packages/typescript/ai-solid/src/use-chat.ts (1)

102-107: ⚠️ Potential issue | 🟠 Major

Remove the eager subscribe/unsubscribe from setup.

This block duplicates the logic in the createEffect below (lines 109-115) which already applies the initial live state on mount. Executing connection side effects during hook setup—outside of Solid's lifecycle system—violates the framework's composable patterns and can cause issues with SSR or strict mode.

🔧 Proposed fix

-  // Apply initial live mode immediately on hook creation.
-  if (options.live) {
-    client().subscribe()
-  } else {
-    client().unsubscribe()
-  }
-
   createEffect(() => {
     if (options.live) {
       client().subscribe()

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-solid/src/use-chat.ts` around lines 102 - 107, Remove
the eager subscribe/unsubscribe call in the hook setup: delete the if
(options.live) { client().subscribe() } else { client().unsubscribe() } block so
connection side-effects are only applied inside the existing createEffect (which
reads options.live and calls client().subscribe()/unsubscribe()); this ensures
client() connection logic runs inside Solid's lifecycle rather than during hook
initialization.

packages/typescript/ai-client/tests/chat-client.test.ts (1)

293-297: ⚠️ Potential issue | 🟡 Minor

Same lint blocker: throw-only async generator mocks.

Line 295 and Line 1280 still define async generators that throw before any yield, so Biome/ESLint will keep failing this file (useYield / require-yield). Return an async iterator object instead, or add a dummy yield before the throw in both tests.

Also applies to: 1280-1282

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-client/tests/chat-client.test.ts` around lines 293 -
297, The async generator mocks throw before yielding (e.g., the
connection.subscribe async function* in the "should expose connectionStatus
error for subscription loop failures" test and the similar async generator at
1280-1282), which trips the require-yield/useYield lint rules; fix by replacing
those async function* mocks with an async iterator object that rejects from
next() (i.e., return { async next() { return Promise.reject(new Error('...')) },
async return() {...}, async throw() {...} } ) or simply add a dummy yield before
the throw so the generator yields once before throwing; update the mock for
connection.subscribe and the other async generator accordingly.

packages/typescript/ai-react/src/types.ts (1)

6-12: Consider re-exporting ConnectionStatus for consumer convenience.

ConnectionStatus is imported and used in UseChatReturn but not included in the re-exports on line 12. Consumers who need to type their own state or callbacks based on this type would need to import directly from @tanstack/ai-client.

💡 Suggested change

 // Re-export types from ai-client
-export type { ChatRequestBody, MultimodalContent, UIMessage }
+export type { ChatRequestBody, ConnectionStatus, MultimodalContent, UIMessage }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-react/src/types.ts` around lines 6 - 12, The file
imports ConnectionStatus but doesn't re-export it, so add ConnectionStatus to
the exported types to make it available to consumers; update the export
statement in types.ts (the export type line that currently exports
ChatRequestBody, MultimodalContent, UIMessage) to also include ConnectionStatus
so types like UseChatReturn can be fully typed by library users.

packages/typescript/ai-react/src/use-chat.ts (1)

115-122: Guard against unnecessary initial unsubscribe() call to avoid triggering state callbacks on mount.

When options.live is initially undefined or false, the else branch calls client.unsubscribe() on mount. This triggers setIsSubscribed(false) and setConnectionStatus('disconnected') callbacks even though the client is already in that state, causing unnecessary state change notifications.

💡 Suggested change

   useEffect(() => {
     if (options.live) {
       client.subscribe()
-    } else {
+    } else if (options.live === false) {
+      // Only explicitly unsubscribe when live transitions from true to false,
+      // not on initial mount when live is undefined
       client.unsubscribe()
     }
   }, [client, options.live])

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-react/src/use-chat.ts` around lines 115 - 122, The
effect calling client.subscribe()/client.unsubscribe() should avoid calling
unsubscribe on mount when the client is already disconnected; update the
useEffect (the block referencing useEffect, options.live, client.subscribe,
client.unsubscribe) to only call client.unsubscribe() when there was a previous
active subscription (e.g., track previous options.live with a ref like
prevLiveRef or check a client.isSubscribed flag) so that on initial mount with
options.live undefined/false you do nothing, but on transitions true→false you
call client.unsubscribe() and trigger setIsSubscribed/setConnectionStatus only
when state actually changes.

packages/typescript/ai-preact/src/types.ts (1)

30-40: Sync the UseChatOptions JSDoc with the new omitted lifecycle callbacks.

The Omit list now excludes onSubscriptionChange, onConnectionStatusChange, and onSessionGeneratingChange, but the comment block still documents only the original four callbacks. Keeping this list complete avoids API confusion.

📝 Suggested doc patch

  * - `onStatusChange` - Managed by Preact state (exposed as `status`)
+ * - `onSubscriptionChange` - Managed by Preact state (exposed as `isSubscribed`)
+ * - `onConnectionStatusChange` - Managed by Preact state (exposed as `connectionStatus`)
+ * - `onSessionGeneratingChange` - Managed by Preact state (exposed as `sessionGenerating`)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-preact/src/types.ts` around lines 30 - 40, Update the
JSDoc for the UseChatOptions type to match the current Omit list: include
documentation entries for the newly omitted lifecycle callbacks
onSubscriptionChange, onConnectionStatusChange, and onSessionGeneratingChange in
addition to the existing onMessagesChange, onLoadingChange, onErrorChange, and
onStatusChange; locate the JSDoc immediately above the UseChatOptions
declaration and add brief descriptions for each callback so the comment
accurately reflects the type's omitted properties.

packages/typescript/ai-svelte/src/create-generate-image.svelte.ts (1)

69-69: Minor: Inconsistent dash character in JSDoc.

