feat(databricks): add Databricks Model Serving + AI Gateway provider#26510
Open
dgokeeffe wants to merge 2 commits intoanomalyco:devfrom
Open
feat(databricks): add Databricks Model Serving + AI Gateway provider#26510dgokeeffe wants to merge 2 commits intoanomalyco:devfrom
dgokeeffe wants to merge 2 commits intoanomalyco:devfrom
Conversation
added 2 commits
May 9, 2026 16:55
Re-synced onto current upstream/dev. Brings the Databricks integration as
a single, cleanly-organised commit on top of upstream:
- packages/opencode/src/auth/index.ts: add DatabricksProfile auth class
(databricks-profile type) to the Auth.Info union for profile-based
Databricks SDK auth.
- packages/opencode/src/provider/provider.ts: register `databricks` as a
custom provider. Auto-discovers serving-endpoints, classifies model
family (Claude / GPT / Gemini / Llama / Qwen / Gemma / Codex), routes
per-family through @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/openai or
the bundled @databricks/ai-sdk-provider as appropriate. Probes
/ai-gateway/anthropic/v1/models at startup; if 200, uses AI Gateway
URLs (anthropic/v1, gemini/v1beta, codex/v1); else falls back to
/serving-endpoints. Override via provider.databricks.options.surface =
"auto" | "ai-gateway" | "model-serving" (default "auto"). Per-request
Databricks SDK auth via dbConfig.authenticate(headers) handles OAuth
token refresh transparently — no background thread needed.
SSE patcher for the OpenAI Responses path (used by GPT on either
surface) addresses two server-side quirks: (a) item IDs up to ~192
chars (OpenAI Responses backend caps at 64) — truncate id/item_id/
call_id deterministically; (b) AI Gateway emits
response.output_item.added with one item id and the subsequent
response.content_part.added/output_text.delta with a different item_id
for the same output_index — track the canonical id per output_index
and rewrite mismatched item_ids on dependent events. Also tool-schema
type:object wrapper patching for the proxy's strict validator.
Optional outgoing-body / incoming-SSE workaround disable via
DATABRICKS_BARE_FETCH=1 env var; useful for verifying which patches
remain load-bearing on a given surface.
AI Gateway path drops the per-endpoint maxTools cap on gpt/codex
family (89 tools verified accepted server-side); model-serving keeps
the cap to avoid the 89-tool rejection.
- packages/opencode/src/provider/transform.ts: strip oversized itemIds
from outgoing assistant messages on Responses paths. Handles both
providerOptions.databricks.itemId (model-serving via bundled provider)
and providerOptions.openai.itemId (AI Gateway via @ai-sdk/openai).
- packages/opencode/src/session/llm.ts: middleware that synthesizes a
tool-input-start chunk before each tool-call and dedupes the bundled
provider's flush() re-emit by toolCallId. Gated on
npm === "@databricks/ai-sdk-provider" so it stays inert when GPT goes
through @ai-sdk/openai on the AI Gateway path.
DELETE-WHEN: drop this middleware once @databricks/ai-sdk-provider
ships a release that (a) emits the AI-SDK v3 tool-streaming lifecycle
on its Responses path and (b) stops setting providerExecuted: true in
flush() when useRemoteToolCalling is false. Currently broken in 0.5.0.
- package.json (catalog): @databricks/ai-sdk-provider 0.5.0 and
@databricks/sdk-experimental 0.16.0. @opentui/{core,solid} pinned to
0.2.0 (latest available on the Databricks internal npm proxy).
- .gitignore: keep local-only notes/reproductions/handover docs out of
the public fork.
Verified: test-databricks-3-classes.sh passes 3/3 against logfood
(model-serving) and aigw (ai-gateway) workspaces — Claude, GPT-5.5,
Gemini all complete tool-use roundtrips on both surfaces.
Co-authored-by: Isaac
- test-databricks-3-classes.sh: drives the full opencode HTTP API end-to-end against three model families (Claude / GPT / Gemini) on whatever Databricks surface is configured. Tests basic response and tool-call execution. - packages/opencode/script/test-databricks-3-classes.ts: SDK-level baseline that uses @ai-sdk/openai directly against /serving-endpoints. Useful for isolating opencode-stack vs raw-API issues during debugging — but note that on AI Gateway it bypasses @databricks/ai-sdk-provider's chunk emission, so a passing run here does NOT validate the bundled provider's path. - .gitignore: ignore models-snapshot.ts (auto-generated by build.ts). Co-authored-by: Isaac
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Issue for this PR
Closes #7983
Type of change
What does this PR do?
Adds
databricksas a custom provider with auto-discovery across both surfaces a Databricks workspace can expose:/serving-endpoints)/ai-gateway)At init the provider probes
/ai-gateway/anthropic/v1/models. If reachable it routes Claude / Gemini / GPT through the official AI-SDK adapters at AI Gateway URLs. If the probe 404s it falls back to/serving-endpointsand routes GPT through@databricks/ai-sdk-provider's bundled Responses adapter. Can be forced viaprovider.databricks.options.surface = "auto" | "ai-gateway" | "model-serving".Auth uses
@databricks/sdk-experimental's standard chain (profile / OAuth / PAT / env). Token refresh is per-request — no background timer.Three Databricks-side quirks are patched client-side and gated so they go inert when not needed:
type: "object"wrapper.databricksFetchpatches outgoing schemas before send. No-op when the SDK already wraps them.id/item_id/call_iddeterministically and strips oversizedproviderOptions.{databricks,openai}.itemIdon outgoing assistant messages. GPT only.response.output_item.addedandresponse.content_part.added/output_text.deltafor the sameoutput_indexon AI Gateway. Patcher tracks the canonical id peroutput_indexand rewrites mismatcheditem_ids on dependent events. GPT/Responses only.session/llm.tssynthesises atool-input-startchunk before eachtool-calland dedupes the bundled provider'sflush()re-emit bytoolCallId. Gated onnpm === "@databricks/ai-sdk-provider"so it stays inert when GPT goes through@ai-sdk/openaion the AI Gateway path. MarkedDELETE-WHENwith the exact upstream conditions for retirement.bun.lockis intentionally not in this PR —bun installafter merge picks up@databricks/ai-sdk-provider@0.5.0and@databricks/sdk-experimental@0.16.0from the catalog.How did you verify your code works?
End-to-end via
test-databricks-3-classes.sh(full opencode agent loop) against both surfaces with Claude Sonnet 4.6, GPT-5.5, and Gemini 2.5 Pro: 3/3 pass on both.To reproduce:
bun install./test-databricks-3-classes.sh— should pass 3/3A SDK-level baseline at
packages/opencode/script/test-databricks-3-classes.tsexercises@ai-sdk/openaidirectly against/serving-endpoints, useful for isolating opencode-stack vs raw-API issues.Screenshots / recordings
N/A — no UI changes.
Checklist