Skip to main content
​
1.6.9

​
✨ Features

  • Bedrock HTTP/2 PING Keepalives — Opt-in HTTP/2 PING keepalives on the Bedrock provider via the new http2_ping_interval_in_seconds config key (0 = off), keeping quiet streams alive through intermediaries that sever idle connections (thanks @jeremym-tanium!)
  • Expanded OTEL Metric Attributes — Metrics now carry a service instance id plus team, customer, and business-unit ids and names, so exported series can be sliced by tenant without post-processing
  • Adaptive Thinking on Raw Passthrough — Adaptive-only Anthropic models (Opus 4.7+, Opus 4.8, Opus 5, Sonnet 5, Fable 5, Mythos 5) now get legacy thinking.type: "enabled" rewritten to the adaptive form on the raw passthrough body as well as the typed request path

​
🐞 Fixed

  • SSE Heartbeat Mid-Line Corruption — The stream reader now tracks line boundaries under a mutex and refuses to emit a heartbeat mid-line, fixing corrupted data: payloads on raw passthrough streams where a heartbeat could split a JSON line
  • SSE Heartbeat Frame Compatibility — Dropped the trailing blank line from the heartbeat comment frame so non-conforming SSE decoders (e.g. openai-go ssestream before v3.43.0) no longer dispatch an empty event and abort mid-stream with “unexpected end of JSON input”
  • Lost Log Rows on Shared Trace IDs — Concurrent HTTP requests that inherit the same W3C trace id no longer overwrite each other’s pending log entry; the join now uses the per-request internal trace id (thanks @AdityaPainuli!)
  • Budget Counters Reset on Force-Sync — config.json force-sync no longer overwrites live current_usage, last_reset, and the token/request rate-limit counters with the values from the file
  • Transcription Filename Dropped — The client’s multipart filename is now carried through transcription ingress, so non-WAV containers are no longer relabelled audio.mp3 and rejected upstream (thanks @AdityaPainuli!)
  • Anthropic Mid-Conversation System Messages — A role:"system" turn that cannot be forwarded natively is now inlined as a user turn instead of being dropped, the placement check enforces both of Anthropic’s clauses, and Bedrock’s system-reminder converter keeps its cache_control marker
  • Bedrock Streaming Correctness — ConverseStream reports stopReason: tool_use for tool-use turns instead of end_turn (thanks @axelray-dev!); message_start carries an all-zero usage object when figures are unknown so @ai-sdk/anthropic 4.0.6-4.0.32 accepts the frame; and encrypted reasoning is preserved as a replay signature when translating Responses history (thanks @zachgersh!)
  • Encrypted Reasoning Fail-Soft — An upstream 400 caused by unverifiable replayed encrypted_content now strips the reasoning content and retries once instead of failing the request, which matters when a key pool rotates or a request falls back to another provider
  • Server-Side Tool Search — tool_search_tool_* types are normalized on the Responses path, and include_server_side_tool_invocations now reaches the Gemini declaration-drop gate (thanks @AdityaPainuli!)
  • DeepSeek Thinking on Multi-Turn — Thinking is no longer silently disabled for ordinary multi-turn conversations through the OpenAI-compatible surface; the shim is non-mutating and inbound reasoning aliases are normalized
  • Vertex and Gemini Response Fidelity — generateContent keeps candidates[0].safetyRatings and avgLogprobs, and the Vertex cached-content methods honour API-key or context-header auth instead of overwriting Authorization with an OAuth token (thanks @TransactCharlie!)
  • Custom Provider Base Resolution — OpenAI models served through a custom provider now resolve to their built-in base provider before deciding reasoning item-id embedding
  • MCP Tool Errors Replayed as Success — Failed MCP tool executions are now marked as errors instead of being replayed to the model as successful results, covering agent-loop failures, the MCP protocol’s own isError flag, and CodeMode lookup/sandbox failures (thanks @AidanAllchin!)
  • Tool-Result Document Blocks — document blocks in tool results survive the Anthropic to Responses conversion with a synthesized filename, and FileURL/FileType propagate through all three chat/responses conversion paths
  • Stream Termination Edge Cases — A nil delta paired with a non-nil finish reason no longer aborts the stream, and GPT-5-series detection tolerates prefixed model names when resolving reasoning-effort support

​
🗄️ Database Migrations

  • No new database migrations in this release.

