Overview - Ai

aiWorkspace

CI Pipeline Executions
- Filtered
Runs
Demo

Insights

Compare tasks
Analytics

Overview

ai

Loading workspace stats

Loading workspace insights...

Statistics interval

7 days30 days

Latest CI Pipeline Executions

Filter

Fuzzy

Sort by

© 2026 - Nx Cloud

Terms of Service Privacy Policy Changelog Status Docs Contact Nx Cloud Pricing Company @NxDevTools

aiWorkspace

Average execution time

8m 23s

345 CI pipeline executions

Computation time saved

13d 10h

Cache hit rate

62%

Time to Green

36m 40s

Was 40m 38s, improved by 10%

Succeeded
feat/evaluator-typesafe
316afba2 refactor: make decide() standalone and rename adapters to *Decider evaluator({ adapter }).decide({ state, questions }) collapses into one decide({ adapter, state, questions }) call, which matches the shape of rerank(). The two-source debug/middleware override goes away with the factory. The four Jev adapter factories follow the new name: typesafeDecider, openRouterDecider, vercelGatewayDecider, and cloudflareDecider, plus their create* variants. None of this has shipped, so there is no migration path to keep. Co-authored-by: Cursor <cursoragent@cursor.com>
by Jack Herri...
Succeeded
feat/evaluator-typesafe
8016a6e5 fix: send ai-gateway-protocol-version on Vercel Gateway evaluate The gateway rejects POST /v4/ai/evaluation-model with 400 'Unsupported gateway protocol version' when the header is absent, so every vercelGatewayEvaluator() decide() call failed. Co-authored-by: Cursor <cursoragent@cursor.com>
by Jack Herri...
Succeeded
main
8de8242b fix(ai-client): fail fast when a passkey unlock lacks user activation (#1332) Preserve Jack Herrington's passkey activation guard while allowing registration follow-up PRF ceremonies. Add storage lifecycle and browser regression coverage. Co-authored-by: Jack Herrington <jherr@pobox.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
main
8de8242b fix(ai-client): fail fast when a passkey unlock lacks user activation (#1332) Preserve Jack Herrington's passkey activation guard while allowing registration follow-up PRF ceremonies. Add storage lifecycle and browser regression coverage. Co-authored-by: Jack Herrington <jherr@pobox.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
1password-byok-fix
4614584c Merge ff5ae8d8fdc289e8cd6280f3119eabd71148c671 into 0a7b3c4393e9d3337637841aa73000d3fdabc58b
by Jack Herri...
Succeeded
1password-byok-fix
f0c13dca Merge 0994603d7803ef6d133b21be1a5ea6e33491f0ae into c9681f7aacd5c938971c056c1636a08589964f87
by Jack Herri...
Succeeded
1password-byok-fix
eb4dfb03 Merge e56ac80c94d63f89caa6c01dbe6e8c9a80061043 into 9b0db21752785e333916aec83c5467d53c3c33f6
by Jack Herri...
Succeeded
1password-byok-fix
21246df0 fix(ai-client): fail fast when a passkey unlock lacks user activation Passkey-encrypted BYOK stores the keyring in IndexedDB and decrypts it with a WebAuthn get() ceremony. Browsers that gate WebAuthn on transient user activation (Safari, Dia) silently suppress that prompt when get() runs outside a user gesture: no UI, and the call never resolves. An unlock buried in an async send pipeline therefore hangs forever. Check navigator.userActivation before the get() and throw a clear, catchable error instead. Callers should trigger unlock (byok.prepare()/unlock()) from the click handler, before awaiting other work. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
1password-byok-fix
21246df0 Merge 73112e1c7fde5adf4ba5abcfd711d503499bd309 into 9b0db21752785e333916aec83c5467d53c3c33f6
by Jack Herri...
Succeeded
main
49642217 feat(gemini): agentic video understanding (GA) (#1294) Add first-class video understanding to the Gemini text adapter. A video content part with `metadata.processing: 'agentic'` routes the turn through the Interactions API for multi-pass agentic understanding (GA on gemini-3.7-flash, gemini-3.6-flash, and gemini-3.5-flash-lite via a new `agentic_video` capability); omit it for the default single-pass generateContent sampling. - Extend GeminiVideoMetadata with processing/fps/startOffset/endOffset and thread fps/offsets into generateContent's videoMetadata. - Route processing:'agentic' video parts through interactions.create, wrapping parts in a user_input step and re-emitting the result as AG-UI chunks. - Add uploadGeminiFile() (Files API upload + poll-until-ACTIVE) and geminiVideoPart() helpers. - Unit test for the routing, an e2e video-understanding feature, docs, a testing/panel demo, and a changeset. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
main
49642217 feat(gemini): agentic video understanding (GA) (#1294) Add first-class video understanding to the Gemini text adapter. A video content part with `metadata.processing: 'agentic'` routes the turn through the Interactions API for multi-pass agentic understanding (GA on gemini-3.7-flash, gemini-3.6-flash, and gemini-3.5-flash-lite via a new `agentic_video` capability); omit it for the default single-pass generateContent sampling. - Extend GeminiVideoMetadata with processing/fps/startOffset/endOffset and thread fps/offsets into generateContent's videoMetadata. - Route processing:'agentic' video parts through interactions.create, wrapping parts in a user_input step and re-emitting the result as AG-UI chunks. - Add uploadGeminiFile() (Files API upload + poll-until-ACTIVE) and geminiVideoPart() helpers. - Unit test for the routing, an e2e video-understanding feature, docs, a testing/panel demo, and a changeset. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
