fix(webapp,sdk): address PR review nits

Four follow-up nits from the second-pass review on #3542.

.server-changes/sessions-dashboard-and-task-source-filter.md — adds
the missing high-level entry for the webapp surface (Sessions page +
task-source filter on Runs). The two existing changesets only cover
@trigger.dev/sdk and @trigger.dev/core, so the dashboard work
wouldn't have shown up in a future server changelog.

apps/webapp/app/routes/realtime.v1.sessions.$session.$io.records.ts
— switched `const loader = ...; export { loader }` to
`export const loader = ...` to match the sibling
`api.v1.deployments.current.ts` and the rest of the route file
convention. Functionally identical.

packages/core/src/v3/sessionStreams/manager.ts +
packages/core/src/v3/inputStreams/manager.ts — two clarifications:
(1) added a JSDoc to `disconnect()` documenting that it intentionally
leaves handlers and waiters in place, so any registered listener will
trigger an auto-reconnect with backoff. Distinguishes from `reset()`
(full clean state, rejects waiters) and `disconnectStream` (single
key, stays down until fresh `on()`/`once()`). (2) `disconnectStream`
now clears `reconnectAttempts` for the key — an explicit teardown is
not evidence of a broken backend, and a future re-attach should start
the backoff at attempt 0.

Author: ericallam

.server-changes/sessions-dashboard-and-task-source-filter.md

@@ -0,0 +1,6 @@
+---
+area: webapp
+type: feature
+---
+
+New Sessions page in the dashboard for inspecting `chat.agent` Session rows alongside their underlying runs, plus a "Task source" filter on the Runs list (Standard / Scheduled / Agent) so agent runs can be sliced out of mixed workloads at a glance.

apps/webapp/app/routes/realtime.v1.sessions.$session.$io.records.ts

@@ -36,7 +36,7 @@ const SearchSchema = z.object({
 // `session_*` friendlyIds (which must reference a real row). External-id
 // form falls through with `row: null` so the boot path doesn't 404 on a
 // fresh chat that hasn't written its first chunk yet.
-const loader = createLoaderApiRoute(
+export const loader = createLoaderApiRoute(
   {
     params: ParamsSchema,
     searchParams: SearchSchema,
@@ -96,5 +96,3 @@ const loader = createLoaderApiRoute(
     return json({ records });
   }
 );
-
-export { loader };

packages/core/src/v3/inputStreams/manager.ts

@@ -192,12 +192,25 @@ export class StandardInputStreamManager implements InputStreamManager {
       this.tails.delete(streamId);
     }
     this.buffer.delete(streamId);
+    // Reset the backoff counter so a future re-attach starts fresh —
+    // an explicit disconnect is a deliberate teardown, not evidence of
+    // a broken backend.
+    this.reconnectAttempts.delete(streamId);
   }
 
   connectTail(runId: string, _fromSeq?: number): void {
     // No-op: tails are now created per-stream lazily
   }
 
+  /**
+   * Tear down all active tails. Does NOT clear handlers or `onceWaiters`,
+   * so any registered listener will trigger an auto-reconnect (with
+   * backoff) the moment it sees no live tail — by design, so a transient
+   * network blip recovers without the caller re-subscribing. Use
+   * `reset()` if you want a full clean state with no resurrection, or
+   * `disconnectStream(streamId)` for a single stream that should stay
+   * down until a fresh `on()` / `once()` attaches.
+   */
   disconnect(): void {
     for (const [, tail] of this.tails) {
       tail.abortController.abort();

packages/core/src/v3/sessionStreams/manager.ts

@@ -217,6 +217,10 @@ export class StandardSessionStreamManager implements SessionStreamManager {
       this.tails.delete(key);
     }
     this.buffer.delete(key);
+    // Reset the backoff counter so a future re-attach starts fresh —
+    // an explicit disconnect is a deliberate teardown, not evidence of
+    // a broken backend.
+    this.reconnectAttempts.delete(key);
   }
 
   clearHandlers(): void {
@@ -231,6 +235,15 @@ export class StandardSessionStreamManager implements SessionStreamManager {
     }
   }
 
+  /**
+   * Tear down all active tails. Does NOT clear handlers or `onceWaiters`,
+   * so any registered listener will trigger an auto-reconnect (with
+   * backoff) the moment it sees no live tail — by design, so a transient
+   * network blip recovers without the caller re-subscribing. Use
+   * `reset()` if you want a full clean state with no resurrection, or
+   * `disconnectStream(sessionId, io)` for a single channel that should
+   * stay down until a fresh `on()` / `once()` attaches.
+   */
   disconnect(): void {
     for (const [, tail] of this.tails) {
       tail.abortController.abort();