Files
goclaw/cmd
yatulandClaude Opus 5 34a1540c18 fix(teams): stream teammate runs so long generations keep the connection alive
handleTeammateMessage pinned Stream: false, so every team member run — coder,
reviewer, researcher — made a non-streamed provider call. On a slow reasoning
model that means a silent connection for the whole generation, and
ResponseHeaderTimeout eventually kills it:

    iter 0 think: llm call: litellm: request failed:
      Post ".../v1/chat/completions": http2: timeout awaiting response headers

Observed on a member asked to write a large single-file page: 15 minutes, zero
output tokens, run failed. The same model over the same provider is fine on an
ordinary channel run, which streams by default.

Streaming here is for connection liveness, not delivery. A teammate run is never
registered with the channel manager, so HandleAgentEvent returns on its first
line and the chunks are dropped; nothing reaches a user incrementally. The task
result is unchanged — it still comes from the final RunResult, since ChatStream
returns the complete response once the stream ends.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-25 22:03:41 +04:00
..
2026-07-15 09:56:43 +07:00
2026-07-31 19:34:40 +07:00
2026-06-12 09:48:01 +07:00