Backport tracking: bug fixes and docs improvements from main (v7) not yet on release-v6.0

[gr2m]

Goal

Track backporting of bug fixes and documentation improvements that were merged into main (AI SDK 7) but are not yet on release-v6.0 (AI SDK 6).

Method: Compared all 375 fix/docs commits on main since the branches diverged (b145dce, 2026-03-05) against release-v6.0 (matching backport commit titles and PR references), then assessed each remaining candidate's applicability by reading the diff and checking whether the affected code/docs — and the bug — exist on release-v6.0.

Snapshot (2026-07-06): 158 already backported · 4 in-flight · 213 candidates assessed below.

✅ Recommended for backport

These fix problems that exist on release-v6.0 and apply without v7-only dependencies.

Bug fixes

• fix(security): prevent streaming tool calls from finalizing on parsable partial JSON #13137 — fix(security): prevent streaming tool calls from finalizing on parsable partial JSON
  • Security fix (VULN-6861): v6's openai, openai-compatible, groq, deepseek, and alibaba streaming code still finalizes tool calls on parsable partial JSON via isParsableJson, and no backport exists on release-v6.0.
• fix(provider/xai): extract reasoning text from content in responses doGenerate #13372 — fix(provider/xai): extract reasoning text from content in responses doGenerate
  • v6 xai responses doGenerate still reads reasoning only from part.summary and its schema lacks the content field, so reasoning returned via content is silently dropped in generateText; fix applies with minor adaptation to v6's evolved condition.
• fix(ai): include tool input on tool result for provider executed dynamic tools #13605 — fix(ai): include tool input on tool result for provider executed dynamic tools
  • v6 run-tools-transformation.ts breaks before toolInputs.set for provider-executed dynamic tools, so their tool results get input undefined; v6 already has the toolCallsByToolCallId map the fix uses.
• fix(ui): export isDynamicToolUIPart from ai package #13961 — fix(ui): export isDynamicToolUIPart from ai package
  • v6 also defines isDynamicToolUIPart in ui-messages.ts but omits it from the ui/index.ts export list, so the same one-line export fix applies directly.
• fix: more precise default message for tool execution denial #14656 — fix: more precise default message for tool execution denial
  • v6 still uses the 'Tool execution denied.' default message that caused GPT-5 models to refuse further calls to the same tool; the one-word message fix applies across the same converter files (plus two extra v6 google files with the same string).
• fix(openai-compatible): buffer tool call deltas until function.name arrives #14760 — fix(openai-compatible): buffer tool call deltas until function.name arrives
  • v6 throws InvalidResponseDataError when the first tool-call delta lacks function.name (real failure with some OpenAI-compatible providers); the buffering fix needs adaptation since v6 lacks StreamingToolCallTracker, but the bug and concept apply.
• fix(ai): tag step/chunk timeout aborts with TimeoutError reason #14943 — fix(ai): tag step/chunk timeout aborts with TimeoutError reason
  • v6 has the same untagged stepAbortController.abort()/chunkAbortController.abort() timeout calls, and the fix's self-contained setAbortTimeout util (TimeoutError DOMException reason) applies without v7-only APIs.
• fix(provider-utils): include object type in default tool schema #15249 — fix(provider-utils): include object type in default tool schema
  • v6 asSchema(undefined) still produces the default tool schema without type: 'object', so strict OpenAI-compatible providers (e.g. DeepSeek, OpenAI-compatible providers fail with tools - missing 'type: object' in JSON Schema #7924) reject tools; one-line version-agnostic fix plus regression test.
• fix(langchain): reset streamEvents text between model invocations #15254 — fix(langchain): reset streamEvents text between model invocations
  • v6 langchain adapter's on_chat_model_start does not reset textStarted/reasoningStarted, so multi-invocation streamEvents runs append later text to the first text part and break part ordering; bug present and fix self-contained.
• fix(langchain): register key mapping for tool calls emitted via messages mode for HITL interrupt matching #15793 — fix(langchain): register key mapping for tool calls emitted via messages mode for HITL interrupt matching
  • v6 langchain utils.ts lacks the else-if that registers emittedToolCallsByKey for tool calls already emitted via messages mode, so the LangGraph HITL interrupt-matching bug (orphaned approval IDs, duplicate events) exists on v6 and the 5-line fix applies.
• fix(ai): return validated elements from array output parseCompleteOutput #15734 — fix(ai): return validated elements from array output parseCompleteOutput
  • v6 output.ts array parseCompleteOutput has the same 'return outerValue.elements as Array' that discards validated values, silently dropping schema transforms/coercions/defaults.
• fix(openai): throw on early stream error events #15922 — fix(openai): throw on early stream error events
  • v6 OpenAI chat/completions/responses models also surface pre-output stream error frames (e.g. insufficient_quota on HTTP 200) as stream error parts instead of throwing retryable APICallError, so the bug exists on v6.
• fix(ai): enforce timeout.stepMs for the whole step in streamText #15840 — fix(ai): enforce timeout.stepMs for the whole step in streamText
  • v6 stream-text has the identical buggy finally { clearStepTimeout(); clearChunkTimeout(); } that disarms the step timer immediately, so timeout.stepMs (which v6 supports) never fires on stalled streams; user-impacting and directly applicable.
• fix(gateway): accept deprecated warnings in image, speech, transcription, and video responses #16030 — fix(gateway): accept deprecated warnings in image, speech, transcription, and video responses
  • v6 gateway image/speech/transcription/video warning schemas lack the 'deprecated' variant, so the shared gateway backend emitting deprecated warnings would fail response parsing on v6.
• fix(ai): omit empty assistant message for data-only blocks in convertToModelMessages #15040 — fix(ai): omit empty assistant message for data-only blocks in convertToModelMessages
  • v6 convert-to-model-messages.ts pushes the assistant message unconditionally (no content.length guard), so data-only blocks produce empty assistant messages on v6 too.
• fix(react): preserve chat when id is undefined #16484 — fix(react): preserve chat when id is undefined
  • v6 use-chat.ts (line 104) has the exact buggy guard that recreates the Chat instance when id is explicitly undefined, wiping messages; the nullish-guard fix applies directly.
• fix(react): use the latest transport in useChat instead of a stale one #16546 — fix(react): use the latest transport in useChat instead of a stale one
  • v6 useChat still passes the transport directly into a once-created Chat instance (only callbacks are proxied), so the stale-transport bug (In useChat in @ai-sdk/react, state sent via transport#body is always stale #7819) exists on v6 and the fix only uses APIs (ChatTransport, DefaultChatTransport) available there.
• fix: MCP callTool hangs and leaks its response handler when an in-flight request is aborted #16595 — fix: MCP callTool hangs and leaks its response handler when an in-flight request is aborted
  • v6 mcp-client.ts has the identical pattern (abort only checked inside the response handler, no abort listener/cleanup), so aborted in-flight callTool also hangs and leaks its handler on v6.
• fix: Google provider marks Gemini external HTTPS file URLs as unsupported #16757 — fix: Google provider marks Gemini external HTTPS file URLs as unsupported
  • v6 google-provider.ts has the same limited supportedUrls (only generativelanguage/youtube), so external HTTPS file URLs are downloaded instead of passed as fileUri; fix is additive and applies cleanly.
• fix: preserve signed tool approval metadata across approval responses #16592 — fix: preserve signed tool approval metadata across approval responses
  • v6 chat.ts has the identical buggy line (approval: { id, approved, reason }) that drops the signature/metadata, and v6 supports signed tool approvals, so the one-line fix applies directly.

Docs (version-agnostic)

• docs(content): harmonize SAP AI Core provider documentation #12922 — docs(content): harmonize SAP AI Core provider documentation
  • Version-agnostic SAP AI Core community provider correction (v6 foundations doc still lists the wrong @mymediset package and the old provider description; the docs explicitly target AI SDK 5.x/6.x).
• docs: Enhance 'tool.toModelOutput' description in js docs and docs website #13481 — docs: Enhance 'tool.toModelOutput' description in js docs and docs website
  • Version-agnostic clarification that toModelOutput is invoked by convertToModelMessages; the v6 jsdoc and tool reference doc lack this note and the API is identical in v6.
• docs: add model mappings to BYOK provider options #14310 — docs: add model mappings to BYOK provider options
  • BYOK model mappings are an AI Gateway service feature equally available to v6 (byok is an untyped record there), and the v6 gateway doc has no mention of model mappings.
• docs: add example to prepareCall jsdoc #15098 — docs: add example to prepareCall jsdoc
  • The prepareCall jsdoc example is version-agnostic; v6's tool-loop-agent-settings.ts has the identical prepareCall option with the same jsdoc lacking the example.
• docs(gateway): clarify BYOK behavior under ZDR filter #15395 — docs(gateway): clarify BYOK behavior under ZDR filter
  • Gateway ZDR/BYOK behavior is a service-level fact; the v6 gateway doc still claims BYOK credentials are always skipped and lacks the 'ZDR-marked BYOK keys are tried first' clarification (text needs light adaptation to v6's wording).
• docs(content): add MongoDB memory provider to agents memory page #15088 — docs(content): add MongoDB memory provider to agents memory page
  • The MongoDB memory provider explicitly targets AI SDK v6 (peerDependency ai ^6.0.0, uses v6 APIs), and the v6 agents memory page lacks the section.
• docs: switch gpt-image-1 references to gpt-image-2 in docs and examples #14684 — docs: switch gpt-image-1 references to gpt-image-2 in docs and examples ⚠️ superseded — v6 docs/examples were separately updated to gpt-image-1.5; cherry-pick is a no-op
  • v6 docs and examples still reference gpt-image-1 while the v6 openai package already supports gpt-image-2, so this factual model update applies equally.
• docs: fix MLflow video embed #16496 — docs: fix MLflow video embed
  • v6 has the identical broken mp4-as-image embed in content/providers/05-observability/mlflow.mdx (only the directory number differs from main's 03-observability).
• docs: fix middleware and message id examples #16493 — docs: fix middleware and message id examples
  • v6 middleware doc has the same result.text bug (content array applies to LanguageModelV3 too, needs V3 type adaptation) and the persistence guide has the same misleading server-side ID wording.
• docs: fix http link to https in deployment guide #16510 — docs: fix http link to https in deployment guide
  • Trivial http->https link fix; the v6 deployment guide still contains the identical http://github.com link.

🤔 Maybe — applicable but lower value or needs a closer look

Bug fixes (10)

• fix(provider/azure): large fixture test instability #13483 — fix(provider/azure): large fixture test instability — Test-only flakiness fix (timeout bump for a large-fixture azure test) and the same flaky test exists on v6; applicable but dev-only value.
• fix(provider): preserve error type prefix in getErrorMessage #14019 — fix(provider): preserve error type prefix in getErrorMessage — v6 getErrorMessage still returns error.message (dropping the error type prefix), but the fix changes error message strings across packages, a cosmetic behavior change that could break user snapshots on a stable branch.
• fix(gateway): use import.meta.url instead of __dirname #14443 — fix(gateway): use import.meta.url instead of __dirname — v6 has the same __dirname usage in the internal gateway generate-model-settings script, but it is a dev-only tooling script and may work fine under v6's tsx setup.
• fix(examples): fix mathjs dependency vulnerability #14424 — fix(examples): fix mathjs dependency vulnerability — v6 examples still pin vulnerable mathjs 14.0.0, but this is a dev-only examples dependency bump with no published-package impact.
• fix: retain stack traces on async errors #14600 — fix: retain stack traces on async errors — Stack-trace retention (return await) is a debuggability/DX improvement that partially applies to v6, but several touched files (tool-loop-agent, generate-video, unified telemetry) differ or are absent, so it needs adaptation and is low-value.
• fix(ai): disable streamText tests that behave differently in Node 24.15+ #14731 — fix(ai): disable streamText tests that behave differently in Node 24.15+ — Test-only fix skipping streamText tests on Node 24.15+; v6 CI pins Node 22 so low urgency, but useful for contributors running v6 tests on newer Node.
• fix(ai): correct mock array result indexing #15154 — fix(ai): correct mock array result indexing — The off-by-one array indexing bug exists in v6's exported mock models (doGenerate[this.doGenerateCalls.length] after push), but it only affects test utilities (dev-only).
• fix(ci): don't fail backport job when fork comment is blocked #16241 — fix(ci): don't fail backport job when fork comment is blocked — v6 backport.yml still has the unguarded gh pr comment (line 189) that fails the job for fork PRs, but this is CI-only and the backport workflow primarily runs from main.
• fix(cookbook): correct highlight ranges in rag-chatbot example #16421 — fix(cookbook): correct highlight ranges in rag-chatbot example — Docs-only highlight-range fix; the v6 rag-chatbot guide has different code snippets and highlight values, so the exact ranges don't transfer and v6 would need its own verification.
• fix(docs): correct example links and typos #16462 — fix(docs): correct example links and typos — Docs/README typo and link fixes; harmless and mostly applicable (4 of 5 files exist on v6) but low value.

Docs (13)

