[christopher-s]

Summary

• Fix Antigravity model registry to match actual Google Cloud Code API model IDs (gemini-3-pro-high/low, gemini-3-flash, etc.)
• Prevent 429 on one model from locking out entire provider connection — use model-only lockModel() for passthrough providers
• Fix non-streaming requests by always streaming upstream and collecting SSE into JSON response
• Filter internal/special-purpose models from quota display (matching CLIProxyAPI exclusion list)
• Normalize bare Pro model IDs to -low tier (matching OpenClaw convention)
• Address SSE collection timeout: return 504 instead of silent 200 with empty content

Changes

Model Registry (providerRegistry.ts)

• Replace stale/wrong model IDs with correct ones from fetchAvailableModels API
• Add High/Low tier variants for Pro models
• Remove non-existent models (gpt-5, gpt-5-mini, gemini-2.0-flash)
• Remove internal-only models (gemini-2.5-pro — browser subagent)
• Add getPassthroughProviders() helper for provider type detection

429 Cascade Fix (chatCore.ts, auth.ts)

• chatCore: For passthrough providers, use lockModel() instead of connection-wide rateLimitedUntil on 429
• auth.ts: Same pattern as secondary defense in markAccountUnavailable
• Both use 2-minute default cooldown (aligned with COOLDOWN_MS.rateLimit)

Non-Streaming Requests (antigravity.ts)

• Always use streaming endpoint upstream to avoid 400 errors from Cloud Code API
• New collectStreamToResponse() method parses Gemini SSE and assembles OpenAI JSON response
• Returns 504 Gateway Timeout with finish_reason: "length" on timeout instead of silent 200

Quota Display (usage.ts)

• Add ANTIGRAVITY_EXCLUDED_MODELS set matching CLIProxyAPI's hardcoded list
• Filters: chat placeholders, tab-completion, thinking variant, browser subagent

Test plan

• Tested all 11 Antigravity models: 9 working, 2 quota-exhausted (but don't cascade)
• Verified 429 on one model doesn't lock out other models on same connection
• Verified bare gemini-3-pro normalizes to gemini-3-pro-low
• Verified non-streaming requests return proper JSON
• Verified SSE timeout returns 504 instead of silent 200

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the antigravity provider by removing its ag alias, updating the model catalog, and transitioning to an upstream streaming-only execution model where non-streaming requests are handled via local SSE collection. It also introduces logic for 'passthrough' providers to allow model-specific rate limiting and error handling. Feedback includes suggestions to optimize the SSE collection loop for efficiency and robustness, restore the ag alias to prevent breaking changes, and simplify the initialization of the passthrough provider registry.

[gemini-code-assist]

The current implementation of collectStreamToResponse has several areas for improvement:

1. Efficiency: A new Promise and event listener are created for the timeout in every iteration of the while loop. This can be optimized by creating the timeout promise once outside the loop.
2. Robustness: If the stream fails for reasons other than a timeout (e.g., a network error), the catch block logs the error but allows the function to "fall through" and return a 200 status with potentially incomplete or malformed JSON. It would be more robust to rethrow non-timeout errors so they can be handled as upstream failures (e.g., 502 Bad Gateway) by the caller.
3. Timeout Detection: Using timeout.aborted is a more reliable way to detect if the timeout signal was triggered compared to string matching on the error message.

    const collect = async () => {
      const chunks: string[] = [];
      let timedOut = false;
      const timeout = AbortSignal.timeout(SSE_COLLECT_TIMEOUT_MS);
      const timeoutPromise = new Promise<never>((_, reject) =>
        timeout.addEventListener("abort", () => reject(new Error("SSE collection timed out")), { once: true })
      );

      try {
        // eslint-disable-next-line no-constant-condition
        while (true) {
          if (signal?.aborted) throw new Error("Request aborted during SSE collection");
          const { done, value } = await Promise.race([
            reader.read(),
            timeoutPromise,
          ]);
          if (done) break;
          chunks.push(decoder.decode(value, { stream: true }));
        }
      } catch (err) {
        timedOut = timeout.aborted;
        const msg = err?.message || String(err);
        log?.warn?.("SSE_COLLECT", `Error collecting SSE stream: ${msg}`);
        if (!timedOut) throw err;
      }

[gemini-code-assist]

Removing the ag alias for the antigravity provider is a breaking change. Any existing client configurations, environment variables, or scripts relying on the ag/ prefix will stop working. Unless there is a specific conflict or a strong requirement to enforce the full name, it is recommended to keep the alias for backward compatibility.

