[Lucas Kim (kimnamu)]

Motivation and Context

Thanks to the Python team for the Bedrock connector (#9100) and its per-provider modules; OpenAI support follows the chat region of bedrock_mistralai.py.

Fixes #14470

BedrockChatCompletion rejected every OpenAI model ID on Bedrock (GPT-6 Sol/Luna/Astra, GPT-5.6, gpt-oss) with ValueError: ... does not contain a valid model provider name.. With the provider forced, the default stopSequences: [] got a 400.

case (live, us-east-1)	Before (main)	After (this PR)	After (this PR + #14469)
GPT-6 Sol us./global.: plain, streaming, system prompt, tools (Auto, Required, streaming)	❌ ValueError	✅	✅
GPT-6 Sol: image, 2-turn, max_tokens=20	❌ ValueError	❌ reasoningContent block (#14468)	✅ (max_tokens=20 gives LENGTH, empty text)
GPT-6 Luna us./global.: "Say pong." non-streaming / streaming	❌ ValueError	❌ 10/30 / 8/30 OK (3 rounds), the rest hit a reasoningContent block	✅ 20/20 / 20/20 (2 rounds)
GPT-6 Luna: system prompt, tools (Auto, Required, streaming)	❌ ValueError	✅	✅
GPT-6 Astra us./global.: plain, streaming, system prompt, tools (Auto, Required, streaming)	❌ ValueError	✅ ("Say pong." 30/30 / 30/30, 3 rounds)	✅ (20/20 / 20/20, 2 rounds)
GPT-6 Astra: image, 2-turn, max_tokens=20	❌ ValueError	❌ reasoningContent block (#14468)	✅ (max_tokens=20 gives LENGTH, empty text)
GPT-5.6 Sol: plain, streaming, system prompt, tools, image / 2-turn, max_tokens=20	❌ ValueError	✅ / ❌ can hit a reasoningContent block	✅
gpt-oss-120b: all of the above	❌ ValueError	❌ reasoningContent block on every text check	✅ (except image: the model has no image input)
BEDROCK_MODEL_PROVIDER=openai (Sol, Astra)	❌ ValidationError	✅ 'pong'	✅
Default inferenceConfig sent	{"stopSequences": []}	{}	{}
stop=["world"] set by the user	sent	✅ unchanged, sent	✅
temperature set by the user on GPT-6	sent (service returns 400 on Converse)	✅ unchanged	✅
Claude Haiku 4.5 (plain, streaming, tools), Nova Lite (plain, streaming)	✅	✅ unchanged	✅
Public API (only the new OPENAI member), additionalModelRequestFields for other providers	✅	✅ unchanged	✅

Description

• bedrock_model_provider.py: add OPENAI = "openai" and one CHAT_COMPLETION_ADDITIONAL_MODEL_REQUEST_FIELDS_MAPPING entry.
• bedrock_openai.py (new): the chat region of bedrock_mistralai.py, returning None, with the Bedrock OpenAI parameters link.
• bedrock_chat_completion.py: "stopSequences": settings.stop or None, so remove_none_recursively drops the empty default like the other unset fields.
• Tests: openai.gpt in the four provider parametrize lists, two GPT-6 profile IDs in test_inference_profile_with_bedrock_model, and a new test_prepare_settings_for_request_with_stop. The old assert all([parsed_settings["inferenceConfig"].values()]) is always true (a one-item list), so the two prepare_settings tests now assert == {}.

Out of scope: text completion and embeddings for OpenAI stay unsupported, and a reasoning-effort setting would be a separate feature. GPT-6 Luna, gpt-oss and some Sol/Astra prompts need #14469 as well; the two PRs touch different lines and pass the Bedrock unit tests together (120).

Tests (red/green, lint) and live output

New tests kept, source reverted to main:

E       AssertionError: assert {'stopSequences': []} == {}          (x7)
E           ValueError: Model ID openai.gpt does not contain a valid model provider name.          (x4)
E           ValueError: Model ID us.openai.gpt-6-sol does not contain a valid model provider name.
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_prepare_settings_for_request[anthropic.claude]   (and the 5 other providers)
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_prepare_settings_for_request[openai.gpt]
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_prepare_settings_for_request_with_application_inference_profile[...]
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_bedrock_chat_completion[openai.gpt]
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_bedrock_streaming_chat_completion[openai.gpt]
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_chat_completion.py::test_bedrock_streaming_chat_completion_invalid_event[openai.gpt]
FAILED tests/unit/connectors/ai/bedrock/services/test_bedrock_model_provider_utils.py::test_inference_profile_with_bedrock_model
12 failed, 106 passed in 0.17s

With the fix:

