Repository navigation
[Bug] qwen3.7-max/plus on OpenCode Go (Zen API) - frequent 524 timeout behind Cloudflare (related to #32418) #33726
Closed as not planned
Description
Description
qwen3.7-max and qwen3.7-plus on OpenCode Go (Zen API) intermittently fail with timeouts. The root cause is Cloudflare in front of opencode.ai having a 120s proxy read timeout. When qwen3.7 generation (especially with thinking mode) takes longer than 120s, Cloudflare returns a 524 timeout error as HTML, which the AI SDK cannot parse as JSON, resulting in cryptic errors and infinite retries.
This matches the findings in #32418 by @BernhardGruen.
Steps to Reproduce
- Send a POST request to
https://opencode.ai/zen/go/v1/messageswith qwen3.7-max or qwen3.7-plus - Enable thinking mode (e.g.,
thinking: {type: "enabled", budget_tokens: 1000}) - Observe intermittent 524 timeout from Cloudflare after ~120s
Expected Behavior
- Requests should not be killed by Cloudflare timeout at 120s
- If timeout occurs, the API should return a proper JSON error response, not an HTML page
Actual Behavior
Cloudflare returns a 524 HTML error page, which the client SDK interprets as an empty ("") error message, triggering automatic retries that also timeout.
Environment
- Service: OpenCode Go (Zen API) — paid subscription
- Endpoint: https://opencode.ai/zen/go/v1/messages
- Models affected: qwen3.7-max, qwen3.7-plus
- Client: OpenClaw (third-party, using
@ai-sdk/anthropic) - Region: Mainland China (via proxy)
Suggested Fix
- Server-side (long-term): Increase Cloudflare proxy read timeout for qwen3.7 endpoints, or implement streaming responses that keep the connection alive
- Server-side (short-term): Translate 524/5xx errors into proper JSON error responses instead of passing through raw Cloudflare HTML
- Client-side fallback: The AI SDK could detect HTML response bodies and produce human-readable error messages instead of empty strings
Related
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Jun 24, 2026
This issue might be a duplicate of existing issues. Please check:
If this issue provides new technical detail (the Cloudflare 524 root cause) beyond what was already reported, consider adding that information to one of the existing issues rather than opening a new one.
Joelincn commentedon Jun 24, 2026
Additional Analysis and Test Data
Adding more context that was in my original report (#33721) but omitted due to the template format.
Full Test Results (via
curltohttps://opencode.ai/zen/go/v1/messages)Pattern Analysis
The non-deterministic behavior also suggests backend resource contention or insufficient qwen3.7 instance allocation:
The fact that T4 (larger budget) works while T2 (smaller budget) does not is particularly puzzling. One hypothesis: under load, the backend may be deprioritizing or prematurely terminating certain requests. The Cloudflare 524 timeout (from #32418) is one piece of the puzzle, but the wide variance in response times suggests qwen3.7 instances may be under-provisioned.
Other models work fine
As a paying subscriber, I would appreciate transparency on whether this is a known capacity issue and if there are plans to scale qwen3.7 resources. Thanks.
github-actions commentedon Aug 24, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.