• #null — docs(contributing): add npm dist-tag step to pre-release cycle guide — Version-agnostic contributor doc (npm dist-tag step in pre-release-cycle.md, which exists on v6 without it), but low value since pre-release cycles are run from main.
• docs: mention isLoopFinished() stop condition in context of agent loops #14539 — docs: document isLoopFinished() stop condition for agent loops — Only the ToolLoopAgent stopWhen reference hunk applies to v6 (its description lacks the isLoopFinished mention), and it needs adaptation (stepCountIs naming, non-/v7/ docs link); the workflow-agent files do not exist on v6.
• docs(contributing): publishing new packages #14814 — docs(contributing): publishing new packages — Provenance/package-bootstrapping documentation in contributing/ is version-agnostic and releases also run from the v6 branch, but it is internal non-deployed contributor docs of modest value.
• docs: tool types #14946 — docs: tool types — The function/dynamic tool taxonomy rewrite conceptually applies to v6 (dynamicTool exists there and v6 still says 'three types of tools'), but the commit also changes provider-utils JSDoc code and needs wording review for v7-isms.
• docs: Improve prepareStep messages documentation #15111 — docs: Improve prepareStep messages documentation — prepareStep messages and pruneMessages exist on v6, but the new compaction guide and rewritten sections use v7 API names (e.g. isStepCount vs v6's stepCountIs) and only some hunks apply.
• docs(contributing): publishConfig and repository keys are not needed #15524 — docs(contributing): publishConfig and repository keys are not needed — Fixes the package-bootstrap template in contributing/releases.md, but the section it edits was added by docs(contributing): publishing new packages #14814 and does not exist on v6; only applicable if docs(contributing): publishing new packages #14814 is backported, and it is internal contributor docs.
• docs(examples): update examples/ai-functions/.env.example #15549 — docs(examples): update examples/ai-functions/.env.example — The rewritten .env.example is generated from main's provider set and is not deployed docs; a v6 backport would need regeneration against v6's providers rather than a cherry-pick.
• docs(examples): bump @angular/* to patched 20.3.25 #16150 — docs(examples): bump @angular/* to patched 20.3.25 — Security patch bump for the angular example is version-agnostic, but v6 pins ^20.3.2 (caret already allows 20.3.25) and the large pnpm-lock diff would not cherry-pick; would need a fresh bump on v6 if desired.
• docs(skill): add warning about version mismatch #16391 — docs(skill): add warning about version mismatch — Both files exist on v6 and a skill/SDK version-mismatch warning is useful there too, but the added text says 'AI SDK 7' and would need adaptation; internal skill/contributing docs, not deployed.
• docs(examples/node-http-server): improve example documentation #16388 — docs(examples/node-http-server): improve example documentation — The improved README applies verbatim (the v6 example has the same server with /stream-data and the same outdated README), but it is an example README, not deployed versioned docs, so the value is low.
• docs(contributing): clarify when to make changeset #16416 — docs(contributing): clarify when to make changeset — Version-agnostic CONTRIBUTING.md changeset clarification that applies to the v6 branch too, but it is non-deployed contributor docs and low value (and carries a 'behaivor' typo).
• docs(openai): update token usage docs #16528 — docs(openai): update token usage docs — v6 openai docs have the same stale guidance (providerMetadata.openai.reasoningTokens/cachedPromptTokens are not populated by the v6 provider either), but the corrected property paths differ on v6 (usage.outputTokens.reasoning, usage.inputTokens.cacheRead), so the fix needs adaptation rather than a cherry-pick.
• docs(skill): remove version specific info from use-ai-sdk skill #16557 — docs(skill): remove version specific info from use-ai-sdk skill — Evergreen rewrite of the use-ai-sdk skill would fix the same version-drift problem in the v6 skill, but skills are repo tooling (not deployed versioned docs) and the v6 skill's version-specific references would need separate removal.

🚫 Assessed as not applicable to v6

Mostly fixes for v7-only code or bugs introduced on main after the branch point. Listed for completeness — no action needed unless the assessment is wrong.

Bug fixes (110)

