ta
TanStack
GitHub
ai
Workspace
GitHub
CI Pipeline Executions
Filtered
Runs
Demo
Insights
Compare tasks
Analytics
Sign in
Toggle sidebar
Overview
⌘K
ai
Loading workspace stats
Loading workspace insights...
Statistics interval
7 days
30 days
Latest CI Pipeline Executions
Status
Fix filter
Filter
Fuzzy
Filter range
Sort by
Sort by
Start time
Sort ascending
Sort descending
Succeeded
feat/ai-byteplus
5ba6db55 fix(ai-byteplus): review follow-ups — header-case auth clobber, validated poll/response bodies, honoured timeout Three defects where a success path trusted a shape the package's own helpers document as unreliable, plus the coverage that would have caught them. - **Auth headers could be clobbered by casing.** `toHeaderRecord` lowercases names for the `Headers` form of `defaultHeaders`, so `authorization` and `Authorization` both survived the spread as distinct object keys; `fetch`'s `Headers` constructor *appends* rather than replaces, producing `"Bearer wrong, Bearer right"` and 401ing every request with what reads like a bad key — the exact failure the guard's comment claimed to prevent. Reserved names are now dropped case-insensitively on both hosts. - **A malformed video poll response polled forever.** `retrieveTask` cast the 200 body to a task, so an empty or non-JSON body (both documented `readJsonBody` outcomes) became `status: undefined`, which `mapStatus` mapped to `processing` — core then polled to `maxDuration` and threw a generic timeout with the real body discarded. The body is now validated, and an unrecognized status throws naming the value rather than claiming progress. Same cast fixed on the image success path. - **`timeout` was accepted and ignored.** `BytePlusArkConfig` extends the OpenAI `ClientOptions`, and a test asserted `maxRetries` survives, but the fetch-based adapters passed neither to `fetch`. `timeout` is now honoured everywhere via `AbortSignal`; `maxRetries` is documented as chat-path only. Also: keep Ark's code/message on 404s instead of collapsing them to "Job not found"; `describeTaskFailure` always returns a reason; surface unrecognized image `data[]` items and warn on partial group failures; accept a string-typed TTS `code`; distinguish an audio-less TTS success from an envelope error; warn rather than silently drop untimed ASR words/utterances and empty transcripts; make the image reference-count guard's unknown-model skip explicit rather than an accidental `count > undefined`. Tests: +33 (277 total), covering each of the above plus two gaps that ran nowhere — the reasoning `delta` round-trip through `StreamProcessor` (pinning the cross-package short-circuit this adapter depends on) and structured-output gating driven through `chat()` rather than asserted at the hook. Correct the `supportsCombinedToolsAndSchema` docstring, which described a code path `chat({ outputSchema })` cannot reach. Docs: record why `@volcengine/ark-runtime` is deliberately not a dependency; document why byteplus is excluded from the OpenRouter model sync (date-suffixed Ark ids aren't derivable from OpenRouter slugs); note the `size` template divergence when porting a Seedance call between the byteplus and fal adapters. Example: surface file-attach failures in the Seedance Studio instead of swallowing them, and fail loudly when the capability probe returns no resolution tiers rather than falling back to a hardcoded 720p.
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
5ba6db55 Merge 9aed5ae8b4ea16678d5659cb68d79b4f35ce4c81 into 826cfed738ec19e8f416bc5373ad50de0c1824cc
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
c1b00013 feat(examples): Ark playground template presets in the Seedance Studio Replicates the top five BytePlus Ark playground video templates as one-click presets. Each carries the console's official English prompt verbatim, its reference media as hotlinked unsigned URLs from Ark's own presets bucket (nothing downloaded or base64'd), and the console's generation defaults — Seedance 2.0, reference mode, smart length, sound on, 720p, adaptive ratio. Settles the open wire question: Ark accepts the prompts' @Video1 / @Image1 / @Audio1 ordinals as prose alongside ordered role-tagged parts, and honours them. Live-verified through this studio (task cgt-20260801081634-zlhvm, dreamina-seedance-2-0-260128, 720p, two reference videos): the output's first frame reproduces @Video1's framing and its last frame reproduces @Video2's, with generated shots in between. The console's internal <<<video_1_1>>> markers stay unreproduced. Fixes a bug this surfaced: generateSeedanceVideoFn narrowed its prompt with asImagePrompt, so the Seedance path rejected every video and audio part — reference media could never have reached Ark, templates or not. - reference images are capped across template media and uploads together - reference audio alone disables Generate, matching Seedance's own rule - editing the prompt or dropping a chip clears the applied-template highlight - mediaUrlToPart() builds URL-sourced image/video/audio parts
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
c1b00013 Merge 97b73ed9fab106654348efaf08ba31ba0c03f1c8 into 826cfed738ec19e8f416bc5373ad50de0c1824cc
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
b2fbd081 refactor(examples): ts-react-media onto @tanstack/ai-react generation hooks — prompt-part image inputs, SSE server fns, abort propagation; Seedance Studio prompt-first layout
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
b2fbd081 Merge 3f8903145888fd5c7741f010d0ab5df94672acdc into 826cfed738ec19e8f416bc5373ad50de0c1824cc
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
f62662c7 docs(ai-byteplus): point Seedance 2.5 section at the Seedance Studio custom-model field
by Tom Beckenham
T
Succeeded
feat/ai-byteplus
f62662c7 Merge c8a78b76252ffc8b22963a5a3a809966c409c48e into 826cfed738ec19e8f416bc5373ad50de0c1824cc
by Tom Beckenham
T
Previous page
Previous
Next
Next page