openclaw-freerouter 2.0.1 → 2.1.0

package/README.md

@@ -7,6 +7,7 @@
 - 🔥 **Wasted money** — Running Claude Opus ($15/$75 per 1M tokens) for every message, even "what's 2+2?"
 - 🤷 **No control** — Can't switch models mid-conversation without editing config files and restarting
 - 📊 **Blind routing** — Your AI shows "freerouter/auto" instead of telling you which model actually answered
+- 🧠 **No adaptive thinking** — OpenClaw's built-in thinking levels (`/think low`, `/think high`) are just prompt hints. Anthropic deprecated manual thinking for Opus 4.6 — it now requires the native `thinking: { type: "adaptive" }` API parameter, which OpenClaw doesn't send
 - 🔧 **Complex setup** — Existing routers need separate servers, Docker, complex infra
 
 ## The Solution
@@ -14,6 +15,7 @@
 FreeRouter is an OpenClaw plugin that:
 - **Classifies every request in <1ms** using a 14-dimension weighted scorer (no LLM needed for classification)
 - **Routes to the cheapest model that can handle it** — Kimi for "hello", Opus for architecture design
+- **Sends native adaptive thinking** — Automatically passes `thinking: { type: "adaptive" }` to Anthropic's API for Opus 4.6, and `thinking: { type: "enabled", budget_tokens: N }` for Sonnet. No prompt hacks — real API-level thinking control that OpenClaw doesn't support natively
 - **Reports the real model name** — You see `anthropic/claude-opus-4-6`, not `freerouter/auto`
 - **Lets you override anytime** — Just say "use opus" in plain English
 
@@ -100,6 +102,36 @@ This prevents accidental switches when you're just talking *about* a model.
 | `haiku`, `haiku-4`, `haiku-4.5` | anthropic/claude-haiku-4-5 |
 | `kimi`, `kimi-k2`, `k2.5` | kimi-coding/kimi-for-coding |
 
+## Adaptive Thinking (Native API-Level)
+
+This is a key reason FreeRouter exists.
+
+**The problem:** OpenClaw's built-in `/think low|medium|high` commands are prompt-level hints — they add text like "think harder" to your prompt. This is unreliable and doesn't use Anthropic's actual thinking API. Worse, **Anthropic deprecated manual thinking (`type: "enabled"`) for Opus 4.6** — it now only supports `type: "adaptive"`, where the model decides how much to think based on the task.
+
+**What FreeRouter does:** Sends the real `thinking` parameter directly to Anthropic's API:
+
+| Model | Thinking Mode | What's Sent |
+|---|---|---|
+| Claude Opus 4.6 | **Adaptive** (always) | `thinking: { type: "adaptive" }` |
+| Claude Sonnet 4.5 | Enabled with budget | `thinking: { type: "enabled", budget_tokens: 4096 }` |
+| Others (Kimi, Haiku) | Off | No thinking parameter |
+
+**Why this matters:**
+- Adaptive thinking lets Opus 4.6 decide how much reasoning it needs — simple questions get quick answers, complex proofs get deep thinking chains
+- You get the full benefit of Claude's extended thinking without managing budgets
+- The `X-FreeRouter-Thinking` response header tells you exactly which thinking mode was used
+
+Configure in your plugin config:
+```json5
+"thinking": {
+  "adaptive": ["claude-opus-4-6"],           // Models that use adaptive (always-on)
+  "enabled": {
+    "models": ["claude-sonnet-4-5"],          // Models that get explicit thinking
+    "budget": 4096                             // Token budget for thinking
+  }
+}
+```
+
 ## How Routing Works
 
 1. You send a message → OpenClaw forwards to FreeRouter
@@ -154,7 +186,7 @@ After install, add to your `openclaw.json`:
   // 3. Plugin config
   "plugins": {
     "entries": {
-      "freerouter": {
+      "openclaw-freerouter": {
         "enabled": true,
         "config": {
           "port": 18801,
@@ -202,7 +234,7 @@ Then restart: `openclaw gateway restart`
 If port 18801 is in use, change it:
 
 ```json5
-{ "plugins": { "entries": { "freerouter": { "config": { "port": 18802 } } } } }
+{ "plugins": { "entries": { "openclaw-freerouter": { "config": { "port": 18802 } } } } }
 ```
 
