Run details

aiWorkspace

CI Pipeline Executions
- Current executionSucceeded
  feat/websocket-transport
- Runs and Agents
- Configuration
- Self-Healing CI
- Project Graph
- Resource Usage
- Analysis
- Conformance
Runs
- Current runSucceeded
- Tasks
- Resource Usage

Insights

Compare tasks
Analytics

fix(ai,ai-client): address WebSocket transport review findings Server (toWebSocketStream): - Surface turn failures as a live RUN_ERROR frame instead of logging and dropping them — the conversation socket stays open, so a swallowed error left the client hanging forever (parity with the HTTP transports' synthesized terminal). - Register an error listener on run and resume sockets: on ws an unhandled error event is an uncaught exception, and an errored socket leaked its heartbeat/idle timers and never aborted its turns. - Close the register-after-close race: a turn validated while the socket closed no longer starts unabortable work, and an abort frame that races ahead of its run's registration is honored instead of dropped. - buildTurnRequest scrubs a handshake ?offset so a mis-routed resume can't silently take the durability replay branch; the always-null WsRunContext.resumeOffset field is removed. - WebSocketLike: overloaded addEventListener replaces the (ev: any) handler, and the never-read bufferedAmount field is dropped. ws sockets now satisfy the type directly, so the plugins' type-laundering guard is deleted. Client (webSocket()): - openOnce is keyed by run/resume mode: joinRun owns a dedicated replay socket (its ?offset handshake was silently discarded when a conversation socket existed, and a later send() wrote run frames onto a listener-less resume socket); a send() during reconnect backoff supersedes the pending resume instead of racing it. - A same-runId resubmit resets sawTerminal/progressed, so a drop during a client-tool continuation reconnects instead of hanging. - stop() now sends the { type: 'abort', runId } frame so the server cancels the turn instead of generating to completion. - The duplicated subscribe/joinRun iterator pump is extracted into a shared createChunkPipe; body option is Record<string, unknown>. Tests: server multi-turn multiplexing with per-turn durability logs, idle timeout (incl. the in-flight-turn guard), error-frame surfacing, socket-error teardown; client new-runId session reset, same-runId resubmit reconnect, abort-frame emission, run/resume socket isolation, clean joinRun completion; new e2e driving the real webSocket() adapter through a forced mid-run drop and auto-resume. Docs updated to match (abort semantics, idle-timeout exemption, runtime support claims, adapter options), config.json dates fixed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

nx run-many --targets=build --exclude=examples/**,testing/**

Succeeded

4 CPU cores

read-write access token used

487919fa969

2%

Cache hits

1 of 56 tasks used cache.

Get faster results

Learn how to enable distribution.

Atomizer enabled

56 groups of tasks optimized.

© 2026 - Nx Cloud

Terms of Service Privacy Policy Changelog Status Docs Contact Nx Cloud Pricing Company @NxDevTools