# Bug report: agent-gw `generate_video` accepted but never completes — always times out at exactly 3600 s

Product Feedback
1

Service: Kimi For Agent API — video generation (generate_video, version v2) Endpoint: https://agent-gw.kimi.com/coding → POST /v1/tools Date: 2026-09-08 (local time, CEST) Status page note: status.moonshot.cn shows all components operational and reports no incidents today; however, video generation is not a tracked component there, so this failure would not appear on the status page.

Summary

Since this morning, every generate_video request I send is accepted by the gateway but never completes. Each request hangs server-side for exactly one hour and then the connection is closed with a read timeout. Four attempts so far (3 real jobs + 1 minimal diagnostic), all with identical behavior.

Timeline (2026-09-08, local time)

#
Time
Request
Reference image
Result
1
~07:00
7 s, 720p, 16:9, detailed fantasy scene
yes
Read timeout at 3600 s
2
~08:00
5 s, 480p, 16:9, same scene
yes
Read timeout at 3600 s
3
18:20
5 s, 480p, 16:9, same scene
yes
Read timeout at 3600 s
4
19:27
4 s, 480p, 16:9, minimal diagnostic prompt (“a single candle flame flickering…”, static shot)
no
Read timeout at 3600 s

Error message (identical for attempts 3 and 4, captured verbatim):

Error generating video: Timeout calling POST /v1/tools: HTTPSConnectionPool(host='agent-gw.kimi.com', port=443): Read timed out. (read timeout=3600.0)

Isolation done on my side

  • Endpoint connectivity is healthy: GET https://agent-gw.kimi.com/coding returns 200 in ~0.7 s ({"message":"Welcome to the Kimi For Agent API!"}).

  • Credentials and SDK are fine: agent-gw Python SDK 0.2.6, API key resolves from local config.

  • Image upload works: upload_storage (via the same gateway, image-to-url) succeeds and returns a valid signed URL every time.

  • Attempt 4 had no reference media at all and a trivial prompt — so the failure is not related to reference images, prompt content, resolution, or duration.

  • Failures span morning (7–8 am) and evening — does not look like simple peak-hour congestion.

  • No local proxies/VPN involved; other network traffic is normal.

Interpretation

The gateway accepts generate_video requests (no 4xx/5xx, no auth error, no quota error) but the generation backend never returns a result. This looks like either:

  1. the video-generation worker/queue behind generate_video is stuck or down, or

  2. requests are being silently dropped after acceptance (possibly account-level gating that hangs instead of returning an error).

Either way, the client cannot distinguish “queued” from “lost” — a quicker failure (HTTP 5xx/503, or a job-status API) would make this diagnosable from the client side.

Requests

  1. Confirm whether video generation via agent-gw is degraded/broken right now.

  2. ETA for a fix, and where video-generation incidents will be posted (currently invisible on status.moonshot.cn).

  3. Please make the API fail fast (or return a job id + status endpoint) instead of hanging until timeout.

  4. If requests were silently dropped after acceptance, confirm no quota was consumed by the four timed-out attempts.

Happy to provide request IDs, exact request bodies, or run further diagnostics on request.