fix(providers/anthropic): forward and rebuild server-side tool_search on the Responses path - #4786
Merged
akshaydeo merged 94 commits intoJul 12, 2026
Conversation
…dle special characters in URLs (maximhq#4714) ## Summary Array-type query parameters in the logs page (models, providers, etc.) were not using the safe URI-encoding parser, meaning values containing characters like `://` (e.g. model names such as `gpt://host/model`) could be misinterpreted as path or query delimiters by TanStack Router. This introduces `parseAsSafeArrayOf` and applies it consistently across all array filters. ## Changes - Added `parseAsSafeArrayOf` to `queryParamsParser.ts` by composing `parseAsArrayOf` with the existing `parseAsSafeString` parser, ensuring full URI-encoding for comma-separated filter values. - Replaced all usages of `parseAsArrayOf(parseAsString)` in the logs page with `parseAsSafeArrayOf` so that array filters (models, providers, aliases, status, etc.) benefit from the same encoding guarantees as string filters. - Added unit tests for both `parseAsSafeString` and `parseAsSafeArrayOf` to verify round-trip correctness with model names containing `://`. ## Type of change - [x] Bug fix ## Affected areas - [x] UI (React) ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` To manually verify, navigate to the logs page and apply a filter using a model name containing `://` (e.g. `gpt://host/model`). Confirm the URL encodes correctly and the filter persists on page reload without routing errors. ## Screenshots/Recordings N/A ## Breaking changes - [x] No ## Related issues maximhq#4603 ## Security considerations No security implications. This change only affects URL query parameter encoding in the UI. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
…4715) ## Summary Adds a reusable `TruncatedLabel` component that displays truncated text with an automatic tooltip when the content overflows its container. ## Changes - Introduces `TruncatedLabel`, a `<span>`-based component that detects when its text content is truncated via CSS overflow and conditionally renders a `Tooltip` to show the full content - Truncation detection is performed by comparing `scrollWidth` to `clientWidth`, and re-evaluated on window resize or when `children` changes - The tooltip content defaults to the `children` value if it is a string, but accepts an explicit `tooltip` prop for custom content - The tooltip is only rendered when the text is actually truncated, avoiding unnecessary DOM overhead when content fits ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm build || npm run build ``` Render a `TruncatedLabel` inside a constrained-width container with a long string. Verify that hovering over the truncated text shows a tooltip with the full content, and that no tooltip appears when the text is not truncated. ## Screenshots/Recordings N/A ## Breaking changes - [ ] Yes - [x] No ## Related issues N/A ## Security considerations No security implications. The component renders user-provided React nodes, which is consistent with existing UI patterns. ## 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
…ation logic (maximhq#4716) ## Summary Extracts the truncated label logic (truncate text with a tooltip on overflow) into a shared `TruncatedLabel` component and replaces all inline implementations with it. ## Changes - Added a reusable `TruncatedLabel` component in `ui/components/ui/truncatedLabel` that handles text truncation and conditionally renders a tooltip when the content overflows - Removed the local `TruncatedName` component from the providers page, which duplicated this logic using `useRef`, `useState`, and a resize event listener - Replaced inline `<span className="truncate ...">` elements in the logs and MCP filter sidebars with `TruncatedLabel` ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` Verify that truncated labels in the providers list and filter sidebars still show a tooltip on hover when the text overflows, and no tooltip when it does not. ## Screenshots/Recordings Verify the providers sidebar and log/MCP filter sidebars visually behave the same as before — truncated text shows a tooltip, non-truncated text does not. ## Breaking changes - [ ] Yes - [x] No ## Related issues maximhq#4604 ## Security considerations None. ## 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
…ximhq#4719) ## Summary Fixes the layout of the Audit Logs page to correctly fill the viewport and apply the appropriate background and border styles. ## Changes - Replaced `h-[calc(100dvh-1rem)]` with `h-[calc(100vh-16px)]` for consistent viewport height calculation - Swapped `mx-auto flex flex-col p-4` utility classes for `no-border-parent bg-background flex` to align with the layout conventions used elsewhere in the app - Added missing newline at end of file ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test Navigate to the Audit Logs page and verify: - The page fills the full viewport height without overflow or clipping - The background color and border styling match the rest of the workspace layout ```sh cd ui pnpm i || npm i pnpm build || npm run build ``` ## Screenshots/Recordings Add before/after screenshots showing the corrected Audit Logs page layout. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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
## Summary Adds collapsible tag support to the `TagInput` component and applies it to the keyword lists on the Complexity Router page. When a keyword list exceeds a configurable limit, tags beyond that limit are hidden behind a gradient overlay with a "Show more" toggle, keeping the UI compact while still allowing full access to all tags. ## Changes - Added `collapsedTagLimit` and `expandButtonTestId` props to `TagInput`. When `collapsedTagLimit` is provided, the component renders in a collapsible layout: tags beyond the limit are hidden with a fade gradient, and "Show more" / "Show less" buttons toggle the expanded state. The collapsed state auto-resets when the tag count drops back to or below the limit. - Set `KEYWORD_COLLAPSED_LIMIT = 8` on the Complexity Router page and passed it along with a `expandButtonTestId` to each keyword `TagInput`. - Standardized border radius tokens from `rounded-lg`/`rounded-md`/`rounded-full` to `rounded-sm` across the Complexity Router page for visual consistency. - Reformatted `index.html` inline shell skeleton from a single minified line to readable, indented HTML and CSS. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` 1. Navigate to the Complexity Router page. 2. Add more than 8 keywords to any keyword list. 3. Verify that tags beyond 8 are hidden with a gradient overlay and a "Show more" button appears. 4. Click "Show more" and confirm all tags are visible with a "Show less" button. 5. Click "Show less" and confirm the list collapses again. 6. Remove tags until 8 or fewer remain and confirm the list stays expanded without the toggle controls. ## Screenshots/Recordings Before/after screenshots of the keyword lists with collapse behavior recommended. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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
…maximhq#4722) * fix: rebuild token_usage from denormalized columns in hybrid log list * refactor: inline hybrid token usage reconstruction * fix: preserve malformed serialized token usage state --------- Co-authored-by: gexiangdong <xiangdong.ge@pandasofcaribbean.com>
…hatRequestBody` and `AnthropicProviderRequestDefaultsMap` (maximhq#3309) ## Summary This PR consolidates Anthropic-family request building across all providers (Anthropic native, Azure, Vertex, Bedrock) into two shared builder functions — `BuildAnthropicChatRequestBody` and `BuildAnthropicResponsesRequestBody` — eliminating duplicated inline logic and provider-specific wrapper helpers that previously scattered the same field-stripping, beta-header injection, and model-field manipulation across multiple files. ## Changes - Introduced `AnthropicProviderRequestDefaults` and `AnthropicProviderRequestDefaultsMap` to encode static, per-provider request-shaping flags (e.g. `DeleteModelField`, `DeleteStreamField`, `AddAnthropicVersion`, `InjectBetaHeadersIntoBody`) in one place. Callers no longer pass these flags directly; the builder looks them up by `cfg.Provider`. - Renamed `Deployment` to `Model` in `AnthropicRequestBuildConfig` for clarity, since all providers now use the same field for model/deployment overrides. - Added `BuildAnthropicChatRequestBody` as the chat-completion analogue of `BuildAnthropicResponsesRequestBody`, covering both raw-body and typed paths, including field stripping, beta-header injection, streaming flag handling, and `fallbacks` deletion. - Bedrock now routes Anthropic models through the Anthropic Messages API format (`invoke` / `invoke-with-response-stream` endpoints) for both chat and responses, rather than the Bedrock Converse API. This includes proper response parsing via `AcquireAnthropicMessageResponse` and streaming via `AnthropicStreamState` / `AnthropicResponsesStreamState`. - Removed private wrapper functions `getRequestBodyForResponses` (Anthropic), `getRequestBodyForAnthropicResponses` (Azure, Vertex), and the inline `CheckContextAndGetRequestBody` closures for Anthropic models in Vertex and Azure, replacing all call sites with direct `BuildAnthropicChatRequestBody` / `BuildAnthropicResponsesRequestBody` calls. - Exported `AcquireAnthropicResponsesStreamState`, `ReleaseAnthropicResponsesStreamState`, `AcquireAnthropicMessageResponse`, and `ReleaseAnthropicMessageResponse` so Bedrock can reuse the Anthropic stream state pool. - Removed `DefaultVertexAnthropicVersion` constant from the Vertex package; the canonical version string now lives in `AnthropicProviderRequestDefaultsMap`. - Bedrock's `releaseBedrockChatResponse` now zeroes the struct before returning it to the pool. - `stripUnsupportedAnthropicFields` is now called inside `BuildAnthropicResponsesRequestBody` on the typed path, making field stripping symmetric across raw and typed paths and across both APIs. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./core/providers/anthropic/... go test ./core/providers/azure/... go test ./core/providers/bedrock/... go test ./core/providers/vertex/... go test ./... ``` Run integration tests against Anthropic, Azure (Anthropic models), Vertex (Claude models), and Bedrock (Claude models) for chat completion, streaming, responses, responses streaming, and count-tokens endpoints. Verify that raw-body passthrough requests produce the same field stripping and beta-header injection as typed requests. ## Breaking changes - [x] Yes - [ ] No `AnthropicRequestBuildConfig` has a breaking field rename: `Deployment` → `Model`. Any external code constructing this struct directly must update the field name. The static shaping flags (`DeleteModelField`, `DeleteRegionField`, `AddAnthropicVersion`, `AnthropicVersion`, `StripCacheControlScope`, `RemapToolVersions`, `InjectBetaHeadersIntoBody`) have been removed from `AnthropicRequestBuildConfig` and are now looked up internally via `AnthropicProviderRequestDefaultsMap`; callers that set these fields must remove them. ## Related issues ## Security considerations No new auth flows, secrets handling, or PII exposure introduced. Field stripping ensures provider-unsupported fields are not forwarded to external APIs. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
…hropicResponsesRequest` and make `completeRequest` a package-level func shared by Anthropic, Azure, and Bedrock providers (maximhq#4394) ## Summary The Anthropic provider's unary request logic was duplicated across the Anthropic, Azure, Bedrock, and Vertex providers. This PR extracts the core non-streaming request execution into a package-level `completeRequest` function and introduces two exported handler functions — `HandleAnthropicChatCompletionRequest` and `HandleAnthropicResponsesRequest` — that encapsulate the full build → send → parse pipeline for chat completions and the Responses API respectively. Azure and Bedrock now delegate directly to these shared handlers for Anthropic-family models instead of reimplementing request dispatch, response parsing, and raw request/response handling inline. A secondary bug fix is included: the large-response streaming client was being activated for count-tokens requests (which should always be buffered) and skipped for all other requests — the condition was inverted. ## Changes - Extracted `completeRequest` as a package-level function accepting explicit `client`, `headers`, `extraHeaders`, `betaHeaderOverrides`, `providerName`, and `logger` arguments, removing the method receiver dependency so it can be called by other providers. - Added `anthropicRequestHeaders` as a provider method to build the `x-api-key` / `anthropic-version` header map, shared across `TextCompletion`, `ChatCompletion`, `Responses`, and `CountTokens`. - Introduced `HandleAnthropicChatCompletionRequest` and `HandleAnthropicResponsesRequest` as exported functions that perform the full unary request lifecycle (body build, HTTP send, large-response detection, response parse, raw request/response attachment). These are now called by the Anthropic, Azure, and Bedrock providers. - Removed `completeMantleRequest` from Bedrock — its logic is now covered by `completeRequest` inside the shared handlers. - Azure's `ChatCompletion` and `Responses` methods now branch early for Anthropic-family models, calling the shared handlers with Azure-specific auth headers, and fall through to the OpenAI-compatible path otherwise, eliminating the post-response model-family branch. - Fixed the inverted condition in `completeRequest` that caused the large-response streaming client to be used for count-tokens requests instead of being skipped for them. - `AnthropicRequestBuildConfig` now carries `BetaHeaderOverrides` so callers do not need to pass it separately. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` Validate that chat completions and Responses API requests succeed for Anthropic-family models routed through the Azure and Bedrock providers, and that count-tokens requests return buffered responses without triggering large-response mode. ## Breaking changes - [ ] Yes - [x] No ## Security considerations Auth headers (`x-api-key`, Bearer tokens, SigV4-signed headers) are applied last in `completeRequest`, after network-config extra headers, ensuring they cannot be overridden by user-supplied configuration. No new secrets or PII handling paths are introduced. ## 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
