[Rana Singh (ranadeepsingh)]

Why this change

SynapseML's OpenAI text transformers currently require callers to select a deployment, and DataFrame workloads cannot express the new prompt-cache options and input breakpoints. This makes it harder to use gpt-6-luna by default or configure caching for repeated instructions and reference material. Missing cache-write accounting and incorrectly forwarded inference controls also make it difficult to understand cache behavior or reliably configure requests.

This PR defaults OpenAIChatCompletion, OpenAIResponses, and OpenAIPrompt to gpt-6-luna only when the deployment is unset, and adds typed cache options and object-valued input breakpoints. Explicit scalar, column, and global deployments still win; supported Prompt model aliases, embedding models, and Foundry model selection are preserved. No deployment is created or changed.

The notebook includes native SDK and SynapseML examples, explicit and implicit caching, separate cache-read/write usage, prewarming, diagnostics, and the supported inference controls. The service controls eligibility, modes, and lifetimes; cache hits and savings are not guaranteed.

Current text-generation examples use gpt-6-luna, with gpt-6.1-sol documented as the larger-model alternative for document question answering. The remaining older GPT example has been migrated. Archived release documentation and embedding models are intentionally unchanged.

Cache and inference-control audit

The audit found and fixed client-side gaps; it did not establish that every hosted deployment supports every cache option.

Area	Fix
Cache requests	Added scalar/column promptCacheKey and the separate legacy promptCacheRetention control. Existing request options and ordered breakpoint objects retain mixed TTLs and comparison IDs.
Cache accounting	Raw outputs and Prompt usage now retain cache_write_tokens; Responses also retains the ttl_30m_tokens / ttl_12h_tokens breakdown. Reads and writes stay distinct, and absent counters remain null.
Diagnostics and prewarming	Responses retains applied options and cache diagnostics. A completed prewarm with empty output returns null Prompt text without a false content-filter error; usage and stored IDs are retained.
Log probabilities	Fixed Chat's integer-to-boolean wire mismatch. Added logProbsEnabled / topLogProbs; the legacy integer logProbs API remains usable. Responses requests the corresponding include data, and raw schemas retain the returned probabilities.
Other controls	Added serviceTier, safetyIdentifier, metadata, Responses reasoning / include / truncation, and Chat logitBias / stopSequences. Chat storage now works through Prompt, including per-row settings and response IDs.
Explicit choices	Responses no longer silently drops temperature / top_p based on a model-name heuristic. Conflicting controls and known unsupported API-specific parameters fail explicitly. Model-specific limits remain provider-validated.

Generated Python constructors, setParams, column overrides, copy and save/load are covered through the real JVM/HTTP path. Additive wire schemas preserve public case-class constructors and legacy usage mappings.

Follow-up review fixes

The four open findings exposed request failures and compatibility regressions:

• Responses log probabilities now use the SDK-supported message.output_text.logprobs selector, including deduplication and conflict detection.
• Saved Foundry prompts retain the no-storage default. Explicit store=false is also a no-op; actual storage requests remain unsupported on that route.
• Per-row storage flags are read before output or usage columns replace them. True, false and null flags work even when their input column becomes the output, usage or response-ID column, with matching runtime and declared schemas.
• Custom usage mappings again treat dotted keys as literal field names. Responses TTL accounting uses explicit nested paths without changing the public mapping constructor or normalization signature. The newly introduced detailed Responses mapping is package-internal, alongside its nested-field normalizer, rather than exposed as an incompatible option for the generic public normalizer.

Boundary: local HTTP captures prove serialization and response handling. Separate live qualification below now establishes cache reads and write accounting on the tested deployment, but not universal provider support, prewarm effectiveness or cost savings. Streaming, tool orchestration, background polling, arbitrary conversation items, audio output and prediction remain outside these synchronous text transformers.

GPT-6 options and cache breakpoints: end-to-end flow

Yes: both the request-level options and nested cache breakpoint objects reach the HTTP request. The API field is prompt_cache_breakpoint (not a Spark checkpoint). This was verified through installed Python wrappers and real Spark transforms against a loopback HTTP receiver using the exact validated JARs, not just setter/getter assertions.

Python setters + typed DataFrame message content
  -> generated Python wrapper / JVM transformer
  -> ServiceParam resolves scalar values or each row's options column
  -> getOptionalParams maps options to the API's JSON fields
  -> message encoder validates and preserves nested breakpoint objects
  -> prepareEntity + request-body encoder produce application/json
  -> POST /openai/v1/responses or /openai/v1/chat/completions
  -> response/error columns and the API's cached-token usage

Input	Responses JSON	Chat Completions JSON
setPromptCacheOptions({"mode": "explicit", "ttl": "30m"})	Top-level prompt_cache_options object	Same
setPromptCacheOptionsCol("options")	Per-row prompt_cache_options; replaces the scalar binding	Same
Content-part prompt_cache_breakpoint={"mode": "explicit"}	input[].content[].prompt_cache_breakpoint object	messages[].content[].prompt_cache_breakpoint object
setReasoningEffort("low")	reasoning.effort	reasoning_effort
setVerbosity("low")	text.verbosity	Top-level verbosity
setMaxCompletionTokens(n)	max_output_tokens	max_completion_tokens

Example: explicit caching on a stable prefix

Using rows, endpoint and key from the complete runnable example below:

stage = (
    OpenAIResponses()
    .setUrl(endpoint)
    .setSubscriptionKey(key)
    .setMessagesCol("messages")
    .setPromptCacheOptions({"mode": "explicit", "ttl": "30m"})
    .setReasoningEffort("low")
    .setVerbosity("low")
    .setOutputCol("response")
    .setErrorCol("error")
)
result = stage.transform(rows)

With no deployment override or global deployment configured, the outgoing v1 request selects gpt-6-luna. Put the breakpoint on the stable developer-message content, before the changing user question. The captured Responses request body is shown below with only the long catalog text abbreviated:

{
  "prompt_cache_options": {
    "mode": "explicit",
    "ttl": "30m"
  },
  "model": "gpt-6-luna",
  "store": false,
  "text": {
    "verbosity": "low"
  },
  "reasoning": {
    "effort": "low"
  },
  "input": [
    {
      "role": "developer",
      "content": [
        {
          "type": "input_text",
          "text": "<stable product catalog prefix; abbreviated here>",
          "prompt_cache_breakpoint": {
            "mode": "explicit"
          }
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "What is Product 17's warranty?"
        }
      ]
    }
  ]
}

The equivalent Chat request uses messages, content type text, top-level reasoning_effort and verbosity; the two cache fields remain JSON objects in the same relative positions. Unset options or a null options-column value omit the top-level field. Explicit null members inside an options object remain JSON null. A null breakpoint is omitted; a non-null scalar/string breakpoint is rejected locally into the error column rather than sent as a quoted JSON string.

Important boundary: OpenAIPrompt forwards request-level options to its selected Chat/Responses stage, but ordinary prompt templates do not create explicit breakpoint markers. Use implicit mode for those templates, or use OpenAIResponses/OpenAIChatCompletion with typed message content for explicit breakpoints. Options alone do not enable explicit cache reads/writes. The provider still decides eligibility and cache use; successful forwarding does not prove a cache hit.

Wire evidence: three public-transform HTTP captures passed: Chat scalar options, Responses scalar options, and Responses column options overriding a scalar value. Each capture verified the model, reasoning/verbosity placement, complete prefix and question, object-valued breakpoint on the prefix, and absence of a breakpoint on the variable question. Those captures were local serialization checks; the separate live evidence below establishes hosted cache reads.

Implementation trace (current source; the captures above were recorded before this audit): shared option mapping, optional parameter collection, Responses option shaping, Responses JSON body, Chat JSON body, and Prompt delegation.

