Heads-up for anyone who hit 502 Bad Gateway on text calls in that window:
What happened: Every text reverse-RPC (sampling/createMessage, agent/complete) returned HTTP 502, while image/generate succeeded over the exact same auth + reverse-RPC + Cloudflare path. GET / and /api/v1/health were both 200.
Root cause: A brief outage of the text-model upstream on anna.partners. The gateway couldn’t get a valid response from the LLM completion backend → 502. The image backend is a separate upstream and stayed healthy.
Why it’s not your app / auth / SDK: Image generation working over the identical path rules out auth, grants, protocol, the SDK, and Cloudflare-in-general. The errors are upstream signals: agent/complete → -32046 (AGENT_ERR_PROVIDER_ERROR — catch-all incl. /copilot/app/complete returning 5xx), and sampling → -32000 wrapping the raw HTTP 502.
Good news: No quota leak — the throw-away session is always torn down in finally, so retrying is safe.
What to do: Retry (it was a ~4-minute window and likely self-healed). If it persists, escalate to whoever runs the anna.partners text-model gateway with the timestamp window + the “text 502 / image OK” contrast. App correctness (e.g. for hackathon review) can be re-verified once the text upstream is back.
Hi @YuntianBi! Huge thanks for such a sharp, well-structured write-up — the “text 502 / image OK” contrast and the throw-away-session teardown detail are exactly the kind of signal that makes triage fast. This is really helpful for the whole community.
You’re right that a brief text-upstream blip can cause this. To make sure we pin down the exact cause (a short self-healing outage looks identical on the surface to a couple of other issues ), could you share a bit more when you have a moment? A few targeted details go a long way:
What would help us localize it fast
Timestamp window (with timezone)
e.g. 2026-06-18 14:32–14:36 UTC — lets us line it up against upstream/gateway logs for that exact window.
Which model was selected
The specific model name in the picker at failure time (e.g. qwen3.7-plus, google/gemini-3-flash-preview, or the free Gemma 4 E4B-it). This matters a lot — a single-model failure vs. all-models failure points to very different root causes.
Was it all text models, or just one?
Did switching to another model (e.g. from Gemma → a Qwen/Gemini option) also 502, or did it recover? “Every model failed” vs. “only one failed” is the single most useful disambiguator.
The raw error body / code
The exact string you saw, e.g.:
-32000 wrapping a bare HTTP 502 with an empty body, vs.
an application-level message like 404 — The model '...' does not exist
These look similar but mean completely different things.
Path contrast you already noticed
You nailed this one — confirming image.generate succeeded on the same auth/reverse-RPC/Cloudflare path in the same window is gold for ruling out auth/SDK/protocol.
Duration & retry behavior
Was it a short window that self-healed on retry (sounds like ~4 min here), or persistent and reproducible every time? Self-healing → likely a transient upstream blip; persistent + reproducible → more likely something deterministic we can patch.
Example of an ideal report
When: 2026-06-18 14:32–14:36 UTC Model:Gemma 4 E4B-it Scope: only Gemma failed; qwen3.7-plus worked in the same window Error:404 — The model 'gemma-...' does not exist Contrast:image.generate OK on the same path Pattern: persistent, reproduced every retry
That single block usually tells us whether it’s a transient upstream outage, a model that’s been rotated/retired, or a code-path issue — and gets you a precise fix or all-clear much faster.
Thanks again for being so thorough and community-minded — drop the details whenever you can and we’ll dig right in! Excited to keep your builds running smoothly.
Huge thanks for such a sharp, well-structured write-up — the “text 502 / image OK” contrast and the throw-away-session teardown detail are exactly the kind of signal that makes triage fast. 
This is really helpful for the whole community. 
), could you share a bit more when you have a moment? A few targeted details go a long way:
