core/v1.8.2

## Core Release v1.8.2

- fix: forward OpenCode Responses requests directly to /v1/responses [@mohammadrezwankhan](https://github.com/mohammadrezwankhan)
- feat: add the `VideoEdit` operation with `BifrostVideoEditRequest`, `VideoEditInput` and `VideoEditParameters` for prompt-driven edits, upscaling and background removal on an existing video supplied as bytes, a URL or a provider video ID; implemented for OpenAI (`/v1/videos/edits`) and Runware (`videoInference`, `upscale`, `removeBackground`), with the model optional when the source is a video ID and the prompt optional for asset-driven task types (#6270)
- feat: batch accounting: `MergeBifrostLLMUsage` promoted to `schemas`, `Endpoint` on `BifrostBatchResultsResponse`, `BatchResultItem.Failed()`, `BatchRequestCountsFromResults`, `BatchRequestCounts.IsZero()`, raw-JSON Gemini batch result parsing and `custom_id` validation in `ConvertRequestsToJSONL`; the settlement engine (`AccountBatchResults` with runner-ID ownership fencing, idempotent aggregate log writes and governance reporting) and a sweeper that polls due jobs with capped, jittered backoff; aggregate log entries carry a `bifrost/<version>` user agent via `BifrostContextKeyRuntimeVersion` (thanks [@SahilChoudhary22](https://github.com/SahilChoudhary22)!) (#5291, #5294, #6474)
- feat: Claude-on-Vertex batch support: `ToVertexBatchCreateRequest` resolves Anthropic families to `publishers/anthropic/models/...`, `vertexConvertRequestsToJSONL` emits Claude-on-Vertex instances, `custom_id` round-trips through `batchResultsByKey`, and `GeminiBatchGenerateContentRequest` keeps `tools`, `toolConfig`, `cachedContent`, `labels` and the display name (#5368)
- feat: add `HTTPTransportPreAuthHook` to the `HTTPTransportPlugin` interface, a phase that runs before transport authentication; `HTTPTransportPreHook` now runs after it (#6375)
  <Warning>
  Breaking for plugin authors: Go plugins implementing `HTTPTransportPlugin` must add `HTTPTransportPreAuthHook` (`.so` plugins that predate it are skipped for that phase), and any plugin that injected a credential such as `x-bf-vk` or `Authorization` from `HTTPTransportPreHook` must move that work to `HTTPTransportPreAuthHook`, since the pre-hook no longer runs before auth.
  </Warning>
- feat: add `semaphore_size` and `inject_timeout` to `PluginConfig` so observability `Inject` calls are context-bounded per plugin (#6341)
- feat: Runware provider expansion: chat completions, streaming and Responses through its OpenAI-compatible `/v1/chat/completions` endpoint (Responses muxed via `ToChatRequest()`), `ListModels` sweeping the curated `modelSearch` catalog with the AIR as the model ID, image upscale via `/v1/images/edits` (`type=upscale`) and image-to-3D via `/v1/videos` (`type=3d`), a shared `settings` extra-param coercion for multipart and JSON callers, prompt-optional asset-driven operations, and input handling for edit, upscale and video task shapes (#6260, #6372, #6208)
- feat: OpenAI `ultrafast` service tier: `BifrostServiceTierUltrafast`, capability-gated forwarding via `serviceTierForModel` on chat, Responses and compaction, and `ultrafast` preserved through `WithDefaults` (#6396)
- feat: JSON bodies on `/v1/images/edits`: `ImageInput` accepts a bare string or `{ "url", "image" }`, typed extra params reach providers with their real types, and `images` is a known field (#6418)
- feat: `EmbeddingData.EncodingFormat` with typed `int8`, `uint8`, `binary`, `ubinary` and `base64` vectors; Bedrock Titan V2 `embeddingTypes` and Cohere `embedding_types` on Converse, native invoke and LangChain `BedrockEmbeddings` compatibility (#6381)
- feat: rerank: `RerankDocument.Data` for structured documents, `RerankResult.ID`, `RerankParameters.NextToken`, `ReturnDocuments` forwarded to Cohere and Vertex, `ToCohereError` for Cohere-shaped errors, `/genai/v1/rank` served cross-provider via `x-model-provider`, cross-provider responses converted back to the caller's wire shape with `ToBedrockRerankResponse`, `ToCohereRerankResponse` and `ToVertexRankResponse`, and rerank cost accounting for Bedrock and Cohere (#6301, #6328)
- feat: datasheet-backed compatibility flows: Anthropic, Bedrock, Cohere and Gemini request shaping (adaptive-only thinking, adaptive thinking, native effort, disable-reasoning, mid-conversation system turns, computer-use and text-editor tool generations, default max output tokens, tool validation, thinking-budget zeroing) is resolved through `schemas.ResolveModelCaps` instead of hardcoded model-name checks (#6281, #6492)