Implementation and compatibility

• Unset service defaults remain implicit after save/load, while explicit service values and ordinary saved defaults retain their existing semantics. Saved output-column names remain stable across Scala/Python loading and pipelines.
• Java/Scala decimals and SQL decimal option values preserve their exact JSON numbers, including through parameter save/load. Foundry subclasses do not inherit the OpenAI deployment fallback.
• Prompt schema validation resolves globals on a temporary copy. Schema-only calls and JVM Pipeline.fit neither freeze global settings nor materialize global credentials on the original stage for later persistence.
• Model-only Prompt configurations on ordinary legacy Azure Chat endpoints fail with deployment/v1 guidance instead of silently choosing a fallback. Existing custom-URL behavior, explicit deployments, v1 model aliases, Responses, and Foundry routes remain supported.
• Invalid breakpoint shapes fail locally through the error column. Fabric's existing model-availability preflight retains its failure behavior and adds conditional guidance about the OpenAI-as-a-Microsoft-Subprocessor setting.

Validation

Live cache qualification (2026-11-09)

Using the published revision's rebuilt JARs and installed Python wrappers, synthetic prefixes were sent to an existing Azure OpenAI deployment and the separate Foundry model-inference route. Each successful case asserted error-free responses and positive cache reads, rather than relying on displayed output. Interpreter, wrapper origins and JAR hashes were recorded.

Live path	Observed result
Azure Chat Completions	1,865 cache-write tokens on the native request; 1,865 cached tokens through SynapseML
Responses scalar options and copy	1,866 written, then 1,866 cached
Responses per-row 30m / 12h	1,862 / 1,864 written under the corresponding TTL counters, then the same positive cache reads; disjoint write counts sum correctly
Implicit Responses through saved/reloaded OpenAIPrompt	1,872 written, then 1,872 cached; normalized TTL accounting preserved
Foundry model inference with gpt-6-luna	1,868 written natively, then 1,868 cached in both SynapseML requests
Notebook explicit scalar, explicit column and implicit Prompt examples	All six rows passed new assertions with 4,101 cached tokens each
Untouched 30-minute interval	Follow-up completed 30m 44s after the seed: 0 cached tokens and 1,865 new 30-minute write tokens, with no request error; the positive-reuse assertion failed

Remaining qualification gaps: the tested Azure endpoint rejects prompt_cache_options.prewarm with HTTP 400, both natively and through SynapseML, including prewarm=false. The existing Foundry Llama deployment completes inference but reports no cache counters, so its caching is unverified, not passed. Low-level OpenAIResponses save/load hits the Java-UDF deserialization guard even without cache options; the unsafe override was not enabled. The successful persistence case above is specifically OpenAIPrompt, not every transformer.

12h acceptance and accounting do not prove 12 hours of retention. The timed check made no intervening requests and did not establish reuse beyond 30 minutes. Since the follow-up occurred after the requested minimum lifetime, its miss is not proof of a TTL violation. TTL is a minimum lifetime, not a deletion deadline. Foundry qualification uses matching default reasoning settings; it does not qualify every inherited inference-control setter. No cache-hit rate, latency or cost-saving guarantee is claimed.

The new opt-in test module and notebook assertions are local, uncommitted follow-up changes, not part of published commit b51b396704 or its CI run. The modified notebook's cache cells were executed live with OpenAI Python SDK 3.46.0. Provider limitations remain visible as failed qualification cases rather than being suppressed.

Current revision: b51b396704 makes only the new detailed Responses mapping package-internal. Four usage-normalizer tests and the Prompt TTL-accounting regression passed on this revision, along with production/test Scala style and cognitive JAR packaging. Existing public mappings and custom literal-key behavior are unchanged.

Preceding review-fix revision (historical full runtime validation): 880f483bf7. All four reported regressions were demonstrated before the fixes. 76 Scala tests passed; the three affected Foundry tests passed again after an equivalent style simplification (not additional tests). Freshly rebuilt and installed packages passed 28 Python tests and 28 subtests, plus seven loopback HTTP transforms. Production/test Scala style, changed Python formatting, wrapper generation and notebook structure checks passed. These results establish local request, persistence and schema behavior, not hosted cache performance.

Earlier rebase validation (historical):

Rebased onto current master: 2570bc2749 incorporates master 6c2699defc. All 12 original commits retain their authors and messages. Two test-only conflicts were resolved by keeping both sets of regressions. A small follow-up moves unchanged HTTP fixture code into a test base and removes two production blank lines to retain the existing file-length limits.

