Back to Signal Feed
CodeTracked since May 19, 2026

Exclude SSE `values` mode from chat streams to cut payloads

The PR changes DeerFlow’s chat frontend to stop subscribing to `values` SSE stream events, leaving `messages-tuple`, `updates`, and `custom` as the active stream modes. Because `messages-tuple` already carries token-by-token output in real time, removing the full-snapshot `values` mode cuts stream traffic for long tasks from about 6.56 MB to 1.64 MB, roughly a 75% reduction.

SSEstreamModevaluesmessages-tuple

What Happened

  • The PR changes DeerFlow’s chat frontend to stop subscribing to `values` SSE stream events, leaving `messages-tuple`, `updates`, and `custom` as the active stream modes. Because `messages-tuple` already carries token-by-token output in real time, removing the full-snapshot `values` mode cuts stream traffic for long tasks from about 6.56 MB to 1.64 MB, roughly a 75% reduction.
  • The PR changes DeerFlow’s chat frontend to stop subscribing to `values` SSE stream events, leaving `messages-tuple`, `updates`, and `custom` as the active stream modes. Because `messages-tuple` already carries token-by-token output in real time, removing the full-snapshot `values` mode cuts stream traffic for long tasks from about 6.56 MB to 1.64 MB, roughly a 75% reduction.
  • 1 evidence item attached for review.

What is Different

Before

Scattered source updates, isolated context, and manual follow-up across multiple feeds.

Now

Removed redundant full-message snapshot streaming for chat responses by updating the frontend subscription to exclude `values`, and adjusted E2E mock events to the messages-tuple wire format so stream handling matches the new behavior.

Why Track This

Why It Matters

Chat users and operators running long tasks should see much lighter browser/network streaming, reducing practical risks like noticeable lag and flaky streaming under constrained network conditions while preserving token-by-token updates. The change cuts transfer per long task from ~6.56 MB to ~1.64 MB by dropping `values` snapshots after each graph step, while keeping the UI fed by `messages-tuple`, `updates`, and `custom` events. Continue monitoring production traffic and stream consumers for any implicit dependency on `values` snapshots, especially on recovery paths where full-state replay assumptions may exist.

Impact

Chat users and operators running long tasks should see much lighter browser/network streaming, reducing practical risks like noticeable lag and flaky streaming under constrained network conditions while preserving token-by-token updates. The change cuts transfer per long task from ~6.56 MB to ~1.64 MB by dropping `values` snapshots after each graph step, while keeping the UI fed by `messages-tuple`, `updates`, and `custom` events. Continue monitoring production traffic and stream consumers for any implicit dependency on `values` snapshots, especially on recovery paths where full-state replay assumptions may exist.

What To Watch Next

  • Watch whether SSE becomes a repeated pattern.
  • Track follow-up changes around AI Workflow Automation.
  • Compare future signals against this evidence trail.
  • Re-check risk flags: watch_for_values_snapshot_consumer_regression, verify_stream_display_integrity_on_long_tasks.
Open Topic TimelineOpen Technical EventOpen Original Sourcewatch_for_values_snapshot_consumer_regression / verify_stream_display_integrity_on_long_tasks / monitor_network_savings_in_real_world_sessions

Supporting Evidence

GITHUB PULL REQUESTHigh Trust

bytedance/deer-flow PR #3056: fix(frontend): exclude values stream mode to reduce SSE payload by ~75%

before: `values` + `messages-tuple` + `updates` + `custom` (~6.56 MB per long task); after: `messages-tuple` + `updates` + `custom` (~1.64 MB, ~75% reduction).