- feat: Gemini 3 per-model `thinkingLevel` support table (`geminiThinkingLevelSupport`) with `clampThinkingLevel` snapping requested levels to the nearest rung (ties break upward) and `lowestThinkingLevel` for `reasoning_effort: "none"`, so `setThinkingBudgetZeroIfSupported` sets the floor level on Gemini 3+ instead of zeroing `thinkingBudget` (#6280)
- feat: Bifrost overhead latency accounting: `upstream_latency` and `overhead_latency` on `BifrostResponseExtraFields` (`PopulateOverheadLatency`, `BifrostContextKeyRequestStartTime`, `populateLatencyExtraFields` so logging plugins see both at hook time); per-phase overhead spans across the request pipeline (`queue-wait`, `attribute-population`, `convertor`, `request-marshal`, `response-parse`, `handle-setup`, `pipeline-pre`, `pipeline-post`, `worker-setup`, `key-pool`, Bedrock `request-sign` and `credentials-fetch`, `response-finalize`) with `StampWorkerHandoff` on `ChannelMessage.sentAt`; lock-free stream overhead accumulators for per-chunk parse, conversion and backpressure installed via `ResetStreamOverhead`, `StampStreamTransport` for the outbound marshal and client-write time, and `defaultSSEDataReader.ReadDataLine` attributing socket reads to upstream; and `IsOverheadBreakdownSpan`, `WithoutOverheadBreakdownSpans` and the `OverheadSpanConsumer` interface so breakdown spans stay out of connectors that do not opt in (#5533, #6388, #6389, #6433, #6470, #6495)
- feat: input/output/additional cost split (`BifrostCost`) on inference usages, extended to speech, transcription and OCR usages
- feat: `Notification`, `NotificationInput`, `NotificationSeverity`, `NotificationAudience` and the `NotificationPublisher` function type for the dashboard notification center (#6207)
- feat: `BifrostContextKeySkipModelCheck` short-circuits the virtual key model allowlist for evaluate-only requests such as `/inspect` while keeping every other governance rule (#6479)
- feat: `HarnessSessionHeaders` and `MaxSessionIDLength` so Claude Code, Codex CLI and OpenCode session headers can fall back into the session ID (#6333)
- feat: `RedactSensitiveHeaders`, with `IsSensitiveHeader` extended to Cloudflare Access (`cf-access-*`), AWS ALB OIDC (`x-amzn-oidc-*`) and generic `jwt`/`assertion` headers (#6371)
- feat: `ResponsesResponseError.Type` and a shared Responses stream-error normalizer so terminal `error`/`response.failed` events inside an HTTP 200 Azure SSE stream surface as errors with their nested type, code and message on both create-stream and retrieve-stream paths (thanks [@dani29](https://github.com/dani29)!) (#6302)
- feat: `ServiceTier` on `StreamAccumulatorResult`, with Anthropic's `service_tier` from `message_start` latched onto the final chunk of chat and Responses streams (#6236)
- feat: OpenRouter speech and transcription through the OpenAI-compatible audio handlers instead of returning unsupported-operation errors (#5734)
- fix: preserve `max_tokens` for OpenCode-compatible chat endpoints (thanks [@Alex-wangyang](https://github.com/Alex-wangyang)!) (#6458)
- fix: HuggingFace chat streaming completed with zero tokens and therefore zero cost while non-streaming calls on the same models priced correctly, for two reasons: HuggingFace was listed as a provider that omits the `[DONE]` marker (it sends one), which made the shared OpenAI streaming loop `break` on the first `finish_reason` and discard the trailing usage-only chunk that several router inference providers emit; and `stream_options.include_usage` never reached the router because the shared streaming handler returns early when a provider supplies a custom request converter. Both are corrected, and an explicit `stream_options` from the caller still wins (thanks [@elliottrabac](https://github.com/elliottrabac)!) (#6478)
- fix: preserve the caller's JSON Schema key order for structured outputs - `ChatParameters.UnmarshalJSON` holds `response_format` as raw bytes and the new `ChatResponseFormat` reader splices them verbatim into OpenAI, Anthropic, Bedrock, Gemini (unless a union `type` array needs normalizing) and Cohere requests, and `ResponsesTextConfigFormatJSONSchema` re-encodes in the decoded key sequence, because OpenAI structured outputs generate fields in the declared order and a re-sorted schema silently changes model behavior (#6235)