• Scala integration: 76 core and 69 OpenAI tests passed, covering code generation, assigned-session persistence, global parameters, HTTP response cleanup, cancellation, cache controls and defaults. After the fixture move, all 31 tests in the affected multimodal suite passed again; that rerun overlaps the 69 tests.
• Python integration: freshly rebuilt core and cognitive packages passed 28 tests and 28 subtests, plus seven loopback HTTP transforms. Installed wrapper origins and rebuilt JAR hashes were recorded.
• Build/style: core and cognitive production/test style checks, JAR packaging and wrapper generation passed. The fixture move preserves every test body and assertion.
• Boundary: these are local checks on Python 3.11.8, Spark/PySpark 3.5.0, Scala 2.12.18 and Java 11. These rebase results are historical; they do not establish hosted cache behavior.

Review and CI status

Current-revision Azure CI succeeded for merge abd8fdc0d29bb92fa704ae0537fd2479b9b5ad08, with verified parents target 6c2699defc and head b51b396704, after the authorized run. The authorized CPU-3 and geospatial retries both succeeded. Final collection found 215 of 215 selected current-attempt results passing, including all five named review regressions, OpenAI.ipynb and HyperOpt.ipynb. No code, timeout, workflow or approval changes were made to obtain the retries. The original observation window expired; terminal results were reconciled later without restarting it.

Strict evidence remains blocked: successful jobs still contain telemetry/TLS, low-disk and missing-coverage warnings, plus dependency-scan alerts requiring assessment. The collector therefore does not report a passing automatic whole-build gate. The two documented fork-excluded jobs remain intentionally skipped.

Fast checks and review: the dead-links check passed on its one authorized retry. All five reported findings are fixed, answered and resolved. Copilot's current-head review returned a service error rather than a substantive review; its one replacement request timed out without a new review. That is not a clean-review verdict. The preceding canceled/deleted build does not qualify this head.

The owner explicitly approved publishing this locally validated patch for human review without completing the six-lens pre-commit gate. The earlier full-patch broad-sweep review was clean; the next lens was stopped at the owner's request and the remaining lenses were not run. This cache-audit follow-up has not restarted that review cycle. Earlier clean reviews do not qualify this revision. Review evidence and the exception are recorded in the branch.

Azure validation of the preceding revision succeeded: the authorized /azp run comment queued a PR-triggered run, not a manual build. Its merge commit 5b78eb57c8b90a52ebd18dd2ac7dee7207c8ed31 has target d3f3af6bd3e4b7d191f51c9eb29f9aeffe788193 and previous PR head 73f9c93e54b20ddb05f68cd6d09a949de28ec244 as its verified ordered parents. The owner approved a one-time exception to the workflow's reviewed-SHA trigger safeguard; no pipeline settings, permissions, platform approvals or coverage conditions were changed.

The published current-attempt results include 155 selected passing regressions across ten Scala suites and three named Python tests. These cover defaults, explicit bindings, request shaping, breakpoint forwarding, decimal persistence and Python save/load. Coverage reports were published to Azure, and Codecov accepted the upload for that tested revision. Publication is confirmed; no Codecov patch-coverage threshold result is claimed.

Cache-audit amendment before this rebase: 276b80e92b contains the request, schema, accounting, prewarm and inference-control fixes summarized above. The preceding documentation amendment selected gpt-6-luna / gpt-6.1-sol; embedding configuration and archived docs remain unchanged. Earlier CI on 73f9c93e54 remains historical. Current notebook execution and bounded live cache results are documented above; final review remains unverified.

Not merge-ready: human approval, a substantive current-head review, warning dispositions and the stated provider/persistence qualification gaps remain unresolved. A successful CI status does not clear these separate gates. The stopped six-lens review cycle has not been restarted; no merge or release is authorized.

Notebook walkthrough, SDK/SynapseML examples, and API mapping

What the notebook demonstrates

• A stable, substantial developer-message prefix followed by changing questions.
• Explicit request-level options and prompt_cache_breakpoint inside actual input content—not merely an options column.
• Scalar options, row-bound options, and an implicit-caching OpenAIPrompt example.
• Native OpenAI Python SDK and SynapseML requests, provider errors, and cached-token usage.