• fix(google): serialize JSON tool results as strings in multi-step calls #13020 — fix(google): serialize JSON tool results as strings in multi-step calls — Adds example files only (no library code change) demonstrating a gemini-2.5-flash-lite workaround; the example paths do not exist on v6.
• fix(deps): update dependency octokit to v5 #12976 — fix(deps): update dependency octokit to v5 — release-v6.0's notify-released script already depends on octokit ^5.0.0.
• fix(provider): fix v4 spec to not use shared v3 types #13482 — fix(provider): fix v4 spec to not use shared v3 types — Fixes the v4 provider spec, which is v7-only; release-v6.0 uses the v3 spec and has none of these files.
• fix(mcp): setting redirect: error for MCP transport #13561 — fix(mcp): setting redirect: error for MCP transport — Setting redirect: 'error' on MCP transports is a deliberate v7 breaking change documented in the 7.0 migration guide, not a fix to backport to v6.
• fix(provider/openai): remove redundant name from custom tool args #13071 — fix(provider/openai): remove redundant name from custom tool args — Breaking API change (major changeset) removing the name argument from openai.tools.customTool(); not a backportable bug fix.
• fix(ai): refactor to flatten model attributes #13615 — fix(ai): refactor to flatten model attributes — Refactors model attributes in the post-branch events/telemetry system (embed-events.ts etc. do not exist on v6).
• fix: narrow the lint-staged glob to handle only JS/TS files #13630 — fix: narrow the lint-staged glob to handle only JS/TS files — v6's lint-staged config already uses a narrowed JS/TS glob (*.{js,jsx,ts,tsx}), so the fix is effectively present.
• fix(ai): include generated files in OTEL response attributes #12634 — fix(ai): include generated files in OTEL response attributes — packages/ai/src/telemetry/open-telemetry-integration.ts was added on main after the branch point (feat(ai): decouple otel from core functions #13051) and does not exist on release-v6.0.
• fix(ai): force template check on 'kind' param #13627 — fix(ai): force template check on 'kind' param — The 'kind' param (from PR feat(openai): add server side compaction for openai #12647) does not exist on release-v6.0; the feature was added on main after the branch point.
• fix(ai): remove otel Tracer api from telemetry settings #13985 — fix(ai): remove otel Tracer api from telemetry settings — v7 breaking removal of the otel Tracer API from telemetry settings; the option must remain on v6.
• fix(provider/gateway): remove duplicate PR_COUNT assignment + update owner of model settings workflow #13982 — fix(provider/gateway): remove duplicate PR_COUNT assignment + update owner of model settings workflow — v6's update-model-settings.yml is an older version without the duplicate PR_COUNT assignment being fixed.
• fix(otel): change experimental_context to context in test #14205 — fix(otel): change experimental_context to context in test — Test-only tweak in packages/otel, which does not exist on release-v6.0.
• fix: better context type inference #14186 — fix: better context type inference — Fixes type inference for the generic tool context feature (feat(ai): change type of experimental_context from unknown to generic #13102) introduced on main after the branch point; GenerationContext/USER_CONTEXT do not exist on v6.
• fix(ai): skip global telemetry registration when local integration defined #14499 — fix(ai): skip global telemetry registration when local integration defined — Fixes createUnifiedTelemetry integration merging, part of the v7 telemetry system that does not exist on release-v6.0.
• fix(ai): remove deprecated experimental_activeTools option #14483 — fix(ai): remove deprecated experimental_activeTools option — v7 removal of the deprecated experimental_activeTools option; must remain on v6.
• fix(anthropic): remove cacheCreationInputTokens from providerMetadata #14570 — fix(anthropic): remove cacheCreationInputTokens from providerMetadata — v7 breaking removal of cacheCreationInputTokens from anthropic providerMetadata; the field must remain on v6.
• fix(ai): remove deprecated experimental_prepareStep #14548 — fix(ai): remove deprecated experimental_prepareStep — Removes the deprecated experimental_prepareStep API as v7 cleanup; backporting would break v6 users.
• fix(workflow): remove obsolete maxSteps parameter #14541 — fix(workflow): remove obsolete maxSteps parameter — packages/workflow does not exist on release-v6.0.
• fix(deps): update dependency @opentelemetry/api to v1.9.1 #14317 — fix(deps): update dependency @opentelemetry/api to v1.9.1 — Dependency bump in packages/otel, which does not exist on release-v6.0.
• fix(ai): limit tool execution time duration to actual tool execution #14601 — fix(ai): limit tool execution time duration to actual tool execution — The inflated durationMs comes from main's executeToolInTelemetryContext wrapper which does not exist on v6; v6 already measures only the actual tool execution.
• fix(ai): remove deprecated experimental_output option #14590 — fix(ai): remove deprecated experimental_output option — Removes the deprecated experimental_output option, a v7 breaking cleanup that must not be backported to v6.
• fix(workflow): Preserve invalid tool call errors in workflow UI streams #14662 — fix(workflow): Preserve invalid tool call errors in workflow UI streams — packages/workflow does not exist on release-v6.0.
• fix: deprecate needsApproval on Tool #14694 — fix: deprecate needsApproval on Tool — Only adds a deprecation of needsApproval as v7 API direction; needsApproval remains the primary v6 API, so no bug to fix there.
• fix(provider/amazon-bedrock): merge custom reasoning config properties when top-level reasoning parameter is set #14711 — fix(provider/amazon-bedrock): merge custom reasoning config properties when top-level reasoning parameter is set — Fixes merging with the top-level reasoning parameter (isCustomReasoning), a v7-only feature; v6 bedrock only reads providerOptions.bedrock.reasoningConfig so the bug does not exist there.
• fix(ai): harden getMediaTypeFromUrl against prototype-property collision #14751 — fix(ai): harden getMediaTypeFromUrl against prototype-property collision — getMediaTypeFromUrl/URL_EXTENSION_TO_MEDIA_TYPE do not exist anywhere on release-v6.0; the feature was added on main after the branch point.
• fix(ai): enforce callOptionsSchema at runtime in ToolLoopAgent #14750 — fix(ai): enforce callOptionsSchema at runtime in ToolLoopAgent — The fix is already present on release-v6.0: tool-loop-agent.ts there already validates options via validateTypes against callOptionsSchema.
• fix(ai): unify generate text event callbacks #14762 — fix(ai): unify generate text event callbacks — v7 API restructuring that unifies generate-text event callbacks; breaking change, not a bug fix for v6.
• fix(ai): remove stopWhen from onStart event #14759 — fix(ai): remove stopWhen from onStart event — Removes stopWhen from the onStart event to simplify v7 generics/runtime-context work; on v6 the property exists in the (experimental) callback-events API and removing it would be a breaking API change, not a bug fix.
• fix: trigger releases after provenance setup #14795 — fix: trigger releases after provenance setup — Empty commit used to trigger releases after provenance setup; nothing to backport.
• fix(ai): reject system messages in messages or prompt by default (opt-in) #14752 — fix(ai): reject system messages in messages or prompt by default (opt-in) — Intentional v7 breaking behavior change (reject system messages in messages/prompt by default); not a backportable bug fix.
• fix(mcp): prevent prototype pollution by using secureJsonParse #14763 — fix(mcp): prevent prototype pollution by using secureJsonParse — Already backported to release-v6.0 as commit 5fee301 (backport v6: fix(mcp): prevent prototype pollution by using secureJsonParse #14826); v6 mcp transports use parseJSONRPCMessage and merge-objects has the proto guard.
• fix(ai): remove ExtractLiteralUnion export #14845 — fix(ai): remove ExtractLiteralUnion export — Removes an accidental ExtractLiteralUnion export that is not exported from packages/ai/src/util/index.ts on v6 (added on main after the branch point).
• fix(ai): remove experimental_customProvider #14848 — fix(ai): remove experimental_customProvider — v7 breaking removal of experimental_customProvider; the API must remain on v6.
• fix(ai): preserve allowSystemInMessages across streamText retries #14844 — fix(ai): preserve allowSystemInMessages across streamText retries — The retry regression came from main's streamLanguageModelCall refactor (each retry re-standardizes the prompt); v6 standardizes the prompt once before retrying, so the bug never existed there.
• fix(ai): fix auto-complete on provider registry and custom provider #14811 — fix(ai): fix auto-complete on provider registry and custom provider — The autocomplete regression was introduced by the v7 pre-release changes (feat: start v7 pre-release cycle #13001) on main after the branch point; v6's registry typing (ExtractLiteralUnion) is unaffected.
• fix(vertex): use correct import for token generator #14919 — fix(vertex): use correct import for token generator — v6's google-vertex-xai-provider-node.ts already uses createAuthTokenGenerator (fix already present).
