The Wire · Showcase
PHOTON SETUP FIXED, AI SDK PATCHES REASONING TOKEN EDGE CASE
By RepoJournal · Filed · About Vercel
Eve's Photon integration now correctly detects dedicated projects and skips redundant registration, while the AI SDK patches a token-counting bug hitting reasoning models on Baseten.
Photon setup in Eve [1] now inspects for existing dedicated projects before registering operators, displaying the least-loaded line for dedicated projects and keeping the existing flow for shared users. This lands alongside Kybernesis Arcana's addition to Eve's official extension gallery [2], which brings memory and workspace isolation to managed workspaces via `eve add extension/arcana`. The test suite hardened against scenario observation races [3]. Across the AI SDK, the openai-compatible provider [6] patches a critical token-counting bug: when reasoning models report `completion_tokens_details.reasoning_tokens` greater than `completion_tokens` (observed with Baseten serving reasoning workloads that hit the length stop mid-reasoning), the text token share is now clamped to 0 rather than going negative. The fix rolls forward into Vercel [4], Together AI [5], Moonshot AI [7], and Hugging Face [8] adapters. On the chat desk, Google Chat attachment downloads now send `alt=media` to the media endpoint [9], fixing a bug where the adapter returned metadata instead of file bytes and failed with a bare 400. Workflow adds opt-in WebSocket transport for event writes [10], defaulting to HTTP and gated behind `WORKFLOW_EVENTS_TRANSPORT=ws` until confidence builds.
One email a day. Unsubscribe in one click.
Keep up with Vercel in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
One email a day. Unsubscribe in one click. Read a past issue →
Action items
- → Update @ai-sdk/openai-compatible and dependent adapters to 3.0.28 before deploying reasoning model integrations vercel/ai [immediate]
- → Pull Eve to pick up Photon setup fix and Kybernesis Arcana extension vercel/eve [plan]
- → Update vercel/chat to fix Google Chat attachment download failures vercel/chat [plan]
- → Test WebSocket event transport in staging before enabling WORKFLOW_EVENTS_TRANSPORT=ws in production vercel/workflow [monitor]
References
- [1] fix(setup): show dedicated Photon numbers ↗ vercel/eve
- [2] feat(docs): add Kybernesis Arcana memory extension ↗ vercel/eve
- [3] test(eve): remove scenario observation races (#1878) vercel/eve
- [4] @ai-sdk/vercel@3.0.28 ↗ vercel/ai
- [5] @ai-sdk/togetherai@3.0.29 ↗ vercel/ai
- [6] @ai-sdk/openai-compatible@3.0.28 ↗ vercel/ai
- [7] @ai-sdk/moonshotai@3.0.31 ↗ vercel/ai
- [8] @ai-sdk/huggingface@2.0.28 ↗ vercel/ai
- [9] fix(gchat): send alt=media when downloading attachments by resourceName ↗ vercel/chat
- [10] Add WebSocket transport for step-execution event writes (opt-in) ↗ vercel/workflow