feat(v0.7.6): Dollar Savings Tracker, Provider Leaderboard, SSE Live …

CHANGELOG.md

@@ -1,3 +1,49 @@
+# v0.7.6 (2026-07-25)
+
+## Features
+- **Dollar Savings Tracker**: hero card di Overview menampilkan total $ Saved (lifetime) dengan per-mechanism breakdown (RTK/Headroom/Pxpipe/Cache/Caveman/Ponytail) + "With vs Without ExtremeRouter" comparison.
+- **Provider Performance Leaderboard**: sortable table di Overview dengan ranking per-provider: Requests, Tokens, TTFT, P95 Latency, Success Rate, Cost. Period selector (24h/7d/30d). Custom provider names resolved dari providerNodes.
+- **SSE Live Dashboard**: unified `/api/dashboard/stream` SSE endpoint (stats + breaker + health). Real-time KPI updates via `useDashboardStream` hook dengan auto-reconnect.
+- **In-App Notification System**: bell icon di header dengan unread badge, dropdown feed, localStorage history persistence. Push notifications untuk provider down/recovered/health degraded/rate limited.
+- **Cross-Transport Fallback**: OpenAI endpoint timeout/5xx → auto-retry Anthropic endpoint (body re-translated). Fresh AbortController per attempt. Applied ke GLM, hcnsec, Bynara, InxoraStudio, CommandCode, Infron, AgentRouter.
+- **New Provider: Bynara**: multi-model router (OpenAI + Anthropic + Responses + Image gen/edits).
+- **New Provider: InxoraStudio Labs (API)**: OpenAI + Anthropic multi-endpoint.
+- **New Provider: InxoraStudio Labs (Web)**: JWT web chat executor (3-step flow), profile badge, auto model discovery.
+- **New Provider: Infron AI**: 457+ models, OpenAI + Anthropic, quota tracker (credit balance).
+- **New Provider: AgentRouter**: GLM/GPT/Claude, OpenAI + Anthropic, custom pricing model discovery.
+- **Updated Provider: Command Code**: migrated dari custom `/alpha/generate` ke standard OpenAI/Anthropic provider API, live model discovery (47 models).
+- **Quota Tracker: Grok Web (Subscription)**: SSO cookie rate-limits (Fast/Thinking/Heavy tiers).
+- **Quota Tracker: Infron AI**: credit balance via `/v1/balance`.
+
+## Fixes
+- **Kiro `REQUEST_BODY_INVALID`**: normalize `(level)` suffix sebelum resolving synthetic variants. Map explicit levels ke native Kiro effort fields hanya untuk supported families (Claude 5 / GPT-5.6). `thinkingLevels.js` server-side gating.
+- **Responses API accumulator**: shared `ResponsesAccumulator` untuk streaming + forced non-stream. Alias-safe tool reconstruction. `preferComplete` untuk snapshot merge. Exactly-once terminal semantics.
+- **Responses `incomplete`/`cancelled`**: terminal events sekarang di-recognize di `responsesStreamHelpers.js`.
+- **Antigravity "model turn" 400**: strip trailing `role:"model"` turns dari `contents[]` sebelum kirim ke Google Cloud Code.
+- **Antigravity tier routing**: `gemini-3.6-flash-high/medium/low` dengan `upstreamModelId: "gemini-3.6-flash-tiered(high)"` + parser di `getModelUpstreamId`.
+- **Cloud Code endpoint isolation**: gemini-cli→`cloudcode-pa`, antigravity→`daily-cloudcode-pa` via per-provider `CLOUD_CODE_API` map.
+- **Circuit breaker `releaseBreakerProbe`**: `monitors`→`breakers` ReferenceError fix.
+- **Combo `comboStickyLimit`**: wrong key `comboStickyLimit`→`comboStickyRoundRobinLimit`.
+- **Combo `comboStrategies` lost-update**: deep-merge per combo-name + null delete-signal.
+- **Cross-transport abort**: fresh AbortController untuk alternate attempt.
+- **Cross-transport error logging**: alternate failure error tidak lagi di-swallow silently.
+- **Responses output getter**: dedup alias-registered tools + iterate string-keyed items.
+- **Notification spam**: `health_update` dan `usage_update` tidak lagi jadi notification.
+- **Notification raw JSON**: `formatEventMessage` return `null` untuk unknown types.
+- **Notification duplicate**: `lastProcessedTs` useRef guard + dedup.
+- **NotificationBell setState-during-render**: `markAllRead` deferred via `requestAnimationFrame`.
+- **NotificationBell key collision**: `idCounter` seeded dari `Date.now()` + max history id.
+- **hcnsec model discovery**: tambah `"openai"` filter + authenticated suggested-models proxy.
+- **WEB_COOKIE_PROVIDERS modelsFetcher**: page.js sekarang include webCookie providers di model discovery chain.
+- **InxoraStudio profile badge**: `InxoraProfile` component + wiring di ConnectionsCard.
+
+## Improvements
+- **Dead code cleanup**: hapus 12 dead files (1,705 LOC) + dead `convertResponsesApiFormat` function.
+- **Junk dependency removed**: `fs` (0.0.1-security placeholder package) dari dependencies.
+- **Command Code executor**: hapus custom `CommandCodeExecutor`, pakai `DefaultExecutor` (OpenAI/Anthropic native).
+- **Model discovery**: `suggested-models` route sekarang support `connectionId` untuk authenticated discovery.
+- **Validate route**: hcnsec, InxoraStudio-web, inxorastudio (API) validate cases.
+
 # v0.7.5 (2026-07-22)
 
 ## Features

cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rsalmn/extremerouter",
-  "version": "0.7.5",
+  "version": "0.7.6",
   "description": "ExtremeRouter CLI - AI Gateway control plane",
   "bin": {
     "extremerouter": "./cli.js"

open-sse/config/appConstants.js

@@ -132,12 +132,28 @@ export const ANTIGRAVITY_HEADERS = {
   "User-Agent": `antigravity/2.1.1 ${platform()}/${arch()}`
 };
 
-// Cloud Code Assist API
+// Cloud Code Assist API endpoints differ by client ecosystem.
+// gemini-cli uses the production host (cloudcode-pa.googleapis.com) while
+// antigravity uses the sandbox host (daily-cloudcode-pa.googleapis.com) per
+// decolua/9router commit 190020c. Keyed by provider id; projectId.js resolves
+// the right endpoints per-connection.
 export const CLOUD_CODE_API = {
-  loadCodeAssist: "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist",
-  onboardUser: "https://cloudcode-pa.googleapis.com/v1internal:onboardUser",
+  "gemini-cli": {
+    loadCodeAssist: "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist",
+    onboardUser: "https://cloudcode-pa.googleapis.com/v1internal:onboardUser",
+  },
+  antigravity: {
+    loadCodeAssist: "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist",
+    onboardUser: "https://daily-cloudcode-pa.googleapis.com/v1internal:onboardUser",
+  },
 };
 
+// Back-compat flat accessor: callers that haven't been migrated to pass a
+// provider default to the gemini-cli (production) endpoints — the previous
+// behavior before this map was split.
+export const CLOUD_CODE_LOAD_CODEASSIST = CLOUD_CODE_API["gemini-cli"].loadCodeAssist;
+export const CLOUD_CODE_ONBOARD_USER = CLOUD_CODE_API["gemini-cli"].onboardUser;
+
 export const LOAD_CODE_ASSIST_HEADERS = {
   "Content-Type": "application/json",
   "User-Agent": "google-api-nodejs-client/9.15.1",

open-sse/config/kiroConstants.js

@@ -16,6 +16,7 @@
  */
 
 import { extractThinking } from "../translator/concerns/thinkingUnified.js";
+import { parseSuffix } from "../translator/concerns/thinkingUnified.js";
 import { effortToBudget } from "../translator/concerns/thinking.js";
 
 export const KIRO_AGENTIC_SUFFIX = "-agentic";
@@ -234,6 +235,83 @@ export function resolveKiroModel(model) {
   return { upstream, agentic, thinking };
 }
 
