[garlicKim21]

Summary

Fixes #3197

When the chat stream (A2AService/SendStreamingMessage over gRPC-Web) breaks while the task is still running, the UI reports the turn as failed ("The agent could not finish this turn / network error"), even though the task keeps running on the server and completes, with its result in the task history.

This is easy to hit behind a proxy that closes quiet streams. A model reading a long prompt sends nothing for a minute or more, and an idle timeout somewhere on the path (an ingress, a reverse proxy, a CDN) ends the response. In our case the controller logged both streams as Canceled after exactly 60,015 ms. Both tasks finished one to two minutes after the cut, by which time the UI had already shown the error banner.

This change keeps following the task instead of failing the turn:

• Once the stream has named its task (the first statusUpdate carries taskId), a transport failure re-attaches with SubscribeToTask. The gateway already implements it (go/core/internal/a2agateway/interactions.go). It is read-only and returns the stored task when the task has already finished, so the user's message is never sent twice.
• Only transport failures are retried: Unknown (a bare network error, or a body that ends without its trailer) and Unavailable (HTTP 502/503/504). Errors the controller decided, a broken stream before any task id is known, a task that has already reached a terminal or input-required state, and a user abort behave as before.
• Re-attaching gives up after three consecutive attempts that came to nothing, so a dead or flapping path still ends the turn. A re-attached stream opens with the task snapshot, so the count resets only on a frame after it, or when a stream stayed open for 10 s, long enough to have been cut for being idle. Attempts wait 0.5 s, 1 s and 2 s.

The change is confined to A2AGrpcChatClient.send. A task snapshot returned by the subscription goes through the existing task branch, which now records the artifacts it shows, so the next append chunk continues them, and shows artifacts again from every snapshot, so a later snapshot fills in chunks sent while the stream was down.

Testing

• New unit tests in ui/src/api/chat/a2aGrpcChatClient.test.ts:
  • a stream that breaks after naming its task re-attaches once and ends with the completed task and its answer
  • a stream that breaks before naming a task does not re-attach
  • an error with a server-decided code (FailedPrecondition) still surfaces and does not re-attach
  • repeated breaks stop after three re-attach attempts
  • re-attaches that break right after the task snapshot stop after three attempts
  • re-attached streams cut after 60 s of quiet are followed to the end
  • a re-attach that delivers a frame after the snapshot resets the count
  • pressing stop while waiting to re-attach ends the turn without re-attaching
  • a reply first seen in the snapshot continues with the next append chunk instead of being replaced by it
  • a later snapshot fills in chunks a reply missed while the stream was down
• The first and fourth tests fail without the change and pass with it. The fifth fails on e804f21, this PR's first commit, and passes on d3c7100. The ninth fails on d3c7100 and the tenth on 3da7a38; both pass on d98835c.
• yarn typecheck: passes.
• yarn lint: 0 errors; the same 6 existing warnings as on main.
• yarn test: 59 files, 604 tests pass.
• Run in node:24.13.0 (ui/.nvmrc) with Corepack and yarn install --immutable, as in the ui-tests CI job. The repository's Playwright suites were not run locally; the live check below drove a browser separately.

Live check

Against a kagent 1.0.0-alpha13 controller, with this branch's UI in Vite live mode (between alpha13 and this branch's base, neither the chat client nor the A2A protos changed). Playwright drove Chromium through a test proxy placed in front of the browser, which cut the streams. The agent used an Ollama model and took 50 to 85 s per answer.

Proxy behaviour	Before d3c7100	After
SendStreamingMessage cut at 3 s, each re-attached stream cut 0.2 s after its first frame	re-attached 93 times in 47 s, until the task finished	gave up after 3 re-attaches, about 11 s after sending, and showed the error; the finished answer appeared on reload
SendStreamingMessage cut at 3 s, each re-attached stream carries only the snapshot and is cut at 12 s	not run	followed through 4 cuts, showed the full answer, no error
SendStreamingMessage cut at 3 s, re-attached stream passed through	not run	re-attached once and streamed the answer to the end

[garlicKim21]

Right, it wasn't bounded. SubscribeToTask opens with the task snapshot, so a path that broke right after that frame reset the count on every attempt. Behind a test proxy that cut each re-attached stream 0.2 s after its first frame, one turn re-attached 93 times in 47 s, until the task finished.

d3c7100 resets the count only on a frame after the snapshot, or when a stream stayed open for 10 s (long enough to have been cut for being idle), and waits 0.5 s, 1 s and 2 s between attempts. With the same proxy the turn now gives up after three re-attaches. Re-attached streams cut every 12 s while the model is quiet are still followed to the end. The new tests in a2aGrpcChatClient.test.ts cover both cases, and I added the live runs to the description.

[Charlesthebird]

Before/after on a fresh kind cluster. A proxy cuts the chat stream after 12 s of silence while a BYO agent stays quiet for 30 s.

• Chat turn: the turn re-attaches to its task and shows the answer. On main it fails with "network error" and the answer only appears after a reload.

reel-pr3198.mp4

Screenshots

Before: turn fails when the stream is cut

Before: answer only after reload

After: turn follows the task and shows the answer

---

🤖 written by Claude

[Charlesthebird]

This automated review is a first pass. A manual review will follow once the blockers are addressed.

---

Nothing blocks merge. One Medium item is inline: a snapshot artifact can be truncated after a re-attach.

✅ What I checked

• ✅ Retried codes: connect-web maps a missing trailer and fetch TypeError to Unknown, and 502/503/504 to Unavailable; a2a-go defaults server errors to Internal.
• ✅ SubscribeToTask returns the stored task when quiescent (interactions.go:124), so a re-attach never resends the message.
• ✅ input_required and terminal states set settled, so they surface instead of re-attaching.
• ✅ pause clears its timer and listener on both paths; an abort mid-wait or mid-subscribe reaches rethrowIfAborted.
• ✅ Snapshot messages dedupe via delivered; snapshot artifacts reuse their artifactId.
• ✅ New tests use fake timers, including Date.now for the 10 s rule, and pass locally.

---

🤖 written by Claude