 Set `"port": 0` to disable the HTTP proxy entirely.

package/openclaw.plugin.json

@@ -1,5 +1,5 @@
 {
-  "id": "freerouter",
+  "id": "openclaw-freerouter",
   "name": "FreeRouter",
   "version": "2.0.0",
   "description": "Smart LLM router — classify requests and route to the best model using your own API keys. 14-dimension weighted scoring, <1ms classification, configurable tiers.",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-freerouter",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "description": "Smart LLM router plugin for OpenClaw — classify requests and route to the best model using your own API keys. 14-dimension scoring, <1ms classification, per-prompt/session model switching.",
   "type": "module",
   "openclaw": {

package/src/cli.ts

@@ -16,7 +16,7 @@ import { buildPricingMap } from "./models.js";
 export function registerCli(api: any) {
   const logger = api.logger ?? console;
 
-  const getPluginConfig = () => api.config?.plugins?.entries?.freerouter?.config ?? {};
+  const getPluginConfig = () => api.config?.plugins?.entries?.["openclaw-freerouter"]?.config ?? {};
 
   api.registerCli(
     ({ program }: any) => {
@@ -162,8 +162,8 @@ export function registerCli(api: any) {
               enabled: { models: ["claude-sonnet-4-5"], budget: 4096 },
             },
           }, null, 2));
-          console.log("\n  Copy this into plugins.entries.freerouter.config in openclaw.json");
-          console.log("  Or run: openclaw config set plugins.entries.freerouter.config '{...}'");
+          console.log('\n  Copy this into plugins.entries["openclaw-freerouter"].config in openclaw.json');
+          console.log('  Or run: openclaw config set plugins.entries["openclaw-freerouter"].config \'{...}\'');
           console.log();
         });
 
@@ -335,7 +335,7 @@ export function registerCli(api: any) {
 
 Step 1: Add this to your openclaw.json under "plugins.entries":
 
-  "freerouter": {
+  "openclaw-freerouter": {
     "enabled": true,
     "config": ${JSON.stringify(pluginConfig, null, 6).split("\n").map((l, i) => i === 0 ? l : "    " + l).join("\n")}
   }

package/src/index.ts

@@ -22,7 +22,7 @@ export const id = "freerouter";
 export default function register(api: any) {
   const logger = api.logger ?? console;
   const getPluginConfig = () => {
-    const cfg = api.config?.plugins?.entries?.freerouter?.config ?? {};
+    const cfg = api.config?.plugins?.entries?.["openclaw-freerouter"]?.config ?? {};
     return cfg;
   };
 
@@ -55,7 +55,7 @@ export default function register(api: any) {
         server.on("error", (err: any) => {
           if (err.code === "EADDRINUSE") {
             logger.error(`[freerouter] Port ${port} is already in use. Run: openclaw freerouter doctor`);
-            logger.error(`[freerouter] To change port: set plugins.entries.freerouter.config.port in openclaw.json`);
+            logger.error(`[freerouter] To change port: set plugins.entries["openclaw-freerouter"].config.port in openclaw.json`);
           } else {
             logger.error(`[freerouter] Server error: ${err.message}`);
           }

package/src/service.ts

@@ -624,11 +624,15 @@ export function createProxyServer(options: ProxyOptions): { server: Server; stat
     }
 
     // Determine thinking mode
+    // Opus 4.6 REQUIRES adaptive thinking — manual mode is deprecated by Anthropic
+    // Other models can use enabled(budget) for explicit thinking control
     const thinkingCfg = pluginConfig.thinking as any;
     const adaptivePatterns = thinkingCfg?.adaptive ?? ["claude-opus-4-6"];
     const enabledCfg = thinkingCfg?.enabled;
 
-    if (adaptivePatterns.some((p: string) => routedModel.includes(p)) && (tier === "COMPLEX" || tier === "REASONING")) {
+    if (adaptivePatterns.some((p: string) => routedModel.includes(p))) {
+      // Adaptive thinking models (e.g., Opus 4.6) — always use adaptive
+      // Anthropic deprecated manual thinking for Opus 4.6
       thinkingMode = "adaptive";
     } else if (enabledCfg?.models?.some((p: string) => routedModel.includes(p))) {
       thinkingMode = `enabled(${enabledCfg.budget ?? 4096})`;