​
🐙 Closed GitHub Issues

  • #5206 — Bedrock ConverseStream egress reports stopReason=end_turn for tool-use turns (should be tool_use)
  • #5211 — Bedrock streaming can drop with “unexpected EOF” when an intermediary idle timeout severs a quiet stream
  • #5256 — Concurrent HTTP requests sharing a W3C trace ID lose LLM log rows
  • #5279 — OpenAI /v1/responses to Anthropic drops the tool_search_tool_regex type, so server-side tool_search never runs
  • #5670 — Transcription drops the client’s multipart filename, so non-WAV containers are relabelled audio.mp3 and rejected
  • #5679 — Anthropic Messages does not propagate Gemini mixed server/client tool opt-in
  • #5843 — generateContent (Gemini format) drops candidates[0].safetyRatings and avgLogprobs on Vertex AI responses
  • #5874 — SSE heartbeat frame aborts streams for openai-go ssestream consumers (< v3.43.0) with “unexpected end of JSON input”
  • #5885 — v1.6.8 omits message_start.message.usage on Bedrock-backed providers, breaking @ai-sdk/anthropic streaming
  • #5887 — DeepSeek thinking silently lost on ALL multi-turn requests via OpenAI-compat inbound (v1.6.7; regression from v1.6.3)
  • #5890 — chat completions surface drops tool_result is_error, so failed tool calls replay to the model as successful
  • #5902 — service_tier silently dropped for gpt-5.4 family, priority/fast requests downgrade to default
  • #5905 — v1.6.8 raw passthrough heartbeat can split SSE data lines and corrupt JSON
  • #5925 — config.json force-sync overwrites budget current_usage and last_reset on startup

