[Antigravity / Cloud Code] Gemini 4 Argon repeats passages in a single SSE response

Google Antigravity
1

In long-context, multi-turn coding conversations with tool history, Gemini 4 Argon intermittently restarts the same commentary with slightly different wording several times within one response. Reasoning-like markers and text also appear in ordinary visible text parts. This interrupts useful tool work.

We reproduced the behavior by directly calling the Antigravity / Cloud Code native streaming endpoint with reconstructed conversation history and minimal historical tool declarations. No generated tool calls were executed.

Environment and reproduction

  • Observed on 2026-10-22, approximately 01:22-01:25 UTC.
  • Model: gemini-4-argon, mapped to gemini-4-argon-high.
  • Approximately 360k input tokens of multi-turn conversation and tool history.
  • maxOutputTokens=32768. The high setting was selected through the mapped model ID; generationConfig did not contain an explicit thinkingLevel or temperature field.
  • Endpoint: https://daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent?alt=sse.

Native upstream capture

  • Google responseId: iGXZarvnLtzi4t4****.
  • One HTTP 200 SSE response, one observed response ID, finishReason: STOP.
  • 38 JSON frames, 36 text parts, zero separately flagged thinking parts.
  • 2,693 visible characters containing four self-restarted passages, over 92.61 seconds.

The repeated passages already exist in candidates[0].content.parts[].text in the native Google response, before client-side rendering or response conversion. This reproduction used a single successful streaming request; the repeated passages were not assembled from four client retries.

The replay used reconstructed history and minimal historical tool declarations, rather than an identical copy of an earlier HTTP request. The reasoning-like text is an observable output symptom; we are not claiming that it exposes actual hidden reasoning. A second Gemini conversation showed much more extensive repetition, but the native capture above verifies four repetitions only.

Expected behavior

Produce one coherent commentary/final answer or a tool call, without restarting the same passage repeatedly within one response.

Could the Gemini/Antigravity inference team investigate this response ID and advise whether this is a known long-context/high-setting regression? The original SSE capture is retained and can be provided through an appropriate support channel.

2

We’re seeing the same pattern in long-running Claude Code sessions using Gemini 4 Argon through CLIProxyAPI.

On October 22, around 16:38–16:40 UTC, we recorded one assistant response containing a 12,107-character text block. It restarted the same acknowledgement 32 times with variations in wording, then issued one tool call. The saved transcript contains the repetitions within that single response.

In another session, one response contained repeated commentary followed by two distinct tool calls requesting the same action. Those calls had different IDs and arguments, so exact duplicate-ID filtering would not catch them.

We observed both cases after deploying a proxy patch that stops content after finishReason or [DONE] and rejects response-ID changes mid-stream. We have since fixed a separate signature-replay issue in the proxy, but we have not established whether that affects the repetition.

For these examples, we have client transcripts and proxy request logs, not raw Google SSE captures. They corroborate the symptoms but do not provide another direct upstream reproduction.

Could the Google team confirm whether this is a known Argon generation issue and suggest a mitigation for long-context tool-use sessions?

3

I’ve reproduced this on Argon Medium as well around the 130k-160k context mark each time.

4

Hi Google Team! Can we please get some help with this issue, it’s been going on for a week now. Thank you!