• fix(openai, openai-compatible): only send null content for assistant messages with tool calls #14950 — fix(openai, openai-compatible): only send null content for assistant messages with tool calls — v6 already contains the fixed logic (content: toolCalls.length > 0 ? text || null : text) in both openai and openai-compatible converters.
• fix(voyage): correct canary version to 2.0.0 to avoid collision with v6 stable #15060 — fix(voyage): correct canary version to 2.0.0 to avoid collision with v6 stable — v7 canary prerelease versioning housekeeping (.changeset/pre.json and voyage canary version); irrelevant to the v6 branch.
• fix(ai): move onToolExecutionStart and onToolExecutionEnd to stable #15058 — fix(ai): move onToolExecutionStart and onToolExecutionEnd to stable — Promotes experimental onToolExecutionStart/End callbacks to stable, a v7 API stabilization rather than a bug fix.
• fix(ai): carry prepareStep message overrides forward across steps #15093 — fix(ai): carry prepareStep message overrides forward across steps — Intentional behavior change (prepareStep message overrides now carry forward) that relies on v7-only initialMessages/responseMessages prepareStep params; backporting would silently change documented v6 semantics.
• fix(ai): limit response messages in StepResult to messages created in that step #15090 — fix(ai): limit response messages in StepResult to messages created in that step — Deliberate v7 behavior change limiting StepResult.response.messages to per-step messages; backporting would break existing v6 semantics.
• fix(ai): remove deprecated LanguageModelUsage properties #15087 — fix(ai): remove deprecated LanguageModelUsage properties — v7 breaking removal of deprecated LanguageModelUsage properties; those properties must remain on v6.
• fix(ai): exclude request and response bodies from text generation results by default to reduce memory usage #15083 — fix(ai): exclude request and response bodies from text generation results by default to reduce memory usage — Breaking v7 default change (excluding request/response bodies from results); backporting would silently change v6 behavior.
• fix(ai): deprecate properties on result that have moved to finalStep #15122 — fix(ai): deprecate properties on result that have moved to finalStep — Deprecates result properties moved to finalStep, a v7-only API/migration concern that does not exist on v6.
• fix: Make generateText and streamText result usage report total usage across all steps and deprecate totalUsage #15121 — fix: Make generateText and streamText result usage report total usage across all steps and deprecate totalUsage — Deliberate v7 semantic change making result.usage accumulate across steps and deprecating totalUsage; backporting would change v6 behavior.
• fix: oxfmt markdown formatting #15113 — fix: oxfmt markdown formatting — Repo-wide oxfmt markdown formatting churn with no user impact; not worth backporting.
• fix: rename system to instructions #15110 — fix: rename system to instructions — The system-to-instructions rename is the core v7 breaking change and must not be backported to v6.
• fix(workflow): Fix stepNumber on telemetry events #15151 — fix(workflow): Fix stepNumber on telemetry events — Fixes stepNumber in @ai-sdk/workflow telemetry; the workflow package does not exist on release-v6.0 (only examples and lockfile touched).
• fix(workflow): pass step number into stream steps #15179 — fix(workflow): pass step number into stream steps — packages/workflow does not exist on release-v6.0.
• fix(ai): accumulative properties on StreamTextResult, GenerateTextResult #15160 — fix(ai): accumulative properties on StreamTextResult, GenerateTextResult — Breaking v7 behavior change making result properties accumulative (documented in the 7.0 migration guide); backporting would change v6 semantics.
• fix: rename onEmbedFinish to onEmbedEnd #15201 — fix: rename onEmbedFinish to onEmbedEnd — Renames the v7-only onEmbedFinish callback; neither onEmbedFinish nor onEmbedEnd exists on release-v6.0.
• fix: rename onRerankFinish to onRerankEnd #15195 — fix: rename onRerankFinish to onRerankEnd — Breaking rename of onRerankFinish to onRerankEnd made during the v7 prerelease; not a backportable bug fix.
• fix: update jsdoc for instructions #15170 — fix: update jsdoc for instructions — One-line jsdoc documenting v7's 'instructions carries forward' prepareStep behavior, which does not exist on v6 (v6 uses system and does not carry forward).
• fix: make sandbox experimental #15253 — fix: make sandbox experimental — Marks the sandbox feature (added on main after the branch point) as experimental; a v7 API-surface decision, not a v6 bug.
• fix: rename telemetry onFinish to onEnd #15243 — fix: rename telemetry onFinish to onEnd — Breaking rename of the telemetry integration onFinish callback to onEnd made during the v7 prerelease; not a backportable bug fix.
• fix(devtools): tidy ToolItem expand/collapse logic #15240 — fix(devtools): tidy ToolItem expand/collapse logic — Cosmetic devtools ToolItem expand/collapse tweak in a file that does not exist on release-v6.0.
• fix(gateway): encode inline v4 file part bytes as { type: 'data' } instead of a data: URL #15050 — fix(gateway): encode inline v4 file part bytes as { type: 'data' } instead of a data: URL — Fix targets the v7 LanguageModelV4 SharedV4FileData tagged union; v6 gateway uses the v3 spec where the data: URL encoding is the correct wire form.
• fix: update step performance metrics with explicit effective, input, output, and total token throughput fields #15307 — fix: update step performance metrics with explicit effective, input, output, and total token throughput fields — Reworks step performance metric throughput fields, a v7-only feature; v6 step-result/stream-text have no performance metrics.
• fix: support tools with optional context #15303 — fix: support tools with optional context — Fixes type inference for tool context (toolsContext/contextSchema), a v7-only feature absent from release-v6.0.
• fix: rename Sandbox.executeCommand to Sandbox.runCommand #15301 — fix: rename Sandbox.executeCommand to Sandbox.runCommand — The Sandbox type with executeCommand does not exist on release-v6.0; this rename targets a v7-only API.
• fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations #15334 — fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations — The dedupe guard (activeReasoning[event.item_id] == null check) is already present in v6's xai-responses-language-model.ts.
• fix(provider/xai): surface full xai error detail in APICallError.message #15338 — fix(provider/xai): surface full xai error detail in APICallError.message — v6 xai-error.ts already contains the identical union schema and code-prefixed errorToMessage (fix was already backported).
• fix(deps): update dependencies #15389 — fix(deps): update dependencies — Routine repo-wide dependency-update maintenance, not a user-facing bug fix suitable for backporting.
• fix(otel): ensure nested context object creates separate attribute #15655 — fix(otel): ensure nested context object creates separate attribute — The @ai-sdk/otel package does not exist on release-v6.0.
• fix(google): auto-inject skip_thought_signature_validator on Gemini 3 replays #15560 — fix(google): auto-inject skip_thought_signature_validator on Gemini 3 replays — Already on release-v6.0 as f62ffe0 (skip_thought_signature_validator auto-injection is present in v6's convert-to-google-generative-ai-messages.ts).
• fix: consider file outputs and tool calls for time to first output #15620 — fix: consider file outputs and tool calls for time to first output — Fixes time-to-first-output step performance metrics, a feature (otel/workflow/step metrics) added on main after the branch point; no performance metrics or sum-token-counts.ts exist on v6.
• fix(ai): pass all stream text parts to onChunk #15618 — fix(ai): pass all stream text parts to onChunk — Intentional v7 behavior change (documented in the 7.0 migration guide); the chunk-subset behavior was deliberately kept on v6 for backwards compatibility.
• fix(vue): forward args in useChat methods #15429 — fix(vue): forward args in useChat methods — The vue useChat composable (use-chat.ts) was added on main after the branch point (feat(vue): add useChat composable #15387) and does not exist on release-v6.0.
• fix: fix streamText tests #15619 — fix: fix streamText tests — Test-only change fixing CI broken on main; v6 tests are unaffected.
• fix: rename onFinish to onEnd #15245 — fix: rename onFinish to onEnd — Deliberate v7 API rename of onFinish to onEnd with new accumulative semantics, not a bug fix applicable to v6.
• fix(gateway): base64-encode inline Uint8Array data on reasoning-file and tool-result file parts #15851 — fix(gateway): base64-encode inline Uint8Array data on reasoning-file and tool-result file parts — Fixes encoding for reasoning-file parts and Uint8Array tool-result file data, which only exist in the v7 LanguageModelV4 spec; v6's LMv3 spec uses base64 strings and has no reasoning-file part.
• fix: add missing name to run-agent-tui returns table #15937 — fix: add missing name to run-agent-tui returns table — Docs-only fix for the v7-only run-agent-tui documentation page which does not exist on release-v6.0.
