fix(bedrock): keep filtering anthropic-beta headers on the Claude platform messages path

[yuneng-berri]

Summary

BedrockClaudePlatformMessagesConfig silently stopped filtering anthropic-beta headers, so betas Bedrock does not accept are now forwarded verbatim to it.

#42152 added an override to AnthropicMessagesConfig:

def should_filter_anthropic_beta_headers(self) -> bool:
    return self._resolved_provider != "anthropic"

That is correct for its own case — when the upstream really is api.anthropic.com, betas should pass through untouched. But AnthropicMessagesConfig.custom_llm_provider is a hardcoded property returning "anthropic", and _resolved_provider just reads it, so the method returns False for every subclass that does not override the provider.

BedrockClaudePlatformMessagesConfig overrides neither custom_llm_provider nor the filter, and nothing earlier in its MRO supplies a provider:

BedrockClaudePlatformMessagesConfig -> BedrockClaudePlatformMixin -> BaseAWSLLM
  -> SignsRequestsWithAWS -> AnthropicMessagesConfig

so it inherits "anthropic" and reports filter_betas=False. (The custom_llm_provider -> "bedrock" property in litellm/llms/bedrock/claude_platform/transformation.py belongs to the chat/converse config and is not in this MRO.)

This is the exact case the base class documents as needing the filter, in litellm/llms/base_llm/anthropic_messages/transformation.py:

Cross-provider translation paths (bedrock, vertex_ai, ...) need this so unsupported betas are dropped. Configs that forward natively to an Anthropic-compatible endpoint return False to pass betas through verbatim.

Before #42152 the base returned True, so Bedrock filtered. The filter runs at litellm/llms/custom_httpx/llm_http_handler.py:2281.

Sweep of every AnthropicMessages config — Bedrock is the only unintended False:

config	provider	filter_betas
AnthropicMessagesConfig	anthropic	False — intended, native
GithubCopilotAnthropicMessagesConfig	github_copilot	False — intended, explicit override
OpenAILikeAnthropicMessagesConfig	anthropic	False — intended, explicit override
AzureAnthropicMessagesConfig	azure_ai	True
MinimaxMessagesConfig	minimax	True
DeepSeekAnthropicMessagesConfig	deepseek	True
TencentAnthropicMessagesConfig	tencent	True
BedrockClaudePlatformMessagesConfig	anthropic	False — bug

Changes

• BedrockClaudePlatformMessagesConfig opts back into the provider-scoped filter.
• New coverage for Bedrock beta filtering — there was none, which is why this landed unnoticed.
• The two passthrough tests whose contrast assertion caught this now compare against the Bedrock config, since a native-Anthropic False is intended behaviour.

I deliberately did not fix this by giving the config custom_llm_provider -> "bedrock". That would be the deeper fix, but _resolved_provider also feeds six model-capability lookups in AnthropicMessagesConfig (reasoning-effort translation, maybe_drop_disabled_thinking, adaptive-effort and thinking-temperature handling), so changing it would alter Bedrock thinking/reasoning behaviour in a regression fix. Worth doing separately, with its own testing.

Testing

Red/green against the new test, source fix removed then restored:

--- RED (fix removed)
FAILED tests/unit/llms/bedrock/claude_platform/...::test_bedrock_claude_platform_filters_anthropic_beta_headers
FAILED tests/unit/llms/github_copilot/...::test_github_copilot_config_disables_anthropic_beta_filtering
FAILED tests/unit/llms/openai_like/...::test_passthrough_disables_anthropic_beta_filtering
3 failed, 90 passed

--- GREEN (fix restored)
93 passed

Filtering demonstrably active again — an unsupported beta is dropped and a renamed one translated:

in:  code-execution-2025-08-25, advanced-tool-use-2025-11-20, context-management-2025-06-27
out: context-management-2025-06-27, tool-search-tool-2025-10-19

• tests/unit/llms/{bedrock,github_copilot,openai_like} + test_llm_http_handler.py: 618 passed
• Full tests/unit as CI runs it (-n 4 --dist=loadscope, 315 files): 6395 passed
• scripts/pre_commit_lint.sh: PASS

How this was found

Triaging the #circleci-triage backlog. The two passthrough tests fail deterministically on main and CI had not yet run the unit job since #42152 merged, so the next scheduled run would have gone red on them. They read like stale assertions, but in each one the test's own subject still passes — the failing line is the contrast guard on the default, which is what surfaced the Bedrock leak.

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_bedrock_claude_platform_beta_filter (cac3628) with main (51ae7ab)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge, with the Bedrock request path and regression coverage aligned with the intended filtering behavior

Summary

This PR restores Anthropic beta-header filtering for the Bedrock Claude platform Messages path

• Adds an explicit filtering override to the Bedrock configuration
• Adds regression coverage for the override and Bedrock-specific beta removal and renaming
• Updates passthrough-provider tests to use Bedrock as the filtering-enabled contrast

_{Reviews (1) · Last reviewed commit: "fix(bedrock): keep filtering anthropic-b..."}

[mateo-berri]

Superseded by #42275, merged: filtering with the bedrock beta map 400s beta-gated requests on the Claude Platform route, which accepts every Anthropic beta