@madgagarin/pi-agentrouter 1.2.1 → 1.4.0

package/README.md

@@ -3,8 +3,13 @@
 [![npm version](https://img.shields.io/npm/v/@madgagarin/pi-agentrouter.svg?color=blue)](https://www.npmjs.com/package/@madgagarin/pi-agentrouter)
 [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
 [![Pi Plugin](https://img.shields.io/badge/Pi-Extension-purple.svg)](https://pi.dev)
+**`pi-agentrouter`** seamlessly connects 4 powerful flagship AI coding models (**GPT-5.6 Sol**, **Claude Opus 4.8**, **Claude Opus 5**, and **DeepSeek V4F**) to your [Pi Coding Agent](https://pi.dev) using a single, unified API key from [AgentRouter](https://agentrouter.org).
 
-A native, zero-config extension for the [Pi Coding Agent](https://pi.dev) that connects and optimizes [AgentRouter](https://agentrouter.org) models (GPT-5.6 Sol, Claude Opus 4.8 / 5) with built-in WAF rate-limit pacing, client fingerprint preservation, prompt caching, and seamless compaction.
+### 💡 Why use this plugin?
+* 🔑 **All 4 Top Models in One Place:** Instantly switch between GPT-5.6 Sol, Claude Opus 4.8, Claude Opus 5, and DeepSeek V4F without managing separate subscriptions or multiple API keys.
+* 💰 **Save up to 80% on Tokens:** Smart prompt caching and session affinity drastically cut token usage and speed up responses.
+* ⚡ **Zero-Config Setup:** Installs in seconds — all models are automatically registered with native reasoning, tool calling, and thinking support.
+* 🛡️ **Built-in Stability:** Automatic cross-process pacing and authentication guards eliminate 401 and 429 errors during long coding sessions and multi-agent tasks.
 
 ---
 
@@ -12,7 +17,7 @@ A native, zero-config extension for the [Pi Coding Agent](https://pi.dev) that c
 
 AgentRouter is a non-profit AI API gateway providing unified access to cutting-edge models:
 
-* **Increased Trial Credits:** By signing up through the referral link below, you unlock **increased trial credits (up to $175 / extra +$50 bonus)** to explore GPT-5.6 Sol and Claude Opus models.
+* **Increased Trial Credits:** By signing up through the referral link below, you unlock **increased trial credits (up to $175 / extra +$50 bonus)** to explore GPT-5.6 Sol, Claude Opus, and DeepSeek models.
 * **Support the Author:** Using this link directly supports the author and the ongoing maintenance of this open-source plugin!
 
 👉 **[Sign up on AgentRouter (Referral Link with Bonus)](https://agentrouter.org/register?aff=34dc)** 👈
@@ -23,15 +28,18 @@ AgentRouter is a non-profit AI API gateway providing unified access to cutting-e
 
 ## ✨ Features
 
-- 🔑 **Unified API Key**: Use a single `sk-...` (or custom token) for all models—both OpenAI GPT and Anthropic Claude endpoints are authenticated seamlessly.
+- 🔑 **Unified API Key**: Use a single `sk-...` (or custom token) for all models—both OpenAI GPT, DeepSeek, and Anthropic Claude endpoints are authenticated seamlessly.
 - 🚀 **Zero-Config Model Auto-Registration**: Automatically registers:
   - `agentrouter-openai/gpt-5.6-sol` (1M Context Window, Native Reasoning, Session Affinity).
-  - `agentrouter-clode/claude-opus-4-8` (512K Context Window, Adaptive Thinking).
-  - `agentrouter-clode/claude-opus-5` (1M Context Window, Adaptive Thinking).
-- ⚡ **Optimized Prompt Caching**: Bundles and configures session affinity headers and thinking structures, achieving **>80% cache hit rates** on consecutive turns.
-- 🛡️ **WAF & Rate-Limit Pacing**: Smart 2.5s throttling applied specifically to AgentRouter calls to avoid cloud WAF / 405 rate-limit blocks.
+  - `agentrouter-openai/deepseek-v4f` (128K Context Window, Reasoning Support, Session Affinity).
+  - `agentrouter-clode/claude-opus-4-8` (512K Context Window, Adaptive Thinking, Empty Signature Compat).
+  - `agentrouter-clode/claude-opus-5` (1M Context Window, Adaptive Thinking, Empty Signature Compat).
+- 🛡️ **Cross-Process Request Pacing**: Shared file-based rate limiter (`~/.pi/agent/.agentrouter-pacing`) synchronizing delays across concurrent subagents, background workers, and the main Pi process to prevent WAF burst blocks.
+- 🔒 **Transport-Level Root Prompt Guard**: Enforces canonical `pi-code` harness header at `index: 0` on every outbound request (`before_provider_request`), preventing 401 unauthorized client errors across multi-turn tool executions.
+- ⚡ **Optimized Prompt Caching & Dynamic Bypass**: Automatically disables aggressive prompt rewriting on AgentRouter routes (`PI_CACHE_OPTIMIZER_NO_PROMPT_REWRITE=1`) while preserving session affinity headers, achieving **>80% cache hit rates** on consecutive turns.
+- 📋 **Package Priority Guard & Auto-Fix**: Automatically checks package order in `settings.json` on startup and offers interactive one-click placement directly above `pi-cache-optimizer`.
 - 📦 **Seamless Compaction (Fixes 401)**: Solves the `401 unauthorized client` error during `/compact` by ensuring valid Pi client fingerprint headers are passed during summarization.
-- ⌨️ **Interactive Terminal Controls**: Change API keys, adjust throttle pacing, and cycle through models or thinking depths directly in the TUI.
+- ⌨️ **Interactive Terminal Controls**: Change API keys, adjust throttle pacing, fix package priority, and cycle through models or thinking depths directly in the TUI.
 