$ uv run pytest tests/unit/connectors/ai/bedrock
118 passed in 0.13s
$ uv run ruff check <changed files>          -> All checks passed!
$ uv run ruff format --check <changed files> -> 5 files already formatted
$ uv run mypy --config-file mypy.ini semantic_kernel/connectors/ai/bedrock -> Success: no issues found in 18 source files
pyupgrade --py310-plus, bandit -c pyproject.toml -> no changes, no findings

"Say pong.", 5 non-streaming + 5 streaming calls per model per round:

this PR          us.openai.gpt-6-sol       non-streaming 5/5  streaming 5/5
this PR          global.openai.gpt-6-sol   non-streaming 5/5  streaming 5/5
this PR          us.openai.gpt-6-luna      rounds 1/2/3: non-streaming 0/5, 3/5, 2/5  streaming 0/5, 3/5, 1/5  ['ServiceInvalidResponseError'] / ['KeyError']
this PR          global.openai.gpt-6-luna  rounds 1/2/3: non-streaming 0/5, 2/5, 3/5  streaming 2/5, 1/5, 1/5  ['ServiceInvalidResponseError'] / ['KeyError']
this PR + #14469 all four IDs              rounds 1/2: non-streaming 5/5, 5/5  streaming 5/5, 5/5
this PR          us.openai.gpt-6-astra     rounds 1/2/3: non-streaming 5/5, 5/5, 5/5  streaming 5/5, 5/5, 5/5
this PR          global.openai.gpt-6-astra rounds 1/2/3: non-streaming 5/5, 5/5, 5/5  streaming 5/5, 5/5, 5/5
this PR + #14469 both Astra IDs            rounds 1/2: non-streaming 5/5, 5/5  streaming 5/5, 5/5

Wider check through BedrockChatCompletion, this PR + #14469 (us.openai.gpt-6-luna shown; Sol, Astra, the global. IDs and us.openai.gpt-5.6-sol give the same OK/FAIL per check):

- plain: OK text='pong' finish=FinishReason.STOP | request(minus messages)={"modelId": "us.openai.gpt-6-luna", "system": [], "inferenceConfig": {}, "additionalModelRequestFields": null}
- system: OK text='HELLO' system_sent=[{'text': 'Reply in UPPERCASE only.'}]
- stream: OK text_chunks=39 ttft=0.77 total=0.96
- tools_auto: OK text='It’s sunny and 21°C in Paris.' history_roles=['user', 'assistant', 'tool']
- tools_required: OK items=['FunctionCallContent'] toolChoice={'tool': {'name': 'w-get_weather'}}
- tools_stream: OK text='It’s sunny and 21°C in Oslo.'
- image: OK text='Red'
- two_turn: OK t1='391' t2='392'
- max_tokens=20: OK text='' finish=FinishReason.LENGTH
- temperature=0.2(user-set): FAIL ValidationException: ... This model doesn't support the temperature field. Remove temperature and try again.

Streaming speed matches raw boto3 converse_stream (this PR + #14469, "Write about 250 words explaining how a hash map works.", maxTokens=2000, 3 interleaved runs (Astra 5), medians):

model                  path  TTFT s  total s  output tokens
us.openai.gpt-6-sol    raw   0.99    3.37     307
us.openai.gpt-6-sol    SK    1.11    3.49     308
us.openai.gpt-6-luna   raw   1.13    2.78     413
us.openai.gpt-6-luna   SK    0.97    2.78     408
us.openai.gpt-6-astra  raw   2.48    9.29     328
us.openai.gpt-6-astra  SK    1.38    8.69     327

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[Copilot]

Copilot review overview

🟢 Approval recommended

No unresolved issues were identified that would block approval.

Review effort: Lite
Findings: None

What changed in this PR

Adds OpenAI model support to the Python Bedrock chat connector and prevents empty stop sequences from being sent.

Changes:

• Adds OpenAI provider routing and chat request handling.
• Omits default empty stopSequences.
• Extends unit tests for OpenAI IDs and stop settings.

File	Description
python/​tests/​unit/​connectors/​ai/​bedrock/​services/​test_bedrock_model_provider_utils.py	Updated as part of this pull request.
python/​tests/​unit/​connectors/​ai/​bedrock/​services/​test_bedrock_chat_completion.py	Updated as part of this pull request.
python/​semantic_kernel/​connectors/​ai/​bedrock/​services/​model_provider/​bedrock_openai.py	Updated as part of this pull request.
python/​semantic_kernel/​connectors/​ai/​bedrock/​services/​model_provider/​bedrock_model_provider.py	Updated as part of this pull request.
python/​semantic_kernel/​connectors/​ai/​bedrock/​services/​bedrock_chat_completion.py	Updated as part of this pull request.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Eduard van Valkenburg (eavanvalkenburg)]

Thank you for the contribution. We are not adding new features to Semantic Kernel (SK) at the moment, so I am closing this PR. Please consider creating a feature request issue in microsoft/agent-framework.