• fix: bump base versions to 1.0.0 for @ai-sdk/harness, @ai-sdk/sandbox-just-bash, and @ai-sdk/sandbox-vercel #15932 — fix: bump base versions to 1.0.0 for @ai-sdk/harness, @ai-sdk/sandbox-just-bash, and @ai-sdk/sandbox-vercel — Version bump to 1.0.0 for harness/sandbox packages that do not exist on release-v6.0.
• fix(workflow): reuse core tool-approval validation in WorkflowAgent #15987 — fix(workflow): reuse core tool-approval validation in WorkflowAgent — Reworks tool-approval validation inside WorkflowAgent (packages/workflow), which does not exist on release-v6.0.
• fix(harness): fix various bugs with harness skills not being correctly processed by the harness adapters #16051 — fix(harness): fix various bugs with harness skills not being correctly processed by the harness adapters — Harness skills processing is in the v7-only harness packages, which do not exist on release-v6.0.
• fix(harness): guard against invalid resuming a session vs continuing a turn #16047 — fix(harness): guard against invalid resuming a session vs continuing a turn — The harness package does not exist on release-v6.0.
• fix(ci): trigger backport on pull_request closed to avoid indexing race #16045 — fix(ci): trigger backport on pull_request closed to avoid indexing race — release-v6.0's backport.yml already contains the fixed race-free labeled/closed trigger version.
• fix(examples): require host origin for MCP sandbox-proxy resource-ready #15995 — fix(examples): require host origin for MCP sandbox-proxy resource-ready — Fixes a single example file (MCP sandbox-proxy) that does not exist on release-v6.0.
• fix(react): deny MCP App tool calls by default when allowedTools is omitted #15996 — fix(react): deny MCP App tool calls by default when allowedTools is omitted — MCP App tool-call handling in @ai-sdk/react is a v7-only feature; the touched files do not exist on release-v6.0.
• fix(harness): avoid using peer dependencies for underlying harness and sandbox SDKs #16029 — fix(harness): avoid using peer dependencies for underlying harness and sandbox SDKs — The harness-* and sandbox-* packages it modifies do not exist on release-v6.0.
• fix: Promote transcribe and TranscriptionResult to stable #16009 — fix: Promote transcribe and TranscriptionResult to stable — Promotes transcribe from experimental to stable as part of v7 API stabilization; v6 intentionally keeps experimental_transcribe.
• fix: Promote generateSpeech and SpeechResult to stable #16010 — fix: Promote generateSpeech and SpeechResult to stable — Promotes experimental generateSpeech/SpeechResult APIs to stable, a v7 stabilization decision rather than a bug fix backportable to v6.
• fix(harness): make listening for sandbox bridge readiness compatible with Bun #16069 — fix(harness): make listening for sandbox bridge readiness compatible with Bun — Touches only harness packages, which do not exist on release-v6.0.
• fix(docs): fix model message file part reference docs #16065 — fix(docs): fix model message file part reference docs — Documents the v7-only tagged FileData/ProviderReference file-part shape; v6 docs and code still use the plain DataContent | URL type.
• fix: Remove deprecated experimental generateImage exports #16017 — fix: Remove deprecated experimental generateImage exports — Removes deprecated experimental generateImage exports as v7 API cleanup; backporting would break v6 users still on those exports.
• fix(devtools): validate viewer database paths #16049 — fix(devtools): validate viewer database paths — The vulnerable remote-dbPath feature does not exist on v6: reloadDb() takes no arguments and /api/notify does not accept a dbPath, so the path-traversal bug was introduced on main after the branch point.
• fix(examples): memoize realtime sessionConfig to keep websocket connected #16141 — fix(examples): memoize realtime sessionConfig to keep websocket connected — Example-only fix for a v7 realtime example whose file does not exist on release-v6.0.
• fix(harness): reject websocket bind errors #16122 — fix(harness): reject websocket bind errors — packages/harness does not exist on release-v6.0.
• fix(harness): fix incomplete OIDC token support for AI Gateway auth in harness adapters #16212 — fix(harness): fix incomplete OIDC token support for AI Gateway auth in harness adapters — Harness packages (harness, harness-claude-code, harness-codex, harness-pi) do not exist on release-v6.0.
• fix(workflow): reject system messages in WorkflowAgent by default (opt-in) #15024 — fix(workflow): reject system messages in WorkflowAgent by default (opt-in) — WorkflowAgent lives in packages/workflow, which does not exist on release-v6.0.
• fix(workflow): forward provider-executed tool approvals to the provider on resume #16256 — fix(workflow): forward provider-executed tool approvals to the provider on resume — The workflow package does not exist on release-v6.0.
• fix(harness-pi): fix client-side tool approval regression #16253 — fix(harness-pi): fix client-side tool approval regression — packages/harness-pi does not exist on release-v6.0.
• fix(tui): allow all agents #16268 — fix(tui): allow all agents — tui package does not exist on release-v6.0.
• fix(react): fix toolMetadata mismatch #16301 — fix(react): fix toolMetadata mismatch — packages/react/src/mcp-apps does not exist on release-v6.0.
• fix(harness): ensure finish chunk's total usage is actually coming from total usage #16335 — fix(harness): ensure finish chunk's total usage is actually coming from total usage — The harness package does not exist on release-v6.0.
• fix(gateway): include realtime in model metadata #16381 — fix(gateway): include realtime in model metadata — Adds the 'realtime' model type to gateway metadata for v7; v6 gateway has no realtime model support (KNOWN_MODEL_TYPES lacks it and no realtime files exist).
• fix(harness-pi): fix potential path traversal issues in Pi harness adapter #16430 — fix(harness-pi): fix potential path traversal issues in Pi harness adapter — The harness-pi package does not exist on release-v6.0.
• fix(harness): use shellQuote where appropriate in harness adapters #16429 — fix(harness): use shellQuote where appropriate in harness adapters — Harness packages (harness-claude-code, harness-codex, harness-pi) do not exist on release-v6.0.
• fix(harness): fix harness Zod usage to be v3/v4 compatible #16417 — fix(harness): fix harness Zod usage to be v3/v4 compatible — Zod compatibility fix for harness packages, none of which exist on release-v6.0 (only lockfile overlaps).
• fix(workflow): forward reasoning generation setting #16411 — fix(workflow): forward reasoning generation setting — packages/workflow does not exist on release-v6.0.
• fix(ai): type agent tool call callback aliases #16461 — fix(ai): type agent tool call callback aliases — v6's AgentCallParameters has no tool execution callbacks at all (only onStepFinish), so the missing deprecated alias typings are a v7-only gap.
• fix(workflow): repair UI message stream part framing in WorkflowChatTransport #16433 — fix(workflow): repair UI message stream part framing in WorkflowChatTransport — packages/workflow does not exist on release-v6.0.
• fix(harness-codex): fix Codex sometimes stopping after a single text response when using workflow-harness #16441 — fix(harness-codex): fix Codex sometimes stopping after a single text response when using workflow-harness — harness-codex package does not exist on release-v6.0.
• fix(harness-claude-code): declare Claude Code's Monitor tool as built-in tool #16520 — fix(harness-claude-code): declare Claude Code's Monitor tool as built-in tool — packages/harness-claude-code does not exist on release-v6.0.
• fix(harness): fix harness relay auth #16432 — fix(harness): fix harness relay auth — Touches only harness-codex/harness-opencode packages, which do not exist on release-v6.0.
• fix(harness-codex): improve Codex prompt framing and avoid placing relay shim file in session workdir #16485 — fix(harness-codex): improve Codex prompt framing and avoid placing relay shim file in session workdir — packages/harness-codex does not exist on release-v6.0.
• fix(amazon-bedrock): send correct bedrockRerankingConfiguration key for reranking #16501 — fix(amazon-bedrock): send correct bedrockRerankingConfiguration key for reranking — v6's bedrock-reranking code already uses the correct bedrockRerankingConfiguration wire key; the bug was introduced on main during a later rename.
• fix(ai): replace dynamic import() with loadBuiltinModule for diagnostics_channel #16192 — fix(ai): replace dynamic import() with loadBuiltinModule for diagnostics_channel — packages/ai/src/telemetry/tracing-channel-publisher.ts was added on main after the branch point (feat(ai): use tracing channels to track parent-child context #15660) and does not exist on release-v6.0.
• fix: send OpenAI chat image data as data URLs #16655 — fix: send OpenAI chat image data as data URLs — The bug is in v7's LanguageModelV4 file-data handling (part.data.data); v6's convert-to-openai-chat-messages already emits data:;base64 URLs.
• fix: add vitest dependency to workflow #16744 — fix: add vitest dependency to workflow — Adds a vitest devDependency to packages/workflow, which does not exist on release-v6.0.

