fix: anthropic fallbacks and refusal

[TejasGhatte]

Summary

Anthropic's native server-side fallback feature (server-side-fallback-2026-06-01) uses a "fallbacks" request field containing JSON objects ({"model": "..."}) — the same wire key that Bifrost uses for cross-provider fallback strings ("provider/model"). Previously, Bifrost unconditionally deleted the entire "fallbacks" field before forwarding requests to Anthropic, which broke native server-side fallbacks. This PR disambiguates the two by shape, preserving native fallback objects while still stripping Bifrost cross-provider strings.

Changes

• Replaced []string on AnthropicMessageRequest.Fallbacks with []AnthropicFallbackEntry, a union type that unmarshals JSON strings as Bifrost cross-provider fallbacks and JSON objects as Anthropic native server-side fallbacks.
• Added AnthropicNativeFallback struct representing one entry of Anthropic's native "fallbacks" list, with optional max_tokens and thinking overrides.
• Replaced the blanket DeleteJSONField("fallbacks") call in both BuildAnthropicResponsesRequestBody and BuildAnthropicChatRequestBody with stripBifrostFallbacksFromBody, which removes only string entries and deletes the field entirely when no native objects remain.
• Added bifrostFallbackModels() and nativeFallbacks() helpers on AnthropicMessageRequest to separate the two entry kinds cleanly.
• Updated ToBifrostResponsesRequest to route native fallback objects into ExtraParams["fallbacks"] for verbatim forwarding, while Bifrost strings continue to populate BifrostResponsesRequest.Fallbacks.
• Updated ToAnthropicResponsesRequest to reconstruct the typed Fallbacks field from ExtraParams["fallbacks"] when native entries are present.
• Added ServerSideFallback bool to ProviderFeatureSupport, enabled only for the Anthropic provider (not Bedrock, Vertex, or Azure), and wired it into AddMissingBetaHeadersToContext so the server-side-fallback-2026-06-01 beta header is auto-injected when native fallbacks are present.
• Registered AnthropicServerSideFallbackBetaHeaderPrefix in the known-prefix lists used for beta header deduplication and provider-feature gating.

Type of change

• Bug fix
• Feature

Affected areas

• Core (Go)
• Providers/Integrations

How to test

go test ./core/providers/anthropic/...

Key test cases added:

• TestAnthropicFallbackEntry_UnmarshalJSON — verifies string vs. object dispatch and round-trip marshalling.
• TestAnthropicMessageRequest_NativeFallbacksParse — regression for the "Invalid JSON" error when a request carries native {"model": ...} fallback objects.
• TestToBifrostResponsesRequest_FallbacksRouting — verifies native objects go to ExtraParams and Bifrost strings go to Fallbacks.
• TestAddMissingBetaHeadersToContext_ServerSideFallback — verifies the beta header is injected on Anthropic and suppressed on Vertex.
• TestBuildAnthropicResponsesRequestBody_NativeFallbacks — end-to-end body assembly for both raw-passthrough and typed paths, covering native preservation, Bifrost stripping, and beta header injection.

Breaking changes

• No

The Fallbacks []string field on AnthropicMessageRequest is an internal type not part of the public Bifrost API surface. Existing Bifrost cross-provider fallback behaviour is unchanged.

Security considerations

No auth, secrets, or PII implications. The change only affects how the "fallbacks" JSON field is forwarded to the Anthropic API.

Checklist

• I read docs/contributing/README.md and followed the guidelines
• I added/updated tests where appropriate
• I updated documentation where needed
• I verified builds succeed (Go and UI)
• I verified the CI pipeline passes locally if applicable

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

tejas ghatte seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[TejasGhatte]

• fix: anthropic fallbacks and refusal #5367 : 2 dependent PRs (#5373 , #5387 ) 👈 (View in Graphite)
• dev

This stack of pull requests is managed by Graphite. Learn more about stacking.

