Add Agent class extending AbstractAgent that accepts aisdk, tanstack,
or custom configurations via a discriminated union. Routes to the
appropriate converter, handles RUN_STARTED/FINISHED/ERROR lifecycle,
and supports clone/abortRun. Re-export Agent and converters from the
agent barrel.
Add convertTanStackStream async generator that converts TanStack AI
stream chunks (TEXT_MESSAGE_CONTENT, TOOL_CALL_START/ARGS/END) into
AG-UI events. Create converters/index.ts barrel that re-exports both
AI SDK and TanStack converters.
Extract the AI SDK fullStream event-mapping loop into a standalone
async generator (convertAISDKStream) that yields AG-UI BaseEvent
objects. This is a pure converter with no lifecycle events — the
caller handles RUN_STARTED/FINISHED/ERROR.
Extract shared helpers (DEFAULT_LAYOUT_OPTIONS, getLayoutGrid,
renderTypeAndExpectLayout) to eliminate duplication, add explanatory
comments for unavoidable `as any` casts, and add missing test for
canvas.getContext returning null.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Split ResizeObserver into container vs textarea observers so textarea
height changes (typing, manual resize) don't invalidate the dimension
cache unnecessarily — only grid/button resizes invalidate it
- Invalidate cache on self-triggered resizes (compact↔expanded toggle)
even when skipping re-evaluation, preventing stale dimensions
- Validate fontSize and fontFamily before constructing font fallback
string, avoiding malformed CSS font values in Safari
- Add dev-mode console.warn when font resolution or canvas context
fails, making silent measurement path failures diagnosable
- Update test ResizeObserver mock to support multiple observers
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Test canvas text measurement triggers expansion for long single-line text
- Test compact layout preserved when text fits within cached width
- Test cache invalidation on resize produces correct re-measurement
- Test empty font string is not cached (guard prevents invalid canvas state)
- Mock ResizeObserver to exercise cache invalidation path in jsdom
- Extend mockLayoutMetrics with getComputedStyle mocks for grid and textarea
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Store compactWidth (pre-padding) in cache, compute compactInnerWidth
at read time from live measurementsRef to avoid stale padding values
- Move cache invalidation after ignoreResize check so programmatic
layout transitions don't needlessly destroy the cache
- Guard against empty/invalid font strings before writing to cache
- Fix comment inaccuracies: 2x (not 3x) getComputedStyle, lazily
populated (not on mount), invalidated (not updated) on resize
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
evaluateLayout runs on every keystroke via useLayoutEffect. Previously it
performed 3x getComputedStyle + 2x getBoundingClientRect on every call to
measure grid/button dimensions and the textarea font — none of which change
between keystrokes.
Cache these measurements in a ref (containerCacheRef), populated lazily on
first evaluateLayout call and invalidated when the ResizeObserver fires.
Per-keystroke reflows drop from 3-4 to 1 (the unavoidable adjustTextareaHeight).
Also removes a stale console.log("FOOBAR") debug line.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The { ...existing, ...message } spread meant a later streaming chunk with
toolCalls: undefined would silently wipe accumulated tool calls, contradicting
the comment's claim that "latest toolCalls wins". Apply the same ?? recovery
logic to toolCalls as content already uses for the || fallback.
Add a test for the flip-side edge case: first occurrence has toolCalls, second
has non-empty content but undefined toolCalls — toolCalls must survive.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Kill useJsonStable in CopilotChatConfigurationProvider — useShallowStableRef
is now the single stabilization primitive for both messageView props and labels
- Harden useShallowStableRef with isPlainObject guard so arrays, Dates, and
class instances are never shallow-compared (reference-only for non-plain objects)
- Remove redundant stableParentLabels — parentConfig?.labels is already stabilized
by the parent provider's own useShallowStableRef
- Uninstall ts-deepmerge from react-core dependencies (no longer imported anywhere)
- Add unit tests for useShallowStableRef in slots.test.ts (5 tests)
- Add changeset (patch) describing the FOR-75 performance fix
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Extract useShallowStableRef<T> into slots.tsx alongside shallowEqual —
replaces the 20-line inline stabilization block in CopilotChat with two
readable call sites, and covers suggestionView for free
- Apply useShallowStableRef to both messageView and suggestionView in CopilotChat
- Use post-hoc assignment for messageView to avoid an empty object allocation
on every render when messageView is undefined
- Rewrite Test 2 (labels) to use LabelConsumerMessage which calls
useCopilotChatConfiguration() directly — context consumers re-render when
their context changes regardless of parent memo boundaries, making this a
genuine regression guard for the labels fix (previous version tracked
assistantRenderCount which was already protected by messageView stabilization)
- Replace `as any` on messageView with scoped cast to preserve structural
type checking on the surrounding object
- Fix text/testid collision: CountingAssistantMessage now renders data-testid
instead of a hardcoded string that matched the agent reply text
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the manual JSON.stringify dep-key pattern (which required three
eslint-disable-next-line comments) with a useJsonStable utility that
stabilizes object references using JSON comparison. The dep array for
mergedLabels is now honest — it references the exact variables used inside
the useMemo callback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Inline `messageView` and `labels` objects passed to CopilotChat caused all
completed assistant messages to re-render on every keystroke because:
1. `ts-deepmerge.merge()` deep-clones its inputs, producing a new `messageView`
reference on every render even when content is identical — defeating
`MemoizedSlotWrapper`'s shallow equality check.
2. An inline `labels` object (new reference each render) invalidated the
`mergedLabels` useMemo in `CopilotChatConfigurationProvider`, causing
every `useCopilotChatConfiguration()` consumer to re-render on every keystroke.
3. The inline `onAddFile` arrow function was a new reference each render.
Fixes:
- Replace `merge()` with shallow spread; stabilize `messageView` via a ref +
`shallowEqual` guard so the same object reference is returned when props
are shallowly equal across renders.
- Use `useCallback` for `handleAddFile`.
- Add JSON.stringify-based dep keys in `CopilotChatConfigurationProvider` so
`mergedLabels` is only recomputed when label values actually change.
Adds deterministic render-count regression tests (FOR-75) that fail on the
unfixed code and pass after the fix.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1. consumeAttachments: read from a ref mirror instead of side-effecting
out of a setState updater, avoiding reliance on synchronous updater
execution under React concurrent mode.
2. Codemod: skip declaration positions (variable, function, class, type,
interface) and non-reference positions (object keys, member accesses).
When a local declaration shadows the import name, only rename
unambiguous type-position references to avoid corrupting unrelated code.
Reverts the copy-dts.mjs approach from #3612 in favor of typesVersions
which is the standard Node/TS mechanism for resolving subpath types
under legacy moduleResolution: "node".
feat(runtime): add defaults and max caps for thread lock TTL and heartbeat
lockTtlSeconds defaults to 20s (max 3600s / 1 hour) and
lockHeartbeatIntervalSeconds defaults to 15s (max 3000s / 50 minutes).
Both values are always set, so the heartbeat timer now starts
unconditionally and the paired-option validation is removed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The CopyButton setTimeout fires after the jsdom test environment is
torn down, causing an unhandled ReferenceError on Node 20. Track the
timer in a ref and clear it on cleanup.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Extract dedup logic into exported deduplicateMessages() pure function so
it can be unit-tested directly
- Wrap with useMemo([messages]) to avoid allocating a new Map on every frame
- Collapse three-way branch to two-way (the two arms were identical)
- Use { ...existing, ...message, content } for a true merge so fields present
only in an earlier occurrence are not silently dropped
- Expand comment to explain why || treats empty string as falsy
- Add deduplicateMessages unit tests: toolCalls assertion, reverse scenario,
non-assistant keep-last behavior
- Rename "keeping the last occurrence" test — no longer accurate for assistant
messages which now merge rather than keep-last
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace `ReactActivityMessageRenderer<unknown>` + `(content as any).resourceUri`
with `ReactActivityMessageRenderer<z.infer<typeof MCPAppsActivityContentSchema>>`
so the renderer prop is properly typed and the `as any` cast is eliminated.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
After the useRenderActivityMessage fix, MCPAppsActivityRenderer receives the
per-thread clone instead of the registry agent. The mock's clone() did not share
isRunning state, causing waitForAgentIdle(clone) to never resolve (emit() only
updated registry.isRunning). Also, tests that monkey-patch runAgent before
renderWithCopilotKit need the clone to delegate proxied MCP requests to the
registry (to pick up the monkey-patch) while running user-message flows on the
clone itself (so clone.messages is updated and rendered by CopilotKit).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Fingerprint all tool-call args on last message (not just last tool call)
so parallel/non-last tool-call streaming is detected and triggers re-render
- Add dev warning when scroll container has clientHeight=0 so virtualization
silently disabled (e.g. chat inside display:none tab) is surfaced early
- Use callback ref for non-autoScroll path in ScrollView to eliminate
useLayoutEffect + eslint-disable; autoScroll path keeps useLayoutEffect
since StickToBottom manages that ref internally
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a jsdom-compatible test that confirms the virtual code path activates
above VIRTUALIZE_THRESHOLD without requiring a real browser viewport. Mocks
clientHeight and getBoundingClientRect on the scroll element to pass both
CopilotChatMessageView's guard and TanStack Virtual's observeElementRect.
Asserts the virtual container div exists and rendered count < TOTAL.
Drains pending rAF callbacks before unmount to prevent spurious uncaught
exceptions after jsdom teardown.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a minimal demo at /a2ui-demo that uses the demo-button agent to
reproduce and validate the A2UI thread-clone bug fix. The agent renders
an A2UI surface with a Confirm button on first run; clicking it fires a
second run that emits a text confirmation message.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
scrollToIndex was firing on every streaming chunk (deduplicatedMessages.length
dep), forcibly yanking users to the bottom even when scrolled up to read
history. Removed the dep — use-stick-to-bottom handles streaming auto-scroll
via content height growth on the virtualizer's total-size div, same as the
flat path. scrollToIndex now only fires on virtual mode activation and thread
switches.
Also removed a stale comment referencing toolResultMap (removed two commits
ago) and replaced it with an accurate description of the linear scan.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- All callbacks (processFiles, handleFileUpload, handleDragOver, handleDragLeave,
handleDrop, removeAttachment, consumeAttachments) are now wrapped in useCallback
with empty deps — referentially stable across renders
- Config values read from configRef to avoid dep array changes
- consumeAttachments no-ops on empty queue (returns same state reference)
- 9 new tests: referential stability across re-renders, re-render counting
(consumeAttachments on empty queue triggers zero re-renders), initial state,
consumeAttachments behavior, removeAttachment no-op
useRenderActivityMessage was calling copilotkit.getAgent() directly,
always returning the registry agent. When a user clicked an A2UI button,
handleAction → runAgent executed on the registry agent — messages
accumulated there while CopilotChat displayed from the per-thread clone
(created by useAgent), so responses appeared to silently vanish.
Apply the same getThreadClone(registryAgent, threadId) ?? registryAgent
pattern already used in useRenderCustomMessages.
Adds a regression test that asserts the renderer receives the clone,
and confirms the test catches the bug when the fix is reverted.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Alem Tuzlak