Docs (50)

• docs(skills): make ADR skill internal #13381 — docs(skills): make ADR skill internal — Marks the ADR skill internal in a skills/ file that does not exist on release-v6.0.
• docs: ADRs #13343 — docs: ADRs — ADR infrastructure (contributing/decisions, ADR skill, AGENTS.md wiring) governs main-branch development and mostly does not exist on release-v6.0.
• docs: start a migration guide for v7 #13414 — docs: start a migration guide for v7 — Creates the v7 migration guide, which is v7-specific and excluded from backports.
• docs(ai): README update #13651 — docs(ai): README update — The commit only adds a changeset file (.changeset/two-pens-teach.md); there is no doc content to backport.
• docs: document noun-verb naming convention for stream event types #13805 — docs: document noun-verb naming convention for stream event types — Documents a naming convention in a contributing/ file that does not exist on release-v6.0.
• docs: add pricing and reporting API reference to gateway docs #13727 — docs: add pricing and reporting API reference to gateway docs — The v6 ai-gateway doc already documents the reporting API (via getSpendReport) and pricing link, so the gap this commit fills does not exist on v6.
• docs: stream pipeline structure #13738 — docs: stream pipeline structure — Architecture doc describing v7 internals; the file does not exist on release-v6.0.
• docs(ai): add toolMs and per-tool timeout docs #14067 — docs(ai): add toolMs and per-tool timeout docs — Documents toolMs and per-tool timeout overrides, a v7-only feature; v6 timeout config only supports totalMs/stepMs/chunkMs.
• docs: add new documentation for telemetry setup #14014 — docs: add new documentation for telemetry setup — Documents the v7-only registerTelemetryIntegration/@ai-sdk/otel setup; v6 telemetry uses experimental_telemetry without a registration step, so the rewrites would be wrong on v6.
• docs(migration): fix formatting of code block #14191 — docs(migration): fix formatting of code block — Formatting fix in the v7 migration guide, which is excluded from backports and absent on release-v6.0.
• docs(architecture): clarify bookkeeping #13659 — docs(architecture): clarify bookkeeping — Clarifies an architecture/ doc that does not exist on release-v6.0.
• docs(workflow): remove obsolete maxSteps #14555 — docs(workflow): remove obsolete maxSteps — Removes obsolete maxSteps from a workflow doc that does not exist on release-v6.0.
• docs: update telemetry docs to show enabled by default #14538 — docs: update telemetry docs to show enabled by default — Documents telemetry being enabled by default, a v7 behavior change; v6 telemetry remains experimental and opt-in via experimental_telemetry.
• docs(workflow): migrate from DurableAgent #14650 — docs(workflow): migrate from DurableAgent — The DurableAgent-to-WorkflowAgent migration doc is v7-only and does not exist on release-v6.0.
• docs(workflow): replace Vercel Workflow with just Workflow or Workflow SDK #15106 — docs(workflow): replace Vercel Workflow with just Workflow or Workflow SDK — The 'Vercel Workflow' naming appears only in v7-era WorkflowAgent/workflow docs; the two files that exist on v6 do not contain the phrase.
• docs: add tool approval page to agents section #15198 — docs: add tool approval page to agents section — The tool approvals page documents the ToolLoopAgent toolApproval option, which is v7-only; v6 uses the needsApproval API.
• docs: add note about tool execution events #15263 — docs: add note about tool execution events — Adds code comments/renames around the v7-only telemetryDispatcher and onToolExecutionStart plumbing in stream-text.ts, which does not exist in that form on v6.
• docs: add variable and identifier naming guidance #15641 — docs: add variable and identifier naming guidance — The contributing naming-guidance file does not exist on release-v6.0; contributor guidance for main development.
• docs: remove experimental telemetry warning #15611 — docs: remove experimental telemetry warning — Removes the experimental-telemetry warning because telemetry went stable in v7; telemetry is still experimental in v6, so the warning must stay.
• docs(examples): migrate deprecated streamText result helpers to standalone helpers #15741 — docs(examples): migrate deprecated streamText result helpers to standalone helpers — Migrates docs/examples away from streamText result helpers deprecated in v7; those helpers are the correct, non-deprecated API in v6.
• docs(contributing): document the Experimental_ prefix seam convention #15833 — docs(contributing): document the Experimental_ prefix seam convention — contributing/naming-conventions.md does not exist on release-v6.0 and the Experimental_ seam convention is contributor guidance for main development, not deployed docs.
• docs: update devtools docs + readme for v7 #15913 — docs: update devtools docs + readme for v7 — Updates devtools docs/README from v6 beta to v7 canary and the v7 DevToolsTelemetry/registerTelemetry API; v6 docs correctly describe the v6 middleware-based devtools.
• docs: Policy-Based Tool Approvals page for @ai-sdk/policy-opa #15875 — docs: Policy-Based Tool Approvals page for @ai-sdk/policy-opa — Documents the new v7-era @ai-sdk/policy-opa package; the new policy-tool-approvals page does not exist on v6 and the other hunks just link to it.
• docs(provider): add Latitude observability provider (v7) #15918 — docs(provider): add Latitude observability provider (v7) — Explicitly the v7 companion of PR docs(provider): add Latitude observability provider v6 #15816, which already added Latitude docs to release-v6.0 with the v6 telemetry pattern (latitude.mdx exists on v6).
• chore(docs): fix harness docs label #16052 — chore(docs): fix harness docs label — Fixes a label in harness docs that do not exist on release-v6.0.
• docs(harness): clarify harness @canary tags #16073 — docs(harness): clarify harness @canary tags — Harness @canary tag docs are for v7-only harness packages and none of the files exist on release-v6.0.
• docs(harness): clarifications #16070 — docs(harness): clarifications — Harness docs clarifications target v7-only pages that do not exist on release-v6.0.
• docs: add telemetry code snippet for nextjs instrumentation #16149 — docs: add telemetry code snippet for nextjs instrumentation — The Next.js snippet uses the v7 registerTelemetry/@ai-sdk/otel API; the v6 telemetry doc is the older experimental_telemetry version without the section being amended.
• docs: update docs for model generation lifecycle #16128 — docs: update docs for model generation lifecycle — Renames event-listeners to lifecycle-callbacks and reflects v7 API changes (onEnd totalUsage->usage, ToolLoopAgent lifecycle events) that do not match the v6 event-listeners doc.
• docs: change highlighted code lines for example #16136 — docs: change highlighted code lines for example — The v6 building-agents page does not contain the toolsContext/experimental_sandbox code blocks whose highlight line numbers were fixed (v7-only content).
• docs: reorder docs to surface observability integrations #16169 — docs: reorder docs to surface observability integrations — 173-file navigation renumbering/restructure of the v7 docs tree (observability moved to 03, community to 05); a structural reorg, not a version-agnostic fix, and v6 has a different directory layout.
• docs: canary -> beta #16156 — docs: canary -> beta — canary -> beta install instructions are v7 prerelease-channel docs for v7-only packages (harness, devtools, migration guide); not relevant to stable v6 docs.
• docs(harness): add contributor focused architecture docs and agent skill for adding a new harness adapter package #16224 — docs(harness): add contributor focused architecture docs and agent skill for adding a new harness adapter package — Harness architecture docs and the add-harness-package skill document v7-only packages that do not exist on release-v6.0.
• docs(harness): minor fixes regarding auth configuration #16215 — docs(harness): minor fixes regarding auth configuration — Harness docs are for the v7-only harness packages and none of the files exist on release-v6.0.
• docs: sort migration guide sections #16176 — docs: sort migration guide sections — Migration-guide work is excluded from backports and the v7 migration guide file does not exist on release-v6.0.
• docs(workflow): fix message persistence guidance for WorkflowAgent #16259 — docs(workflow): fix message persistence guidance for WorkflowAgent — WorkflowAgent docs are v7-only and the file does not exist on release-v6.0.
• docs: add raindrop as an observability provider #16252 — docs: add raindrop as an observability provider — The Raindrop observability page explicitly documents the v7 telemetry interface (registerTelemetry), which does not exist in v6.
• docs: fix Raindrop observability provider examples #16257 — docs: fix Raindrop observability provider examples — raindrop.mdx does not exist in the v6 observability docs (content/providers/05-observability), so there is nothing to fix.
• docs: add new skill for migrating to v7 #16216 — docs: add new skill for migrating to v7 — Adds the v6-to-v7 migration skill, which is explicitly v7 migration work and its files do not exist on release-v6.0.
• docs(skill): add harness-authoring gotchas to add-harness-package #16310 — docs(skill): add harness-authoring gotchas to add-harness-package — Harness packages and the add-harness-package skill are v7-only and the file does not exist on release-v6.0.
• docs: add realtime Gateway guidance #16314 — docs: add realtime Gateway guidance — Realtime (experimental_realtime/useRealtime) is a v7-only feature; the realtime docs do not exist on v6 and the gateway doc changes are all realtime guidance.
• docs(context): fix line highlights #16336 — docs(context): fix line highlights — The runtime-and-tool-context page is v7-only and does not exist on release-v6.0.
• docs: remove experimental from devtools + deadlink #16375 — docs: remove experimental from devtools + deadlink — The dead-link fix targets 17-mcp-apps.mdx which does not exist on v6, and removing the devtools 'experimental' warning reflects v7 stabilization, so it would be wrong for the v6 docs.
• docs(harness): remove beta tags from harness docs #16369 — docs(harness): remove beta tags from harness docs — Harness docs (content/providers/02-ai-sdk-harnesses) are a v7-only feature and do not exist on release-v6.0.
• docs: update langfuse docs for v7 compat #16333 — docs: update langfuse docs for v7 compat — Updates Langfuse docs for v7 compatibility and the file does not exist on release-v6.0.
• docs: AI SDK 7 #16365 — docs: AI SDK 7 — AI SDK 7 release announcement updates (v7 migration guide, harness docs, devtools beta notes) are v7-specific.
• docs: update onFinish to onEnd #16398 — docs: update onFinish to onEnd — onFinish -> onEnd is a v7 API rename; v6 still uses onFinish, so the v6 docs are correct as-is.
• docs: add Sentry observability guide #16394 — docs: add Sentry observability guide — New Sentry observability guide is built around the v7 registerTelemetry/@ai-sdk/otel API, which does not exist in v6 (experimental_telemetry).
• docs: fix language model middleware spec links #16487 — docs: fix language model middleware spec links — v6 middleware docs already link to the release-v6.0 branch v3 spec (links work) and the v6 reference page consistently documents LanguageModelV3, so the 404/mismatch problem is main-only.
• docs: document xai responses api default in v7 migration guide and provider docs #16542 — docs: document xai responses api default in v7 migration guide and provider docs — Documents the v7-only xAI responses-API default in the v7 migration guide and a 'since AI SDK 7' provider note; v6 keeps chat completions as default.