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features
  • Added Anthropic native server-side fallback support (including fallback boundary handling) and full stop_details propagation across streaming and non-streaming.
  • Added beta support for fallback credit, including request/response round-tripping and per-attempt model usage.
• Bug Fixes
  • Improved fallback request shaping to preserve native fallback objects only when supported, while stripping cross-provider fallback markers/strings elsewhere.
  • Ensured unsupported fallback boundary blocks are dropped when converting to OpenAI, Bedrock, Cohere, and Gemini.
• Tests
  • Expanded unit and end-to-end coverage for fallback shapes, beta header behavior, stop_details, and fallback credit behavior.

Walkthrough

Anthropic now supports native server-side fallback objects, fallback boundary content blocks, stop details, fallback-credit fields, and usage model metadata across request construction and Bifrost conversions. Unsupported providers filter Anthropic-specific fields and boundary markers.

Changes

Anthropic server-side fallback support

Layer / File(s)	Summary
Fallback contracts and feature gating core/providers/anthropic/types.go, core/schemas/responses.go	Schemas define native fallback entries, fallback boundary blocks, stop details, usage model metadata, beta headers, and provider capabilities.
Fallback request routing and assembly core/providers/anthropic/chat.go, core/providers/anthropic/responses.go, core/providers/anthropic/utils.go, core/providers/anthropic/requestbuilder.go	Native fallbacks and fallback-credit tokens are routed through typed and extra parameters, feature-gated headers are added, and request bodies retain supported native objects while removing Bifrost fallback strings.
Response conversion core/providers/anthropic/responses.go	Fallback boundaries, stop details, fallback-credit metadata, and usage model values round-trip through streaming and non-streaming conversions.
Unsupported-provider filtering core/providers/openai/responses.go, core/providers/bedrock/responses.go, core/providers/cohere/responses.go, core/providers/gemini/responses.go	Fallback boundary markers are dropped from provider request or response content when unsupported.
Validation and end-to-end coverage core/providers/anthropic/*_test.go, core/providers/openai/responses_test.go, tests/e2e/api/collections/provider-harness.json	Tests cover fallback shapes, routing, headers, body construction, provider filtering, boundary events, stop details, fallback credit, usage metadata, and API scenarios.

Estimated code review effort: 4 (Complex) | ~60 minutes

Sequence Diagram(s)

sequenceDiagram
  participant AnthropicRequest
  participant BifrostRequest
  participant AnthropicRequestBuilder
  participant AnthropicResponseStream
  participant BifrostResponseStream
  AnthropicRequest->>BifrostRequest: route native fallbacks and credit token
  BifrostRequest->>AnthropicRequestBuilder: pass typed and extra fallback parameters
  AnthropicRequestBuilder->>AnthropicRequestBuilder: inject headers and filter request fields
  AnthropicResponseStream->>BifrostResponseStream: convert fallback boundaries and stop details
  BifrostResponseStream->>AnthropicResponseStream: reconstruct Anthropic stream events

Loading

Suggested reviewers: akshaydeo, danpiths, pratham-mishra04, sammaji, roroghost17

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title concisely describes the main change: Anthropic fallbacks and refusal metadata handling.
Description check	✅ Passed	The description matches the template well and includes summary, changes, testing, breaking changes, security, and checklist sections.
Docstring Coverage	✅ Passed	Docstring coverage is 88.89% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch 07-20-fix_anthropic_fallbacks_and_refusal

Warning

There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure.

🔧 golangci-lint (2.12.2)

level=error msg="[linters_context] typechecking error: pattern ./...: directory prefix . does not contain main module or its selected dependencies"

---

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Confidence Score: 5/5

This looks safe to merge.

• The updated chat path promotes native fallback objects before beta-header detection.
• Unsupported provider paths strip Anthropic-only fallback fields.
• The new request and response behavior has focused test coverage.
• No blocking issues were found in the changed code.