- fix: open reasoning stream items that carry both an encrypted payload and a visible summary as `thinking` blocks instead of `redacted_thinking` on the Anthropic egress, with `isReasoningItem` and `reasoningPayloadAndSummary` shared by the native-reasoning and misclassified-function-call branches (#6292)
- fix: replayed thinking blocks through the Anthropic ingress with a `bedrock/` model prefix: content-less `tool_result` blocks are kept, interleaved text/tool-use/thinking order is preserved by the grouped converter, `incomplete` maps to `error` on Converse `toolResult.status`, and buffered reasoning is consumed by the item that owns it, so multi-turn tool use no longer wedges (#6346)
- fix: Gemini/Vertex HTTP 400s on Claude Code traffic routed through `/anthropic/v1/messages`: trailing assistant prefills are trimmed on both the Responses and chat paths, mid-conversation `system` messages are inlined in place instead of hoisted into `systemInstruction`, and `AnthropicMessageResponse` gains `ExtraFields` (#6363)
- fix: alias Bedrock `toolUseId`/`toolResultId` values longer than 64 characters or outside `[a-zA-Z0-9_.:-]` (such as Gemini thought-signature IDs) with a deterministic hash prefix, applied identically on `tool_use` and `tool_result` in both the Responses and chat converters (#6300)
- fix: route Grok (`xai.`) models through the `openai/v1` Mantle path on Bedrock and Bedrock Mantle, since they have no Converse equivalent (#6022)
- fix: register Bedrock Mantle in `ProviderSendsDoneMarker` so its streams end after `finish_reason` instead of waiting for a `[DONE]` marker (#6021)
- fix: include OpenRouter embedding models from `/v1/embeddings/models` in `ListModels`, merged case-insensitively and best-effort (#6264)
- fix: force `reasoning.effort` to `"none"` for models that reason by default but do not support reasoning with tool calls when they advertise `supports_none_reasoning_effort`, instead of dropping `reasoning` outright (#6293)
- fix: backfill upscale output resolution on Replicate from the `target`/`factor` params and `metrics.resolution_target` bands so resolution-tiered pricing bills the real output size (#6083)
- fix: filter forwarded `Accept-Encoding` to the codecs `CheckAndDecodeBody` can decode (`gzip`, `x-gzip`, `deflate`, `br`, `zstd`, `identity`), restrict streaming endpoints to `gzip`/`identity` via `SetPassthroughHeadersForStreaming`, and decode chained content encodings in reverse order (#6360)
- fix: `tool_sync_interval` handling: negative values are rejected (the "disable sync" semantic is gone now that the connection checker drives discovery and liveness together), `ResolveToolSyncInterval` follows the global setting for sub-second values, a fresh per-call checker starts on `EnableClient` and on a sticky-to-per-call flip, and `MCPManager.UpdateToolSyncInterval`, `ConnectionCheckerManager.SetGlobalInterval`/`ApplyGlobalInterval`/`RetimeClient` and `ClientConnectionChecker.SetHealthyInterval` hot-reload the global cadence and re-time running checkers in place; `GetMCPConfig` carries the stored global interval (#6502)
- fix: `SetClientTools` and `UpdateClientCredentials` replace the MCP tool map instead of `maps.Copy`-merging into it, so a tool removed upstream is evicted from memory once the database has dropped it (#6484)
- fix: per-call shared-credential MCP clients (`oauth`, `headers`, `none`) refresh tools synchronously on credential update instead of returning `ErrMCPReconnectNotApplicable`; disabled per-call clients and per-user auth types keep the sentinel (#6483)
- fix: park a failed `EnableClient` dial at `Disabled` instead of `Unstable`, add `ErrMCPEnableConnectFailed` so callers do not roll back the persisted `disabled` flag, and guard `isEnableable` on both state and config so the admin can retry (#6431)
- fix: `output_item.done` replaces server-side tool item shells (`web_search_call`, `code_interpreter_call`, `image_generation_call`) in the Responses streaming accumulator so their full payload survives (#6475)
