openclaw - 💡(How to fix) Fix github-copilot: Gemini models route to unsupported /responses endpoint, and openai-completions transport bypasses Copilot IDE auth headers [1 comments, 2 participants]

Summary

github-copilot/gemini-3-flash-preview (and likely all Gemini//chat/completions-only Copilot models) fail 100% of the time and silently fall back to the configured non-Copilot fallback (e.g. anthropic/claude-sonnet-4-6), generating substantial unintended Anthropic API spend.

Two bugs in series in dist/extensions/github-copilot/:

Bug 1 — resolveCopilotTransportApi routes Gemini to the wrong endpoint

dist/models-CZL3_C6Y.js:

function resolveCopilotTransportApi(modelId) {
  return (normalizeOptionalLowercaseString(modelId) ?? "").includes("claude")
    ? "anthropic-messages"
    : "openai-responses";
}

Any non-Claude Copilot model is routed through POST /responses. GitHub Copilot rejects this for Gemini:

400 unsupported_api_for_model: model gemini-3-flash-preview is not supported via Responses API.

Direct verification (curl):

• POST https://api.githubcopilot.com/responses with model: gemini-3-flash-preview → 400 unsupported_api_for_model
• POST https://api.githubcopilot.com/chat/completions with same model → 200 OK

Bug 2 — openai-completions transport bypasses Copilot IDE-header injection

dist/stream-Dy0kutKG.js:

function wrapCopilotProviderStream(ctx) {
  return wrapCopilotOpenAIResponsesStream(wrapCopilotAnthropicStream(ctx.streamFn));
}

Only anthropic-messages and openai-responses are wrapped to inject Copilot's required Editor-Version / Copilot-Integration-Id / User-Agent headers (via buildCopilotRequestHeaders → buildCopilotDynamicHeaders). openai-completions is not wrapped, so any model routed through chat-completions hits:

400 bad request: missing Editor-Version header for IDE auth

Confirmed by setting "api": "openai-completions" override on gemini-3-flash-preview in openclaw.json — request now goes to /chat/completions (Bug 1 bypassed) but immediately fails with "missing Editor-Version header for IDE auth" (Bug 2 hits).

This is the same class of issue as #71538 (closed 2026-04-25) but for the openai-completions transport path, which that PR did not cover.

Environment

• OpenClaw: 2026.4.24 (cbcfdf6)
• Provider: github-copilot (device login, Pro+ plan with Gemini access verified)
• Model affected: github-copilot/gemini-3-flash-preview (also: gemini-3.1-pro-preview, gemini-2.5-pro, and any non-Claude/non-responses-supporting model)
• Available models confirmed via GET https://api.githubcopilot.com/models — Gemini IDs are listed for the account.

Reproduction

1. Set github-copilot/gemini-3-flash-preview as a heartbeat model or run:

   openclaw capability model run --model github-copilot/gemini-3-flash-preview --prompt "ping"

2. Observe in logs:

   [model-fallback/decision] decision=candidate_failed
     requested=github-copilot/gemini-3-flash-preview
     reason=format
     detail=400 building HTTP client: selecting model endpoint:
            scheduler enqueue for model endpoint selection:
            no model endpoints available given user constraints

3. Add "api": "openai-completions" to the model definition in openclaw.json and retry. Error changes to:

   400 bad request: missing Editor-Version header for IDE auth

Impact

• Silent cost amplification. Gemini was configured as the heartbeat model on three agents (1h cadence × 24h × 3 agents = ~72 calls/day intended free). All 72 calls/day fell through to anthropic/claude-sonnet-4-6 fallback. Plus subagent spawns. Combined: ~287 fallback events/day on this model alone, materially contributing to a multi-day Anthropic bill spike.
• The primary→fallback flow makes this invisible by design: the request "succeeds" (just not on the requested provider), so users only notice via the bill.

Proposed Fix

1. resolveCopilotTransportApi should return "openai-completions" for known Gemini model IDs (or any model ID not on Copilot's Responses-API allowlist).
2. wrapCopilotProviderStream should also wrap openai-completions with buildCopilotRequestHeaders.
3. Optional: surface a fast-path warning when model-fallback/decision shows reason=format for the same primary provider repeatedly (signal: "your primary is broken, you're paying for fallback").

Happy to test a patched build locally.