OpenAI documentation

• Prompt caching: modes, breakpoints, eligibility, lifetimes, and usage
• Responses request reference
• Chat Completions reference
• Python SDK and verified SDK request types

Python SDK versus SynapseML

Run these snippets in order with a Spark session, a SynapseML build containing this change, and the OpenAI SDK installed. Set OPENAI_BASE_URL to your Azure OpenAI v1 URL ending in /openai/v1/, OPENAI_API_KEY to your key, and OPENAI_DEPLOYMENT to your chosen deployment. No deployment is created or changed.

Shared input and options

import os

endpoint = os.environ["OPENAI_BASE_URL"]
key = os.environ["OPENAI_API_KEY"]
deployment = os.environ["OPENAI_DEPLOYMENT"]
options = {"mode": "explicit", "ttl": "30m"}
prefix = "Answer using this product catalog.\n" + "\n".join(
    f"Product {n}: warranty {1 + n % 3} years; returns within 30 days with a receipt. "
    "Report shipping damage within seven days; repairs require a serial number."
    for n in range(1, 81)
)
messages = [
    {
        "role": "developer",
        "content": [{
            "type": "input_text",
            "text": prefix,
            "prompt_cache_breakpoint": {"mode": "explicit"},
        }],
    },
    {
        "role": "user",
        "content": [{"type": "input_text", "text": "What is Product 17's warranty?"}],
    },
]

OpenAI Python SDK

from openai import OpenAI

with OpenAI(api_key=key, base_url=endpoint) as client:
    response = client.responses.create(
        model=deployment,
        input=messages,
        prompt_cache_options=options,
        reasoning={"effort": "low"},
        text={"verbosity": "low"},
        store=False,
    )
print(response.output_text)
print(response.usage.input_tokens_details.cached_tokens)

SynapseML

from pyspark.sql import SparkSession
from synapse.ml.services.openai import OpenAIResponses

spark = SparkSession.builder.getOrCreate()
schema = (
    "messages array<struct<role:string,content:array<struct<type:string,text:string,"
    "prompt_cache_breakpoint:map<string,string>>>>>"
)
rows = spark.createDataFrame([(messages,)], schema)
transformer = (
    OpenAIResponses()
    .setUrl(endpoint)
    .setSubscriptionKey(key)
    .setDeploymentName(deployment)
    .setMessagesCol("messages")
    .setPromptCacheOptions(options)
    .setReasoningEffort("low")
    .setVerbosity("low")
    .setOutputCol("response")
    .setErrorCol("error")
)
transformer.transform(rows).select(
    "response.output", "response.usage.input_tokens_details.cached_tokens", "error"
).show(truncate=False)

Mapping and behavior

OpenAI field	SynapseML surface
Top-level prompt_cache_options	setPromptCacheOptions(dict) or setPromptCacheOptionsCol(column)
Responses input	OpenAIResponses.setMessagesCol(column)
input[].content[].prompt_cache_breakpoint	The same nested DataFrame field on existing input_text, input_image, or input_file content
Chat messages[].content[].prompt_cache_breakpoint	The same field on supported Chat content; text uses type="text"
Responses cached tokens	response.usage.input_tokens_details.cached_tokens
Chat cached tokens	response.usage.prompt_tokens_details.cached_tokens
OpenAIPrompt normalized usage column	usage.input_token_details["cached_tokens"] for either API; this is a map, unlike the raw response structs

An unset parameter or null options-column value omits the field, preserving service defaults. Selecting the column binding replaces the scalar binding. JSON values, including booleans, retain their types. Explicit null object members remain JSON null, distinct from absent keys. Breakpoint values must be objects in nested struct content; legacy string-valued content maps cannot represent a breakpoint object. Use a Spark struct column for mixed string/Boolean objects; string-valued maps are sufficient for mode and ttl. The inherited Foundry serialization path also forwards the option rather than silently dropping it.

