embeddings. An embedding request must preserve an exact embedding endpoint id rather than fuzzy-routing it to a chat endpoint.

Modules

• backend - profile-resolved workspace client, endpoint cache, and model resolution;
• models — Responses-only endpoint routing policy;
• provider — LiteLLM CustomLLM adapter and exported dbx_provider singleton; owns the Chat Completions message pipeline (trailing-assistant repair, JSON nudge, Claude prompt-cache marking) and the rate-limit-aware streaming retry;
• payload_guard — pre-call hook that downscales oversize base64 images to keep requests under the 32 MiB serving limit;
• reasoning — opt-in effort classification and TTL-backed follow-up context;
• routing — model-only proxy hook for native Responses-only calls;
• access_log — one-line per-request dbx-access telemetry;
• patches - version-specific LiteLLM startup guards;
• cli - profile-resolving launcher for the packaged LiteLLM proxy config.

For standalone Python endpoint resolution and invocation helpers, use dbx-tools-model. For the TypeScript local proxy, use @dbx-tools/cli-model-proxy.