plat 268

← Pulse platform index

PLAT-268 — Workflow chat could run normally in tmux while showing no tools or response

Coordination	Value
Assigned agent	Codex
Ticket state	implemented; restarted UI reverify pending
Last synchronized	2026-09-01

• Priority: P0 reliability. The user repeatedly saw a submitted Workflow Builder message remain alone in an otherwise blank chat while the retained tmux showed normal assistant output and successful tool execution.
• Live evidence: session b445627b-deae-407c-a05e-02fcd7343d7c, Social Media / twitter-automation, submitted at 21:38 on 2026-09-01.

Root cause

The request path was healthy. The frontend's retained live-input probe returned 404 and correctly fell back to POST /api/query; the server created the same session, emitted assistant chunks, and executed multiple tools. The browser nevertheless showed none of them.

ChatArea.tsx normally consumes those events over a long-lived EventSource. Browsers cap concurrent HTTP/1.1 connections per origin, and AgentWorks can have several restored tabs holding SSE connections. A connection waiting for a browser socket does not reliably emit onerror, so the existing “start polling only after SSE errors repeatedly” fallback never activates.

The code already contained a 750 ms REST catch-up loop specifically because of this connection-cap behavior, but startProductEventCatchUp immediately returned unless fullTurnStreaming was enabled. Product surfaces were protected; ordinary Workflow Builder chat was not. This exact turn called the function after its successful fallback query, but the guard disabled it.

This is distinct from PLAT-250: PLAT-250 resets a stale cursor after a backend restart; this incident used a fresh session and generated a healthy current event stream that the browser transport never consumed.

Fix

• Generalized the product-only loop into startForegroundEventCatchUp and run it for every submitted foreground turn, including retained live-input, ordinary query fallback, and live_input_delivered responses.
• Active streaming/restored tabs also start the same loop on mount, so a turn already in progress when the frontend reloads or hot-updates self-heals without requiring another user message.
• The catch-up reads the same session event store as SSE. Durable event IDs already make simultaneous SSE and REST delivery idempotent in the frontend store, so this adds no duplicate chat rows.
• Stop polling at completed, error, stopped, or inactive once no background agents remain; replacing the old completed/error-only condition also prevents a stopped retained session from leaking a timer.

Verification

• Root cause confirmed from the exact live prompt stream and server log: tmux response/tool activity continued while the browser remained blank.
• Frontend TypeScript, focused lint (zero errors; one unrelated existing hooks warning), production build, and bundle-budget checks pass.

Reverify

Restart the frontend/backend, open several workflow chats, then submit a new Workflow Builder message. Its assistant response and tool calls must appear without refreshing the page even when the EventSource connection is delayed.