- Call closeReasoningIfOpen() once before the switch instead of in every case branch
- Remove as-casts from closeReasoningIfOpen by using typed const variables
- Remove as-casts from new test code; use typed imports and unknown-safe casts
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add cardinality assertions (toHaveLength(1)) to all 5 auto-close tests
so duplicate REASONING_MESSAGE_END / REASONING_END events are caught
- Add missing RUN_FINISHED final-event assertion to abort auto-close test
- Add test for consecutive reasoning blocks with no reasoning-end between
them (exercises closeReasoningIfOpen inside reasoning-start)
- Add test for exception thrown mid-stream (exercises catch block path)
- Add test for stream exhaustion without terminal event while reasoning is
open (exercises !terminalEventEmitted fallback)
- Expand changeset description to cover all three behavioral fixes
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- reasoning-end case now delegates to closeReasoningIfOpen() so a late
SDK-emitted reasoning-end is a no-op when auto-close already fired,
preventing duplicate REASONING_MESSAGE_END + REASONING_END events
- Hoist isInReasoning / reasoningMessageId / closeReasoningIfOpen before
the try block so the catch block can call closeReasoningIfOpen() --
exceptions thrown mid-reasoning now emit proper lifecycle events before
RUN_ERROR instead of leaking an unclosed reasoning state
- Remove dead (part as any).delta fallback in reasoning-delta case; the
SDK always provides part.text in this branch
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When closeReasoningIfOpen() auto-closes a prior block and a second
reasoning-start arrives with id="0" or no id, the handler was silently
reusing the previous reasoningMessageId. Mirror the text-start pattern:
always call randomUUID() when the provider id is falsy or "0".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Call closeReasoningIfOpen() in the reasoning-start handler before opening
a new block, guarding against consecutive reasoning blocks without an
intervening reasoning-end (e.g. Anthropic extended thinking with tools)
- Call closeReasoningIfOpen() in the !terminalEventEmitted fallback block
so streams that end without a finish/abort/error event still close properly
- Add abort() test-helper and tests for abort + error cases with open reasoning
- Strengthen auto-close ordering assertions to explicitly verify
REASONING_MESSAGE_END precedes REASONING_END in all three auto-close tests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two fixes for Anthropic model stalls (fixes#3323):
1. Skip reasoning-delta events with empty text — EventSchemas.parse()
rejects delta: "" and kills the RxJS Observable before any other
events can fire.
2. Auto-close open reasoning lifecycle at every phase transition —
@ai-sdk/anthropic never emits reasoning-end, leaving the state
machine stuck.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Incorporates the V1/V2 flat package consolidation (96885b595). All our
thread-isolation changes auto-merged to the new package paths. Three
content conflicts resolved (import paths updated from @/ aliases to
relative paths and @copilotkitnext/* to @copilotkit/*):
- packages/react-core/src/v2/hooks/use-agent.tsx
- packages/react-core/src/v2/hooks/use-render-custom-messages.tsx
- packages/react-core/src/v2/components/chat/CopilotChatMessageView.tsx
File location conflict resolved: use-agent-thread-isolation.test.tsx
moved from packages/v2/react/... to packages/react-core/src/v2/...
Two bugs found in CR:
1. unsubscribeFromAgent() was a no-op: subscribeToAgent() always used the
composite agentId:threadId key (since agent.threadId defaults to a UUID),
but unsubscribeFromAgent() looked for the bare agentId key, which was
never written. Registry agents subscribed via onAgentsChanged were never
cleaned up on replacement. Fix: add isClone flag to subscribeToAgent() —
registry agents (isClone=false, default) use bare key; per-thread clones
(isClone=true, via subscribeAgentToStateManager) use composite key.
2. Error-state provisional agent was not cached in provisionalAgentCache,
unlike the Disconnected/Connecting path. Any dependency change while in
Error state produced a new agent reference, triggering spurious
re-subscriptions. Fix: cache the Error-state provisional using the same
cacheKey and headers-refresh pattern as the other provisional paths.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
With per-thread agent clones, the registry agent has empty messages.
reloadSuggestions was checking messageCount on the registry agent, so
dynamic suggestions (which require messageCount > 0) never fired after
the first message. Fix by passing the thread clone (with actual messages)
through reloadSuggestions → generateSuggestions.
Also updates the demo to use claude-3-7-sonnet (extended thinking support)
when ANTHROPIC_API_KEY is set.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add afterEach with vi.clearAllTimers() to prevent the 300ms setTimeout in
WebInspectorElement (src/index.ts:2065) from firing after the jsdom environment
tears down, which caused intermittent "document is not defined" errors in CI.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- StateManager now uses composite `agentId:threadId` subscription keys so
per-thread clones get independent subscriptions without stomping the
registry agent's subscription.