…atible providers from `schemas.Key` to `map[string]string` auth header param (maximhq#4425) ## Summary This PR standardizes how Bearer token authentication headers are constructed across all OpenAI-compatible providers. Previously, each call site independently built the `Authorization: Bearer <token>` header map with duplicated inline logic. A new `BearerAuthHeader(key)` helper is introduced in the OpenAI package and used uniformly everywhere. Additionally, the Azure provider's private `completeRequest` method is removed. Its non-Anthropic request paths (text completion, chat completion, responses, embedding, compaction) are now delegated directly to the shared `Handle*` functions in the OpenAI package, consistent with how other providers already work. The `Handle*` functions themselves are updated to accept a pre-built `authHeader map[string]string` instead of a raw `schemas.Key`, making them provider-agnostic and compatible with non-Bearer auth schemes (e.g., Azure API key headers, SigV4). ## Changes - Added `BearerAuthHeader(key schemas.Key) map[string]string` to the OpenAI provider package, which returns an `Authorization: Bearer <token>` header map, or an empty map when the key carries no value. - Updated all `Handle*Request` and `handleOpenAILargePayloadPassthrough` function signatures to accept `authHeader map[string]string` instead of `schemas.Key`, applying the map directly to request headers. - Replaced all inline `var authHeader map[string]string` + conditional assignment blocks across Cerebras, Fireworks, Groq, HuggingFace, Mistral, Nebius, Ollama, Opencode, OpenRouter, Parasail, Perplexity, SGL, VLLM, xAI, and Bedrock with calls to `openai.BearerAuthHeader(key)`. - Removed the Azure provider's `completeRequest` method and replaced its usage in `TextCompletion`, `ChatCompletion`, `Responses`, `Embedding`, and `Compaction` with direct calls to the corresponding shared OpenAI `Handle*` functions, passing Azure-specific auth headers and pre-resolved endpoint URLs. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` Verify that all OpenAI-compatible providers (OpenAI, Azure, Cerebras, Fireworks, Groq, HuggingFace, Mistral, Nebius, Ollama, Opencode, OpenRouter, Parasail, Perplexity, SGL, VLLM, xAI, Bedrock Mantle) continue to authenticate correctly and that requests succeed for text completion, chat completion, responses, embeddings, and compaction endpoints. ## Breaking changes - [ ] Yes - [x] No ## Security considerations The `BearerAuthHeader` helper preserves the existing behavior of omitting the `Authorization` header when the key value is empty, which is intentional for providers that supply auth via other mechanisms (e.g., extra headers or SigV4 signing). No secrets are logged or exposed. ## 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
… closure passed through request handlers (maximhq#4735) ## Summary Replaces the pre-build-and-sign approach for Bedrock Mantle SigV4 authentication with a `BodySigner` callback that is invoked after the request handler has marshaled the body. This ensures the signature always covers the exact bytes sent on the wire, eliminating the previous double-marshal pattern where the body was built once for signing and again inside the handler. ## Changes - Introduces a new `BodySigner` type (`func(jsonData []byte) (map[string]string, *schemas.BifrostError)`) in `core/providers/utils/bodysigner.go`. Handlers call it after building the request body and apply the returned headers to the outgoing request. - Adds the `signer` parameter to `HandleOpenAIChatCompletionRequest`, `HandleOpenAIChatCompletionStreaming`, `HandleOpenAIResponsesRequest`, `HandleOpenAIResponsesStreaming`, `HandleAnthropicChatCompletionRequest`, `HandleAnthropicChatCompletionStreaming`, `HandleAnthropicResponsesRequest`, and `HandleAnthropicResponsesStream`. All existing callers pass `nil`. - Rewrites Bedrock Mantle's SigV4 paths (`mantleChatCompletions`, `mantleChatCompletionsStream`, `mantleResponses`, `mantleResponsesStream`) to construct a `BodySigner` closure when no API key is present, instead of pre-building the body, signing it, and merging the signature headers into `extraHeaders`. The Bearer path no longer needs a separate early-return branch. - Removes the now-unnecessary `maps` import and the intermediate `extraHeaders` map copies in the Mantle code paths. ## Type of change - [ ] Bug fix - [x] Refactor - [ ] Feature - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` For Bedrock Mantle with SigV4 (empty key value), verify that requests to chat completions, streaming chat completions, responses, and streaming responses are signed correctly and accepted by the Bedrock endpoint. For Bearer key paths, confirm that no signing is attempted and the `Authorization` header is set as expected. ## Breaking changes - [ ] Yes - [x] No ## Security considerations The `BodySigner` callback signs the exact serialized bytes that are placed on the wire. Previously, the body was serialized twice (once for signing, once inside the handler), which could in theory produce a signature mismatch if marshaling were non-deterministic. This change closes that gap by signing after the final body is set. ## 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
…ropic and OpenAI-compatible routing (maximhq#4736) ## Summary Introduces `bedrock_mantle` as a first-class, standalone provider that owns the Bedrock Mantle surface (`bedrock-mantle.{region}.api.aws`). Previously, Mantle routing was handled as an internal routing decision inside the existing `bedrock` provider. The new provider gives operators a dedicated configuration surface for Claude (native Anthropic Messages API), OpenAI-compatible models (gpt-*), and Gemma models served through Mantle, without requiring a full Bedrock setup. ## Changes - Added `schemas.BedrockMantle` (`"bedrock_mantle"`) as a new `ModelProvider` constant and registered it in `StandardProviders`, `dynamicallyConfigurableProviders`, `CanProviderKeyValueBeEmpty`, and `isKeySkippingAllowed`. - Added `BedrockMantleKeyConfig` to the `Key` struct, carrying AWS credentials and region for SigV4 auth against the `bedrock-mantle` service. The existing `BedrockKeyConfig` is unchanged. - Introduced the `core/providers/bedrockmantle` package implementing the full `Provider` interface. Chat, streaming chat, Responses, and streaming Responses dispatch by model family: Anthropic-family models use the native Anthropic Messages surface (`/anthropic/v1/messages`); all others use the OpenAI-compatible surface (`/v1` or `/openai/v1`). All other operations return unsupported-operation errors. - Refactored `signAWSRequest` in the `bedrock` package to accept a `*BedrockKeyConfig` instead of individual credential fields, eliminating the now-redundant `signAWSRequestFromKey` wrapper. All call sites updated accordingly. - Exported `SignMantleV4Headers` (previously `mantleSigV4Headers`, a method on `BedrockProvider`) so the new `bedrockmantle` package can sign requests without depending on the internal Bedrock provider struct. The function now supports both `BedrockKeyConfig` and `BedrockMantleKeyConfig` by mapping the latter into a synthetic `BedrockKeyConfig` for signing, and correctly handles GET requests (nil body) for the list-models path. - Extended the Anthropic chat and Responses request builders to convert native structured outputs to tool calls for `BedrockMantle`, matching the existing `Vertex` workaround. - Added `BedrockMantle` to the comprehensive LLM test harness (`ComprehensiveTestAccount`) with key config, provider config, and a full test file covering the supported scenarios (chat, streaming, tool calls, vision, structured outputs, prompt caching, reasoning, list models) and explicitly disabling unsupported ones. - Marked `isMantleModel` in `bedrock/mantle.go` as deprecated in favour of the new provider. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Set AWS credentials and run the new provider test: ```sh export AWS_ACCESS_KEY_ID=... export AWS_SECRET_ACCESS_KEY=... export AWS_SESSION_TOKEN=... # optional, for temporary credentials export AWS_REGION=us-east-1 go test ./core/providers/bedrockmantle/... -v -run TestBedrockMantle ``` To run the full suite (skips Bedrock Mantle automatically when credentials are absent): ```sh go test ./... ``` Configure a `bedrock_mantle` provider by supplying a `BedrockMantleKeyConfig` (or a Bearer API key in `Value`) with the desired region. The region can also be embedded as a prefix in the model ID (e.g. `us-west-2/anthropic.claude-haiku-4-5`) or set at the alias level via `AliasConfig.Region`. ## Breaking changes - [ ] Yes - [x] No The `signAWSRequest` signature change is internal to the `bedrock` package and does not affect any public API. The `isMantleModel` function is deprecated but not removed. ## Security considerations AWS credentials for `BedrockMantleKeyConfig` follow the same `SecretVar` resolution pattern used by `BedrockKeyConfig` (env-var references, never inlined literals). SigV4 signing is performed per-request on the exact body bytes that are sent, so the signature always covers what is transmitted. When a Bearer API key is present it takes precedence and no AWS credentials are required. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
…on, and UI support (maximhq#4737) ## Summary Adds `bedrock_mantle` as a first-class provider, enabling Bifrost to route requests to AWS Bedrock through a Mantle proxy endpoint. The provider supports the same SigV4 credential options as the existing Bedrock provider (inherited IAM role, explicit access/secret key, session token, AssumeRole) as well as a Bearer API key authentication mode. ## Changes - Added `BedrockMantle` to the Anthropic passthrough allowlist in `clearAnthropicPassthroughForNonNativeProvider` so raw request bodies are preserved when routing through Bedrock Mantle. - Added `BedrockMantleKeyConfig` redaction logic in `clientconfig.go`, mirroring the existing Bedrock redaction pattern. - Added a new `migrationAddBedrockMantleKeyColumns` database migration that introduces seven `bedrock_mantle_*` SigV4 credential columns to the `config_keys` table. - Extended `TableKey` with the seven Bedrock Mantle credential fields, along with `BeforeSave` serialization and `AfterFind` reconstruction hooks. - Updated `mergeUpdatedKey` in the HTTP handler to correctly restore redacted Bedrock Mantle credential fields during key updates. - Fixed `isClaudeModel` in the Anthropic integration to recognize `bedrock_mantle` (previously incorrectly matched `bedrock`) as a provider that can serve Claude models. - Included `BedrockMantleKeyConfig` in the key hash inputs used by `mergeProviderKeys` and `reconcileProviderKeys` for config file/DB reconciliation. - Added Bedrock Mantle credential redaction to `GetAllKeys`. - Extended `config.schema.json` with `bedrock_mantle_key` and `provider_with_bedrock_mantle_config` definitions and registered `bedrock_mantle` as a valid provider name throughout the schema. - Added UI support: provider icon (reusing the Bedrock SVG mark with a distinct gradient ID), model placeholder text, `isKeyRequiredByProvider` entry, label, form schema (`BedrockMantleKeyConfigSchema`), type definitions (`BedrockMantleKeyConfig`, `DefaultBedrockMantleKeyConfig`), and a full authentication method tab UI (IAM Role / Explicit Credentials / API Key) matching the Bedrock provider UX. - Added `bedrock_mantle` to the Anthropic beta-headers provider family and the provider config sheet's Anthropic family list. - Stripped the internal `_auth_type` field from `bedrock_mantle_key_config` before submitting the form payload. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [x] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh # Core/Transports go version go test ./... # UI cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` Configure a `bedrock_mantle` provider in `config.json` or via the UI with one of the three auth methods: - **IAM Role (Inherited):** set only `region`; leave access/secret key empty. - **Explicit Credentials:** set `access_key`, `secret_key`, and `region`; optionally set `session_token`, `role_arn`, `external_id`, and `session_name`. - **API Key:** set `region` and provide a Bearer token as the key `value`. Send a request targeting a Claude model through the `bedrock_mantle` provider and verify the response is returned correctly and that credentials are redacted in the UI and API responses. ## Screenshots/Recordings _Add before/after screenshots of the new Bedrock Mantle provider form and icon in the UI._ ## Breaking changes - [ ] Yes - [x] No ## Related issues _Link related issues and discussions._ ## Security considerations - All seven Bedrock Mantle credential fields (`access_key`, `secret_key`, `session_token`, `region`, `role_arn`, `external_id`, `role_session_name`) are stored as `SecretVar` and are redacted in API responses and the UI, consistent with the existing Bedrock provider handling. - The `_auth_type` discriminator field is stripped from the payload before it is persisted or transmitted. ## 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
## Summary
Adds documentation for the AWS Bedrock Mantle provider, a distinct AWS endpoint (`bedrock-mantle.{region}.api.aws`) that exposes Claude models via the native Anthropic Messages API and OpenAI-family/Gemma models via an OpenAI-compatible API — all addressable through a single `bedrock_mantle/<model>` prefix in Bifrost.
