Repository navigation

Issues

New issue

How AI SDK maintenance and community facilitation is changing
#17243 · gr2m opened on Jul 14, 2026
1

openai-compatible: make the assistant reasoning replay field configurable (reasoning_content vs reasoning) — Cerebras 400s on the hardcoded default
Bug
vercel/ai#24127 · opened on Oct 27, 2026
2
generateObject and streamObject publish nothing on the ai:telemetry tracing channel
Bug
vercel/ai#24125 · opened on Oct 27, 2026
2
@nikolas-sapa A note on your audit-generated issues
vercel/ai#24118 · opened on Oct 27, 2026
Opt-in: execute tools as soon as their input is complete in streamText
Feature
vercel/ai#24108 · opened on Oct 27, 2026
1
Add a Tako Contents tool to the AI Gateway provider
Feature
vercel/ai#24079 · opened on Oct 26, 2026
2
@ai-sdk/harness: a denied host tool call reaches the tool's own toModelOutput, which can fail the turn
Bug
vercel/ai#24057 · opened on Oct 26, 2026
3
fix(gateway): an aborted request becomes a GatewayResponseError with an invented statusCode 500
Bug
vercel/ai#24056 · opened on Oct 26, 2026
5
docs: add Bifrost to the OpenAI-compatible providers section
Documentation
vercel/ai#24054 · opened on Oct 26, 2026
1
Registry entry: MrScraper, web page retrieval with browser rendering and Google search
Documentation
vercel/ai#24050 · opened on Oct 26, 2026
2
google: GoogleJSONAccumulator leaves stringOpen true on final string continuation and string terminator chunks
Bug
vercel/ai#24044 · opened on Oct 25, 2026
2
docs: add Phala to the OpenAI-compatible providers section
Documentation
vercel/ai#24043 · opened on Oct 25, 2026
1
sandbox-vercel: adding request transformations silently drops response rules from the network policy
Bug
vercel/ai#24035 · opened on Oct 25, 2026
4
tool-web-fetch throws on a malformed redirect Location header
Bug
vercel/ai#23971 · opened on Oct 25, 2026
4
Registry entry: Tracekit, policy checks and approvals for tool calls, with a signed record
Feature
vercel/ai#23965 · opened on Oct 25, 2026
1
@ai-sdk/react: MCP App does not implement baseUriDomains
Bug
vercel/ai#23962 · opened on Oct 25, 2026
2
harness-claude-code: let hosts mark host tools read-only and always-loaded
Feature
vercel/ai#23961 · opened on Oct 25, 2026
1
harness-claude-code: host tool schemas lose their limits and defaults
Bug
vercel/ai#23960 · opened on Oct 25, 2026
3
docs: middleware cache example replays responses across models and providers
Documentation
vercel/ai#23956 · opened on Oct 24, 2026
4
@ai-sdk/react: MCPAppRenderer can not load script from CDN
Bug
vercel/ai#23946 · opened on Oct 24, 2026
3
Non-base64 data URLs in file parts throw InvalidCharacterError or send the wrong bytes
Bug
vercel/ai#23943 · opened on Oct 24, 2026
3
@ai-sdk/openai: GPT-5.x has no supportedReasoningEfforts, so unsupported reasoning efforts are sent unchecked and fail with 400
Bug
vercel/ai#23929 · opened on Oct 23, 2026
3
harness-codex: Codex subagent activity never reaches the stream (forward it as raw parts, like harness-claude-code)
Feature
vercel/ai#23875 · opened on Oct 23, 2026
1
Google Vertex: support Google Auth Library 11
Feature
vercel/ai#23870 · opened on Oct 23, 2026
1
@ai-sdk/open-responses: streamText drops the server's response id, model and timestamp, generateText keeps them
Bug
vercel/ai#23869 · opened on Oct 23, 2026
4
amazon-bedrock: named toolChoice removes the other tools for Anthropic models and invalidates the prompt cache
Bug
vercel/ai#23849 · opened on Oct 22, 2026
2

Navigation Menu

Repository navigation

Issues

Issues

openai-compatible: make the assistant reasoning replay field configurable (reasoning_content vs reasoning) — Cerebras 400s on the hardcoded default

`generateObject` and `streamObject` publish nothing on the `ai:telemetry` tracing channel

@nikolas-sapa A note on your audit-generated issues

Opt-in: execute tools as soon as their input is complete in streamText

Add a Tako Contents tool to the AI Gateway provider

@ai-sdk/harness: a denied host tool call reaches the tool's own `toModelOutput`, which can fail the turn

fix(gateway): an aborted request becomes a GatewayResponseError with an invented statusCode 500

docs: add Bifrost to the OpenAI-compatible providers section

Registry entry: MrScraper, web page retrieval with browser rendering and Google search

google: GoogleJSONAccumulator leaves stringOpen true on final string continuation and string terminator chunks

docs: add Phala to the OpenAI-compatible providers section

sandbox-vercel: adding request transformations silently drops `response` rules from the network policy

tool-web-fetch throws on a malformed redirect Location header

Registry entry: Tracekit, policy checks and approvals for tool calls, with a signed record

@ai-sdk/react: MCP App does not implement baseUriDomains

harness-claude-code: let hosts mark host tools read-only and always-loaded

harness-claude-code: host tool schemas lose their limits and defaults

docs: middleware cache example replays responses across models and providers

@ai-sdk/react: MCPAppRenderer can not load script from CDN

Non-base64 data URLs in file parts throw InvalidCharacterError or send the wrong bytes

@ai-sdk/openai: GPT-5.x has no supportedReasoningEfforts, so unsupported reasoning efforts are sent unchecked and fail with 400

harness-codex: Codex subagent activity never reaches the stream (forward it as raw parts, like harness-claude-code)

Google Vertex: support Google Auth Library 11

@ai-sdk/open-responses: streamText drops the server's response id, model and timestamp, generateText keeps them

amazon-bedrock: named toolChoice removes the other tools for Anthropic models and invalidates the prompt cache