RetrieX SSE reconnect replay fix

Patch-only fix for fragile browser streaming after the job-based EventSource flow.

Problem

EventSource is allowed to reconnect automatically when the browser, proxy or PHP/Nginx connection is interrupted. The previous job guard treated a second /ask-sse/{jobId} connection while the job was still running as an application error. In the UI this could append messages like:

event: error
data: Der Antwort-Stream läuft bereits oder wurde nach einem Verbindungsabbruch erneut geöffnet...

This happened especially in slower Shopware/search flows, but it could also happen in pure RAG answers.

Change

Streamed answer chunks now receive monotonically increasing SSE id: values.
Each streamed chunk is additionally written to a per-job replay buffer under var/stream_jobs/*.stream.ndjson.
If EventSource reconnects to a job that is already running, the controller no longer emits the misleading duplicate-stream error.
Instead, the reconnecting request reads Last-Event-ID, replays missing chunks and tails the still-running job until it completes, fails or is marked interrupted.
Completed jobs can replay any missing chunks and then emit done.
The frontend tracks event.lastEventId and ignores duplicate/replayed chunks defensively.
Expired job cleanup also removes the replay buffer sidecar file.

Changed files

src/Controller/AskSseController.php
public/assets/js/base.js

Safety

This patch does not change retrieval, PromptBuilder, AgentRunner, scoring, intent detection, Shopware query generation or RAG behavior. It only changes the transport/reconnect layer for SSE.

After installing

php bin/console cache:clear
php bin/console mto:agent:config:validate
php bin/console mto:agent:regression:test

Hard-refresh the browser or clear browser cache because public/assets/js/base.js changed.

1.9 KiB Raw Blame History