Repository navigation
Fix: false rejections no longer kill models
- Upstream
sensitive_words_detected500s are prompt content-filter rejections, not outages: passed through with no retry, no model lock, no breaker penalty (IsContentFilterRejection). - Recovery probe carries no API key, so any 4xx (keyless 401) now clears the model lock instead of extending it forever. Previously one bad prompt hid deepseek-v4-flash from /v1/models for 10+ hours.
availableModelsnow counts healthy models, matching /v1/models.
Feat: default safe persona + gpt-6-astra
- Unset
INJECT_SYSTEM_PROMPTinjects a benign engineering-framing persona into every request (custom text overrides, empty disables). - Added
gpt-6-astra(1.05M ctx, 128K out) to the default model list; specs researched via OpenRouter.
Found while mapping: the provider flags the canonical 'You are a helpful assistant' string in user content — reword to 'coding/useful assistant'.