 ---
 
@@ -83,6 +91,7 @@ To enable quick model cycling with `Ctrl+P` and set default thinking levels, add
   "defaultThinkingLevel": "medium",
   "enabledModels": [
     "agentrouter-openai/gpt-5.6-sol",
+    "agentrouter-openai/deepseek-v4f",
     "agentrouter-clode/claude-opus-4-8",
     "agentrouter-clode/claude-opus-5"
   ]
@@ -95,7 +104,7 @@ To enable quick model cycling with `Ctrl+P` and set default thinking levels, add
 
 | Shortcut | Action | Description |
 | :--- | :--- | :--- |
-| `Ctrl + P` | 🔄 **Next Model** | Cycles to the next model (`gpt-5.6-sol` ➔ `claude-opus-4-8` ➔ `claude-opus-5`). |
+| `Ctrl + P` | 🔄 **Next Model** | Cycles to the next model (`gpt-5.6-sol` ➔ `deepseek-v4f` ➔ `claude-opus-4-8` ➔ `claude-opus-5`). |
 | `Shift + Ctrl + P` | 🔄 **Previous Model** | Cycles to the previous model. |
 | `Shift + Tab` | 🧠 **Cycle Thinking Level** | Toggles reasoning depth: `off` ➔ `minimal` ➔ `low` ➔ `medium` ➔ `high`. |
 | `Ctrl + T` | 👁 **Toggle Thinking Visibility** | Collapses or expands thinking/reasoning blocks on screen. |
@@ -105,9 +114,10 @@ To enable quick model cycling with `Ctrl+P` and set default thinking levels, add
 
 ## 🛠️ In-Chat Commands
 
-* `/agentrouter` — View current plugin status, active model, masked key, and pacing interval.
+* `/agentrouter` — View current plugin status, active model, priority position, masked key, and pacing interval.
 * `/agentrouter key <your-key>` — Update API key for all AgentRouter models on the fly.
-* `/agentrouter pacing <ms>` — Adjust the minimum delay between consecutive requests (default: `2500` ms).
+* `/agentrouter pacing <ms>` — Adjust the minimum delay between consecutive requests (default: `3500` ms).
+* `/agentrouter fix-order` — Automatically reorder `settings.json` packages to place this plugin directly before `pi-cache-optimizer`.
 * `/compact` — Compress conversation history safely without 401 authorization errors.
 
 ---
@@ -117,7 +127,7 @@ To enable quick model cycling with `Ctrl+P` and set default thinking levels, add
 #### Q: Why does `/compact` fail on raw proxy configurations?
 AgentRouter performs client fingerprint verification. Raw summarization requests without Pi's standard prompt signatures get rejected with `401 unauthorized client`. This plugin intercepts the `session_before_compact` event and automatically injects proper authentication signatures.
 
-#### Q: Does the 2.5s pacing delay affect local or other cloud models?
+#### Q: Does the 3.5s pacing delay affect local or other cloud models?
 No. The pacing logic specifically filters for AgentRouter endpoints (`isAgentRouter`). Native OpenAI, Anthropic, Gemini, or local models run at full speed without delay.
 