Explicit mode needs input breakpoints; options alone do not enable explicit cache reads or writes. Top-level Responses instructions cannot hold a breakpoint, which is why the example uses developer-message content. Ordinary OpenAIPrompt templates do not insert markers; its notebook example therefore uses implicit mode.

Keep the reusable prefix stable and place variable input after it. The service decides eligibility and support; cache hits, latency improvements, and savings are not guaranteed. The inline snippets in this description are illustrative. The local notebook follow-up now asserts positive cached-token counts, with live results documented above.

The examples use 30m and a mode-only breakpoint. The guide documents additional Responses lifetime behavior, while some generated references and provider deployments differ. SynapseML forwards serializable values and exposes provider errors through the error column; it does not enforce a 30m-only rule. prompt_cache_options.ttl is not translated into the older prompt_cache_retention policy. Raw and normalized usage now expose cached_tokens and cache_write_tokens separately. Responses additionally preserves the write breakdown in cache_write_tokens_details.ttl_30m_tokens and cache_write_tokens_details.ttl_12h_tokens; these dotted names are the keys in Prompt's normalized input-details map. Missing counters remain null.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[github-actions]

Hey Rana Singh (@ranadeepsingh) 👋!
Thank you so much for contributing to our repository 🙌.
Someone from SynapseML Team will be reviewing this pull request soon.

We use semantic commit messages to streamline the release process.
Before your pull request can be merged, you should make sure your first commit and PR title start with a semantic prefix.
This helps us to create release messages and credit you for your hard work!

Examples of commit messages with semantic prefixes:

• fix: Fix LightGBM crashes with empty partitions
• feat: Make HTTP on Spark back-offs configurable
• docs: Update Spark Serving usage
• build: Add codecov support
• perf: improve LightGBM memory usage
• refactor: make python code generation rely on classes
• style: Remove nulls from CNTKModel
• test: Add test coverage for CNTKModel

To test your commit locally, please follow our guild on building from source.
Check out the developer guide for additional guidance on testing your change.

[Rana Singh (ranadeepsingh)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🔵 Needs a closer look

Current-head service-backed CI, branch-pinned Python validation, and final CI-qualified review remain pending.

No open issues 🎉

What changed in this PR

Adds opt-in prompt-cache controls to SynapseML’s OpenAI transformers while preserving provider defaults and caller-selected deployments.

Changes:

• Adds validated scalar/per-row cache options, serializer forwarding, and persistence coverage.
• Updates current examples and test selections to gpt-6-luna.
• Adds conditional Fabric administrator guidance and records validation evidence.

