fix(antigravity): stabilize streaming and usage refresh

[dhaern]

Summary

This is an important Antigravity stability fix. It addresses two related production pain points:

• Antigravity/Gemini SSE payloads wrapped as { response: { candidates: [...] } } are now recognized by stream readiness, so valid Antigravity streams are not incorrectly treated as empty/zombie streams.
• Antigravity usage and model metadata refreshes are deduplicated with singleflight, filtered to user-callable public models, and manual Limits & Quotas refreshes bypass the short-lived automatic caches.
• The Antigravity usage behavior now lives on the real open-sse/services/usage.ts path, avoiding duplicated legacy fetcher logic.

Why this matters

Before this change, Antigravity could produce valid upstream streaming responses that failed the generic readiness gate because their useful content was nested under response.candidates. That made healthy requests look like empty streams and could cause unnecessary retries/fallback behavior.

The usage/model refresh path could also create avoidable churn by probing credits and model metadata repeatedly, especially when the dashboard or scheduled sync asked for fresh provider limits. This PR keeps automatic refreshes efficient while preserving fresh data for explicit manual refreshes.

Details

• Treats nested response, choices, and candidates containers as useful stream-content candidates.
• Adds a shared isUserCallableAntigravityModelId() filter based on the public Antigravity catalog and aliases.
• Filters Antigravity remote model discovery so internal/non-callable upstream models are not exposed as selectable models.
• Adds singleflight + short TTL behavior for automatic Antigravity metadata/credit probes.
• Ensures manual provider-limit refreshes pass forceRefresh through to bypass automatic Antigravity caches.
• Moves Antigravity usage coverage to the real usage service path and removes the duplicate legacy fetcher test.

Validation

• node --import tsx/esm --test tests/unit/usage-service-hardening.test.ts tests/unit/stream-readiness.test.ts tests/unit/antigravity-model-aliases.test.ts
• node --import tsx/esm --test tests/unit/provider-models-route.test.ts tests/unit/model-sync-route.test.ts
• npm run typecheck:core
• npm run typecheck:noimplicit:core
• npm run lint (passes with existing warnings only)
• git diff --check

No build or service restart was performed as part of this PR preparation.

[gemini-code-assist]

Code Review

This pull request implements caching for Antigravity model discovery, usage quotas, and credit probes to optimize performance, alongside a forceRefresh mechanism for manual usage updates. It refactors model filtering by centralizing user-callable logic and expands stream readiness checks to support additional response keys. Feedback identifies missing accessToken validation in the new caching functions that could lead to runtime errors and points out redundant logic in the stream readiness utility that can be simplified.

[gemini-code-assist]

The accessToken should be validated at the start of this function. If accessToken is null or undefined, the call to buildAntigravityUsageCacheKey on the next line will throw a runtime error when it attempts to call .substring(0, 16). Adding a guard clause here ensures the service handles missing credentials gracefully, consistent with the patterns used for other providers in this file.

Suggested change

	async function fetchAntigravityAvailableModelsCached(
	accessToken: string,
	projectId?: string | null,
	options: AntigravityUsageOptions = {}
	): Promise<unknown> {
	async function fetchAntigravityAvailableModelsCached(
	accessToken: string,
	projectId?: string | null,
	options: AntigravityUsageOptions = {}
	): Promise<unknown> {
	if (!accessToken) throw new Error("Access token is required");

[gemini-code-assist]

The accessToken should be validated before use. If it's missing, buildAntigravityUsageCacheKey will throw an error when calling .substring(). Since this credit probe is intended to be a best-effort background task, returning null early is a safe way to handle missing credentials without interrupting the main usage fetch flow.

Suggested change

	async function probeAntigravityCreditBalance(
	accessToken: string,
	accountId: string,
	projectId?: string | null,
	options: AntigravityUsageOptions = {}
	): Promise<number | null> {
	async function probeAntigravityCreditBalance(
	accessToken: string,
	accountId: string,
	projectId?: string | null,
	options: AntigravityUsageOptions = {}
	): Promise<number | null> {
	if (!accessToken) return null;

[gemini-code-assist]

The if block preceding this line (lines 64-72 in the full file) has become redundant. Since hasUsefulValue(payload) is now called unconditionally and it has been updated to recursively check for all relevant keys (including choices, candidates, and parts), the specific check for type property matches is no longer necessary. Removing that block would simplify the function while maintaining the same logic.

[dhaern]

@gemini-code-assist addressed in 0c60201.

• Added missing Antigravity access-token guards so usage refreshes fail gracefully instead of reaching substring/cache-key code with an undefined token.
• Kept the credit probe best-effort by returning null when credentials are missing.
• Simplified stream readiness by removing the now-redundant type-gated block and relying on the recursive useful-value detection.
• Added regression coverage for missing Antigravity tokens and OpenAI Responses text deltas.

Validation rerun:

• node --import tsx/esm --test tests/unit/usage-service-hardening.test.ts tests/unit/stream-readiness.test.ts tests/unit/antigravity-model-aliases.test.ts
• node --import tsx/esm --test tests/unit/provider-models-route.test.ts tests/unit/model-sync-route.test.ts
• npm run typecheck:core
• npm run typecheck:noimplicit:core
• npm run lint (existing warnings only)
• git diff --check

[gemini-code-assist]

Thank you for the update and for providing the validation results. The changes to stabilize Antigravity streaming, deduplicate usage refreshes, and implement proper access-token guards look solid. The added regression tests for missing tokens and text deltas are also appreciated. I have no further concerns regarding this implementation.

[diegosouzapw]

Thanks @dhaern for this great contribution! 🎉 The singleflight usage logic and stream content handling are a massive improvement for the Antigravity integration. The PR has been successfully integrated and will be included in the v3.7.5 release!