This file uses -- (double hyphen) while other files in this PR (e.g., use-generate-image.ts in React/Vue) use — (em dash) for the same transport mode description pattern.

🔧 Suggested fix for consistency

-* - **ConnectConnectionAdapter** -- Streaming transport (SSE, HTTP stream, custom)
+* - **ConnectConnectionAdapter** — Streaming transport (SSE, HTTP stream, custom)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-svelte/src/create-generate-image.svelte.ts` at line
69, Update the JSDoc line in create-generate-image.svelte.ts to use an em dash
for consistency: replace the double hyphen in the description for
ConnectConnectionAdapter with an em dash (match the style used in
use-generate-image.ts and other files) so the line reads
"ConnectConnectionAdapter — Streaming transport (SSE, HTTP stream, custom)";
locate the JSDoc that references ConnectConnectionAdapter to make this
single-character change.

packages/typescript/ai/tests/stream-processor.test.ts (1)

2209-2244: Add one regression test for tool-call routing after pending ID reconciliation.

Current dedup tests validate text routing after startAssistantMessage() ID swap, but not TOOL_CALL_START (before swap) followed by TOOL_CALL_ARGS/END (after swap). Adding that case would lock in the map-rewrite behavior.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai/tests/stream-processor.test.ts` around lines 2209 -
2244, Add a regression test in the same suite that mirrors the existing
TEXT_MESSAGE_START dedup case but for tool calls: instantiate StreamProcessor,
call startAssistantMessage() to create the pending manual tool-call message,
then process a TOOL_CALL_START chunk with a different messageId (e.g.,
'server-tool-1') and tool name, then send TOOL_CALL_ARGS and TOOL_CALL_END
chunks using the server messageId, finalize the stream, and assert there is only
one message whose id was updated to the server id and whose parts include the
tool-call args and correct tool metadata; use the StreamProcessor methods
startAssistantMessage, processChunk, getMessages, and finalizeStream and chunk
types TOOL_CALL_START, TOOL_CALL_ARGS, TOOL_CALL_END to locate where to add the
test.

packages/typescript/ai-svelte/src/create-chat.svelte.ts (1)

98-100: ⚠️ Potential issue | 🟠 Major

Expose unsubscribe() for live subscription teardown.

When live: true, client.subscribe() starts a durable subscription, but the returned interface only provides stop(). Per chat-client.ts, stop() cancels request-local work while unsubscribe() tears down the live subscription. Svelte consumers have no way to release the connection on component destroy.

🔧 Proposed fix

In create-chat.svelte.ts:

  const stop = () => {
    client.stop()
  }

+ const unsubscribe = () => {
+   client.unsubscribe()
+ }

  // In return object:
  return {
    // ... existing properties
    stop,
+   unsubscribe,
    // ...
  }

In types.ts, add to CreateChatReturn:

  stop: () => void
+ /**
+  * Unsubscribe from the live subscription loop.
+  * Call this on component destroy when using `live: true`.
+  */
+ unsubscribe: () => void

Also applies to: 102-105, 177-186

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai-svelte/src/create-chat.svelte.ts` around lines 98 -
100, When options.live is true, client.subscribe() starts a durable live
subscription but the returned object only exposes stop(), so callers cannot
fully tear down the live connection; update create-chat.svelte.ts to capture the
subscription result from client.subscribe() and expose its unsubscribe() method
alongside stop(), and update the CreateChatReturn type in types.ts to include
unsubscribe(): void (or nullable) so Svelte components can call unsubscribe() on
destroy; ensure you also add unsubscribe to the other similar branches mentioned
(lines ~102-105 and ~177-186) where client.subscribe() is used.

packages/typescript/ai/src/activities/chat/stream/processor.ts (1)

1376-1391: Content/thinking aggregation across messages may yield unexpected results.

getResult() concatenates totalTextContent and thinkingContent from all message states. For multi-message sessions, this produces a single aggregated string without message boundaries. Consider if this is the intended behavior, or if consumers expect per-message results.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai/src/activities/chat/stream/processor.ts` around lines
1376 - 1391, getResult() currently flattens all messageStates by concatenating
each state's totalTextContent and thinkingContent into single strings, which
loses per-message boundaries; change getResult() (and its return shape) to
preserve per-message results by returning a map or array keyed by message id
(e.g., messageStates.forEach -> build perMessageContents: Record<messageId,
{content, thinking}>) and keep the existing aggregated content only if strictly
needed (or add it as an optional field) so callers can access both per-message
and aggregated data; update any consumers of getResult() to read
perMessageContents (or handle the added optional aggregated field) and reference
messageStates, totalTextContent, and thinkingContent when making the change.

packages/typescript/ai/tests/stream-processor.test.ts (1)

2470-2508: Test uses raw chunk without runId - verify this is intentional for backward compat.

This test sends RUN_FINISHED without runId, which tests backward compatibility for streams that don't include run tracking. The test description could clarify this is testing backward compat behavior.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/typescript/ai/tests/stream-processor.test.ts` around lines 2470 -
2508, Update the test description for the "double onStreamEnd guard" case to
explicitly state it's intentionally sending a RUN_FINISHED chunk without a runId
to verify backward-compatibility for streams that don't include run tracking;
locate the test using StreamProcessor and processChunk (the it block that sends
RUN_FINISHED then TEXT_MESSAGE_END) and change the it string to mention
"backward-compat: RUN_FINISHED without runId" so future maintainers understand
this is intentional.