- feat: send `s3://` image and document references to Bedrock Converse as the `s3Location` source member instead of downloading the bytes and re-uploading them - Converse resolves the object itself, which skips a round trip and the 25 MiB inline cap entirely. Image format is derived from the object extension since nothing is fetched and there is no `Content-Type` to read, and an extension-less object is rejected up front rather than producing an opaque 400 (#6239)
- feat: resolve Vertex URL sources per model family rather than inlining everything - a `gs://` URI is now forwarded to Gemini/Gemma as `fileData.fileUri` (the documented form, resolved under the caller's own project IAM, and the only thing that keeps multi-hundred-MB video inputs viable) and read from Cloud Storage with the request key's own Google credentials for Claude-on-Vertex, which accepts base64 sources only. `http(s)` is still always fetched: forwarding one was measured against the harness and Vertex rejected every endpoint shape with `URL_REJECTED-REJECTED_FC_TOO_MANY_PENDING` (#6239)
- fix: always emit a Gemini candidate carrying its finish reason on `generateContent`, even when nothing visible was generated - a thinking model that spends its whole output budget before emitting a token is a successful 200 with an empty answer, but `Candidates` is `omitempty`, so dropping that candidate produced a body with no `candidates` key at all and left a lone `usageMetadata` object that every Gemini-shaped client dereferences blind (#6239)
- fix: drop payload-free Gemini parts when assembling a candidate - every `Part` field is `omitempty`, so such a part marshals to exactly `{}`; the harness observed one on the wire when a transcription request for an unintelligible tone came back as `parts:[{}]`, where it is noise a client will try to read and it masks the contentless case by making the parts slice look non-empty (#6239)
- fix: accept a bare model identifier on Bedrock rerank by synthesizing the foundation-model ARN from the resolved region - Rerank is the one Bedrock surface that names its model by ARN rather than by bare ID, so all three rerank drop-ins in the provider harness 400'd on `amazon.rerank-v1:0`. The partition is derived from the region (`aws`, `aws-cn`, `aws-us-gov`) so GovCloud and China build a correct ARN, and an explicit ARN still passes through untouched (#6239)
- fix: stop stripping `file_url` from OpenAI-shaped chat file blocks on marshal - dropping it produced `{"type":"file","file":{}}` and an upstream complaint about a missing `file_id`, which hid the fact that a source had been discarded. Providers that cannot take a URL now say so by name, and any OpenAI-compatible endpoint that does accept one keeps working without a Bifrost change (#6239)
- fix: leave URL content sources Bifrost cannot download in place on the OpenAI and native-Anthropic paths instead of failing the request - only `http(s)` is fetched, and whether a `gs://`, `s3://` or scheme-less reference is usable is the provider's call, so the source now travels as `{"type":"url"}` and the platform answers for itself (#6239)
- perf: JSON serialization on the hot path: shared MCP tools cache their serialized bytes on `ChatTool` (`EnsureSerialized`, `precomputeToolSerialization`) so catalog tools are marshalled once, and `OrderedMap.MarshalJSON` writes compact JSON directly into a buffer with an inline HTML-safe string escaper instead of re-routing every nested map through `MarshalSorted`, pinned by a byte-identity fuzz harness (#6241, #6242)
- perf: allocation and tracing reductions on the request path: `StartSpanID` and `SpanFromHandle` on the tracer with `Span.SetAttributes` for bulk writes, a resolved-once attribute block in `executeRequestWithRetries`, reusable worker delivery timers, `Span.Reset` keeping map capacity, `reservedKeys` as a set, pre-sized `userValues`, logging context reads deferred to the final chunk, no redundant `fmt.Sprintf` in logger calls, cached plugin span names, compact JSON request bodies, `math/rand/v2` in `GetRandomString`, and a `HasPluginLogs` guard before draining plugin logs (#5657, #5956, #5957, #6211)
- chore: remove the legacy `gen_ai.*`-namespaced Bifrost-internal attribute constants, `AttrPromptTokens`/`AttrCompletionTokens`, `AttrLegacyRetryCount` and the nanosecond `AttrTimeToFirstToken` in favor of the canonical `bifrost.*` keys (#6403)
- chore: build with Go 1.26.6 (#6269)

### Installation

```bash
go get github.com/maximhq/bifrost/core@v1.8.2
```

---
_This release was automatically created from version file: `core/version`_