fix(combo): skip retries on all-rate-limited 429, fix thinking signature cross-provider 400

[benzntech]

Summary

Fixes combo routing when providers hit rate limits, and fixes cross-provider thinking signature errors.

Bug 1: Combo retries dead provider on 429 cooling down

When Claude Code hits its token limit, the provider returns 429 with model_cooldown/cooling down message. The combo treated this as a normal transient 429 and retried the same model 2x (wasting ~3s) before falling back. isAllAccountsRateLimitedResponse only matched 503 + "unavailable" — now also matches 429 + "cooling down"/"model_cooldown".

Fix: Expand pattern matching and skip transient retries when isAllAccountsRateLimited is true (both priority and round-robin paths).

Bug 2: lastStatus not set on quality-check failure

When a provider returns 200 but fails validateResponseQuality, lastStatus was never set, causing ALL_ACCOUNTS_INACTIVE instead of proper fallback to the next model.

Fix: Set lastStatus = 422 and lastError on quality-check failure path.

Bug 3: SSE event: prefix not detected

Anthropic SSE responses start with event: lines, not data:. validateResponseQuality rejected valid SSE as "response is not valid JSON".

Fix: Also accept text.startsWith("event:") and text.includes("\ndata:").

Bug 4: Invalid thinking signature across providers

When combo fallback switches providers, thinking block signatures from provider A are invalid for provider B, causing 400 "Invalid signature in thinking block".

Fix: Convert thinking blocks to redacted_thinking in outgoing requests — accepted without signature validation.

Bug 5: Clipboard copy broken on HTTP

window.isSecureContext guard prevented navigator.clipboard on HTTP origins, even though localhost is treated as secure context.

Fix: Remove isSecureContext guard, try clipboard API directly.

Test plan

• 6 new test cases covering all fixes
• All existing tests pass

[gemini-code-assist]

Code Review

This pull request enhances the combo routing engine by improving rate-limit detection, handling Claude thinking blocks across provider boundaries, and refining response quality validation. Specifically, it adds support for 429 'cooling down' status codes to trigger immediate fallbacks, converts Claude thinking blocks to 'redacted_thinking' to avoid signature validation errors, and relaxes SSE response validation to accept 'event:' prefixes. Feedback was provided regarding the assignment of lastError in the quality check failure paths; it is recommended to update lastError unconditionally to ensure the final error response reflects the most recent failure reason, maintaining consistency with other error handling paths in the service.

[gemini-code-assist]

The assignment to lastError here uses the logical OR operator (||), which means it will only be updated if lastError is currently null or empty. However, in the main error handling path (line 1713), lastError is updated unconditionally to reflect the most recent failure. To maintain consistency and ensure the client receives the most relevant error message when all models fail, this should be a direct assignment.

Suggested change

	lastError = lastError || `quality check failed: ${quality.reason}`;
	lastError = `quality check failed: ${quality.reason}`;

[gemini-code-assist]

Similar to the priority strategy, lastError should be updated unconditionally here to ensure the final error response reflects the most recent failure reason, consistent with the behavior at line 2061.

Suggested change

	lastError = lastError || `quality check failed: ${quality.reason}`;
	lastError = `quality check failed: ${quality.reason}`;

[diegosouzapw]

Manually merged valuable changes into release/v3.7.0 to avoid combo.ts conflicts with PR #1449

[diegosouzapw]

Code already fully integrated. Unable to retroactively merge due to branch ancestry conflicts, but all your code is in production. Thank you @benzntech!