[Bug]: Antigravity Gemini 4 Argon weekly limit shows 100% used when cloudcode-pa overwrites daily-cloudcode-pa #32419

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Sign in to Antigravity in T3 Code with a personal Google account (oauth-personal, non-GCP-ToS consumer tier such as Google AI Pro) where daily-cloudcode-pa.googleapis.com has remaining gemini-argon-wtu-weekly quota, while cloudcode-pa.googleapis.com reports remainingFraction: 0 for gemini-argon-wtu-weekly.
2. Refresh provider status / inspect ~/.t3/caches/antigravity.json (usageLimits.windows) or Usage → Limits.
3. Start or continue an Antigravity thread using gemini-4-argon (Gemini 4 Argon).

Expected behavior

T3 Code reports the gemini-argon-wtu-weekly usage and reset timestamp from the endpoint that the Antigravity ACP runtime actually uses for the account (https://daily-cloudcode-pa.googleapis.com for personal/consumer accounts, e.g. 13.77% used / 86.23% remaining), matching the active session's real quota.

Actual behavior

T3 Code reports gemini-argon-wtu-weekly as 100% used (usedPercent: 100, resetsAt: "2026-11-08T18:27:14.000Z"), even though gemini-4-argon turns succeed normally and have 86.23% weekly quota remaining on the active endpoint (resetsAt: "2026-11-14T08:41:35Z").

Live comparison on the same account

Calling POST /v1internal:retrieveUserQuotaSummary with T3's exact header (User-Agent: antigravity/hub/2.24.0 (aidev_client; os_type=windows; arch=x64; cl=963137146)) and { "project": "aicode-consumers" } against both hosts in ENDPOINTS:

Moreover, calling POST /v1internal:generateContent or streamGenerateContent with model: "gemini-4-argon-low" directly against https://cloudcode-pa.googleapis.com returns HTTP 429 RESOURCE_EXHAUSTED, whereas the active Antigravity ACP session succeeds because the ACP server routes consumer OAuth sessions (usesGcpTos === false) to https://daily-cloudcode-pa.googleapis.com:

# google3/cloud/developer_experience/antigravity_extensions/acp_server/ccpa_connection/onboard.py
def _determine_endpoint(load_response: Mapping[str, Any]) -> str:
  """Determines the correct CCPA endpoint based on user's tier.

  Consumer tiers (Free, Google AI Pro/Ultra) use Prod Daily.
  Enterprise/Workspace tiers use Prod.
  """
  paid_tier = load_response.get("paidTier", {}) or {}
  uses_gcp_tos = paid_tier.get("usesGcpTos", False)
  return _PROD_ENDPOINT if uses_gcp_tos else _PROD_DAILY_ENDPOINT

Root cause

In apps/server/src/provider/antigravityUsageLimits.ts#L18-L21 and #L280-L308, fetchWindows queries both https://daily-cloudcode-pa.googleapis.com and https://cloudcode-pa.googleapis.com in parallel and deduplicates windows by taking whichever host reports the higher usedPercent:

const windows = new Map<string, ServerProviderUsageWindow>();
for (const window of responses.flat()) {
  const previous = windows.get(window.id);
  if (!previous || window.usedPercent > previous.usedPercent) windows.set(window.id, window);
}

That max(usedPercent) merge was introduced in #20081 when the probe still sent User-Agent: antigravity/hub/1.0.0, which caused daily-cloudcode-pa to omit the Gemini group on some accounts. Now that #28090 updated the probe User-Agent to antigravity/hub/2.24.0, daily-cloudcode-pa returns all three quota groups (Gemini Models, Claude and GPT models, and Gemini 4 Argon) with the account's live consumer quota state. However, because cloudcode-pa is still queried in parallel and merged via window.usedPercent > previous.usedPercent, an exhausted or stale gemini-argon-wtu-weekly bucket on cloudcode-pa overwrites the live remaining quota on daily-cloudcode-pa.

Impact

Minor bug or occasional failure — Gemini 4 Argon · Weekly is shown as 100% exhausted in T3 Code with an earlier/stale reset time even when the user has 86%+ weekly Argon quota remaining and active Argon threads work normally.

Version or commit

T3 Code Desktop 0.1.7 (serverVersion: "0.1.7", Antigravity ACP 1.4.1); verified unchanged on main @ 83a3674ffd84ffb33de768210c806cb1ac53fd80.

Environment

Windows 11 x64 (win32 x64), T3 Code 0.1.7, Antigravity provider (oauth-personal, Google AI Pro tier).

Suggested fix direction

1. Route oauth-personal probes to https://daily-cloudcode-pa.googleapis.com first (matching onboard.py's _determine_endpoint for consumer accounts, or respecting endpoint_url in acp_token.json if present) and only fall back to https://cloudcode-pa.googleapis.com for buckets missing from the primary endpoint, rather than letting cloudcode-pa overwrite buckets already returned by daily-cloudcode-pa via window.usedPercent > previous.usedPercent.
2. Note that PR fix(usage): show Antigravity Gemini limits when one quota endpoint omits them #20081's "Not checked" section explicitly anticipated this edge case: "Hosts disagreeing right after a reset: the more-used (pre-reset) reading would show until both hosts agree."

Filed by Gemini 4 Argon (gemini-4-argon) in the Antigravity harness through T3 Code at the user's request.