🚧 In-flight — open backport PRs against release-v6.0

• [v6.0] fix: Raw JSON.parse used in production provider code (prototype pollution risk) #16763 — [v6.0] fix: Raw JSON.parse used in production provider code (prototype pollution risk)
• [v6.0] fix: extractJsonMiddleware removes valid leading space from final streamed suffix #16762 — [v6.0] fix: extractJsonMiddleware removes valid leading space from final streamed suffix
• [v6.0] fix: normalize bare Anthropic API base URL to avoid missing /v1 path #16761 — [v6.0] fix: normalize bare Anthropic API base URL to avoid missing /v1 path
• [v6.0] fix: embedMany over-batches @ai-sdk/google embedding models #16760 — [v6.0] fix: embedMany over-batches @ai-sdk/google embedding models
• Backport: feat(alibaba): support wan2.7 text-to-video and reference-to-video models #16660 — Backport: feat(alibaba): support wan2.7 text-to-video and reference-to-video models
• Backport pr 16328 to release v6.0 #16571 — Backport pr 16328 to release v6.0
• ci: bump the github-actions group across 1 directory with 4 updates #16537 — ci: bump the github-actions group across 1 directory with 4 updates
• docs: backport MLflow video embed fix #16497 — docs: backport MLflow video embed fix
• backport: feat(ai): optimize text accumulation runtime to O(N) #15906 — backport: feat(ai): optimize text accumulation runtime to O(N)
• chore(examples): reproduction scripts for Gemini 3 thoughtSignature issues #15548 — chore(examples): reproduction scripts for Gemini 3 thoughtSignature issues
• feat(mistral): add Voxtral transcribe model support #15076 — feat(mistral): add Voxtral transcribe model support
• feat: add typed approval data schema support for tool definitions #13306 — feat: add typed approval data schema support for tool definitions

---

Assessments were made by inspecting each commit's diff against release-v6.0; individual verdicts may need human review.

[gr2m]

In-flight PR triage (2026-07-06)

Reviewed all open PRs against release-v6.0 from the in-flight section above.

Merged

• [v6.0] fix: embedMany over-batches @ai-sdk/google embedding models #16760 — fix: embedMany over-batches @ai-sdk/google embedding models. Minimal port of embedMany over-batches @ai-sdk/google embedding models #16101; Gemini batch API limit is 100.
• [v6.0] fix: normalize bare Anthropic API base URL to avoid missing /v1 path #16761 — fix: normalize bare Anthropic API base URL to avoid missing /v1 path. Port of @ai-sdk/anthropic: ANTHROPIC_BASE_URL convention differs from official Anthropic SDK, breaks under Claude Code / Cursor #15542 with tests for both the option and env-var paths.
• [v6.0] fix: extractJsonMiddleware removes valid leading space from final streamed suffix #16762 — fix: extractJsonMiddleware removes valid leading space from final streamed suffix. Port of extractJsonMiddleware removes valid leading space from final streamed suffix #15921 with regression test.
• [v6.0] fix: Raw JSON.parse used in production provider code (prototype pollution risk) #16763 — fix: raw JSON.parse in production provider code (prototype pollution risk). Port of Raw JSON.parse used in production provider code (prototype pollution risk) #15812; secureJsonParse already existed on v6, this wires it into the anthropic/google/gateway sync-parse paths.
• ci: bump the github-actions group across 1 directory with 4 updates #16537 — ci: dependabot GitHub Actions bumps. All four new SHAs verified against their upstream release tags (checkout v7.0.0, pnpm/action-setup v6.0.9, cache v6.1.0, changesets/action v1.9.0).
• docs: backport MLflow video embed fix #16497 — docs: MLflow video embed fix (admin-merged, gr2m-authored, docs-only).
• chore(examples): reproduction scripts for Gemini 3 thoughtSignature issues #15548 — chore(examples): Gemini 3 thoughtSignature repro scripts (admin-merged, gr2m-authored, examples-only; was in draft — marked ready first).
• Backport: feat(alibaba): support wan2.7 text-to-video and reference-to-video models #16660 — Backport: feat(alibaba) wan2.7 t2v/r2v. The bot had committed literal merge-conflict markers; resolved by porting main's changes to the v3 video spec (V4→V3 type renames, added media/ratio to v6's inline options schema, dropped the accidentally cherry-picked separate options file). Verified the resolved file matches main modulo the spec rename; alibaba tests 124/124 green locally.

Follow ups for later

• backport: feat(ai): optimize text accumulation runtime to O(N) #15906 — backport: feat(ai) optimize text accumulation to O(N). The original PR feat(ai): optimize text accumulation runtime to O(N) #15897 is still open on main — this isn't a backport yet. The lazy getter/setter approach on UI message parts is also subtle enough (serialization, structuredClone, React rendering) that it should land and soak on main first.
• Backport pr 16328 to release v6.0 #16571 — Backport PR support video reference input for r2v #16328 (video reference input for r2v). support video reference input for r2v #16328 is also still open on main, the backport has merge conflicts across 52 files, and it modifies the frozen v3 provider spec. Should wait until the original merges, then be regenerated.
• feat(mistral): add Voxtral transcribe model support #15076 — feat(mistral): Voxtral transcribe model support. Feature PR from a fork targeting release-v6.0 directly, conflicting since May with failing CI. Needs a decision on whether new transcription support should land on v6 at all, plus a rebase.
• feat: add typed approval data schema support for tool definitions #13306 — feat: typed approval data schema for tool definitions. Large public-API change (±1000 lines, new generic on tool definitions) whose own description says it targets main, yet it's based against release-v6.0. Likely needs retargeting to main.

Not touched: #16765 (Version Packages) — will regenerate after the merges above; handle via the usual release review.