Important Files Changed

Filename	Overview
core/providers/anthropic/chat.go	Promotes native fallbacks and fallback credit tokens from extra parameters into typed Anthropic chat requests.
core/providers/anthropic/requestbuilder.go	Preserves supported native fallback objects while stripping Bifrost strings and unsupported fields.
core/providers/anthropic/responses.go	Carries fallback boundaries, refusal details, credit data, and usage models through response conversions.
core/providers/anthropic/types.go	Adds typed wire representations for native fallbacks, refusal details, and related provider features.
core/providers/anthropic/utils.go	Adds provider-aware fallback filtering and beta-header injection.
core/providers/anthropic/serversidefallback_test.go	Tests fallback routing, provider gating, headers, refusal metadata, credits, and stream conversions.

_{Reviews (4): Last reviewed commit: "fix: anthropic fallbacks and refusal" | Re-trigger Greptile}

[coderabbitai]

Actionable comments posted: 1

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

core/providers/anthropic/responses.go (1)

3438-3446: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Mirror native fallback handling in core/providers/anthropic/chat.go. ToAnthropicChatRequest should carry ExtraParams["fallbacks"] into AnthropicMessageRequest.Fallbacks like the responses path does, so Chat requests preserve Anthropic-native fallbacks and trigger the beta header.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@core/providers/anthropic/responses.go` around lines 3438 - 3446, The chat
request conversion in ToAnthropicChatRequest does not propagate native
fallbacks. Mirror the responses conversion by reading ExtraParams["fallbacks"],
assigning it to AnthropicMessageRequest.Fallbacks, and preserving the existing
behavior that triggers the beta header when fallbacks are present.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@core/providers/anthropic/utils.go`:
- Around line 1285-1317: Gate native fallback preservation on the target
provider’s ServerSideFallback capability. Update stripBifrostFallbacksFromBody
in core/providers/anthropic/utils.go and both call sites in
core/providers/anthropic/requestbuilder.go at lines 379-384 and 602-607 to pass
or apply cfg.Provider, ensuring native JSON-object fallbacks are stripped when
!ProviderFeatures[cfg.Provider].ServerSideFallback while retaining current
behavior for supported providers.

---

Outside diff comments:
In `@core/providers/anthropic/responses.go`:
- Around line 3438-3446: The chat request conversion in ToAnthropicChatRequest
does not propagate native fallbacks. Mirror the responses conversion by reading
ExtraParams["fallbacks"], assigning it to AnthropicMessageRequest.Fallbacks, and
preserving the existing behavior that triggers the beta header when fallbacks
are present.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 80227888-d6a2-45ef-900a-99c806c457bb

📥 Commits

Reviewing files that changed from the base of the PR and between e1550bf and 83571c1.

📒 Files selected for processing (5)

• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/types.go
• core/providers/anthropic/utils.go
• core/providers/anthropic/utils_test.go

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

core/providers/anthropic/responses.go (1)

3315-3334: 🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win

Strip Anthropic-native fallbacks before cross-provider fallback. Params.ExtraParams is reused on fallback attempts, and providers like Bedrock forward ExtraParams verbatim, so this raw Anthropic field can reach a non-Anthropic upstream and break the request. core/providers/anthropic/responses.go:3315-3334

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@core/providers/anthropic/responses.go` around lines 3315 - 3334, Update
AnthropicMessageRequest.ToBifrostResponsesRequest so Anthropic-native fallbacks
are not retained in Params.ExtraParams when cross-provider fallback requests may
reuse it; only forward the native “fallbacks” field for the Anthropic upstream,
preventing providers such as Bedrock from receiving it verbatim.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@core/providers/anthropic/responses.go`:
- Around line 3315-3334: Update
AnthropicMessageRequest.ToBifrostResponsesRequest so Anthropic-native fallbacks
are not retained in Params.ExtraParams when cross-provider fallback requests may
reuse it; only forward the native “fallbacks” field for the Anthropic upstream,
preventing providers such as Bedrock from receiving it verbatim.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: a138295d-cc1b-44f6-bd0e-a8e3a0bcf9b2

📥 Commits

Reviewing files that changed from the base of the PR and between 83571c1 and db9bd25.