video-understanding
4533e43b feat(gemini): agentic video understanding (GA) Add first-class video understanding to the Gemini text adapter. A video content part with `metadata.processing: 'agentic'` routes the turn through the Interactions API for multi-pass agentic understanding (GA on gemini-3.7-flash, gemini-3.6-flash, and gemini-3.5-flash-lite via a new `agentic_video` capability); omit it for the default single-pass generateContent sampling. - Extend GeminiVideoMetadata with processing/fps/startOffset/endOffset and thread fps/offsets into generateContent's videoMetadata. - Route processing:'agentic' video parts through interactions.create, wrapping parts in a user_input step and re-emitting the result as AG-UI chunks. - Add uploadGeminiFile() (Files API upload + poll-until-ACTIVE) and geminiVideoPart() helpers. - Unit test for the routing, an e2e video-understanding feature, docs, a testing/panel demo, and a changeset. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
video-understanding
4533e43b Merge e5216030baf5560bdb836764d3280931327435e2 into 4852c78900d80547aae64f9a24f95f25d2621a4d
by Jack Herri...
Succeeded
feat/create-chat-hook
7429d77b feat(ui): group createChatHook options and pass layout slots as components (#1282) * feat(ui)!: hand layout and message components, not render thunks `layout` now receives `Messages`, `Interrupts`, and `Input`; `message` receives `Parts`. `renderMessages`, `renderInterrupts`, `renderInput`, and `renderParts` are gone — this API ships first in this PR, so there is nothing released to migrate. React and Solid only. Vue and Svelte compose through slots and snippets and already had empty `LayoutProps`. `Messages` and `Interrupts` were already factory-closure declarations, so passing them directly is strictly more stable than the `useCallback` thunks they replace. `Parts` was the hard part: it closes over the message, so the naive version builds a new component per render and React remounts every part on each stream chunk. It now reads a message-scoped context that `MessageView` provides, which makes it one component for the factory's lifetime. `create-ui-stability.test.tsx` pins that: the same `Parts` identity must survive a re-render with a new message object. `Input` is on the layout props only when the config registers an `input`, so rendering one you never registered is a compile error rather than a silent no-op. This is inferred from a `TInput` parameter on the `input` property; a self-referential config constraint was tried first and silently collapses to `false`. Two inference caveats, both found by making the examples compile: - Referencing the factory's own result inside its config (`UI.useChatContext()`) is circular and blocks inference. The examples and the e2e route now use `createChatHookContexts()`, already the documented cycle-breaker. - An inline `function` expression for `input` also blocks it. Arrows and named references infer fine, so the examples hoist the component and both UI docs call the trap out. The default falls back to a union rather than `undefined` so that when inference does degrade, `input` stays assignable and only `Input` goes missing — a false negative instead of a rejected config. Two doc snippets rendered an input they never registered; the new types caught them. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(ui): warn instead of crashing on an unregistered Input The conditional `Input` type hides the prop when no `input` is registered, but that inference degrades on some config shapes, and the runtime had no component to render in that case — React threw "Element type is invalid" on an undefined element. Always supply a component and warn once instead, matching the existing "Missing tools.<name> component" warning. The floor is now consistent: you find out at build time where inference works, and on the first dev render where it doesn't. The inference trigger turned out to be narrower than the previous commit claimed. It is source order, not the `function` keyword: an inline `function` expression for `input` infers fine as long as it is listed before `layout`. Hoisting the component out of the config bought nothing — the config literal is evaluated once at module scope, so an inline component is already a stable reference — so the examples and the e2e route go back to inline named components, ordered `input` first. Both UI docs now describe the ordering rule rather than telling people to hoist. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * feat(ui)!: split createChatHook options into named component groups `chatComponents` is gone. Registration is now flat on the factory call, one group per kind: createChatHook({ options, context: { chatContext, partContext, interruptContext }, components: { input, layout, message }, toolsComponents: { ... }, interruptsComponents: { tools: { ... }, generic: { ... } }, partsComponents: { ... }, }) `createChatUI` takes the same shape as its second argument, so both entry points register widgets identically. All four adapters (React, Solid, Vue, Svelte) move together. Vue and Svelte keep their flat internal runtime shape; the new config is mapped onto it at the factory boundary, so nothing downstream of that changes. `partsComponents` is plural, unlike the original sketch — the registries are `parts`/`tools`/`interrupts` and the maps hold many entries; `PartProps` and `ChatUIPartKey` stay singular because they describe one part. This also removes the `Input` inference fragility rather than documenting around it. `layout` and `input` are now siblings inside `components`, so TypeScript resolves that inner literal as a unit and fixes `TInput` before contextually typing `layout`. Declaration order no longer matters and an inline `function` expression works, both of which broke inference when these keys sat at the config's top level. The ordering caveat is out of both UI docs, and a type test pins the case that used to fail so a future flattening cannot regress it silently. Per instruction, no migration guidance was added anywhere; the existing `docs/migration/create-ui.md` page covers the older package move only and has been updated to the new shape rather than extended. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: drop migration guidance for the unreleased chat UI API `createChatHook` and `createChatUI` have never been published, so teaching them as something to migrate to was wrong. I checked before cutting: `@tanstack/ai-react-ui@0.8.22` (and Solid 0.7.21, Vue 0.2.40) export only the older `Chat` / `ChatMessages` / `ChatInput` compound components, with no `createChatHook` anywhere. So the page keeps the one migration that is real — swapping `@tanstack/ai-*-ui` for the framework `/ui` subpath, which those published packages do need — and loses the Before/After, Steps, and Gotchas sections that walked through adopting the new factory. The new API is documented in the UI guides as a new API. The `Chat` component's `@deprecated` notices pointed at that page for `createChatHook` guidance, so they now point at the matching UI guide. The `docs/api/*` pages no longer advertise the migration page as reading for the typed chat UI. The `*-ui` shim packages still link there, which is still exactly what that page covers. Also brings this branch in line with repo rules I had been working without, having only picked up the worktree's CLAUDE.md late: - Removed the em dashes the docs skill forbids from text added here. - Bumped `updatedAt` in `docs/config.json` for the five pages with real content changes. The four `docs/api/*` pages only lost a stale link, which the rule classes as a factual fix that must not bump. - Rewrote the `create-chat-hook` changeset to describe the shipping option groups instead of the old `chatComponents` shape, and aligned `typed-headless-chat-ui` with the new names. Since none of this has released, these describe the API rather than a change to it. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * test(e2e): pin the Parts subtree against remounting mid-stream `CLAUDE.md` requires E2E coverage for a behavior change and this stack had none. The behavior worth guarding is the reason `Parts` reads a message-scoped context instead of closing over the message: a `Parts` rebuilt per render makes React remount every part on each stream chunk, throwing away local state in tool and part widgets. The purchase tool widget on `/headless-ui` now stamps a per-mount sequence into `data-mount-seq`, and the spec asserts it stays at 1 across the tool-call chunks and again across the resume stream after approval. Verified the test actually catches the regression rather than just passing: reverting `Parts` to a per-render closure and rebuilding takes the widget to 7 mounts, and the assertion fails on the first check. Full suite: 638 passed, 10 failed. Both failing specs (`multi-turn-structured` across providers, and the flaky `durable-takeover`) fail identically on the PR base `ad5c6c28d` with none of this stack's changes present, verified in a scratch worktree. They run against `examples/ts-react-chat`, which uses `useChat` and never touches the `/ui` surface. `multi-turn-structured` is a fixture-playback problem: turn 3 receives turn 2's recipe. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(ui): add five chat UI recipes The framework guides open with one example that registers tools, interrupts, structured output, input and parts at once. A reader wiring their first chat box has to skip past all of it. These five pages each do one thing, in order: 1. `basic-chat` — layout, message, text part. No tools. 2. `format-a-tool` — one tool, branching on `part.state`. 