## Changes
- Added a new `bedrock-mantle.mdx` provider page covering model ID formats, supported operations, all three authentication modes (SigV4 with explicit credentials, IAM role/inherited credentials, and Bearer API key), IAM role assumption via `role_arn`, and usage examples.
- Added the Bedrock Mantle configuration block to the `providers.mdx` config reference, with tabs for Static Credentials, IAM Role, and API Key (Bearer) auth modes.
- Added Bedrock Mantle to the provider capability matrix in `overview.mdx`.
- Registered `bedrock-mantle` in `docs.json` so it appears in the sidebar navigation.
## Type of change
- [ ] Bug fix
- [ ] Feature
- [ ] Refactor
- [x] Documentation
- [ ] Chore/CI
## Affected areas
- [ ] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (React)
- [x] Docs
## How to test
Navigate to the Bedrock Mantle provider page and config reference in the rendered docs and verify:
- The sidebar entry for `bedrock-mantle` appears between `bedrock` and `cerebras`.
- All three auth tabs (Static Credentials, IAM Role, API Key) render correctly in both the provider page and the config reference.
- The capability matrix row for `bedrock_mantle/<model>` is present and accurate.
- Cross-links between the provider page and the config reference resolve correctly.
## Breaking changes
- [x] No
## Security considerations
Authentication credentials (`access_key`, `secret_key`, `session_token`, API keys) are documented using the `env.*` indirection pattern, consistent with how other providers handle secrets. No credentials are hardcoded in examples.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
…ment config (maximhq#4739) ## Summary Adds E2E test configuration and capability definitions for the `bedrock_mantle` provider, enabling it to be tested through the Bifrost V1 API test suite. ## Changes - Added `bedrock_mantle` to `provider-capabilities.json` with `chat_completions`, `chat_completions_with_tools`, `responses`, `responses_with_tools`, and `list_models` enabled - Added a new Postman environment file (`bifrost-v1-bedrock-mantle.postman_environment.json`) configured to use `anthropic.claude-opus-4-8` as the default model and `us-east-1` as the default region, with secret placeholders for API key, access key, secret key, and session token ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Run the E2E test suite targeting the `bedrock_mantle` provider using the new Postman environment: ```sh # Ensure the Bifrost server is running locally on port 8080 # Load the environment file and run the collection against bedrock_mantle newman run tests/e2e/api/bifrost-v1.postman_collection.json \ -e tests/e2e/api/provider_config/bifrost-v1-bedrock-mantle.postman_environment.json \ --env-var "bedrock_mantle_api_key=<your_api_key>" \ --env-var "bedrock_mantle_access_key=<your_access_key>" \ --env-var "bedrock_mantle_secret_key=<your_secret_key>" ``` Expected outcome: chat completions, tool-use, responses, and model listing tests pass; all unsupported capability tests are skipped or return expected errors. ## Screenshots/Recordings N/A ## Breaking changes - [ ] Yes - [x] No ## Related issues N/A ## Security considerations The Postman environment file stores API key, access key, secret key, and session token as `secret` type fields with empty default values, ensuring credentials are not committed to the repository. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
…d toggle when budget is alignable (maximhq#4724) ## Summary The calendar alignment toggle in the provider governance form was previously shown whenever any budget existed. This PR restricts its visibility and submission to only when at least one budget uses a calendar-alignable reset period (day, week, month, or year). ## Changes - Introduced a `showCalendarAlignment` derived boolean that checks whether any configured budget has a reset duration supported by `supportsCalendarAlignment`. - Replaced the previous condition (`watchedBudgets.length > 0`) with `showCalendarAlignment` to control rendering of the calendar alignment toggle. - Updated the form submission payload so that `calendar_aligned` is only set to `true` when at least one budget actually supports calendar alignment — preventing the flag from being submitted for incompatible budget configurations. ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test 1. Navigate to a provider's governance settings in the UI. 2. Add a budget with a reset duration that does **not** support calendar alignment (e.g., hourly). Verify the calendar alignment toggle does **not** appear. 3. Add or change a budget to use a calendar-alignable period (e.g., daily, weekly, monthly, yearly). Verify the toggle **does** appear. 4. Enable the toggle and save. Confirm `calendar_aligned: true` is included in the submitted payload. 5. Remove all calendar-alignable budgets and save. Confirm `calendar_aligned` is not set to `true` in the payload. ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` ## Screenshots/Recordings _Before:_ Calendar alignment toggle appears whenever any budget is present, regardless of reset period. _After:_ Calendar alignment toggle only appears when at least one budget uses a day/week/month/year reset period. ## Breaking changes - [x] No ## Related issues ## Security considerations None. ## 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
…hq#4725) ## Summary Fixes a bug where updating or deleting provider-level governance for a custom provider whose name contains a space (e.g. `"OpenRouter Base"`) would return a 404. The UI percent-encodes the provider name in the URL path (`OpenRouter%20Base`), but the handler was comparing the raw encoded string directly against the stored provider name, causing the lookup to fail. Closes maximhq#4689 ## Changes - `updateProviderGovernance` and `deleteProviderGovernance` now call `url.PathUnescape` on the `provider_name` path parameter before using it, matching the decoded name against what is stored in the config store. - Returns a `400` if the path parameter contains an invalid percent-encoding sequence. - Added a regression test (`TestProviderGovernance_DecodesEncodedProviderName`) that seeds a provider with a space in its name, issues a PUT and DELETE using the percent-encoded path param, and asserts both succeed and persist correctly. - Added a guard test (`TestProviderGovernance_UnknownProviderStill404`) to confirm that a genuinely unknown provider still returns 404 after the decode change. ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./transports/bifrost-http/handlers/... -run TestProviderGovernance ``` Expected output: all three `TestProviderGovernance_*` tests pass. Specifically: - `TestProviderGovernance_DecodesEncodedProviderName` — PUT and DELETE with `OpenRouter%20Base` return `200`. - `TestProviderGovernance_UnknownProviderStill404` — PUT with an unknown encoded name returns `404`. ## Breaking changes - [ ] Yes - [x] No ## Related issues - Resolves maximhq#4689 ## Security considerations `url.PathUnescape` is used rather than `url.QueryUnescape` to correctly handle path-encoded characters. Invalid encoding sequences are rejected with a `400` rather than passed through, preventing malformed input from reaching the config store. ## Checklist - [x] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
… (maximhq#4568) Signed-off-by: Akshay Deo <akshay@akshaydeo.com> Co-authored-by: Akshay Deo <akshay@akshaydeo.com>
* Fix mcp reconnect failure on startup * test: assert failed MCP client cleanup --------- Co-authored-by: Gowtham <692171+HackToHell@users.noreply.github.com>
…put items (maximhq#4121) * fix(responses): preserve codex tool_search_call/tool_search_output input items Bifrost's Responses input deserializer rejected codex's tool-search follow-up request with HTTP 400 "openai responses request input is neither a string nor an array of responses messages", which hung/failed the agent turn. The fix teaches ResponsesMessage about the two tool_search item types and round-trips them verbatim. Background, since tool_search is non-obvious: How codex's tool_search works (the path that hits this bug) ----------------------------------------------------------- codex normally sends every MCP tool inline in the request `tools[]` as `{type:"function", ...}`. But when a model's catalog has `supports_search_tool: true` AND the tool count crosses DIRECT_MCP_TOOL_EXPOSURE_THRESHOLD (= 100) — e.g. an agent wired to several MCP servers — codex stops sending them inline and "defers" them behind a discovery tool: should_defer = supports_search_tool && (ToolSearchAlwaysDeferMcpTools || n >= 100) The deferred flow is a two-request round-trip: 1. Request 1: codex hides the deferred tools and instead declares one tool: {"type":"tool_search","execution":"client","description":"...", "parameters":{query, limit}} `execution:"client"` means the model does NOT run the search — codex does. 2. The model emits a `tool_search_call` with `arguments` = {query, limit}. 3. codex runs the search CLIENT-SIDE: a BM25 index over the deferred tool metadata (codex's ToolSearchHandler, core/src/tools/handlers/tool_search.rs, using the `bm25` crate). It picks the top-N matching tools. 4. Request 2 (follow-up): codex appends two items to `input[]`: - {"type":"tool_search_call", "call_id":..., "execution":"client", "arguments":{...}} - {"type":"tool_search_output", "call_id":..., "status":"completed", "execution":"client", "tools":[ {type:"function", ...the matches} ]} and also surfaces the discovered tools in `tools[]`. The model can now call them. This repeats as the model needs more tools. Root cause ---------- ResponsesMessage (the element type of the Responses `input` array AND the response `Output` array) doesn't model `tool_search_call` / `tool_search_output`: - The call's `arguments` is a JSON OBJECT, whereas function_call's `arguments` is a JSON STRING. So it cannot decode into ResponsesToolMessage.Arguments (*string) -> sonic.Unmarshal of the whole []ResponsesMessage errors -> OpenAIResponsesRequestInput.UnmarshalJSON falls through to the "neither a string nor an array" 400. The entire request dies before reaching OpenAI. - The output's `tools` array is also unmodeled (would be dropped/mangled, which OpenAI then rejects with "Missing input[N].tools[0].type"). OpenAI's Responses API supports both items natively (verified end-to-end against the gateway: the tool_search tool spec is accepted and echoed; OpenAI validates arguments-as-object and tools[].type). So this is purely a Bifrost modelling gap, in the same family as the tool-type allowlist that already lists ResponsesToolTypeToolSearch / ResponsesToolTypeNamespace — just a different code path (input-item deserialization vs the request tools[] allowlist). Fix --- Add ResponsesMessageTypeToolSearchCall / ResponsesMessageTypeToolSearchOutput and give ResponsesMessage custom (Un)MarshalJSON that preserves these two item types verbatim (original bytes in, original bytes out), so the object `arguments` and the `tools` array survive intact. Every other item type defers to the default struct (de)coding, unchanged. One change covers both directions because request input and response Output are both []ResponsesMessage. Impact: unblocks codex tool-search deferral (multi-MCP-server / >=100-tool agents) through Bifrost. Verified with a round-trip test reproducing the exact follow-up payload, plus the existing providers/openai and schemas suites (no regressions). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(responses): reset ResponsesMessage receiver in UnmarshalJSON Clear the receiver at the top of ResponsesMessage.UnmarshalJSON so a reused instance never retains a stale rawToolSearch (or other field) from a prior decode. Without this, unmarshalling a tool_search item and then a normal message into the same value would leave the preserved bytes in place, and MarshalJSON would re-emit them. Not reachable via the array-decode path (each element starts zero), but a cheap, defensive correctness fix. Addresses CodeRabbit review on PR maximhq#4121. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(openapi): correct prompt_cache_retention enum to in_memory The chat schema declared the enum as [in-memory, 24h], but OpenAI's actual accepted values are in_memory (underscore) and 24h. The hyphenated form was a typo from when the enum was first added and never matched OpenAI, so spec-generated clients produced Literal['in-memory', '24h'] and rejected the valid value with a pydantic literal_error. The Go runtime treats prompt_cache_retention as a pass-through *string, so no behavior changes — only the spec enum, the regenerated openapi.json, and the doc comment are corrected. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: Akshay Deo <akshay@akshaydeo.com> Co-authored-by: Suresh Kumar Ponnusamy <suresh@atomicwork.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Akshay Deo <akshay@akshaydeo.com>
## Summary Fixes a bug in the Vertex provider where the code path for handling Gemini/Gemma model families was duplicated, with the non-streaming branch incorrectly using `ToGeminiChatCompletionRequest` (without image URL scheme support) while the streaming branch used `ToGeminiChatCompletionRequestWithImageURLSchemes`. This consolidates the logic so both paths use the image URL scheme-aware converter. ## Changes - Replaced `ToGeminiChatCompletionRequest` with `ToGeminiChatCompletionRequestWithImageURLSchemes` in the non-streaming Gemini/Gemma branch, making it consistent with the streaming branch - Removed the duplicate non-streaming Gemini/Gemma and OpenAI handler blocks that had been incorrectly separated from the streaming path, consolidating them into a single unified code path ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Send a chat completion request to the Vertex provider using a Gemini or Gemma model with image URL content. Verify that image URLs are correctly processed in both streaming and non-streaming modes. ```sh go test ./core/providers/vertex/... ``` ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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
## Summary
Consolidates the two separate `UnmarshalJSON` implementations on `ResponsesMessage` into a single method that handles both the verbatim `tool_search` preservation and the `arguments` normalization logic. Previously, the file contained a duplicate `UnmarshalJSON` definition — the first handled `tool_search` items and fell back to a plain `sonic.Unmarshal`, while the second (the correct one) handled argument normalization. The duplicate caused the normalization path to be unreachable for non-`tool_search` items, meaning `tool_search_call` items with object-typed `arguments` would silently fail mid-stream and hang streaming clients.
