Skip to content

[agent][inference] Base inference stream stalls mid-generation, no keepalive/recovery → orphaned turn + empty reply #4761

Description

@M3gA-Mind

Base inference stream stalls mid-generation with no keepalive/timeout recovery → orphaned turn, empty reply (no terminal event)

Evidence

g7-29 "Explain, with full reasoning visible, how you'd triage an inbound Telegram message" (pure reasoning, no tools, no sub-agents):

  • events: ready, inference_start, iteration_start, 192 text_delta, 2 inference_heartbeat. First token @19s (poor TTFT).
  • streamed 826 chars of coherent step-by-step reasoning, then FROZE mid-sentence ("…urgent, asap, help, can you"), emitted only heartbeats, and the server dropped the stream at 250.9s with NO chat_done and NO chat_error → empty final reply.

Ask

This is a DISTINCT root cause from the sub-agent reduce-hang (no sub-agent here) and from turn-interference — the base inference stream itself stalls and the turn is orphaned with no terminal event. Add mid-stream keepalive / idle-timeout recovery on the core↔backend inference connection, and ensure a terminal event (chat_done or chat_error) always fires when the upstream stream stalls/drops. Also investigate the ~19s TTFT. (Some upstream slowness on staging is infra, but the missing recovery + missing terminal event is a code gap.) Core inference streaming path.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Fields

No fields configured for issues without a type.

Projects

Status
In progress

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions