Repository Issue Activity (beta)

#19320 AI SDK's provider abstraction as an execution layer for portable EvalPort eval suites

#19311 Negative outputTokens.text: the #18614 clamp is missing from sibling usage converters

#8994 MockLanguageModelV2 break down when using provider-utils v3.0.10

#19313 Security: no working private channel to report a tool-approval signing issue

#19302 harness: turn deadlocks when one model step emits two host tool calls that both need approval

#19300 `Output.object()`: `generateText` and `streamText` throw different error classes for the same unparseable truncated output; `generateText`'s error carries no diagnostics

#19241 `useAIState(key)` drops updates when two setters fire in the same tick. Wrapped server actions patch against garbage too.

#17871 Zod 4 conversion erases z.record() value schemas by overwriting additionalProperties with false

#19236 `@ai-sdk/openai-compatible`: support video file parts for providers that accept `video_url` content

#19157 Stdio MCP transport cannot launch npx-based servers on Windows (shell: false is hardcoded)

#19068 harness-claude-code: finish-step reports the whole turn's cumulative usage on a turn's last step, not that step's own

#19156 harness-pi drops approved builtin tool results from continued streams

#9300 Add text splitting support

#10781 [@ai-sdk/anthropic | @ai-sdk/google] Duplicate text-start/text-end IDs in multi-step flow

#19243 `@ai-sdk/vue` `useCompletion` can't clear a non-empty `initialCompletion`. Empty model responses never render.

#19229 Image model spec: advertise file-input (editing/mask) support so callers can detect it before doGenerate

#19274 @ai-sdk/openai: normalize malformed replayed tool arguments for strict Chat Completions backends

#19238 validateApprovedToolApprovals throws InvalidToolInputError after the human approves, with no repairToolCall path

#19205 `@ai-sdk/mcp` SSE transport swallows non-2xx HTTP responses. `tools()` and `callTool()` hang forever.

#14143 Branded message id not allowed

#18999 Unparseable tool-call inputs are double-encoded on replay (convertToModelMessages rawInput + openai serializeToolCallArguments) → 400 on strict OpenAI-compatible backends

#17988 Image token usage: @ai-sdk/bytedance

#17987 Image token usage: @ai-sdk/openai-compatible

#19250 @ai-sdk/otel: embed/embedMany emit two identical "embeddings" spans

#14360 AI Gateway silently ignores mask parameter for OpenAI inpainting (gpt-image-1)