## Changes
- Removed the redundant first `UnmarshalJSON` that short-circuited to `sonic.Unmarshal` without normalizing `arguments`, leaving only the correct implementation that handles both the `rawToolSearch` early-return and the `arguments` object-to-string normalization.
- Relocated `MarshalJSON` to follow `UnmarshalJSON` for logical grouping.
- The fix ensures `tool_search_call` items whose `arguments` field is a JSON object (e.g. `{}` while in-progress, `{"query":"...","limit":10}` when completed) are correctly stringified into the `*string` field expected by `ResponsesToolMessage`, preventing decode failures that previously dropped items silently.
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (React)
- [ ] Docs
## How to test
```sh
go test ./core/schemas/...
```
Validate by sending a request that triggers `tool_search_call` streaming events and confirming that items with both `{}` (in-progress) and `{"query":"...","limit":10}` (completed) `arguments` values are decoded without error and do not hang the streaming client.
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
## Security considerations
None.
## 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
## Summary Adds a `server` (client ID) filter to the MCP clients list endpoint and UI, allowing users to filter the MCP clients table to a specific server. The filter state is persisted in the URL via query parameters, enabling shareable and bookmarkable filtered views. ## Changes - Added `ClientID` field to `MCPClientsQueryParams` and applied it as a `WHERE client_id = ?` clause in `GetMCPClientsPaginated` - Exposed the filter via a new `server` query parameter on `GET /api/mcp/clients` - Migrated the MCP registry page from local `useState` to `nuqs` `useQueryStates`, storing `search`, `server`, and `offset` in the URL - Added `server` prop and `onServerFilterClear` callback to `MCPClientsTable`, rendering a dismissible "Server filter" badge/button when the filter is active - Extended `GetMCPClientsParams` type and the RTK Query API call to pass the `server` parameter through to the backend - `hasActiveFilters` now accounts for both `debouncedSearch` and `server`, preventing the empty state from showing while a server filter is active ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test 1. Navigate to the MCP Registry page. 2. Confirm that `search`, `server`, and `offset` appear in the URL and survive a page refresh. 3. Set a `server` query param (e.g. `?server=<client_id>`) directly in the URL and verify the table filters to only that client. 4. Click the "Server filter" dismiss button and confirm the filter clears and the URL updates. 5. Verify that the empty state is not shown when a server filter is active but returns no results. ```sh # Core/Transports go test ./framework/configstore/... ./transports/bifrost-http/... # UI cd ui pnpm i pnpm build ``` ## Screenshots/Recordings _Add before/after screenshots showing the server filter badge and URL state._ ## Breaking changes - [ ] Yes - [x] No ## Related issues _Link related issues here._ ## Security considerations The `client_id` filter is applied as a parameterised query (`WHERE client_id = ?`), so there is no SQL injection risk. No secrets or PII are exposed through the new filter parameter. ## 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
…mhq#4770) ## Summary Error messages returned on JSON decode failures were leaking internal decoder details (e.g., field names, Go type information, and `cannot unmarshal` messages) back to API callers. This replaces all such messages with a single, generic `"Invalid request payload"` string to avoid exposing implementation internals. ## Changes - Replaced all `fmt.Sprintf("invalid/Invalid request format: %v", err)` and similar patterns across handlers (`config`, `featureflags`, `governance`, `inference`, `mcp`, `mcp_per_user_headers`, `mcpinference`, `provider_keys`, `providers`, `session`) with the static string `"Invalid request payload"`. - Removed now-unused `fmt` import from `mcpinference.go`. - Added `requestpayload_test.go` with two tests that assert the generic message is returned and that decoder internals (`cannot unmarshal`, field names, Go struct details) are not present in the response body or error string. - Updated the existing `governance_test.go` assertion for the unknown-field case to expect `"Invalid request payload"` instead of `"unknown field"`. ## Type of change - [ ] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./transports/bifrost-http/handlers/... ``` Confirm that: - `TestSessionLoginInvalidPayloadDoesNotExposeDecoderDetails` passes and the response body contains `"Invalid request payload"` with no decoder internals. - `TestPrepareRequestInvalidPayloadDoesNotExposeDecoderDetails` passes and the returned error is exactly `"invalid request payload"`. - `TestComplexityAnalyzerConfigPutRejectsInvalidPayloads` passes with the updated `"Invalid request payload"` expectation for the unknown-field case. ## Breaking changes - [ ] Yes - [x] No ## Security considerations Decoder error messages from Go's `encoding/json` and `sonic` can expose internal struct field names, type information, and value details. Returning these verbatim in HTTP responses constitutes an information disclosure risk. This change ensures all parse-failure responses return a fixed, opaque message regardless of the underlying decode error. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable
Anthropic's server-side tool_search emits server_tool_use(tool_search) -> tool_search_tool_result(tool_references) -> tool_use(discovered tool) in one streamed response. The responses stream converter had no handler for these blocks, so it dropped the tool_references and emitted orphan function_call argument deltas (args with no parent item) that desync the client parser. Mirror the existing web_search_tool_result path: recognize the tool_search server_tool_use + tool_search_tool_result blocks, emit a tool_search_call item (added on the query block, completed on the result block carrying the discovered tool references), suppress the query arg deltas, and skip the spurious generic done. The follow-up tool_use is forwarded unchanged by the generic path. Adds ResponsesMessageTypeToolSearchCall + ResponsesToolSearchCall to the schema. Only touches tool_search branches; web_search/web_fetch/advisor logic unchanged. Refs maximhq#4780 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Charlie Chen <charliec@zetachain.com>
|
|
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (2)
🚧 Files skipped from review as they are similar to previous changes (2)
📝 WalkthroughSummary by CodeRabbit
WalkthroughAdds Anthropic ChangesAnthropic tool_search streaming
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant Anthropic
participant ToBifrostResponsesStream
participant OutputItems
Anthropic->>ToBifrostResponsesStream: Send tool_search server_tool_use
ToBifrostResponsesStream-->>ToBifrostResponsesStream: Track tool-search state
Anthropic->>ToBifrostResponsesStream: Send tool_search_tool_result
ToBifrostResponsesStream->>OutputItems: Persist tool_search_call with tool_references
ToBifrostResponsesStream-->>Anthropic: Emit completed tool_search_call
Possibly related PRs
Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
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" Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Greptile P1: after setToolArguments initializes ResponsesToolMessage for a JSON-decoded tool_search_call, CallID/ID can both be nil, so the reverse path would emit server_tool_use / tool_search_tool_result blocks with a nil id that Anthropic rejects. Guard toolUseID == nil and skip, matching the existing nil-ResponsesToolMessage contract. Covered by a new test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
2 tasks
# Conflicts: # core/schemas/responses.go
Contributor
Author
|
Hey @akshaydeo @TejasGhatte — bumping this one 🙏 Rebased on latest |
Signed-off-by: Akshay Deo <akshay@akshaydeo.com>
akshaydeo
approved these changes
Jul 12, 2026
| if t := gjson.GetBytes(data, "type").String(); isRawPreservedItem(t) { | ||
| mt := ResponsesMessageType(t) | ||
| m.Type = &mt | ||
| m.rawToolSearch = append([]byte(nil), data...) |
Contributor
There was a problem hiding this comment.
m.rawToolSearch references a field that does not exist on *ResponsesMessage. The struct only declares rawPreserved []byte (line 1101); there is no rawToolSearch []byte declaration anywhere in the struct or its embedded types. A go build ./core/schemas/ against this commit will fail with m.rawToolSearch undefined. This line appears to be a leftover from an earlier design where tool_search raw bytes were meant to be stored separately from rawPreserved; the field declaration was never added. Since rawPreserved is already set two lines later, this line can simply be removed.