​
🔧 Maintenance

  • Dependency Upgrades — Bumped core to v1.7.7 and framework to v1.5.7; all plugins bumped to pick up the cascade (compat v0.1.33, governance v1.6.11, jsonparser v1.5.34, logging v1.6.7, maxim v1.6.34, mocker v1.5.34, modelcatalogresolver v1.0.15, otel v1.4.6, prompts v1.0.34, semanticcache v1.5.34, telemetry v1.5.34)
​
1.7.7
  • feat: add BifrostContextKeyCompatDroppedParams so the compat plugin can carry its dropped-parameter list per request instead of on the shared plugin struct (#5902)
  • feat: rewrite thinking.type: "enabled" to the adaptive form for adaptive-only Anthropic models (Opus 4.7+, Opus 4.8, Opus 5, Sonnet 5, Fable 5, Mythos 5) on both the typed request path and the raw passthrough body, and gate stripUnsupportedAnthropicFields on capModel so custom deployment names resolve correctly (#5934)
  • feat: add service instance id to OTEL attributes (#5849)
  • feat: add team / customer / business-unit ids and names to OTEL metrics (#5848)
  • feat: opt-in HTTP/2 PING keepalives on the Bedrock provider via a configurable interval (0 = off) @jeremym-tanium
  • fix: always emit a well-formed message_start on the Anthropic surface - the frame now carries a message object with a usage object and an array content even when the upstream created event has no payload, and the chat-completions converter no longer panics on a content-less assistant message
  • fix: carry tool-result is_error across the chat/Responses mux in both directions, and stop the unmarshal reattach gate from dropping a tool message that marks a failure without a tool_call_id (#5890)
  • fix: mark failed MCP tool executions as errors instead of replaying them to the model as successful results - covers agent-loop execution errors, the MCP protocol’s own isError flag which was previously discarded, and CodeMode lookup/sandbox failures (#5890)
  • fix: carry tool-result is_error through the chat completions surface so Anthropic replay and Bedrock Converse status reflect failed tool calls @AidanAllchin
  • fix: preserve encrypted reasoning as a Bedrock replay signature when translating Responses history @zachgersh
  • fix: include_server_side_tool_invocations now reaches the Gemini declaration-drop gate, so mixed server/client tool opt-in propagates from the Anthropic Messages surface (#5935) @AdityaPainuli
  • fix: inline mid-conversation role:"system" messages as a user turn when the native system form is unavailable (unsupported model or wrong provider), tighten the Anthropic placement check to both of its clauses, and stop convertBifrostSystemReminderToBedrockUserMessage from dropping cache_control (#5931)
  • fix: map Bedrock ConverseStream stopReason to tool_use for tool-use turns instead of reporting end_turn (#5209) @axelray-dev
  • fix: fail soft on invalid_encrypted_content by stripping replayed reasoning encrypted_content and retrying once, cloning reasoning structs rather than mutating them in place, and reporting the true retry ceiling in retry logs (#5892)
  • fix: normalize tool_search_tool_* types on the Responses path so server-side tool search actually runs (#5891)
  • fix: stop disabling DeepSeek thinking on ordinary multi-turn conversations - the disable gate is narrowed, the thinking shim is now non-mutating, reasoning_content is stripped except on tool calls, and inbound reasoning aliases are normalized (#5888)
  • fix: emit an all-zero usage object on message_start for both the native Anthropic and Bedrock invoke streaming paths so @ai-sdk/anthropic 4.0.6-4.0.32 no longer rejects the frame as missing a required field (#5886)
  • fix: carry document tool-result blocks through the Anthropic to Responses converters with a synthesized filename when the block has no title, and propagate FileURL and FileType across all three chat/responses conversion paths (#5884)
  • fix: resolve the base provider for custom providers via the new BifrostContextKeyBaseProviderType and ResolveBaseProvider helper, so reasoning item-id embedding is decided from the built-in provider rather than the custom provider name (#5897)
  • fix: a nil stream delta with a non-nil finish reason no longer bails out of the stream (#5878)
  • fix: keep candidates[0].safetyRatings and avgLogprobs on generateContent (Gemini format) responses from Vertex AI (#5877)
  • fix: support API-key and context-header auth in the Vertex cached content methods instead of unconditionally overwriting the Authorization header with an OAuth token (#5462) @TransactCharlie
  • fix: detect the GPT-5 series with strings.Contains instead of strings.HasPrefix so prefixed model names still resolve reasoning-effort support (#5867)
  • fix: normalize and whitespace-trim attribute values when exporting metrics (#5847)
  • fix: carry the client’s multipart filename through transcription ingress so non-WAV containers are no longer relabelled audio.mp3 and rejected (#5823) @AdityaPainuli
  • fix: concurrent requests sharing a W3C trace id no longer lose LLM log rows (#5356) @AdityaPainuli
  • fix: clamp http2_ping_interval_in_seconds to avoid int64 overflow on conversion to time.Duration @jeremym-tanium
  • perf: avoid copying known request fields on the Anthropic path (#5809) @zachgersh
​
1.5.7
  • feat: expose SetSupportedParamsForTest on the datasheet Store and plumb supports_service_tier through it, so packages outside the datasheet package can seed capability data (#5928)
  • fix: UpdateBudget and UpdateRateLimit now carry live runtime counters (current_usage, last_reset, token_current_usage, token_last_reset, request_current_usage, request_last_reset) forward from the existing row instead of overwriting them, and mergeGovernanceConfig adopts the persisted counters on config force-sync (#5932)
  • fix: concurrent requests sharing a W3C trace id no longer lose log rows in the tracing store (#5356) @AdityaPainuli
  • fix: normalize tool_search_tool_* on the Responses path (#5891)
​
0.1.33
  • fix: scope the compat plugin’s dropped-parameter list to the request that produced it - the list was held on the process-wide plugin struct, so concurrent requests raced on it and extra_fields.dropped_compat_plugin_params could report another request’s dropped params (#5902)
  • feat: log parameters scrubbed by should_drop_params at Debug, and a dropped service_tier at Warn - a silently stripped service_tier downgrades the request to the provider’s default tier with no upstream error, which was previously untraceable (#5902)
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.6.11
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.5.34
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.6.7
  • fix: join pending log entries on the trace’s InternalID instead of its TraceID, so concurrent requests that inherit the same W3C trace id no longer overwrite each other’s pending entry and lose log rows (#5356) @AdityaPainuli
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.6.34
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.5.34
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.0.15
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.4.6
  • feat: add service instance id to the exported OTEL resource attributes (#5849)
  • feat: add team / customer / business-unit ids and names to OTEL metrics (#5848)
  • fix: normalize and whitespace-trim attribute values when exporting metrics (#5847)
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.0.34
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.5.34
  • chore: upgraded core to v1.7.7 and framework to v1.5.7
​
1.5.34
  • feat: add service instance id to the exported telemetry attributes (#5849)
  • feat: add team / customer / business-unit ids and names to exported metrics (#5848)
  • fix: normalize and whitespace-trim attribute values when exporting metrics (#5847)
  • chore: upgraded core to v1.7.7 and framework to v1.5.7