File	Description
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-6-gpt-6.1-sol.md	Records hardening review and validation limitations.
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-5-gpt-6.1-sol.md	Records test-coverage review.
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-4-gpt-6.1-sol.md	Records correctness review.
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-3-gpt-6.1-sol.md	Records robustness review.
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-2-gpt-6.1-sol.md	Records architecture review.
reviews/​pre-pr-luna-defaults-cache/​task-unknown-attempt-2-stage-attempt-1-review-1-gpt-6.1-sol.md	Records initial source review.
docs/​Explore Algorithms/​OpenAI/​Quickstart - Understand and Search Forms.ipynb	Updates example deployment.
docs/​Explore Algorithms/​OpenAI/​Quickstart - Multimodal OpenAI Prompter with Responses API.ipynb	Updates Responses example deployment.
docs/​Explore Algorithms/​OpenAI/​OpenAI.ipynb	Documents caching and updates reasoning examples.
docs/​Explore Algorithms/​OpenAI/​Langchain.ipynb	Updates LangChain example deployment.
core/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​fabric/​OpenAIFabricSettingSuite.scala	Tests conditional guidance and existing status gates.
core/​src/​main/​scala/​com/​microsoft/​azure/​synapse/​ml/​fabric/​OpenAIFabricSetting.scala	Extends missing-model administrator guidance.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIV1EndpointSuite.scala	Updates endpoint-test model fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIResponsesSuite.scala	Updates Responses fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIResponsesMultimodalSuite.scala	Updates multimodal Responses fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIPromptSuite.scala	Updates prompt model selections and assertions.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIPromptResponsesSuite.scala	Updates live Responses selection and reasoning settings.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIPromptMultimodalRequestSuite.scala	Tests Foundry cache forwarding; updates fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIPromptCacheSuite.scala	Tests cache validation, omission, and serialization.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIEmbeddingsSuite.scala	Updates non-embedding default fixture.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIChatCompletionSuite.scala	Updates chat test selections and fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIChatCompletionMultimodalSuite.scala	Updates multimodal chat fixtures.
cognitive/​src/​test/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIAPIKey.scala	Updates shared live-test deployment.
cognitive/​src/​test/​python/​synapsemltest/​services/​test_ServiceParamPythonBridge.py	Tests generated cache accessors, copying, and persistence.
cognitive/​src/​test/​python/​synapsemltest/​services/​openai/​test_StructuredOutput.py	Updates structured-output deployment.
cognitive/​src/​test/​python/​synapsemltest/​services/​openai/​test_OpenAIResponsesMultimodal.py	Updates Responses model fixture.
cognitive/​src/​test/​python/​synapsemltest/​services/​openai/​test_OpenAIMultimodalRequest.py	Updates request and response fixtures.
cognitive/​src/​test/​python/​synapsemltest/​services/​openai/​test_OpenAIDefaults.py	Updates defaults-test deployment.
cognitive/​src/​test/​python/​synapsemltest/​services/​openai/​test_OpenAIChatCompletionMultimodal.py	Updates chat model fixture.
cognitive/​src/​test/​python/​synapsemltest/​services/​langchain/​test_LangchainTransform.py	Updates LangChain model fixtures.
cognitive/​src/​main/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIResponses.scala	Includes cache options in Responses payloads.
cognitive/​src/​main/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAIChatCompletion.scala	Includes cache options in chat payloads.
cognitive/​src/​main/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​openai/​OpenAI.scala	Defines cache parameters and shared validation.
cognitive/​src/​main/​scala/​com/​microsoft/​azure/​synapse/​ml/​services/​aifoundry/​AIFoundryChatCompletion.scala	Includes cache options in Foundry payloads.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Supplied evidence leaves current-head service-backed CI and branch-pinned Python validation pending for the cross-language API changes.

No open issues 🎉

_{Review effort: Balanced}

[Rana Singh (ranadeepsingh)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🔵 Needs a closer look

The supplied evidence leaves current-head CI, branch-pinned Python validation, and the final CI-qualified review outstanding for these cross-language, provider-facing changes.

No open issues 🎉

_{Review effort: Balanced}

[Rana Singh (ranadeepsingh)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🔵 Needs a closer look

Supplied evidence does not establish completed current-head CI and final qualification for the cross-service Scala/Python API changes.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Shared persistence changes affect models beyond OpenAI, requiring maintainer assessment of backward compatibility and CI evidence because Spark-dependent validation could not be independently reproduced.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

An invalid log-probability selector causes Responses validation failures, while saved-and-loaded Foundry prompts forward an unsupported storage default and fail before making HTTP requests.

2 open issues

• Use the correct Responses logprobs include selector · New
• Ignore Foundry's default store value after Prompt reload · New

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Active defects include an unsupported log-probability include value, a compatibility regression in literal-key usage mappings, and storage-flag column collisions that cause analysis failures.

4 open issues

• Ignore Foundry's default store value after Prompt reload
• Use the correct Responses logprobs include selector
• Preserve the storage flag before replacing input columns · New
• Preserve literal-key handling in usage mappings · New

_{Review effort: Balanced}

[Rana Singh (ranadeepsingh)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🟡 Changes recommended

The public ResponsesDetailed mapping fails to normalize Responses usage because dotted keys are treated as literal fields, while nested-path handling is available only through an internal normalizer.

1 open issue

• ResponsesDetailed mapping treats dotted keys as literal fields · New

4 resolved since last review

• Ignore Foundry's default store value after Prompt reload
• Use the correct Responses logprobs include selector
• Preserve literal-key handling in usage mappings
• Preserve the storage flag before replacing input columns

_{Review effort: Balanced}

[Rana Singh (ranadeepsingh)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.