+/**
+ * Resolve the per-family effort path for a Kiro upstream model id.
+ *
+ * Kiro upstream (CodeWhisperer) accepts native effort fields ONLY for specific
+ * model families. Other families (legacy Claude 4.5, GLM, DeepSeek, Qwen,
+ * MiniMax) reject native effort as REQUEST_BODY_INVALID — their reasoning flows
+ * solely via the `<thinking_mode>` system tag injection.
+ *
+ * Returns one of:
+ *   - "kiro-claude"  → Claude 5 family (Sonnet 5, Haiku 5, Opus 5)
+ *   - "kiro-gpt"     → GPT-5.6 family (Sol, Terra, Luna)
+ *   - null           → not supported (tag-only reasoning)
+ *
+ * @param {string} upstreamModel - clean upstream id (after resolveKiroModel)
+ * @returns {string|null}
+ */
+export function resolveKiroEffortPath(upstreamModel) {
+  if (typeof upstreamModel !== "string" || !upstreamModel) return null;
+  const lower = upstreamModel.toLowerCase();
+  // Claude 5 family (NOT Claude 4.x — those are legacy and reject native effort).
+  // Match claude-{opus,sonnet,haiku}-5 explicitly to exclude 4.5/4.6/etc.
+  if (/^claude-(opus|sonnet|haiku)-5(\b|$|-)/.test(lower)) return "kiro-claude";
+  // GPT-5.6 family.
+  if (/^gpt-5\.6-/.test(lower)) return "kiro-gpt";
+  return null;
+}
+
+/**
+ * Resolve a Kiro model id after consuming the generic `model(level)` suffix
+ * that the dashboard thinking-level picker appends.
+ *
+ * The suffix is an ExtremeRouter request override (parsed by parseSuffix),
+ * NOT part of Kiro's upstream model id. We strip it before resolving synthetic
+ * variants (-thinking / -agentic) so the upstream modelId stays clean.
+ *
+ * @param {string} model
+ * @returns {{ model: string, upstream: string, agentic: boolean, thinking: boolean, thinkingOverride: object|null }}
+ */
+export function resolveKiroModelIntent(model) {
+  const { cleanModel, override } = parseSuffix(model);
+  return {
+    model: cleanModel,
+    ...resolveKiroModel(cleanModel),
+    thinkingOverride: override,
+  };
+}
+
+/**
+ * Apply a parsed `model(level)` override to a request body without mutating the
+ * caller's body. Translates the override into the appropriate thinking shape so
+ * downstream consumers (resolveKiroThinkingBudget, the system-tag injector) see
+ * a consistent intent regardless of whether it came from the body fields or the
+ * model-name suffix.
+ *
+ * @param {object} body
+ * @param {object|null} override - { mode: "level"|"budget"|"auto"|"none", ... }
+ * @returns {object} new body with thinking fields set
+ */
+export function applyKiroThinkingOverride(body, override) {
+  if (!override) return body;
+
+  const next = { ...body };
+  if (override.mode === "budget") {
+    delete next.output_config;
+    delete next.reasoning_effort;
+    delete next.reasoning;
+    next.thinking = { type: "enabled", budget_tokens: override.budget };
+    return next;
+  }
+
+  next.output_config = {
+    ...(body.output_config || {}),
+    effort: override.mode === "level" ? override.level : override.mode,
+  };
+  return next;
+}
+
 /**
  * Build the magic system-prompt prefix that turns Kiro reasoning on.
  * Same shape as CLIProxyAPIPlus.

open-sse/config/providerModels.js

@@ -48,7 +48,24 @@ export function getModelType(aliasOrId, modelId) {
 export function getModelUpstreamId(aliasOrId, modelId) {
   const models = PROVIDER_MODELS[aliasOrId];
   const found = models?.find(m => m.id === modelId);
-  if (found?.upstreamModelId) return found.upstreamModelId;
+  if (found?.upstreamModelId) {
+    // Tier-preset suffix support: a model entry can carry a preset tag in its
+    // upstreamModelId, e.g. "gemini-3.6-flash-tiered(high)" for tiered routing.
+    // When the caller appended an effort/tier suffix (e.g. "-high"), we merge
+    // the caller's suffix with the preset; otherwise the preset wins on its own.
+    // Port of decolua/9router commit 190020c (isolates tiered model routing).
+    const resolvedId = found.upstreamModelId;
+    const presetMatch = resolvedId.match(/\([^()]+\)\s*$/);
+    const presetSuffix = presetMatch?.[0] || "";
+    if (presetSuffix) {
+      const resolvedBase = resolvedId.slice(0, presetMatch.index).trim();
+      // Caller suffix (passed as modelId tail) takes precedence if present.
+      const callerSuffixMatch = typeof modelId === "string" ? modelId.match(/-(high|medium|low)$/i) : null;
+      const callerTier = callerSuffixMatch?.[0] ? `(${callerSuffixMatch[0].slice(1).toLowerCase()})` : "";
+      return resolvedBase + (callerTier || presetSuffix);
+    }
+    return resolvedId;
+  }
   if (aliasOrId === "cx" && typeof modelId === "string" && modelId.endsWith(CODEX_REVIEW_SUFFIX)) {
     return modelId.slice(0, -CODEX_REVIEW_SUFFIX.length);
   }

open-sse/executors/antigravity.js

@@ -195,6 +195,17 @@ export class AntigravityExecutor extends BaseExecutor {
       return c;
     });
 
+    // Google Cloud Code rejects requests where the last content turn is from
+    // the model ("Requests ending with a model turn are not supported"). This
+    // happens when the client sends a conversation history that ends with an
+    // assistant message (e.g. Claude Code sending the previous response as
+    // context). Strip trailing model turns so the last entry is always user.
+    if (Array.isArray(contents) && contents.length > 0) {
+      while (contents.length > 1 && contents[contents.length - 1]?.role === "model") {
+        contents.pop();
+      }
+    }
+
     // Sanitize tool schemas and function names before sending to Antigravity.
     let tools = body.request?.tools;
 

open-sse/executors/index.js

@@ -15,7 +15,6 @@ import { OpenCodeGoExecutor } from "./opencode-go.js";
 import { GrokWebExecutor } from "./grok-web.js";
 import { PerplexityWebExecutor } from "./perplexity-web.js";
 import { OllamaLocalExecutor } from "./ollama-local.js";
-import { CommandCodeExecutor } from "./commandcode.js";
 import { XiaomiTokenplanExecutor } from "./xiaomi-tokenplan.js";
 import { MimoFreeExecutor } from "./mimo-free.js";
 import { CodeBuddyExecutor } from "./codebuddy-cn.js";
@@ -30,6 +29,7 @@ import { BlackboxWebExecutor } from "./blackbox-web.js";
 import { ZenmuxFreeExecutor } from "./zenmux-free.js";
 import { ApiAirforceExecutor } from "./api-airforce.js";
 import { FreeBuffWebExecutor } from "./freebuff-web.js";
+import { InxorastudioWebExecutor } from "./inxorastudio-web.js";
 import { PerplexityAgentExecutor } from "./perplexity-agent.js";
 import { QwenCloudExecutor } from "./qwencloud.js";
 import { T3ChatWebExecutor } from "./t3-web.js";
@@ -73,7 +73,7 @@ const executors = {
   "grok-web": new GrokWebExecutor(),
   "perplexity-web": new PerplexityWebExecutor(),
   "ollama-local": new OllamaLocalExecutor(),
-  commandcode: new CommandCodeExecutor(),
+  // commandcode now uses DefaultExecutor (OpenAI/Anthropic native endpoints)
   "xiaomi-tokenplan": new XiaomiTokenplanExecutor(),
   "mimo-free": new MimoFreeExecutor(),
   mmf: new MimoFreeExecutor(), // Alias for mimo-free
@@ -103,6 +103,7 @@ const executors = {
   "zenmux-free": new ZenmuxFreeExecutor(),
   "api-airforce": new ApiAirforceExecutor(),
   "freebuff-web": new FreeBuffWebExecutor(),
+  "inxorastudio-web": new InxorastudioWebExecutor(),
   "perplexity-agent": new PerplexityAgentExecutor(),
   "qwencloud": new QwenCloudExecutor(),
   lmarena: new LMArenaExecutor(),
@@ -142,7 +143,7 @@ export { OpenCodeGoExecutor } from "./opencode-go.js";
 export { GrokWebExecutor } from "./grok-web.js";
 export { PerplexityWebExecutor } from "./perplexity-web.js";
 export { OllamaLocalExecutor } from "./ollama-local.js";
-export { CommandCodeExecutor } from "./commandcode.js";
+// CommandCodeExecutor removed — provider now uses standard OpenAI/Anthropic endpoints (DefaultExecutor)
 export { XiaomiTokenplanExecutor } from "./xiaomi-tokenplan.js";
 export { MimoFreeExecutor } from "./mimo-free.js";
 export { CodeBuddyExecutor } from "./codebuddy-cn.js";
@@ -171,6 +172,7 @@ export { HuggingChatExecutor } from "./huggingchat.js";
 export { ZenmuxFreeExecutor } from "./zenmux-free.js";
 export { ApiAirforceExecutor } from "./api-airforce.js";
 export { FreeBuffWebExecutor } from "./freebuff-web.js";
+export { InxorastudioWebExecutor } from "./inxorastudio-web.js";
 export { PerplexityAgentExecutor } from "./perplexity-agent.js";
 export { QwenCloudExecutor } from "./qwencloud.js";
 export { LMArenaExecutor } from "./lmarena.js";