Selecting Gemini 4 Argon in agy CLI drains standard Gemini quota instead of Argon weekly limit

When explicitly setting the active model to Gemini 4 Argon (/model gemini-4-argon), prompt executions appear to route through or deduct usage from the standard Gemini model quota pool rather than the dedicated Argon weekly quota pool.

The dedicated Argon weekly limit remains completely untouched at 100%, while the standard Gemini quota drains significantly faster than normal.

Environment / Setup

  • OS: Windows WSL
  • CLI Version: 1.5.0
  • Account Tier: Google AI Pro

Steps to Reproduce

  1. Launch agy in any working repository.
  2. Select or switch to Argon via /model (e.g., gemini-4-argon / gemini-4-argon-medium).
  3. Verify the model using /btw "What model are you?" (it correctly reports Argon).
  4. Run several regular agentic development tasks / prompts.
  5. Check quota status via /usage.

Expected Behavior

  • Prompt usage while on Argon should consume and decrease the dedicated Gemini 4 Argon Weekly Limit.
  • Standard Gemini quotas (e.g., 5-hour / Flash baseline) should not be impacted or drained by Argon prompts.

Observed Behavior

  • Argon Weekly Limit: Remains completely frozen at 100%.
  • Standard Gemini Quota: Depletes at an accelerated rate, rapidly exhausting the general session/window limits.
  • Model Identity: Running /btw "What model are you?" returns that it is Argon, indicating the session metadata or system prompt recognizes Argon, but billing/token routing at the backend or proxy level appears misrouted to the standard Gemini quota pool.

Hi @honlam,

Could you please upgrade to the Antigravity latest CLI version v1.5.1. Let me know is the issue resolves or the issue still persists?

I have updated it the v1.5.1. Here are my findings: when setting /model Argon with thinking effort set to /high, requests appear to route to the general Gemini model pool rather than the dedicated Argon pool. When thinking effort is configured to low or medium, requests route to the Argon pool as expected.

Steps to Reproduce

  1. In agy, set the model: /model Argon
  2. Set thinking effort to high: /high
  3. Submit a query to inspect system identity (e.g., “Who are you?” or “What model are you?”)
  4. Note the returned identity and pool characteristics.
  5. Change thinking effort to /low or /medium and repeat the prompt.

Expected Behavior

Both high and low/medium thinking configurations for /model Argon should consistently route to the Argon pool, maintaining model-specific parameters and weights.

Actual Behavior

  • Effort = low / medium: Successfully routes to the Argon pool.

  • Effort = high: Diverts to the general Gemini pool. The model explicitly identifies itself as:

    “I am Gemini 4 Argon (configured at High reasoning effort), a large language model built by Google, running as Antigravity.”

Suggested Investigation

  • Check the backend router mapping for Argon when the thinking_effort = "high" flag is appended.
  • Verify whether the high-effort profile for Argon is missing a target endpoint and falling back to default Gemini endpoints.