fix(bedrock): resolve realtime AWS credentials from config, not just env

[devin-ai-integration]

Relevant issues

Follow-up to #31924. The Nova Sonic realtime handler built its smithy client with EnvironmentCredentialsResolver, which only reads AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY from the process environment. Any aws_access_key_id, aws_secret_access_key, aws_session_token, aws_role_name, aws_profile_name, aws_web_identity_token, etc. set in the proxy config's litellm_params were silently ignored, so realtime sessions failed unless ambient env credentials happened to exist

This change routes the params (already plumbed through realtime_api/main.py into BedrockRealtime.async_realtime) through BaseAWSLLM.get_credentials, the same boto3-based resolution every other Bedrock endpoint uses (static keys, session tokens, role assumption, profiles, web identity), and adapts the resulting boto3 credentials to the smithy client via a small BotoCredentialsResolver. When no config params are set, get_credentials falls back to the standard boto3 chain, which still covers env vars, so existing env-based setups keep working. The resolver freezes credentials via run_in_executor so a refreshable-credentials STS call never blocks the event loop

Linear ticket

Pre-Submission checklist

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

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

No AWS credentials were available in this session, so a live Bedrock run wasn't possible. To verify against the real API, put keys only in the config (none in env) and run the same qa_client.py flow from #31924:

model_list:
  - model_name: "bedrock-sonic"
    litellm_params:
      model: bedrock/amazon.nova-sonic-v1:0
      aws_region_name: us-east-1
      aws_access_key_id: <key>
      aws_secret_access_key: <secret>
    model_info:
      mode: realtime

general_settings:
  master_key: "sk-qa-1234"

env -u AWS_ACCESS_KEY_ID -u AWS_SECRET_ACCESS_KEY .venv/bin/litellm --config qa_config.yaml --port 49781 --detailed_debug > proxy.log 2>&1 &
.venv/bin/python qa_client.py 49781

Before this PR the session dies with SmithyIdentityError: AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY are required; after it, the client receives response.text.delta events and a final response.done

Regression tests pin both halves: BotoCredentialsResolver maps frozen boto3 credentials to a smithy AWSCredentialsIdentity, and async_realtime forwards all config aws_* params into get_credentials and wires the resulting resolver into the client config

Type

🐛 Bug Fix

Changes

litellm/llms/bedrock/realtime/handler.py: replace EnvironmentCredentialsResolver with BotoCredentialsResolver(self.get_credentials(aws_access_key_id=..., aws_role_name=..., ...))

tests/test_litellm/llms/bedrock/realtime/test_bedrock_realtime_handler.py: regression tests for the credential path

Link to Devin session: https://app.devin.ai/sessions/9c6563e5a0834358aa1db302935bb36f
Requested by: @mateo-berri

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR replaces the EnvironmentCredentialsResolver in the Bedrock Nova Sonic realtime handler with a new BotoCredentialsResolver that reads credentials from the full get_credentials() boto3-based resolution chain, fixing silent failures when aws_access_key_id and related params were set in the proxy config rather than process environment variables.

• handler.py: Adds BotoCredentialsResolver (adapts a frozen boto3 credentials object to the smithy AWSCredentialsIdentity interface) and wires all aws_* params through BaseAWSLLM.get_credentials() before constructing the smithy Config.
• test_bedrock_realtime_handler.py: Two new tests — one unit-tests BotoCredentialsResolver.get_identity() with a fake boto3 credentials object, the other verifies that async_realtime forwards all config credentials into get_credentials() and stores the resulting resolver in the smithy Config.

Confidence Score: 4/5

Safe to merge; the credential fix is correct and all aws_* config params are properly forwarded through the existing boto3 resolution chain.

The core fix is sound: BotoCredentialsResolver correctly adapts the boto3 frozen credentials object to the smithy identity interface, and get_credentials() already handles every auth flow (static keys, session tokens, role assumption, web identity, profiles) with env-var fallback. The only concern is that get_frozen_credentials() inside the async get_identity() method is a synchronous botocore call that can block the event loop when refreshable credentials expire mid-session and a network round-trip to STS is needed.

litellm/llms/bedrock/realtime/handler.py — specifically the get_identity method in BotoCredentialsResolver

Important Files Changed

Filename	Overview
litellm/llms/bedrock/realtime/handler.py	Replaces EnvironmentCredentialsResolver with BotoCredentialsResolver; get_frozen_credentials() in get_identity() is a synchronous blocking call that could stall the event loop for refreshable credentials (role assumption, web identity) when mid-session token refresh occurs.
tests/test_litellm/llms/bedrock/realtime/test_bedrock_realtime_handler.py	Adds two well-scoped regression tests: one for BotoCredentialsResolver mapping and one for end-to-end credential forwarding through async_realtime; all mocked, no real network calls.

_{Reviews (1): Last reviewed commit: "fix(bedrock): resolve realtime AWS crede..." | Re-trigger Greptile}

[greptile-apps]

get_frozen_credentials() is a synchronous botocore call inside an async method. For RefreshableCredentials (produced by role assumption, web identity, or profile auth), an expired token triggers a synchronous STS HTTP request here, which will block the asyncio event loop for the duration of that network round-trip. Static key credentials are unaffected since get_frozen_credentials() is instant for them, but long-running WebSocket sessions with role-assumed credentials will hit this on every credential refresh cycle.

Suggested change

	async def get_identity(self, *, properties: Any) -> Any:
	from smithy_aws_core.identity import AWSCredentialsIdentity
	frozen = self._credentials.get_frozen_credentials()
	return AWSCredentialsIdentity(
	access_key_id=frozen.access_key,
	secret_access_key=frozen.secret_key,
	session_token=frozen.token,
	)
	async def get_identity(self, *, properties: Any) -> Any:
	import asyncio
	from smithy_aws_core.identity import AWSCredentialsIdentity
	loop = asyncio.get_event_loop()
	frozen = await loop.run_in_executor(None, self._credentials.get_frozen_credentials)
	return AWSCredentialsIdentity(
	access_key_id=frozen.access_key,
	secret_access_key=frozen.secret_key,
	session_token=frozen.token,
	)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[devin-ai-integration]

Closing: superseded by #32275, which landed the same config-credential fix