Summary

I consume the Command Code provider API from a third-party OpenAI-compatible client. With model=deepseek/deepseek-v4.1-flash and stream=true, time-to-first-token is the clear weak point: it is not a stable cost — it varies widely between requests, but in my day-to-day use it is long far more often than not. 20 seconds or more is common, and I have measured it as high as 146.8 seconds.

Once the first token finally arrives, generation is fast. Here is one instrumented request from my client (screenshot attached):

Those numbers are the whole report: of 167.9 seconds total, 146.8 seconds (87%) elapsed before the first token arrived, while the generation phase afterwards ran at 175.9 tok/s — a perfectly healthy rate. Nothing is wrong with how fast the model generates; the problem is entirely in front of the first token.

Expected Behavior

• On a fast model such as V4.1 Flash, a streaming request should normally begin returning content within a few seconds, and should not routinely sit silent for 20s+.
• Even when the upstream model needs a long time to think (V4.1 Flash is a reasoning model with thinking on by default), the gateway should flush SSE response headers and/or emit periodic SSE comment heartbeats so clients and intermediaries know the connection is alive rather than stalled.
• A client should not have to distinguish "model is thinking" from "edge is buffering" by guesswork; a server-timing/TTFB header would make this trivial.

Actual Behavior

• Time-to-first-token is long far more often than not, commonly 20s+, with a measured worst case of 146.8s. It is not every single request — the number swings quite a bit between requests, and some requests do come back considerably faster.
• While waiting, the client receives nothing to render: the prompt sits there with no content. (I have not captured a raw packet trace, so I can't state whether the HTTP response headers themselves arrive early; what I can confirm is that no token content reaches the client during that window.)
• After the first token, throughput is normal — 175.9 tok/s in the measurement above.

Steps to reproduce the issue

1. Configure an OpenAI-compatible client (I use DSH) with the Command Code provider base URL and API key.
2. Set model=deepseek/deepseek-v4.1-flash, stream=true.
3. Send a series of ordinary prompts and record the client's time-to-first-token for each.
4. Observe that TTFT is frequently 20s or more, varies widely from request to request, and in at least one case reached 146.8s.
5. Note that once generation begins, throughput is normal (~175.9 tok/s) — total wall-clock time is dominated by the pre-token wait.

Command Code Version

n/a — consumed via the hosted provider API endpoint, not the CLI.

Operating System

Windows

Terminal/IDE

DSH (third-party OpenAI-compatible API client)

Shell

n/a — HTTP API client

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

Client-side request timing for one streaming call (screenshot attached):

请求计时 (request timing)
  开始时间 (start)      2026-10-03 08:18:27.224
  总时长 (total)        167.9 s
  首 token 延迟 (TTFT)  146.8 s
  生成 (generation)      21.1 s
  吞吐量 (throughput)    175.9 tok/s

• Network: mainland China, direct connection, no proxy/VPN. Dynamic requests fail from Celerity/PuntoNet Ecuador (Cloudflare UIO) but work via Movistar #790 documents a region-specific relay/edge-node failure (Cloudflare UIO), so the edge node serving mainland China is a plausible contributor and worth checking.
• Model: deepseek/deepseek-v4.1-flash (reasoning model, thinking on by default), not the default deepseek/deepseek-v4-flash. Think-token looping / raw (unrepaired) tool-call failures on DeepSeek V4.1 Flash & GLM 5.3 #967 reports think-token looping on this same model, which would also present as high TTFT. I can't fully rule that out — a pure API client doesn't expose reasoning tokens — but the healthy 175.9 tok/s afterwards and the large request-to-request variance both argue against a model that is simply reasoning slowly every time.
• Related issues: API 524 on non-streaming requests that take too long (~180s); streaming works fine #696 (relay buffers non-streaming responses → 180s 524 — buffering on the same relay), Bug: First prompt after suspend/resume hangs for seconds to minutes #819 (first prompt stalls — different trigger), Think-token looping / raw (unrepaired) tool-call failures on DeepSeek V4.1 Flash & GLM 5.3 #967 (think-token looping on V4.1 Flash / GLM 5.3), Dynamic requests fail from Celerity/PuntoNet Ecuador (Cloudflare UIO) but work via Movistar #790 (region-specific relay failure), deepseek V4 flash connot close thinking mode #776 (thinking mode cannot be disabled).
• Why I suspect the gateway rather than the model: generation after the first token runs at 175.9 tok/s, so the model is fast. And the TTFT is not a stable, prompt-dependent cost — it swings widely between otherwise ordinary requests, which is characteristic of queueing, cold connections or edge-node scheduling. Model inference time tends to be comparatively steady and to show up as reduced throughput, not as a silent wait followed by a fast stream.