feat: add usage tracking with live throughput rates and analytics dashboard

fix: route numbered OpenAI-compatible instance model refs

feat: numbered OpenAI-compatible endpoints admin page

feat: add generic OpenAI-compatible endpoint provider

feat: admin dedicated pages, pool key management, keypool quotas (#1350)

feat: add key management and provider keypool (# new feature)

Make OpenCode Zen mappings explicit (#1342)

Restore OpenCode reasoning replay for tool turns (#1328)

Restore OpenAI subscription generation compatibility (#1326)

Keep tool names portable across OpenAI transports (#1309)

Add Azure OpenAI provider (#1284)

fix: detect NIM 500 context overflow as context_window_exceeded (#1280)

Add ChatGPT OAuth as an OpenAI Codex provider (#1279)

Make Kilo discovery and reasoning protocol-correct (#1274)

feat: add Kilo.ai provider (#1273)

Close interrupted tool turns before new user input (#1271)

Preserve gateway model identity in Anthropic responses (#1270)

Contain MiniMax native tool markup across NIM streams (#1259)

Report provider permission denials as HTTP 403 (#1249)

Recover MiniMax-M3 native tool calls on NVIDIA NIM (#1246)

Make ProviderModelInfo the sole model-catalog contract (#1222)

Make provider model discovery the sole catalog owner (#1221)

Recover Claude sessions from provider context overflow (#1215)

Give Google reasoning controls one owner (#1211)

Coordinate provider recovery under concurrent load (#1205)

Add Google Vertex AI with renewable ADC (#1193)

Add Amazon Bedrock Mantle support (#1192)

Add Kimi Code subscription support (#1183)

Honor reasoning tiers on numeric-budget providers (#1150)

Make reasoning policy provider-neutral and client-aware (#1148)

Fix Cerebras reasoning serialization (#1138)

Preserve DeepSeek cache prefixes across tool turns (#1126)

Add searchable model selection to the Admin UI (#1121)

Add Ollama Cloud as a first-class provider (#1106)

Consolidate OpenAI-compatible provider adapters (#1069)

Remove the native Anthropic provider transport (#1067)

Retry degraded NVIDIA NIM functions as overloads (#1064)

Make response stream lifetimes explicit (#1060)

Make provider stream ownership exact and typed (#1059)

Make provider backoff admission monotonic (#1053)

Make provider capabilities semantic and typed (#1047)

Canonicalize provider failure and retry ownership (#1046)

Introduce a typed application boundary for provider execution (#1045)

Move protocol models to their protocol owners (#1044)

Replace global runtime resources with explicit ownership (#1042)

Establish single-owner runtime with stream-safe provider hot swaps (#1036)

Restore protocol-correct provider errors without client retry loops (#1033)

Move runtime packages under src namespace (#1029)