 #### Q: How to use custom subagents with AgentRouter?

package/index.ts

@@ -89,7 +89,7 @@ export function fixPackagePriorityInSettings(): boolean {
 
 const initialConfig = loadConfig();
 let currentApiKey = normalizeApiKey(process.env.AGENTROUTER_API_KEY || initialConfig.apiKey || "");
-let minIntervalMs = initialConfig.minIntervalMs ?? 2500;
+let minIntervalMs = initialConfig.minIntervalMs ?? 3500;
 const PACING_FILE = path.join(process.env.HOME || "", ".pi/agent/.agentrouter-pacing");
 
 export function getLastRequestEndTime(): number {
@@ -118,6 +118,53 @@ export function isAgentRouter(providerName?: string, baseUrl?: string): boolean
   return false;
 }
 
+export const CANONICAL_PI_HEADER =
+  "You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files.";
+
+/**
+ * Enforces that the canonical pi-code system prompt signature is strictly at index 0.
+ * If another plugin or wrapper prepended text before the canonical header, it reorders
+ * the header to the very top and shifts the injected prefix right after it.
+ * If the header is missing entirely (e.g. from replace mode), it prepends the canonical header.
+ * This guarantees both WAF client authentication and stable prompt cache prefix matching.
+ */
+export function enforceCanonicalRootPrompt(systemPrompt: string | any[] | undefined): string | any[] {
+  if (!systemPrompt) {
+    return CANONICAL_PI_HEADER;
+  }
+
+  if (typeof systemPrompt === "string") {
+    const text = systemPrompt.trim();
+    const piHeaderRegex = /(?:You are [^\n\r]*operating inside pi[^\n\r]*\n?|You are (?:pi|Pi)[^\n\r]*\n?)/i;
+
+    if (piHeaderRegex.test(text)) {
+      const match = text.match(piHeaderRegex);
+      if (match && match.index !== undefined && match.index > 0) {
+        const header = match[0].trim();
+        const prefix = text.slice(0, match.index).trim();
+        const rest = text.slice(match.index + match[0].length).trim();
+        return `${header}\n\n${prefix}${rest ? "\n\n" + rest : ""}`;
+      }
+      return text;
+    } else {
+      return `${CANONICAL_PI_HEADER}\n\n${text}`;
+    }
+  }
+
+  if (Array.isArray(systemPrompt)) {
+    if (systemPrompt.length === 0) {
+      return [{ type: "text", text: CANONICAL_PI_HEADER }];
+    }
+    const firstBlock = systemPrompt[0];
+    if (firstBlock && typeof firstBlock.text === "string") {
+      firstBlock.text = enforceCanonicalRootPrompt(firstBlock.text) as string;
+    }
+    return systemPrompt;
+  }
+
+  return systemPrompt;
+}
+
 export default function (pi: ExtensionAPI) {
   function registerAgentRouterProviders(apiKey: string): void {
     pi.registerProvider("agentrouter-openai", {
@@ -141,6 +188,18 @@ export default function (pi: ExtensionAPI) {
             sendSessionAffinityHeaders: true,
           },
         },
+        {
+          id: "deepseek-v4f",
+          name: "deepseek-v4f",
+          reasoning: true,
+          input: ["text"],
+          contextWindow: 131072,
+          maxTokens: 65536,
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+          compat: {
+            sendSessionAffinityHeaders: true,
+          },
+        },
       ],
     });
 
@@ -149,18 +208,11 @@ export default function (pi: ExtensionAPI) {
       baseUrl: "https://agentrouter.org",
       apiKey,
       api: "anthropic-messages",
-      headers: {
-        "User-Agent": "claude-cli/1.0.108 (external, cli)",
-        "anthropic-beta": "claude-code-20250219,interleaved-thinking-2025-05-14,context-management-2025-06-27",
-        "x-stainless-lang": "js",
-        "x-stainless-package-version": "0.38.0",
-        "x-stainless-os": "Linux",
-        "x-stainless-arch": "x64",
-        "x-stainless-runtime": "node"
-      },
       compat: {
         forceAdaptiveThinking: true,
+        allowEmptySignature: true,
         sendSessionAffinityHeaders: true,
+        supportsEagerToolInputStreaming: false,
       },
       models: [
         {
@@ -173,7 +225,9 @@ export default function (pi: ExtensionAPI) {
           cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
           compat: {
             forceAdaptiveThinking: true,
+            allowEmptySignature: true,
             sendSessionAffinityHeaders: true,
+            supportsEagerToolInputStreaming: false,
           },
         },
         {
@@ -186,7 +240,9 @@ export default function (pi: ExtensionAPI) {
           cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
           compat: {
             forceAdaptiveThinking: true,
+            allowEmptySignature: true,
             sendSessionAffinityHeaders: true,
+            supportsEagerToolInputStreaming: false,
           },
         },
       ],
@@ -203,6 +259,63 @@ export default function (pi: ExtensionAPI) {
     }
   }
 