Suggested change

	alias: undefined,
	alias: "ag",

[gemini-code-assist]

The initialization of _passthroughProviderIds can be simplified. The try-catch block is unnecessary as REGISTRY is defined in the same module and Object.values is a safe operation here. Additionally, if the initialization were to fail and return null, the current implementation would create a new Set on every call to getPassthroughProviders(), which is inefficient.

Suggested change

	const _passthroughProviderIds: Set<string> | null = (() => {
	try {
	const ids = new Set<string>();
	for (const entry of Object.values(REGISTRY)) {
	if (entry.passthroughModels) ids.add(entry.id);
	}
	return ids;
	} catch { return null; }
	})();
	export function getPassthroughProviders(): Set<string> {
	return _passthroughProviderIds ?? new Set<string>();
	}
	/** Set of provider IDs with passthroughModels enabled — 404s are model-specific, not account-level. */
	const _passthroughProviderIds = new Set(
	Object.values(REGISTRY)
	.filter((entry) => entry.passthroughModels)
	.map((entry) => entry.id)
	);
	export function getPassthroughProviders(): Set<string> {
	return _passthroughProviderIds;
	}

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	0

Changes Since Previous Review (3fad847 → ff15828)

This commit continues the cleanup of non-viable Antigravity models:

1. Image Registry - Removed gemini-2.5-flash-preview-image-generation and gemini-3.1-flash-image-preview from the Antigravity provider's model list. These models return empty responses and are not usable for image generation.

2. Model Specs - Reduced maxOutputTokens for gemini-3.1-pro-high and gemini-3.1-pro-low from 131072 to 65535. This aligns with the actual API limits for these models.

Both changes are appropriate cleanup based on findings from earlier commits in this PR.

Previous Issues Carried Forward

File	Line	Status
open-sse/executors/antigravity.ts	247	Open - SSE collection optimization suggestion
open-sse/config/providerRegistry.ts	362	Open - alias removal breaking change note
open-sse/config/providerRegistry.ts	1561	Open - getPassthroughProviders simplification suggestion

These remain unresolved from previous reviews but are not blockers for this commit.

Files Reviewed (2 files)

• open-sse/config/imageRegistry.ts - Removed non-viable image models from Antigravity
• src/shared/constants/modelSpecs.ts - Corrected maxOutputTokens limits for Gemini 3.1 models

[christopher-s]

Model Cleanup — Tested and Verified

After testing all models against a live Antigravity account, several models were found to be non-viable and have been removed:

Removed from available models list

These either return empty responses, have permanently exhausted quota, or are not functional for chat completions:

• gemini-3-pro-high/low — returns empty responses, quota unusable
• gemini-2.5-flash/flash-lite — quota always exhausted on free tier accounts
• gemini-3.1-flash-image-preview — preview variant, not functional

Hidden from quota UI (in addition to above)

• gemini-3-flash-agent — internal agent model, not user-callable
• gemini-3.1-flash-lite — not usable for chat

Final viable model list (7 models)

Model	Confirmed Working
claude-opus-4-6-thinking	Yes
claude-sonnet-4-6	Yes
gemini-3-flash	Yes
gemini-3.1-flash-image	Yes
gemini-3.1-pro-high	Yes
gemini-3.1-pro-low	Yes
gpt-oss-120b-medium	Yes

[diegosouzapw]

Merged manually into release/v3.4.2. Thanks for the contribution!