This was referenced Jul 13, 2026
akshaydeo
added a commit
that referenced
this pull request
Jul 14, 2026
… on the Responses path (#4786) * fix domains * docs update * fix: use `parseAsSafeArrayOf` for logs page array query params to handle special characters in URLs (#4714) ## Summary Array-type query parameters in the logs page (models, providers, etc.) were not using the safe URI-encoding parser, meaning values containing characters like `://` (e.g. model names such as `gpt://host/model`) could be misinterpreted as path or query delimiters by TanStack Router. This introduces `parseAsSafeArrayOf` and applies it consistently across all array filters. ## Changes - Added `parseAsSafeArrayOf` to `queryParamsParser.ts` by composing `parseAsArrayOf` with the existing `parseAsSafeString` parser, ensuring full URI-encoding for comma-separated filter values. - Replaced all usages of `parseAsArrayOf(parseAsString)` in the logs page with `parseAsSafeArrayOf` so that array filters (models, providers, aliases, status, etc.) benefit from the same encoding guarantees as string filters. - Added unit tests for both `parseAsSafeString` and `parseAsSafeArrayOf` to verify round-trip correctness with model names containing `://`. ## Type of change - [x] Bug fix ## Affected areas - [x] UI (React) ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` To manually verify, navigate to the logs page and apply a filter using a model name containing `://` (e.g. `gpt://host/model`). Confirm the URL encodes correctly and the filter persists on page reload without routing errors. ## Screenshots/Recordings N/A ## Breaking changes - [x] No ## Related issues https://github.com/maximhq/bifrost/issues/4603 ## Security considerations No security implications. This change only affects URL query parameter encoding in the UI. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * feat: add TruncatedLabel component with tooltip on overflow (#4715) ## Summary Adds a reusable `TruncatedLabel` component that displays truncated text with an automatic tooltip when the content overflows its container. ## Changes - Introduces `TruncatedLabel`, a `<span>`-based component that detects when its text content is truncated via CSS overflow and conditionally renders a `Tooltip` to show the full content - Truncation detection is performed by comparing `scrollWidth` to `clientWidth`, and re-evaluated on window resize or when `children` changes - The tooltip content defaults to the `children` value if it is a string, but accepts an explicit `tooltip` prop for custom content - The tooltip is only rendered when the text is actually truncated, avoiding unnecessary DOM overhead when content fits ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm build || npm run build ``` Render a `TruncatedLabel` inside a constrained-width container with a long string. Verify that hovering over the truncated text shows a tooltip with the full content, and that no tooltip appears when the text is not truncated. ## Screenshots/Recordings N/A ## Breaking changes - [ ] Yes - [x] No ## Related issues N/A ## Security considerations No security implications. The component renders user-provided React nodes, which is consistent with existing UI patterns. ## 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 * refactor: extract `TruncatedLabel` component and replace inline truncation logic (#4716) ## Summary Extracts the truncated label logic (truncate text with a tooltip on overflow) into a shared `TruncatedLabel` component and replaces all inline implementations with it. ## Changes - Added a reusable `TruncatedLabel` component in `ui/components/ui/truncatedLabel` that handles text truncation and conditionally renders a tooltip when the content overflows - Removed the local `TruncatedName` component from the providers page, which duplicated this logic using `useRef`, `useState`, and a resize event listener - Replaced inline `<span className="truncate ...">` elements in the logs and MCP filter sidebars with `TruncatedLabel` ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` Verify that truncated labels in the providers list and filter sidebars still show a tooltip on hover when the text overflows, and no tooltip when it does not. ## Screenshots/Recordings Verify the providers sidebar and log/MCP filter sidebars visually behave the same as before — truncated text shows a tooltip, non-truncated text does not. ## Breaking changes - [ ] Yes - [x] No ## Related issues https://github.com/maximhq/bifrost/issues/4604 ## Security considerations None. ## 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 * fix: update audit logs page layout classes and add newline at EOF (#4719) ## Summary Fixes the layout of the Audit Logs page to correctly fill the viewport and apply the appropriate background and border styles. ## Changes - Replaced `h-[calc(100dvh-1rem)]` with `h-[calc(100vh-16px)]` for consistent viewport height calculation - Swapped `mx-auto flex flex-col p-4` utility classes for `no-border-parent bg-background flex` to align with the layout conventions used elsewhere in the app - Added missing newline at end of file ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test Navigate to the Audit Logs page and verify: - The page fills the full viewport height without overflow or clipping - The background color and border styling match the rest of the workspace layout ```sh cd ui pnpm i || npm i pnpm build || npm run build ``` ## Screenshots/Recordings Add before/after screenshots showing the corrected Audit Logs page layout. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * feat: add collapsible tag limit to TagInput (#4730) ## Summary Adds collapsible tag support to the `TagInput` component and applies it to the keyword lists on the Complexity Router page. When a keyword list exceeds a configurable limit, tags beyond that limit are hidden behind a gradient overlay with a "Show more" toggle, keeping the UI compact while still allowing full access to all tags. ## Changes - Added `collapsedTagLimit` and `expandButtonTestId` props to `TagInput`. When `collapsedTagLimit` is provided, the component renders in a collapsible layout: tags beyond the limit are hidden with a fade gradient, and "Show more" / "Show less" buttons toggle the expanded state. The collapsed state auto-resets when the tag count drops back to or below the limit. - Set `KEYWORD_COLLAPSED_LIMIT = 8` on the Complexity Router page and passed it along with a `expandButtonTestId` to each keyword `TagInput`. - Standardized border radius tokens from `rounded-lg`/`rounded-md`/`rounded-full` to `rounded-sm` across the Complexity Router page for visual consistency. - Reformatted `index.html` inline shell skeleton from a single minified line to readable, indented HTML and CSS. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` 1. Navigate to the Complexity Router page. 2. Add more than 8 keywords to any keyword list. 3. Verify that tags beyond 8 are hidden with a gradient overlay and a "Show more" button appears. 4. Click "Show more" and confirm all tags are visible with a "Show less" button. 5. Click "Show less" and confirm the list collapses again. 6. Remove tags until 8 or fewer remain and confirm the list stays expanded without the toggle controls. ## Screenshots/Recordings Before/after screenshots of the keyword lists with collapse behavior recommended. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * fix: rebuild token_usage from denormalized columns in hybrid log list (#4722) * fix: rebuild token_usage from denormalized columns in hybrid log list * refactor: inline hybrid token usage reconstruction * fix: preserve malformed serialized token usage state --------- Co-authored-by: gexiangdong <xiangdong.ge@pandasofcaribbean.com> * refactor: centralize Anthropic request building into `BuildAnthropicChatRequestBody` and `AnthropicProviderRequestDefaultsMap` (#3309) ## Summary This PR consolidates Anthropic-family request building across all providers (Anthropic native, Azure, Vertex, Bedrock) into two shared builder functions — `BuildAnthropicChatRequestBody` and `BuildAnthropicResponsesRequestBody` — eliminating duplicated inline logic and provider-specific wrapper helpers that previously scattered the same field-stripping, beta-header injection, and model-field manipulation across multiple files. ## Changes - Introduced `AnthropicProviderRequestDefaults` and `AnthropicProviderRequestDefaultsMap` to encode static, per-provider request-shaping flags (e.g. `DeleteModelField`, `DeleteStreamField`, `AddAnthropicVersion`, `InjectBetaHeadersIntoBody`) in one place. Callers no longer pass these flags directly; the builder looks them up by `cfg.Provider`. - Renamed `Deployment` to `Model` in `AnthropicRequestBuildConfig` for clarity, since all providers now use the same field for model/deployment overrides. - Added `BuildAnthropicChatRequestBody` as the chat-completion analogue of `BuildAnthropicResponsesRequestBody`, covering both raw-body and typed paths, including field stripping, beta-header injection, streaming flag handling, and `fallbacks` deletion. - Bedrock now routes Anthropic models through the Anthropic Messages API format (`invoke` / `invoke-with-response-stream` endpoints) for both chat and responses, rather than the Bedrock Converse API. This includes proper response parsing via `AcquireAnthropicMessageResponse` and streaming via `AnthropicStreamState` / `AnthropicResponsesStreamState`. - Removed private wrapper functions `getRequestBodyForResponses` (Anthropic), `getRequestBodyForAnthropicResponses` (Azure, Vertex), and the inline `CheckContextAndGetRequestBody` closures for Anthropic models in Vertex and Azure, replacing all call sites with direct `BuildAnthropicChatRequestBody` / `BuildAnthropicResponsesRequestBody` calls. - Exported `AcquireAnthropicResponsesStreamState`, `ReleaseAnthropicResponsesStreamState`, `AcquireAnthropicMessageResponse`, and `ReleaseAnthropicMessageResponse` so Bedrock can reuse the Anthropic stream state pool. - Removed `DefaultVertexAnthropicVersion` constant from the Vertex package; the canonical version string now lives in `AnthropicProviderRequestDefaultsMap`. - Bedrock's `releaseBedrockChatResponse` now zeroes the struct before returning it to the pool. - `stripUnsupportedAnthropicFields` is now called inside `BuildAnthropicResponsesRequestBody` on the typed path, making field stripping symmetric across raw and typed paths and across both APIs. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./core/providers/anthropic/... go test ./core/providers/azure/... go test ./core/providers/bedrock/... go test ./core/providers/vertex/... go test ./... ``` Run integration tests against Anthropic, Azure (Anthropic models), Vertex (Claude models), and Bedrock (Claude models) for chat completion, streaming, responses, responses streaming, and count-tokens endpoints. Verify that raw-body passthrough requests produce the same field stripping and beta-header injection as typed requests. ## Breaking changes - [x] Yes - [ ] No `AnthropicRequestBuildConfig` has a breaking field rename: `Deployment` → `Model`. Any external code constructing this struct directly must update the field name. The static shaping flags (`DeleteModelField`, `DeleteRegionField`, `AddAnthropicVersion`, `AnthropicVersion`, `StripCacheControlScope`, `RemapToolVersions`, `InjectBetaHeadersIntoBody`) have been removed from `AnthropicRequestBuildConfig` and are now looked up internally via `AnthropicProviderRequestDefaultsMap`; callers that set these fields must remove them. ## Related issues ## Security considerations No new auth flows, secrets handling, or PII exposure introduced. Field stripping ensures provider-unsupported fields are not forwarded to external APIs. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * refactor: extract `HandleAnthropicChatCompletionRequest` / `HandleAnthropicResponsesRequest` and make `completeRequest` a package-level func shared by Anthropic, Azure, and Bedrock providers (#4394) ## Summary The Anthropic provider's unary request logic was duplicated across the Anthropic, Azure, Bedrock, and Vertex providers. This PR extracts the core non-streaming request execution into a package-level `completeRequest` function and introduces two exported handler functions — `HandleAnthropicChatCompletionRequest` and `HandleAnthropicResponsesRequest` — that encapsulate the full build → send → parse pipeline for chat completions and the Responses API respectively. Azure and Bedrock now delegate directly to these shared handlers for Anthropic-family models instead of reimplementing request dispatch, response parsing, and raw request/response handling inline. A secondary bug fix is included: the large-response streaming client was being activated for count-tokens requests (which should always be buffered) and skipped for all other requests — the condition was inverted. ## Changes - Extracted `completeRequest` as a package-level function accepting explicit `client`, `headers`, `extraHeaders`, `betaHeaderOverrides`, `providerName`, and `logger` arguments, removing the method receiver dependency so it can be called by other providers. - Added `anthropicRequestHeaders` as a provider method to build the `x-api-key` / `anthropic-version` header map, shared across `TextCompletion`, `ChatCompletion`, `Responses`, and `CountTokens`. - Introduced `HandleAnthropicChatCompletionRequest` and `HandleAnthropicResponsesRequest` as exported functions that perform the full unary request lifecycle (body build, HTTP send, large-response detection, response parse, raw request/response attachment). These are now called by the Anthropic, Azure, and Bedrock providers. - Removed `completeMantleRequest` from Bedrock — its logic is now covered by `completeRequest` inside the shared handlers. - Azure's `ChatCompletion` and `Responses` methods now branch early for Anthropic-family models, calling the shared handlers with Azure-specific auth headers, and fall through to the OpenAI-compatible path otherwise, eliminating the post-response model-family branch. - Fixed the inverted condition in `completeRequest` that caused the large-response streaming client to be used for count-tokens requests instead of being skipped for them. - `AnthropicRequestBuildConfig` now carries `BetaHeaderOverrides` so callers do not need to pass it separately. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` Validate that chat completions and Responses API requests succeed for Anthropic-family models routed through the Azure and Bedrock providers, and that count-tokens requests return buffered responses without triggering large-response mode. ## Breaking changes - [ ] Yes - [x] No ## Security considerations Auth headers (`x-api-key`, Bearer tokens, SigV4-signed headers) are applied last in `completeRequest`, after network-config extra headers, ensuring they cannot be overridden by user-supplied configuration. No new secrets or PII handling paths are introduced. ## 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 * refactor: introduce `BearerAuthHeader` helper and migrate OpenAI-compatible providers from `schemas.Key` to `map[string]string` auth header param (#4425) ## Summary This PR standardizes how Bearer token authentication headers are constructed across all OpenAI-compatible providers. Previously, each call site independently built the `Authorization: Bearer <token>` header map with duplicated inline logic. A new `BearerAuthHeader(key)` helper is introduced in the OpenAI package and used uniformly everywhere. Additionally, the Azure provider's private `completeRequest` method is removed. Its non-Anthropic request paths (text completion, chat completion, responses, embedding, compaction) are now delegated directly to the shared `Handle*` functions in the OpenAI package, consistent with how other providers already work. The `Handle*` functions themselves are updated to accept a pre-built `authHeader map[string]string` instead of a raw `schemas.Key`, making them provider-agnostic and compatible with non-Bearer auth schemes (e.g., Azure API key headers, SigV4). ## Changes - Added `BearerAuthHeader(key schemas.Key) map[string]string` to the OpenAI provider package, which returns an `Authorization: Bearer <token>` header map, or an empty map when the key carries no value. - Updated all `Handle*Request` and `handleOpenAILargePayloadPassthrough` function signatures to accept `authHeader map[string]string` instead of `schemas.Key`, applying the map directly to request headers. - Replaced all inline `var authHeader map[string]string` + conditional assignment blocks across Cerebras, Fireworks, Groq, HuggingFace, Mistral, Nebius, Ollama, Opencode, OpenRouter, Parasail, Perplexity, SGL, VLLM, xAI, and Bedrock with calls to `openai.BearerAuthHeader(key)`. - Removed the Azure provider's `completeRequest` method and replaced its usage in `TextCompletion`, `ChatCompletion`, `Responses`, `Embedding`, and `Compaction` with direct calls to the corresponding shared OpenAI `Handle*` functions, passing Azure-specific auth headers and pre-resolved endpoint URLs. ## Type of change - [ ] Bug fix - [ ] Feature - [x] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` Verify that all OpenAI-compatible providers (OpenAI, Azure, Cerebras, Fireworks, Groq, HuggingFace, Mistral, Nebius, Ollama, Opencode, OpenRouter, Parasail, Perplexity, SGL, VLLM, xAI, Bedrock Mantle) continue to authenticate correctly and that requests succeed for text completion, chat completion, responses, embeddings, and compaction endpoints. ## Breaking changes - [ ] Yes - [x] No ## Security considerations The `BearerAuthHeader` helper preserves the existing behavior of omitting the `Authorization` header when the key value is empty, which is intentional for providers that supply auth via other mechanisms (e.g., extra headers or SigV4 signing). No secrets are logged or exposed. ## 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 * refactor: replace pre-built SigV4 body signing with lazy `BodySigner` closure passed through request handlers (#4735) ## Summary Replaces the pre-build-and-sign approach for Bedrock Mantle SigV4 authentication with a `BodySigner` callback that is invoked after the request handler has marshaled the body. This ensures the signature always covers the exact bytes sent on the wire, eliminating the previous double-marshal pattern where the body was built once for signing and again inside the handler. ## Changes - Introduces a new `BodySigner` type (`func(jsonData []byte) (map[string]string, *schemas.BifrostError)`) in `core/providers/utils/bodysigner.go`. Handlers call it after building the request body and apply the returned headers to the outgoing request. - Adds the `signer` parameter to `HandleOpenAIChatCompletionRequest`, `HandleOpenAIChatCompletionStreaming`, `HandleOpenAIResponsesRequest`, `HandleOpenAIResponsesStreaming`, `HandleAnthropicChatCompletionRequest`, `HandleAnthropicChatCompletionStreaming`, `HandleAnthropicResponsesRequest`, and `HandleAnthropicResponsesStream`. All existing callers pass `nil`. - Rewrites Bedrock Mantle's SigV4 paths (`mantleChatCompletions`, `mantleChatCompletionsStream`, `mantleResponses`, `mantleResponsesStream`) to construct a `BodySigner` closure when no API key is present, instead of pre-building the body, signing it, and merging the signature headers into `extraHeaders`. The Bearer path no longer needs a separate early-return branch. - Removes the now-unnecessary `maps` import and the intermediate `extraHeaders` map copies in the Mantle code paths. ## Type of change - [ ] Bug fix - [x] Refactor - [ ] Feature - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./... ``` For Bedrock Mantle with SigV4 (empty key value), verify that requests to chat completions, streaming chat completions, responses, and streaming responses are signed correctly and accepted by the Bedrock endpoint. For Bearer key paths, confirm that no signing is attempted and the `Authorization` header is set as expected. ## Breaking changes - [ ] Yes - [x] No ## Security considerations The `BodySigner` callback signs the exact serialized bytes that are placed on the wire. Previously, the body was serialized twice (once for signing, once inside the handler), which could in theory produce a signature mismatch if marshaling were non-deterministic. This change closes that gap by signing after the final body is set. ## 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 * feat: add `bedrock_mantle` as a first-class provider with native-Anthropic and OpenAI-compatible routing (#4736) ## Summary Introduces `bedrock_mantle` as a first-class, standalone provider that owns the Bedrock Mantle surface (`bedrock-mantle.{region}.api.aws`). Previously, Mantle routing was handled as an internal routing decision inside the existing `bedrock` provider. The new provider gives operators a dedicated configuration surface for Claude (native Anthropic Messages API), OpenAI-compatible models (gpt-*), and Gemma models served through Mantle, without requiring a full Bedrock setup. ## Changes - Added `schemas.BedrockMantle` (`"bedrock_mantle"`) as a new `ModelProvider` constant and registered it in `StandardProviders`, `dynamicallyConfigurableProviders`, `CanProviderKeyValueBeEmpty`, and `isKeySkippingAllowed`. - Added `BedrockMantleKeyConfig` to the `Key` struct, carrying AWS credentials and region for SigV4 auth against the `bedrock-mantle` service. The existing `BedrockKeyConfig` is unchanged. - Introduced the `core/providers/bedrockmantle` package implementing the full `Provider` interface. Chat, streaming chat, Responses, and streaming Responses dispatch by model family: Anthropic-family models use the native Anthropic Messages surface (`/anthropic/v1/messages`); all others use the OpenAI-compatible surface (`/v1` or `/openai/v1`). All other operations return unsupported-operation errors. - Refactored `signAWSRequest` in the `bedrock` package to accept a `*BedrockKeyConfig` instead of individual credential fields, eliminating the now-redundant `signAWSRequestFromKey` wrapper. All call sites updated accordingly. - Exported `SignMantleV4Headers` (previously `mantleSigV4Headers`, a method on `BedrockProvider`) so the new `bedrockmantle` package can sign requests without depending on the internal Bedrock provider struct. The function now supports both `BedrockKeyConfig` and `BedrockMantleKeyConfig` by mapping the latter into a synthetic `BedrockKeyConfig` for signing, and correctly handles GET requests (nil body) for the list-models path. - Extended the Anthropic chat and Responses request builders to convert native structured outputs to tool calls for `BedrockMantle`, matching the existing `Vertex` workaround. - Added `BedrockMantle` to the comprehensive LLM test harness (`ComprehensiveTestAccount`) with key config, provider config, and a full test file covering the supported scenarios (chat, streaming, tool calls, vision, structured outputs, prompt caching, reasoning, list models) and explicitly disabling unsupported ones. - Marked `isMantleModel` in `bedrock/mantle.go` as deprecated in favour of the new provider. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Set AWS credentials and run the new provider test: ```sh export AWS_ACCESS_KEY_ID=... export AWS_SECRET_ACCESS_KEY=... export AWS_SESSION_TOKEN=... # optional, for temporary credentials export AWS_REGION=us-east-1 go test ./core/providers/bedrockmantle/... -v -run TestBedrockMantle ``` To run the full suite (skips Bedrock Mantle automatically when credentials are absent): ```sh go test ./... ``` Configure a `bedrock_mantle` provider by supplying a `BedrockMantleKeyConfig` (or a Bearer API key in `Value`) with the desired region. The region can also be embedded as a prefix in the model ID (e.g. `us-west-2/anthropic.claude-haiku-4-5`) or set at the alias level via `AliasConfig.Region`. ## Breaking changes - [ ] Yes - [x] No The `signAWSRequest` signature change is internal to the `bedrock` package and does not affect any public API. The `isMantleModel` function is deprecated but not removed. ## Security considerations AWS credentials for `BedrockMantleKeyConfig` follow the same `SecretVar` resolution pattern used by `BedrockKeyConfig` (env-var references, never inlined literals). SigV4 signing is performed per-request on the exact body bytes that are sent, so the signature always covers what is transmitted. When a Bearer API key is present it takes precedence and no AWS credentials are required. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * feat: add `bedrock_mantle` provider with SigV4 key config, DB migration, and UI support (#4737) ## Summary Adds `bedrock_mantle` as a first-class provider, enabling Bifrost to route requests to AWS Bedrock through a Mantle proxy endpoint. The provider supports the same SigV4 credential options as the existing Bedrock provider (inherited IAM role, explicit access/secret key, session token, AssumeRole) as well as a Bearer API key authentication mode. ## Changes - Added `BedrockMantle` to the Anthropic passthrough allowlist in `clearAnthropicPassthroughForNonNativeProvider` so raw request bodies are preserved when routing through Bedrock Mantle. - Added `BedrockMantleKeyConfig` redaction logic in `clientconfig.go`, mirroring the existing Bedrock redaction pattern. - Added a new `migrationAddBedrockMantleKeyColumns` database migration that introduces seven `bedrock_mantle_*` SigV4 credential columns to the `config_keys` table. - Extended `TableKey` with the seven Bedrock Mantle credential fields, along with `BeforeSave` serialization and `AfterFind` reconstruction hooks. - Updated `mergeUpdatedKey` in the HTTP handler to correctly restore redacted Bedrock Mantle credential fields during key updates. - Fixed `isClaudeModel` in the Anthropic integration to recognize `bedrock_mantle` (previously incorrectly matched `bedrock`) as a provider that can serve Claude models. - Included `BedrockMantleKeyConfig` in the key hash inputs used by `mergeProviderKeys` and `reconcileProviderKeys` for config file/DB reconciliation. - Added Bedrock Mantle credential redaction to `GetAllKeys`. - Extended `config.schema.json` with `bedrock_mantle_key` and `provider_with_bedrock_mantle_config` definitions and registered `bedrock_mantle` as a valid provider name throughout the schema. - Added UI support: provider icon (reusing the Bedrock SVG mark with a distinct gradient ID), model placeholder text, `isKeyRequiredByProvider` entry, label, form schema (`BedrockMantleKeyConfigSchema`), type definitions (`BedrockMantleKeyConfig`, `DefaultBedrockMantleKeyConfig`), and a full authentication method tab UI (IAM Role / Explicit Credentials / API Key) matching the Bedrock provider UX. - Added `bedrock_mantle` to the Anthropic beta-headers provider family and the provider config sheet's Anthropic family list. - Stripped the internal `_auth_type` field from `bedrock_mantle_key_config` before submitting the form payload. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [x] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh # Core/Transports go version go test ./... # UI cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` Configure a `bedrock_mantle` provider in `config.json` or via the UI with one of the three auth methods: - **IAM Role (Inherited):** set only `region`; leave access/secret key empty. - **Explicit Credentials:** set `access_key`, `secret_key`, and `region`; optionally set `session_token`, `role_arn`, `external_id`, and `session_name`. - **API Key:** set `region` and provide a Bearer token as the key `value`. Send a request targeting a Claude model through the `bedrock_mantle` provider and verify the response is returned correctly and that credentials are redacted in the UI and API responses. ## Screenshots/Recordings _Add before/after screenshots of the new Bedrock Mantle provider form and icon in the UI._ ## Breaking changes - [ ] Yes - [x] No ## Related issues _Link related issues and discussions._ ## Security considerations - All seven Bedrock Mantle credential fields (`access_key`, `secret_key`, `session_token`, `region`, `role_arn`, `external_id`, `role_session_name`) are stored as `SecretVar` and are redacted in API responses and the UI, consistent with the existing Bedrock provider handling. - The `_auth_type` discriminator field is stripped from the payload before it is persisted or transmitted. ## 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 * feat: add AWS Bedrock Mantle provider docs (#4738) ## Summary Adds documentation for the AWS Bedrock Mantle provider, a distinct AWS endpoint (`bedrock-mantle.{region}.api.aws`) that exposes Claude models via the native Anthropic Messages API and OpenAI-family/Gemma models via an OpenAI-compatible API — all addressable through a single `bedrock_mantle/<model>` prefix in Bifrost. ## Changes - Added a new `bedrock-mantle.mdx` provider page covering model ID formats, supported operations, all three authentication modes (SigV4 with explicit credentials, IAM role/inherited credentials, and Bearer API key), IAM role assumption via `role_arn`, and usage examples. - Added the Bedrock Mantle configuration block to the `providers.mdx` config reference, with tabs for Static Credentials, IAM Role, and API Key (Bearer) auth modes. - Added Bedrock Mantle to the provider capability matrix in `overview.mdx`. - Registered `bedrock-mantle` in `docs.json` so it appears in the sidebar navigation. ## Type of change - [ ] Bug fix - [ ] Feature - [ ] Refactor - [x] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [x] Docs ## How to test Navigate to the Bedrock Mantle provider page and config reference in the rendered docs and verify: - The sidebar entry for `bedrock-mantle` appears between `bedrock` and `cerebras`. - All three auth tabs (Static Credentials, IAM Role, API Key) render correctly in both the provider page and the config reference. - The capability matrix row for `bedrock_mantle/<model>` is present and accurate. - Cross-links between the provider page and the config reference resolve correctly. ## Breaking changes - [x] No ## Security considerations Authentication credentials (`access_key`, `secret_key`, `session_token`, API keys) are documented using the `env.