📒 Files selected for processing (8)

• core/providers/anthropic/chat.go
• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/serversidefallback_test.go
• core/providers/anthropic/types.go
• core/providers/anthropic/utils.go
• core/providers/anthropic/utils_test.go
• core/schemas/responses.go

🚧 Files skipped from review as they are similar to previous changes (2)

• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/utils.go

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/e2e/api/collections/provider-harness.json`:
- Around line 40994-41044: Update the request model in the test named
“anthropic/claude-fable-5 · native object fallbacks forwarded with
server-side-fallback beta” to use a currently live Anthropic model documented
elsewhere in the harness, such as claude-opus-4-8 or claude-sonnet-5. Keep the
fallback payload and beta-header assertions unchanged.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 43d38f3d-ef17-490b-b5e5-935004ed23e8

📥 Commits

Reviewing files that changed from the base of the PR and between db9bd25 and 42ad071.

📒 Files selected for processing (14)

• core/providers/anthropic/chat.go
• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/serversidefallback_test.go
• core/providers/anthropic/types.go
• core/providers/anthropic/utils.go
• core/providers/anthropic/utils_test.go
• core/providers/bedrock/responses.go
• core/providers/cohere/responses.go
• core/providers/gemini/responses.go
• core/providers/openai/responses.go
• core/providers/openai/responses_test.go
• core/schemas/responses.go
• tests/e2e/api/collections/provider-harness.json

🚧 Files skipped from review as they are similar to previous changes (5)

• core/providers/anthropic/chat.go
• core/schemas/responses.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/types.go
• core/providers/anthropic/utils_test.go

[coderabbitai]

🧹 Nitpick comments (1)

core/providers/anthropic/utils.go (1)

504-512: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Minor: inconsistent error handling for the fallback_credit_token strip.

Every other strip in this function wraps the error with descriptive context (fmt.Errorf("strip raw <field>: %w", err)) and reuses the outer err. This branch shadows the outer err with a fresh var err error and returns it unwrapped, losing the "which strip failed" context.

♻️ Align with sibling strips

 	if !features.FallbackCredit {
-		var err error
 		jsonBody, err = providerUtils.DeleteJSONField(jsonBody, "fallback_credit_token")
 		if err != nil {
-			return nil, err
+			return nil, fmt.Errorf("strip raw fallback_credit_token: %w", err)
 		}
 	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@core/providers/anthropic/utils.go` around lines 504 - 512, Update the
fallback_credit_token removal branch in the surrounding function to reuse the
existing outer err variable instead of declaring a shadowing local error, and
wrap DeleteJSONField failures with descriptive context consistent with the
sibling field-stripping branches before returning.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@core/providers/anthropic/utils.go`:
- Around line 504-512: Update the fallback_credit_token removal branch in the
surrounding function to reuse the existing outer err variable instead of
declaring a shadowing local error, and wrap DeleteJSONField failures with
descriptive context consistent with the sibling field-stripping branches before
returning.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: f0a88200-000a-49f6-bd7b-4239e44ac7f1

📥 Commits

Reviewing files that changed from the base of the PR and between 42ad071 and b587562.

📒 Files selected for processing (14)

• core/providers/anthropic/chat.go
• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/serversidefallback_test.go
• core/providers/anthropic/types.go
• core/providers/anthropic/utils.go
• core/providers/anthropic/utils_test.go
• core/providers/bedrock/responses.go
• core/providers/cohere/responses.go
• core/providers/gemini/responses.go
• core/providers/openai/responses.go
• core/providers/openai/responses_test.go
• core/schemas/responses.go
• tests/e2e/api/collections/provider-harness.json

🚧 Files skipped from review as they are similar to previous changes (10)

• core/providers/anthropic/chat.go
• core/providers/bedrock/responses.go
• core/providers/openai/responses.go
• core/providers/openai/responses_test.go
• tests/e2e/api/collections/provider-harness.json
• core/providers/cohere/responses.go
• core/schemas/responses.go
• core/providers/anthropic/requestbuilder.go
• core/providers/anthropic/responses.go
• core/providers/anthropic/types.go