Repository navigation
github-copilot: gemini-3.8-flash Responses API failure recurs on 2.0.25 #53961
Description
Description
GitHub Copilot gemini-3.8-flash failed in an OpenCode 2.0.25 subagent session with:
model gemini-3.8-flash does not support Responses API.
This appears to be a recurrence of #53712, which was closed by its reporter after the problem stopped occurring, without a linked fix. The same error is still observable on 2.0.25. This report does not assume a dropped variant or claim a confirmed root cause.
OpenCode version
2.0.25 (release/latest channel)
Operating System
Linux x64 under WSL2, kernel 6.18.40.1-microsoft-standard-WSL2; shell /usr/bin/bash.
Terminal
OpenChamber web UI, using a managed opencode serve instance. Terminal environment variables are not set.
Plugins
[email protected], [email protected], OpenChamber agent-tool integration, and an opencode-mem loader. The configuration also contains -opencode.browser; its loaded status was not independently verified. A plugin-disabled reproduction has not been performed.
Steps to reproduce
Observed workflow:
- Run OpenCode 2.0.25 as an OpenChamber-managed server with GitHub Copilot authenticated.
- Configure an oh-my-opencode-slim custom subagent to use
github-copilot/gemini-3.8-flash. - Dispatch a background task to that subagent from a primary session.
- The child task terminates with the error above, before returning a task result.
The delegated task was read-only configuration troubleshooting. Its private prompt and session contents are intentionally omitted. Only one occurrence was observed in this workflow; failure frequency has not been established. A replacement task using another provider/model completed, but that is not a controlled Copilot comparison.
Expected Behavior
The model should be routed to an endpoint that GitHub Copilot supports for it, and the child session should return a response.
Actual Behavior
The child task failed with model gemini-3.8-flash does not support Responses API. Retrieving its terminal result confirmed the same failure.
Additional Context
In #53712, the reporter found that the Copilot language hook honors options.endpoint when present, but the fallback heuristic can select Responses for gemini-* when live endpoint metadata is missing. They reported that Copilot's live model list consistently advertised /chat/completions for gemini-3.8-flash.
That is a relevant hypothesis, not independently confirmed for this failed request. We have not captured its outgoing endpoint or live supported-endpoints metadata. Please investigate whether the fallback remains unsafe when model endpoint metadata is unavailable, and whether a documented per-model endpoint override is an appropriate workaround. Forcing every Copilot model onto chat completions could affect models that require Responses, so a model-specific workaround would be preferable.
No service restart, plugin-disabled run, or paid model-call reproduction has been performed specifically for this report.
Activity
opencode-agent commentedon Oct 8, 2026
Thanks for the detailed report and for linking #53712.
I couldn't reproduce the failure itself: the error text comes from GitHub Copilot's API, and this sandbox has no Copilot account. I did read the code on 2.0.25 and on the latest
v2(a67cf4c), and the path suspected in #53712 is there and unchanged:failed to sync GitHub Copilot modelsand carries on without endpoint information.gemini-*model to the Responses API. That fallback was added on purpose in fix(core): route responses-only Copilot models to the Responses API #51024, which assumed Copilot serves Gemini only on Responses.If Copilot lists only
/chat/completionsforgemini-3.8-flash, as #53712 observed, any request made while the live list is missing or out of date would fail exactly like this. That would also explain why it comes and goes. Confirming it needs someone with Copilot access to check the live model data, so a maintainer will need to look at it.Things that would help if it happens again:
failed to sync GitHub Copilot modelswarning around the time of the failure.A possible model-specific workaround, which I haven't tested against Copilot: set the endpoint for just this model in your config:
{ "provider": { "github-copilot": { "models": { "gemini-3.8-flash": { "settings": { "endpoint": "chat" } } } } } }Related: #53712 (same error and model), #51024 (added the Responses fallback for Gemini, Grok and MAI Code).