- Added `activeRun` map so `handleNewMessage` (called without `input` by
ag-ui's addMessage path) can still associate messages with the correct run.
- Added `revoked` flag + `subRunId`/`runFinished` run-isolation logic to
handle the timing race where a new run's events arrive through the previous
pipeline before `subscribeAgentToStateManager` has set up a new one —
detected via `runFinished && input.runId === subRunId`, resolved by minting
a fresh UUID so both runs get distinct state entries.
- Empty state `{}` is no longer persisted; `getStateByRun` returns `undefined`
so renderers can distinguish "no snapshot received yet" from a real state.
- `useAgent`: per-thread clones are shared via a module-level WeakMap so all
components on the same (agentId, threadId) see the same instance.
- Tests updated to cover sequential-run state isolation and the empty-state
semantics change.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
After the V1/V2 consolidation merge, several changesets still referenced
old @copilotkitnext/* package names that no longer exist. Fixed package
references and ran changeset version to apply pending version bumps.
https://claude.ai/code/session_01A5aahBWFTmLPR3xZSeh4SD
Flatten all packages from packages/v1/* and packages/v2/* into packages/* —
every package now lives directly under the @copilotkit/ scope with no v1/v2
subdirectories.
- Move all v1 packages (react-core, react-ui, runtime, shared, etc.) from
packages/v1/* to packages/*
- Absorb v2 react code into packages/react-core/src/v2/ (exported via /v2 subpath)
- Absorb v2 agent code into packages/runtime/src/agent/ (exported via /v2 subpath)
- Move v2 packages (core, angular, demo-agents, etc.) to packages/*
- Replace all @copilotkitnext/* imports with @copilotkit/* equivalents
- Keep @copilotkitnext/angular as the sole exception (angular remains on next)
- Update CI workflows, renovate config, release scripts for flat structure
- No public API surface changes — all exports fields are preserved
Co-authored-by: Alem Tuzlak <t.zlak@hotmail.com>
Signed-off-by: Tyler Slaton <tyler@copilotkit.ai>
- Fix 1: store {source, clone} in threadAgentCache to invalidate stale
clones when registry agent is replaced (reconnect/hot-reload)
- Fix 2: add lastConnectedAgentRef in V1 wrapper to prevent connectAgent
firing on every Connected render; ref reset in cleanup for StrictMode
- Fix 3: throw if clone() returns same instance to prevent silent state
destruction via setMessages([]) / setState({}) on the source agent
- Fix 4: clear bare agentId key from provisionalAgentCache on threadId
transition to eliminate phantom provisionals
- Fix 5: clear threadAgentCache on unmount to release clones holding
listeners, WebSocket connections, or timers
- Fix 6: move HttpAgent headers sync out of useMemo into useEffect to
avoid side effects in concurrent-mode pure computation
Tests: T-1 registry replacement, T-2 threadId switching, T-3 provisional
→ real transition, T-4 V1 threadId passthrough, T-5 non-vacuous isolation
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
## Summary
- `{ ...defaultComponents, ...components }` in `Markdown.tsx` created a
new object reference on every render
- `MemoizedReactMarkdown` has a custom memo comparator checking
`prevProps.components === nextProps.components` — referential equality
always failed, so every prior message re-rendered on every streaming
chunk
- Fix: wrap the merge in `useMemo([components])` so the reference is
stable
Fixes CPK-7156 / closes#2543
## Test plan
- [x] Verified locally with the form-filling example: completed messages
no longer log new renders while a new message streams
- [x] `nx build @copilotkit/react-ui` passes
- [x] `nx test @copilotkit/react-ui` passes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
{ ...defaultComponents, ...components } created a new object reference on
every render, causing MemoizedReactMarkdown's custom comparator to always
fail and re-render every prior message on each new streaming chunk.
Wrapping the merge in useMemo([components]) gives a stable reference so
the memo comparator works as intended.
Fixes CPK-7156
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat(runtime-v2): wire createLicenseChecker into CopilotRuntime
feat(runtime-v2): wire license checks into route handlers
feat(runtime-v1): pass licenseToken through to v2 runtime
When resolve() is called, the interrupt event value is now included
alongside the user's response in forwardedProps.command.interruptEvent.
This enables agent adapters (e.g. Mastra) to round-trip interrupt
context (toolCallId, runId, type) without storing state between runs.
Uses a ref to avoid changing the resolve callback's dependency array,
preventing double-invocation of the handler effect.
The LangGraph adapter is unaffected — it ignores the extra field.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Multiple CopilotChat components with the same agentId but different
threadIds were sharing a single agent instance, causing message state
to bleed between them. The useAgent hook now creates per-thread clones
keyed by agentId:threadId, keeping messages and state isolated.
Closes#2957
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Summary
`<CopilotSidebar defaultOpen={false} />` was opening anyway because the
prop was not being forwarded through the provider chain. This PR fixes
that - and also addresses the architectural trade-off flagged in review.
### Who is affected
**User A** just uses `<CopilotSidebar />` with no outer provider
Before the fix, their sidebar opened by default. After the fix, it still
opens by default. Nothing changed for them.
**User B** uses `<CopilotSidebar defaultOpen={false} />`
Before: broken (opened anyway). After: works.
**User C** wraps their app in a bare `CopilotChatConfigurationProvider`
and reads `isModalOpen` from it.
Before our fix `isModalOpen` was stuck at true forever. After: it
reflects the sidebar's state.
### What changed
- The `defaultOpen` prop is forwarded correctly through the provider
chain to both `CopilotSidebar` and `CopilotPopup`.
- Outer `useCopilotChatConfiguration()` hooks now reflect sidebar state.
Fixed via bidirectional state sync between nested providers.
### Files changed
- `packages/v2/react/src/components/chat/CopilotChat.tsx` - accept and
forward `isModalDefaultOpen`
- `packages/v2/react/src/components/chat/CopilotSidebar.tsx` - pass
`isModalDefaultOpen={defaultOpen}`
- `packages/v2/react/src/components/chat/CopilotPopup.tsx` - same
- `packages/v2/react/src/providers/CopilotChatConfigurationProvider.tsx`
- fix resolution logic + bidirectional sync
-
`packages/v2/react/src/providers/__tests__/CopilotChatConfigurationProvider.test.tsx`
- add tests
### Test plan
- [x] 843 tests pass (including 4 new bidirectional sync tests)
- [x] New tests cover all three reproduction scenarios from
[tkt-modal-default-open](https://github.com/CopilotKit/deep-agent-cpk-experiments/tree/main/app/client/src/tickets/tkt-modal-default-open)
**Linear:**
[CPK-7152](https://linear.app/copilotkit/issue/CPK-7152/fix-defaultopenfalse-prop-ignored-in-copilotsidebar)
**GitHub Issue:** #3475
**Related PRs:** #3477, #3496 (stalled community PRs addressing the same
issue)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
PR #3513 fixed Behavior A (defaultOpen={false} now respected) but broke
Behavior B: users who read isModalOpen from an outer
CopilotChatConfigurationProvider no longer see state changes originating
in the inner provider (e.g. CopilotSidebarView's provider).
Root cause: with hasExplicitDefault=true, the inner provider used its own
setter exclusively, so open/close events never reached ancestor providers.
Fix: when a provider owns its state (hasExplicitDefault=true), wrap
setModalOpen to also call parentConfig.setModalOpen — propagating changes
upward through the provider chain. Add a useEffect to sync parent→child
when an ancestor's state is changed externally (e.g. via an outer hook).
The mount-skip guard (isMounted ref) ensures the child's own
isModalDefaultOpen is honoured on first render and not overwritten.
Three new tests mirror the reproduction scenarios from:
https://github.com/CopilotKit/deep-agent-cpk-experiments/tree/main/app/client/src/tickets/tkt-modal-default-open
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@ag-ui/client 0.0.47 had a bug where connectAgent() hung forever
for agents that don't implement connect(), causing detachActiveRun()
and runAgent() to deadlock. Fixed in 0.0.48.
## Summary
Fixes a race condition where `detachActiveRun()` was called without
`await`, allowing a new run to overwrite pipeline state before cleanup
finished — silently dropping the previous run.
- **`run-handler.ts`**: `void agent.detachActiveRun()` → `await
agent.detachActiveRun()`. The original deadlock that motivated
fire-and-forget (ConnectNotImplementedError cleanup) has been resolved
in the current `@ag-ui/client`. Confirmed via git history: added as
fire-and-forget in `5b3037a` on Mar 13 by Max Korp.
- **`use-copilot-chat_internal.ts`**: Added AbortController cleanup in
the `connectAgent` useEffect, mirroring the V2 `CopilotChat.tsx`
pattern. Guards state updates against aborted signals.
- Updated test mocks to include `detachActiveRun`
**Linear:**
[CPK-7151](https://linear.app/copilotkit/issue/CPK-7151/fix-race-condition-in-runhandlerrunagent-causes-dropped-runs)
**GitHub Issue:** #3482
## Files changed
- `packages/v2/core/src/core/run-handler.ts`
- `packages/v1/react-core/src/hooks/use-copilot-chat_internal.ts`
-
`packages/v1/react-core/src/hooks/__tests__/use-copilot-chat-internal-connect.test.tsx`
-
`packages/v1/react-core/src/hooks/__tests__/use-coagent-state-render.e2e.test.tsx`
## Test plan
- [x] All tests pass (pre-commit hook ran full suite)
- [x] Manual: rapid sequential message sends don't drop runs
- [x] Manual: verify no deadlock on initial connectAgent → immediate
runAgent
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Martha Kelly Schumann