3. `tool-approval` — `needsApproval`, inline or in the list. 4. `custom-interrupt` — `defineInterrupt` with your own schemas. 5. `request-context` — `forwardedProps` from screen to route. Two gaps these close: there was no starting point smaller than the full example, and nothing documented sending per-request context from the chat UI. `docs/advanced/runtime-context.md` covers only the server half, so the new page shows both and says plainly that `forwardedProps` is client-controlled and must not be trusted for identity. React only. The Solid, Vue and Svelte guides already point at the React page for the full map, and each now links here too. Named `recipes`, not `examples`: `scripts/verify-links.ts` treats any path containing `/examples/` as a directory link for the site's app examples, so `docs/ui/examples/*.md` reported every internal link broken. While writing these I checked the API instead of trusting the existing prose, and corrected two things in my own drafts: the terminal tool state is `complete`, not `output-available`, and `ToolCallPart` has no `errorMessage` field. I also confirmed the grouped `ToolProps` snippet is really type-checked by breaking it on purpose and watching kiira fail. All 1183 doc snippets typecheck. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * ci: apply automated fixes * feat(ui): add Queue layout slot and QueueProps Layout receives a stable Queue component that maps pending sends. Each item has cancelQueued bound. Register the row with components.queue and QueueProps. * ci: apply automated fixes --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com> Co-authored-by: Alem Tuzlak <t.zlak@hotmail.com>
by Jack Herri...
Succeeded
feat/ui-layout-component-slots
37b97757 Merge 3310d683437feb1362d661740120e407a2ed8b20 into ad5c6c28d9d4a27ca8a19d06072d726aa7933b2a
by Jack Herri...
Succeeded
feat/ui-layout-component-slots
6e20c402 Merge f2e46b8bb3aa4ed0415486bae95df21544b15700 into ad5c6c28d9d4a27ca8a19d06072d726aa7933b2a
by Jack Herri...
Succeeded
feat/ui-layout-component-slots
11c22f2b Merge a326703a9aa8fc820ea3c86d87ce42e8b9010524 into ad5c6c28d9d4a27ca8a19d06072d726aa7933b2a
by Jack Herri...
Succeeded
feat/ui-layout-component-slots
ed7ca5df docs(ui): add five chat UI recipes The framework guides open with one example that registers tools, interrupts, structured output, input and parts at once. A reader wiring their first chat box has to skip past all of it. These five pages each do one thing, in order: 1. `basic-chat` — layout, message, text part. No tools. 2. `format-a-tool` — one tool, branching on `part.state`. 3. `tool-approval` — `needsApproval`, inline or in the list. 4. `custom-interrupt` — `defineInterrupt` with your own schemas. 5. `request-context` — `forwardedProps` from screen to route. Two gaps these close: there was no starting point smaller than the full example, and nothing documented sending per-request context from the chat UI. `docs/advanced/runtime-context.md` covers only the server half, so the new page shows both and says plainly that `forwardedProps` is client-controlled and must not be trusted for identity. React only. The Solid, Vue and Svelte guides already point at the React page for the full map, and each now links here too. Named `recipes`, not `examples`: `scripts/verify-links.ts` treats any path containing `/examples/` as a directory link for the site's app examples, so `docs/ui/examples/*.md` reported every internal link broken. While writing these I checked the API instead of trusting the existing prose, and corrected two things in my own drafts: the terminal tool state is `complete`, not `output-available`, and `ToolCallPart` has no `errorMessage` field. I also confirmed the grouped `ToolProps` snippet is really type-checked by breaking it on purpose and watching kiira fail. All 1183 doc snippets typecheck. `pnpm test:ci` and `pnpm test:dts` pass across all 86 projects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
by Jack Herri...
Succeeded
feat/ui-layout-component-slots
ed7ca5df Merge ec9289f6f8b0cf2e8b9334e694e5d37c5bfdb2bc into ad5c6c28d9d4a27ca8a19d06072d726aa7933b2a
by Jack Herri...
Succeeded
main
e04ff6ab feat(ai-compaction): pluggable context-window compaction (#1235) * feat(ai-compaction): add context-window compaction middleware Add @tanstack/ai-compaction — withCompaction() rewrites messages via the chat() onConfig hook before each model call: keeps the recent tail verbatim, replaces the older head with a summary (when a summarize callback is given) or an eviction marker, and preserves tool-call/result pairing. Includes a panel demo (/compaction) and an e2e wire test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: apply automated fixes * docs: add compaction guide Document @tanstack/ai-compaction under Advanced > Middleware: the problem it solves, evict vs summarize wiring, the options table, and what it keeps safe. Add the nav entry and cross-link from the Middleware guide. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(ai-compaction): pluggable compaction strategies Refactor withCompaction around a pluggable CompactionStrategy (mirroring AgentLoopStrategy). Ship three built-in strategies: evictOldest (default), summarizeOldest, and clearToolResults (observation masking for agent loops). Update the docs guide, README, panel demo (strategy selector), and add an e2e case for clearToolResults. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: apply automated fixes * feat(ai-compaction): add composeStrategies combinator composeStrategies runs strategies in order and escalates: it stops once the transcript is back under maxTokens. Lets you clear old tool output first and fall back to evicting old messages only when that isn't enough. Docs, README, and unit tests included. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs: document compaction + persistence interaction and test the seam Server-side `withPersistence` and `withCompaction` share the run's message array. Compaction rewrites it in `onConfig`, and `withPersistence.onFinish` saves that array with a full-overwrite `saveThread`, so the stored thread becomes the compacted one. This was undocumented and untested. - Add a "Compaction and persistence" section to the compaction guide, plus a callout on the chat-persistence page, with the ways to keep a full transcript. - Add a with-persistence unit test that drops a message in `onConfig` and asserts the saved thread is the compacted set. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(ai-persistence): stamp stable ids on persisted messages The chat engine ids assistant messages but leaves incoming user messages, engine-created tool messages, and compaction-injected messages without one. `withPersistence` now fills in an id for any message that lacks one, in place, before each `saveThread`. The same message keeps its id across a run's saves and, when the server owns the thread, across the next turn's reload. This lets a row-keyed persistence adapter reconcile by id (SELECT id, version then delete/insert/update) instead of rewriting the whole transcript. Order and version (content hash) stay the adapter's to own; see the new "Storing messages per row" section in the store reference. - ensureMessageIds() at all four save points (start, streaming snapshot, interrupt boundary, finish) plus the pending-turn seam. - Tests: every persisted message has an id, and earlier ids stay stable across a continuation turn. - Existing verbatim-transcript assertions relaxed to tolerate the added id. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(ai-compaction): preserve history with persisted checkpoints (#1250) * feat: preserve history during compaction * ci: apply automated fixes --------- Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com> * fix(ai-compaction): keep trailing tool turns and skip init * feat(ai-compaction): show compaction stats in AI DevTools * fix(examples): use BYOK on compaction chat and expand DevTools rows * fix(ai-devtools): selecting a hook leaves the dashboard * fix(examples): load compaction chat keys from .env The /compaction page no longer asks you to paste a key. The API reads provider keys from examples/ts-react-chat/.env. Vite loads that file into process.env so server routes can use it. * feat(ai-devtools): add a Compaction tab with before/after previews The AI DevTools panel now has a Compaction tab. Each compact lists when it ran, token and message counts, the budget, dropped messages, and the transcript sent to the model. The /compaction example also shows a small banner from the same compaction:state event. * fix(ai-devtools): render compaction with existing timeline and message UI The Compaction tab now uses the same step row, JsonTree, and User View message cards as the rest of the AI panel. Message previews keep up to 4000 characters so the text wraps instead of clipping. * feat(ai-compaction): emit started, state, and ended events withCompaction now injects compaction:started, compaction:state, and compaction:ended CUSTOM events. The chat client re-emits all three. DevTools shows them as Started, State, and Ended rows. The /compaction example banner listens to the same sequence. * feat(ai): add ctx.emitCustomEvent for live middleware events Middleware can push CUSTOM chunks while a hook is still running. Compaction emits compaction:started before the strategy finishes. * fix(ai-devtools): pad compaction before/after columns Give the dropped vs sent-to-model grid the same left padding it already had on the right. * fix(ai): accept ISO createdAt strings after JSON hydrate reconstructChat returns createdAt as a string. uiMessagesToWire then called Date.toISOString and threw on the next send. * feat(examples): persist compaction chat with SQLite withPersistence saves the full transcript. Compaction still only rewrites providerMessages. Reload keeps every message. * chore: drop createdAt hydrate changeset now that main has the fix * fix: sherif dep order and oxlint prefer-readonly after main merge * test: compaction skips init and keeps trailing tool groups --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com> Co-authored-by: Alem Tuzlak <t.zlak@hotmail.com>
by Jack Herri...