*` indirection pattern, consistent with how other providers handle secrets. No credentials are hardcoded in examples. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [ ] I added/updated tests where appropriate - [x] I updated documentation where needed - [ ] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * tests: add `bedrock_mantle` provider capabilities and Postman environment config (#4739) ## Summary Adds E2E test configuration and capability definitions for the `bedrock_mantle` provider, enabling it to be tested through the Bifrost V1 API test suite. ## Changes - Added `bedrock_mantle` to `provider-capabilities.json` with `chat_completions`, `chat_completions_with_tools`, `responses`, `responses_with_tools`, and `list_models` enabled - Added a new Postman environment file (`bifrost-v1-bedrock-mantle.postman_environment.json`) configured to use `anthropic.claude-opus-4-8` as the default model and `us-east-1` as the default region, with secret placeholders for API key, access key, secret key, and session token ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Run the E2E test suite targeting the `bedrock_mantle` provider using the new Postman environment: ```sh # Ensure the Bifrost server is running locally on port 8080 # Load the environment file and run the collection against bedrock_mantle newman run tests/e2e/api/bifrost-v1.postman_collection.json \ -e tests/e2e/api/provider_config/bifrost-v1-bedrock-mantle.postman_environment.json \ --env-var "bedrock_mantle_api_key=<your_api_key>" \ --env-var "bedrock_mantle_access_key=<your_access_key>" \ --env-var "bedrock_mantle_secret_key=<your_secret_key>" ``` Expected outcome: chat completions, tool-use, responses, and model listing tests pass; all unsupported capability tests are skipped or return expected errors. ## Screenshots/Recordings N/A ## Breaking changes - [ ] Yes - [x] No ## Related issues N/A ## Security considerations The Postman environment file stores API key, access key, secret key, and session token as `secret` type fields with empty default values, ensuring credentials are not committed to the repository. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * chunking_strategy as extra params for openai models (#4741) * fix: fix the provider governance form UI to only show calendar aligned toggle when budget is alignable (#4724) ## Summary The calendar alignment toggle in the provider governance form was previously shown whenever any budget existed. This PR restricts its visibility and submission to only when at least one budget uses a calendar-alignable reset period (day, week, month, or year). ## Changes - Introduced a `showCalendarAlignment` derived boolean that checks whether any configured budget has a reset duration supported by `supportsCalendarAlignment`. - Replaced the previous condition (`watchedBudgets.length > 0`) with `showCalendarAlignment` to control rendering of the calendar alignment toggle. - Updated the form submission payload so that `calendar_aligned` is only set to `true` when at least one budget actually supports calendar alignment — preventing the flag from being submitted for incompatible budget configurations. ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test 1. Navigate to a provider's governance settings in the UI. 2. Add a budget with a reset duration that does **not** support calendar alignment (e.g., hourly). Verify the calendar alignment toggle does **not** appear. 3. Add or change a budget to use a calendar-alignable period (e.g., daily, weekly, monthly, yearly). Verify the toggle **does** appear. 4. Enable the toggle and save. Confirm `calendar_aligned: true` is included in the submitted payload. 5. Remove all calendar-alignable budgets and save. Confirm `calendar_aligned` is not set to `true` in the payload. ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` ## Screenshots/Recordings _Before:_ Calendar alignment toggle appears whenever any budget is present, regardless of reset period. _After:_ Calendar alignment toggle only appears when at least one budget uses a day/week/month/year reset period. ## Breaking changes - [x] No ## Related issues ## Security considerations None. ## 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 * fix: custom providers with space in names could not set budget (#4725) ## Summary Fixes a bug where updating or deleting provider-level governance for a custom provider whose name contains a space (e.g. `"OpenRouter Base"`) would return a 404. The UI percent-encodes the provider name in the URL path (`OpenRouter%20Base`), but the handler was comparing the raw encoded string directly against the stored provider name, causing the lookup to fail. Closes #4689 ## Changes - `updateProviderGovernance` and `deleteProviderGovernance` now call `url.PathUnescape` on the `provider_name` path parameter before using it, matching the decoded name against what is stored in the config store. - Returns a `400` if the path parameter contains an invalid percent-encoding sequence. - Added a regression test (`TestProviderGovernance_DecodesEncodedProviderName`) that seeds a provider with a space in its name, issues a PUT and DELETE using the percent-encoded path param, and asserts both succeed and persist correctly. - Added a guard test (`TestProviderGovernance_UnknownProviderStill404`) to confirm that a genuinely unknown provider still returns 404 after the decode change. ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./transports/bifrost-http/handlers/... -run TestProviderGovernance ``` Expected output: all three `TestProviderGovernance_*` tests pass. Specifically: - `TestProviderGovernance_DecodesEncodedProviderName` — PUT and DELETE with `OpenRouter%20Base` return `200`. - `TestProviderGovernance_UnknownProviderStill404` — PUT with an unknown encoded name returns `404`. ## Breaking changes - [ ] Yes - [x] No ## Related issues - Resolves #4689 ## Security considerations `url.PathUnescape` is used rather than `url.QueryUnescape` to correctly handle path-encoded characters. Invalid encoding sequences are rejected with a `400` rather than passed through, preventing malformed input from reaching the config store. ## Checklist - [x] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * fix: pass through gs:// image URLs on Vertex Gemini closes #4402 (#4568) Signed-off-by: Akshay Deo <akshay@akshaydeo.com> Co-authored-by: Akshay Deo <akshay@akshaydeo.com> * Fix mcp reconnect failure on startup (#4316) * Fix mcp reconnect failure on startup * test: assert failed MCP client cleanup --------- Co-authored-by: Gowtham <692171+HackToHell@users.noreply.github.com> * fix(responses): preserve codex tool_search_call/tool_search_output input items (#4121) * fix(responses): preserve codex tool_search_call/tool_search_output input items Bifrost's Responses input deserializer rejected codex's tool-search follow-up request with HTTP 400 "openai responses request input is neither a string nor an array of responses messages", which hung/failed the agent turn. The fix teaches ResponsesMessage about the two tool_search item types and round-trips them verbatim. Background, since tool_search is non-obvious: How codex's tool_search works (the path that hits this bug) ----------------------------------------------------------- codex normally sends every MCP tool inline in the request `tools[]` as `{type:"function", ...}`. But when a model's catalog has `supports_search_tool: true` AND the tool count crosses DIRECT_MCP_TOOL_EXPOSURE_THRESHOLD (= 100) — e.g. an agent wired to several MCP servers — codex stops sending them inline and "defers" them behind a discovery tool: should_defer = supports_search_tool && (ToolSearchAlwaysDeferMcpTools || n >= 100) The deferred flow is a two-request round-trip: 1. Request 1: codex hides the deferred tools and instead declares one tool: {"type":"tool_search","execution":"client","description":"...", "parameters":{query, limit}} `execution:"client"` means the model does NOT run the search — codex does. 2. The model emits a `tool_search_call` with `arguments` = {query, limit}. 3. codex runs the search CLIENT-SIDE: a BM25 index over the deferred tool metadata (codex's ToolSearchHandler, core/src/tools/handlers/tool_search.rs, using the `bm25` crate). It picks the top-N matching tools. 4. Request 2 (follow-up): codex appends two items to `input[]`: - {"type":"tool_search_call", "call_id":..., "execution":"client", "arguments":{...}} - {"type":"tool_search_output", "call_id":..., "status":"completed", "execution":"client", "tools":[ {type:"function", ...the matches} ]} and also surfaces the discovered tools in `tools[]`. The model can now call them. This repeats as the model needs more tools. Root cause ---------- ResponsesMessage (the element type of the Responses `input` array AND the response `Output` array) doesn't model `tool_search_call` / `tool_search_output`: - The call's `arguments` is a JSON OBJECT, whereas function_call's `arguments` is a JSON STRING. So it cannot decode into ResponsesToolMessage.Arguments (*string) -> sonic.Unmarshal of the whole []ResponsesMessage errors -> OpenAIResponsesRequestInput.UnmarshalJSON falls through to the "neither a string nor an array" 400. The entire request dies before reaching OpenAI. - The output's `tools` array is also unmodeled (would be dropped/mangled, which OpenAI then rejects with "Missing input[N].tools[0].type"). OpenAI's Responses API supports both items natively (verified end-to-end against the gateway: the tool_search tool spec is accepted and echoed; OpenAI validates arguments-as-object and tools[].type). So this is purely a Bifrost modelling gap, in the same family as the tool-type allowlist that already lists ResponsesToolTypeToolSearch / ResponsesToolTypeNamespace — just a different code path (input-item deserialization vs the request tools[] allowlist). Fix --- Add ResponsesMessageTypeToolSearchCall / ResponsesMessageTypeToolSearchOutput and give ResponsesMessage custom (Un)MarshalJSON that preserves these two item types verbatim (original bytes in, original bytes out), so the object `arguments` and the `tools` array survive intact. Every other item type defers to the default struct (de)coding, unchanged. One change covers both directions because request input and response Output are both []ResponsesMessage. Impact: unblocks codex tool-search deferral (multi-MCP-server / >=100-tool agents) through Bifrost. Verified with a round-trip test reproducing the exact follow-up payload, plus the existing providers/openai and schemas suites (no regressions). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(responses): reset ResponsesMessage receiver in UnmarshalJSON Clear the receiver at the top of ResponsesMessage.UnmarshalJSON so a reused instance never retains a stale rawToolSearch (or other field) from a prior decode. Without this, unmarshalling a tool_search item and then a normal message into the same value would leave the preserved bytes in place, and MarshalJSON would re-emit them. Not reachable via the array-decode path (each element starts zero), but a cheap, defensive correctness fix. Addresses CodeRabbit review on PR #4121. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(openapi): correct prompt_cache_retention enum to in_memory The chat schema declared the enum as [in-memory, 24h], but OpenAI's actual accepted values are in_memory (underscore) and 24h. The hyphenated form was a typo from when the enum was first added and never matched OpenAI, so spec-generated clients produced Literal['in-memory', '24h'] and rejected the valid value with a pydantic literal_error. The Go runtime treats prompt_cache_retention as a pass-through *string, so no behavior changes — only the spec enum, the regenerated openapi.json, and the doc comment are corrected. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: Akshay Deo <akshay@akshaydeo.com> Co-authored-by: Suresh Kumar Ponnusamy <suresh@atomicwork.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Akshay Deo <akshay@akshaydeo.com> * fix: build fix in vertex (#4765) ## Summary Fixes a bug in the Vertex provider where the code path for handling Gemini/Gemma model families was duplicated, with the non-streaming branch incorrectly using `ToGeminiChatCompletionRequest` (without image URL scheme support) while the streaming branch used `ToGeminiChatCompletionRequestWithImageURLSchemes`. This consolidates the logic so both paths use the image URL scheme-aware converter. ## Changes - Replaced `ToGeminiChatCompletionRequest` with `ToGeminiChatCompletionRequestWithImageURLSchemes` in the non-streaming Gemini/Gemma branch, making it consistent with the streaming branch - Removed the duplicate non-streaming Gemini/Gemma and OpenAI handler blocks that had been incorrectly separated from the streaming path, consolidating them into a single unified code path ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [x] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test Send a chat completion request to the Vertex provider using a Gemini or Gemma model with image URL content. Verify that image URLs are correctly processed in both streaming and non-streaming modes. ```sh go test ./core/providers/vertex/... ``` ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * fix: responses.go build fix (#4766) ## Summary Consolidates the two separate `UnmarshalJSON` implementations on `ResponsesMessage` into a single method that handles both the verbatim `tool_search` preservation and the `arguments` normalization logic. Previously, the file contained a duplicate `UnmarshalJSON` definition — the first handled `tool_search` items and fell back to a plain `sonic.Unmarshal`, while the second (the correct one) handled argument normalization. The duplicate caused the normalization path to be unreachable for non-`tool_search` items, meaning `tool_search_call` items with object-typed `arguments` would silently fail mid-stream and hang streaming clients. ## Changes - Removed the redundant first `UnmarshalJSON` that short-circuited to `sonic.Unmarshal` without normalizing `arguments`, leaving only the correct implementation that handles both the `rawToolSearch` early-return and the `arguments` object-to-string normalization. - Relocated `MarshalJSON` to follow `UnmarshalJSON` for logical grouping. - The fix ensures `tool_search_call` items whose `arguments` field is a JSON object (e.g. `{}` while in-progress, `{"query":"...","limit":10}` when completed) are correctly stringified into the `*string` field expected by `ResponsesToolMessage`, preventing decode failures that previously dropped items silently. ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./core/schemas/... ``` Validate by sending a request that triggers `tool_search_call` streaming events and confirming that items with both `{}` (in-progress) and `{"query":"...","limit":10}` (completed) `arguments` values are decoded without error and do not hang the streaming client. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * feat: add server/client_id filter to MCP clients list (#4767) ## Summary Adds a `server` (client ID) filter to the MCP clients list endpoint and UI, allowing users to filter the MCP clients table to a specific server. The filter state is persisted in the URL via query parameters, enabling shareable and bookmarkable filtered views. ## Changes - Added `ClientID` field to `MCPClientsQueryParams` and applied it as a `WHERE client_id = ?` clause in `GetMCPClientsPaginated` - Exposed the filter via a new `server` query parameter on `GET /api/mcp/clients` - Migrated the MCP registry page from local `useState` to `nuqs` `useQueryStates`, storing `search`, `server`, and `offset` in the URL - Added `server` prop and `onServerFilterClear` callback to `MCPClientsTable`, rendering a dismissible "Server filter" badge/button when the filter is active - Extended `GetMCPClientsParams` type and the RTK Query API call to pass the `server` parameter through to the backend - `hasActiveFilters` now accounts for both `debouncedSearch` and `server`, preventing the empty state from showing while a server filter is active ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [x] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test 1. Navigate to the MCP Registry page. 2. Confirm that `search`, `server`, and `offset` appear in the URL and survive a page refresh. 3. Set a `server` query param (e.g. `?server=<client_id>`) directly in the URL and verify the table filters to only that client. 4. Click the "Server filter" dismiss button and confirm the filter clears and the URL updates. 5. Verify that the empty state is not shown when a server filter is active but returns no results. ```sh # Core/Transports go test ./framework/configstore/... ./transports/bifrost-http/... # UI cd ui pnpm i pnpm build ``` ## Screenshots/Recordings _Add before/after screenshots showing the server filter badge and URL state._ ## Breaking changes - [ ] Yes - [x] No ## Related issues _Link related issues here._ ## Security considerations The `client_id` filter is applied as a parameterised query (`WHERE client_id = ?`), so there is no SQL injection risk. No secrets or PII are exposed through the new filter parameter. ## 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 * fix: redact decoder details from invalid request payload errors (#4770) ## Summary Error messages returned on JSON decode failures were leaking internal decoder details (e.g., field names, Go type information, and `cannot unmarshal` messages) back to API callers. This replaces all such messages with a single, generic `"Invalid request payload"` string to avoid exposing implementation internals. ## Changes - Replaced all `fmt.Sprintf("invalid/Invalid request format: %v", err)` and similar patterns across handlers (`config`, `featureflags`, `governance`, `inference`, `mcp`, `mcp_per_user_headers`, `mcpinference`, `provider_keys`, `providers`, `session`) with the static string `"Invalid request payload"`. - Removed now-unused `fmt` import from `mcpinference.go`. - Added `requestpayload_test.go` with two tests that assert the generic message is returned and that decoder internals (`cannot unmarshal`, field names, Go struct details) are not present in the response body or error string. - Updated the existing `governance_test.go` assertion for the unknown-field case to expect `"Invalid request payload"` instead of `"unknown field"`. ## Type of change - [ ] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [x] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [ ] UI (React) - [ ] Docs ## How to test ```sh go test ./transports/bifrost-http/handlers/... ``` Confirm that: - `TestSessionLoginInvalidPayloadDoesNotExposeDecoderDetails` passes and the response body contains `"Invalid request payload"` with no decoder internals. - `TestPrepareRequestInvalidPayloadDoesNotExposeDecoderDetails` passes and the returned error is exactly `"invalid request payload"`. - `TestComplexityAnalyzerConfigPutRejectsInvalidPayloads` passes with the updated `"Invalid request payload"` expectation for the unknown-field case. ## Breaking changes - [ ] Yes - [x] No ## Security considerations Decoder error messages from Go's `encoding/json` and `sonic` can expose internal struct field names, type information, and value details. Returning these verbatim in HTTP responses constitutes an information disclosure risk. This change ensures all parse-failure responses return a fixed, opaque message regardless of the underlying decode error. ## Checklist - [ ] I read `docs/contributing/README.md` and followed the guidelines - [x] I added/updated tests where appropriate - [ ] I updated documentation where needed - [x] I verified builds succeed (Go and UI) - [ ] I verified the CI pipeline passes locally if applicable * feat(anthropic): forward server-side tool_search streaming results Anthropic's server-side tool_search emits server_tool_use(tool_search) -> tool_search_tool_result(tool_references) -> tool_use(discovered tool) in one streamed response. The responses stream converter had no handler for these blocks, so it dropped the tool_references and emitted orphan function_call argument deltas (args with no parent item) that desync the client parser. Mirror the existing web_search_tool_result path: recognize the tool_search server_tool_use + tool_search_tool_result blocks, emit a tool_search_call item (added on the query block, completed on the result block carrying the discovered tool references), suppress the query arg deltas, and skip the spurious generic done. The follow-up tool_use is forwarded unchanged by the generic path. Adds ResponsesMessageTypeToolSearchCall + ResponsesToolSearchCall to the schema. Only touches tool_search branches; web_search/web_fetch/advisor logic unchanged. Refs maximhq/bifrost#4780 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Charlie Chen <charliec@zetachain.com> * test(anthropic): cover tool_search bm25 + completed response; set Name on done item Address review feedback on the tool_search streaming patch: - carry the tool name on the tool_search_call output_item.done (parity with the added item and the advisor_call done path) via a new ToolSearchToolName state field - table-drive the tool_references test over both regex and bm25 variants - assert the terminal response.completed Output includes the tool_search_call with its tool_references (covers the OutputItems persistence path) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Charlie Chen <charliec@zetachain.com> * feat(anthropic): rebuild tool_search_call on the Bifrost to Anthropic path The forward streaming path now emits a tool_search_call carrying the discovered tool_references, but the request builder (ConvertBifrostMessagesToAnthropicMessages) had no case for it, so on a follow-up turn the item hit default: continue and was dropped — losing the server-side tool_search context (unlike web_search/advisor, which round-trip). Add convertBifrostToolSearchCallToAnthropicBlocks + a ResponsesMessageTypeToolSearchCall case mirroring the web_search/advisor pattern: rebuild server_tool_use(tool_search_*) followed by its tool_search_tool_result carrying the tool_references, so the pair stays in the assistant message on subsequent turns. Adds a pure-function reverse-path test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Charlie Chen <charliec@zetachain.com> * fix: update audit logs page layout classes and add newline at EOF (#4719) ## Summary Fixes the layout of the Audit Logs page to correctly fill the viewport and apply the appropriate background and border styles. ## Changes - Replaced `h-[calc(100dvh-1rem)]` with `h-[calc(100vh-16px)]` for consistent viewport height calculation - Swapped `mx-auto flex flex-col p-4` utility classes for `no-border-parent bg-background flex` to align with the layout conventions used elsewhere in the app - Added missing newline at end of file ## Type of change - [x] Bug fix - [ ] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test Navigate to the Audit Logs page and verify: - The page fills the full viewport height without overflow or clipping - The background color and border styling match the rest of the workspace layout ```sh cd ui pnpm i || npm i pnpm build || npm run build ``` ## Screenshots/Recordings Add before/after screenshots showing the corrected Audit Logs page layout. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * feat: add collapsible tag limit to TagInput (#4730) ## Summary Adds collapsible tag support to the `TagInput` component and applies it to the keyword lists on the Complexity Router page. When a keyword list exceeds a configurable limit, tags beyond that limit are hidden behind a gradient overlay with a "Show more" toggle, keeping the UI compact while still allowing full access to all tags. ## Changes - Added `collapsedTagLimit` and `expandButtonTestId` props to `TagInput`. When `collapsedTagLimit` is provided, the component renders in a collapsible layout: tags beyond the limit are hidden with a fade gradient, and "Show more" / "Show less" buttons toggle the expanded state. The collapsed state auto-resets when the tag count drops back to or below the limit. - Set `KEYWORD_COLLAPSED_LIMIT = 8` on the Complexity Router page and passed it along with a `expandButtonTestId` to each keyword `TagInput`. - Standardized border radius tokens from `rounded-lg`/`rounded-md`/`rounded-full` to `rounded-sm` across the Complexity Router page for visual consistency. - Reformatted `index.html` inline shell skeleton from a single minified line to readable, indented HTML and CSS. ## Type of change - [ ] Bug fix - [x] Feature - [ ] Refactor - [ ] Documentation - [ ] Chore/CI ## Affected areas - [ ] Core (Go) - [ ] Transports (HTTP) - [ ] Providers/Integrations - [ ] Plugins - [x] UI (React) - [ ] Docs ## How to test ```sh cd ui pnpm i || npm i pnpm test || npm test pnpm build || npm run build ``` 1. Navigate to the Complexity Router page. 2. Add more than 8 keywords to any keyword list. 3. Verify that tags beyond 8 are hidden with a gradient overlay and a "Show more" button appears. 4. Click "Show more" and confirm all tags are visible with a "Show less" button. 5. Click "Show less" and confirm the list collapses again. 6. Remove tags until 8 or fewer remain and confirm the list stays expanded without the toggle controls. ## Screenshots/Recordings Before/after screenshots of the keyword lists with collapse behavior recommended. ## Breaking changes - [ ] Yes - [x] No ## Related issues ## Security considerations None. ## 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 * fix: rebuild token_usage from denormalized columns in hybrid log list (#4722) * fix: rebuild token_usage from denormalized columns in hybrid log list * refactor: inline hybrid token usage reconstruction * fix: preserve malformed serialized token usage state --------- Co-authored-by: gexiangdong <xiangdong.ge@pandasofcaribbean.com> * refactor: centralize Anthropic request building into `BuildAnthropicChatRequestBody` and `AnthropicProviderRequestDefaultsMap` (#3309) ## Summary This PR consolidates Anthropic-family request building across all providers (Anthropic native, Azure, Vertex, Bedrock) into two…
This was referenced Jul 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
13 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Saw #4780 is assigned to @TejasGhatte — sharing a working patch with tests in case it's useful; happy to adjust or hand off.
Summary
The Anthropic provider advertises
tool_searchas supported (ToolSearch: truefor Claude models), but server-side tool_search isn't actually handled on the/v1/responsespath — in either direction:tool_search_callin history isn't rebuilt into Anthropic blocks, so it's dropped on the way back to Anthropic.Only happens through Bifrost — going straight to Anthropic works. Filed as #4780.
Changes
core/providers/anthropic/responses.go:ToBifrostResponsesStream(response/streaming): addtool_searchhandling alongside the existingweb_search/web_fetch/code_execution/advisor server-tool paths — recognizeserver_tool_use(tool_search_*)andtool_search_tool_resultat content_block_start/stop, and on result block-stop emit atool_search_calloutput_item.donecarrying the discovered tool names (tool_references). The discoveredtool_useis forwarded by the existing generic path.ConvertBifrostMessagesToAnthropicMessages(request): add aResponsesMessageTypeToolSearchCallcase +convertBifrostToolSearchCallToAnthropicBlocks, rebuilding theserver_tool_use(tool_search)+tool_search_tool_resultblocks from history — mirrors the existingweb_search/advisor request handling.core/schemas/responses.go: add aToolReferencescarrier (ResponsesToolSearchCall) onResponsesToolMessage(mirrors the existing advisor-call shape). Thetool_search_callitem type already exists.web_search/web_fetch/advisor/code_executionpaths are untouched.AnthropicContentBlockTypeToolSearchToolResult) and theToolReferencesfield already existed intypes.gobut were never read/written — tool_search was never wired up on either direction.Type of change
Affected areas
How to test
core/providers/anthropic/toolsearch_test.gocovers both directions:server_tool_use(tool_search_tool_regex)→tool_search_tool_result(tool_references=[...])→tool_use(<found tool>)through the converter. Before this change the same fixture dropstool_referencesand emits an orphanfunction_call_arguments.delta; after, it emits a cleantool_search_call(added→done carrying the discovered tool) followed by the discovered tool call. Both regex and bm25 variants and the completed-response path are covered.tool_search_calland asserts it is rebuilt into the correct Anthropicserver_tool_use+tool_search_tool_resultblocks.Breaking changes
Related issues
Closes #4780
Security considerations
None — response/request shape only; no auth/secrets/PII.
Checklist