+  // Inter-request pacing to respect AgentRouter WAF rate limits and enforce canonical root prompt
+  pi.on("before_provider_request", async (event, ctx) => {
+    const provider = ((event as any)?.model?.provider || ctx?.model?.provider || "").toLowerCase();
+    const baseUrl = ((event as any)?.model?.baseUrl || (ctx?.model as any)?.baseUrl || "");
+
+    if (isAgentRouter(provider, baseUrl)) {
+      // 1. Cross-process request pacing
+      const lastEnd = getLastRequestEndTime();
+      const now = Date.now();
+      const elapsed = now - lastEnd;
+      if (lastEnd > 0 && elapsed < minIntervalMs) {
+        const waitMs = minIntervalMs - elapsed;
+        await new Promise((resolve) => setTimeout(resolve, waitMs));
+      }
+
+      // 2. Enforce canonical root system prompt in payload (Anthropic & OpenAI formats)
+      const payload = event.payload;
+      if (payload) {
+        if (payload.system !== undefined) {
+          payload.system = enforceCanonicalRootPrompt(payload.system);
+        }
+        if (Array.isArray(payload.messages) && payload.messages.length > 0) {
+          const firstMsg = payload.messages[0];
+          if (firstMsg && firstMsg.role === "system") {
+            if (typeof firstMsg.content === "string") {
+              firstMsg.content = enforceCanonicalRootPrompt(firstMsg.content);
+            } else if (Array.isArray(firstMsg.content)) {
+              firstMsg.content = enforceCanonicalRootPrompt(firstMsg.content);
+            }
+          }
+        }
+      }
+    }
+    return undefined;
+  });
+
+  pi.on("message_end", async (_event, ctx) => {
+    const model = ctx?.model;
+    if (isAgentRouter(model?.provider, (model as any)?.baseUrl)) {
+      setLastRequestEndTime(Date.now());
+    }
+  });
+
+  pi.on("turn_end", async (_event, ctx) => {
+    const model = ctx?.model;
+    if (isAgentRouter(model?.provider, (model as any)?.baseUrl)) {
+      setLastRequestEndTime(Date.now());
+    }
+  });
+
+  pi.on("agent_end", async (_event, ctx) => {
+    const model = ctx?.model;
+    if (isAgentRouter(model?.provider, (model as any)?.baseUrl)) {
+      setLastRequestEndTime(Date.now());
+    }
+  });
+
   pi.on("session_start", async (_event, ctx) => {
     updatePromptRewriteEnvForModel(ctx.model);
     setLastRequestEndTime(Date.now());
@@ -251,20 +364,6 @@ export default function (pi: ExtensionAPI) {
     }
   });
 
-  pi.on("turn_end", async (_event, ctx) => {
-    const model = ctx.model;
-    if (isAgentRouter(model?.provider, (model as any)?.baseUrl)) {
-      setLastRequestEndTime(Date.now());
-    }
-  });
-
-  pi.on("agent_end", async (_event, ctx) => {
-    const model = ctx.model;
-    if (isAgentRouter(model?.provider, (model as any)?.baseUrl)) {
-      setLastRequestEndTime(Date.now());
-    }
-  });
-
   pi.on("session_before_compact", async (event, ctx) => {
     const model = ctx.model;
     if (!model || !isAgentRouter(model.provider, (model as any).baseUrl)) {

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@madgagarin/pi-agentrouter",
-  "version": "1.2.1",
-  "description": "Pi coding agent extension for AgentRouter (GPT-5.6 Sol & Claude Opus 5) with rate-limit pacing and caching.",
+  "version": "1.4.0",
+  "description": "Pi coding agent extension for AgentRouter (GPT-5.6 Sol, Claude Opus 5, DeepSeek V4F) with rate-limit pacing, prompt caching, and WAF protection.",
   "publishConfig": {
     "access": "public"
   },
@@ -17,6 +17,8 @@
     "gpt-5.6-sol",
     "claude-opus-5",
     "claude-opus-4-8",
+    "deepseek-v4f",
+    "deepseek",
     "prompt-caching",
     "openai",
     "claude",