You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
The upstream provider did not respond before the proxy timeout. #86
changed the title [-]The upstream provider did not respond before the proxy timeout. You must retry with a smaller request or select a faster provider.[/-][+]The upstream provider did not respond before the proxy timeout. [/+]on Jul 10, 2026
came from the old web-search loop applying the global connectTimeoutMs (default: 200,000 ms) to the entire non-streaming routed-model iteration, including response generation.
Starting with v2.7.6, routed-model web-search iterations are streamed internally, and the timeouts are separated:
connectTimeoutMs — DNS/TCP/TLS and response headers only
webSearchSidecar.routedModelStallTimeoutMs — continuous inactivity between routed-model response-body bytes
webSearchSidecar.timeoutMs — one hosted web-search request
Every non-empty response chunk resets routedModelStallTimeoutMs, so a slow LM Studio generation may run longer than 200 seconds as long as it keeps producing bytes.
For a local model that can stay silent for more than 200 seconds during prefill/thinking, merge the following into ~/.opencodex/config.json:
Restart opencodex after editing the configuration.
connectTimeoutMs only needs raising if LM Studio takes more than 200 seconds to return response headers; routedModelStallTimeoutMs is the important setting when headers arrive but the model produces no body bytes for an extended period.
Please update to v2.7.6 and reopen if the timeout still reproduces with these settings.
Activity
[-]The upstream provider did not respond before the proxy timeout. You must retry with a smaller request or select a faster provider.[/-][+]The upstream provider did not respond before the proxy timeout. [/+]lidge-jun commentedon Jul 10, 2026
Resolved in v2.7.6 by commit
1f146adf.The reported error:
came from the old web-search loop applying the global
connectTimeoutMs(default: 200,000 ms) to the entire non-streaming routed-model iteration, including response generation.Starting with v2.7.6, routed-model web-search iterations are streamed internally, and the timeouts are separated:
connectTimeoutMs— DNS/TCP/TLS and response headers onlywebSearchSidecar.routedModelStallTimeoutMs— continuous inactivity between routed-model response-body byteswebSearchSidecar.timeoutMs— one hosted web-search requestEvery non-empty response chunk resets
routedModelStallTimeoutMs, so a slow LM Studio generation may run longer than 200 seconds as long as it keeps producing bytes.For a local model that can stay silent for more than 200 seconds during prefill/thinking, merge the following into
~/.opencodex/config.json:{ "connectTimeoutMs": 600000, "webSearchSidecar": { "routedModelStallTimeoutMs": 600000 } }Restart opencodex after editing the configuration.
connectTimeoutMsonly needs raising if LM Studio takes more than 200 seconds to return response headers;routedModelStallTimeoutMsis the important setting when headers arrive but the model produces no body bytes for an extended period.Please update to v2.7.6 and reopen if the timeout still reproduces with these settings.
lidge-jun commentedon Jul 10, 2026
Closing — this was resolved in v2.7.6 (commit 1f146ad). See the earlier comment for config guidance.