pi-gemini-empty-continuation

A Pi coding agent extension that auto-recovers from Gemini thinking truncation and zero-token empty responses, ported directly from Google Antigravity (agy / jetski) internal continuation architecture.

English | 中文说明

---

English

The Problem

Google Gemini (notably Gemini 3.8 Flash) occasionally outputs several thousand tokens of internal reasoning (thought: true), but stops before generating visible text (thought: false with text: "") without calling any tools.

In standard Pi agent runs, this causes the agent loop to interpret the turn as complete (stopReason: "stop"), leaving the user with an empty message bubble and stalled workflow. Generic retry extensions like pi-retry-empty fail to catch this because they only check for zeroTokens (whereas Gemini consumed hundreds or thousands of reasoning tokens).

The Solution: Antigravity Architecture

In Google's official Antigravity CLI (agy), this is solved through a dedicated post-invocation hook: empty_output_continuation_check.go.

This extension faithfully ports that architecture into Pi:

1. Dual-Scenario Detection (diagnoseEmptyAssistantResponse):
   • Case A (Pure Thinking Truncation): The model generated thoughts, but omitted text and tool calls.
   • Case B (Zero-Token Blank): The model cleanly returned zero tokens without text or tool calls.
2. Auto-Continuation Step (addContinueStep):
   • Automatically injects an adaptive continuation request via pi.sendUserMessage(..., { deliverAs: "followUp" }).
   • Bounded by a retry counter (maxRetries = 2) to prevent infinite token depletion.
   • Visible in TUI with clear chat messages, status bar progress, and notifications.

Installation

Option 1: Via npm (Recommended)

pi install npm:pi-gemini-empty-continuation

Option 2: Via Git Repository

pi install git:github.com/sususu98/pi-gemini-empty-continuation

Option 3: Local Global Extension

Download extensions/index.ts directly into your global Pi extensions directory:

mkdir -p ~/.pi/agent/extensions
curl -sL https://raw.githubusercontent.com/sususu98/pi-gemini-empty-continuation/main/extensions/index.ts \
  -o ~/.pi/agent/extensions/gemini-empty-continuation.ts

Then run /reload in your active Pi session.

Commands

• /gemini-continuation — View current status and retry count.
• /gemini-continuation on / /gemini-continuation off — Enable or disable the extension.
• /gemini-continuation <N> — Set maximum consecutive retries (default: 2).

---

中文说明

背景痛点

Google Gemini 系列模型（尤其在 Gemini 3.8 Flash 等开启双通道思考推理的模型上）在长上下文或复杂任务中，偶尔会出现耗费数千 token 进行了深入思考（thought: true），但在输出正文时戛然而止（thought: false 为空字符串），且未触发任何工具调用的现象。

在标准 Pi 会话中，宿主判定该轮次正常结束（stopReason: "stop"），导致界面上只显示思考折叠块，正文区完全空白，需要人工手动打字输入“继续”。而现存的重试插件（如 pi-retry-empty）因检测的是 zeroTokens，会因为 Gemini 产生了思考 token 而彻底漏判。

核心实现：对齐 Google Antigravity (agy)

通过逆向分析 Google 官方 CLI 二进制文件中的 empty_output_continuation_check.go，本插件将官方的双模态防护机制完整移植到 Pi Agent：

1. 双模态精准诊断 (diagnoseEmptyAssistantResponse)：
   • 场景 A（思考截断）：检测到 hasThinking === true 但 hasText === false && hasToolCall === false，精准识别只思考不说话的情况。
   • 场景 B（完全 0 token 空白）：检测到模型未思考也未输出任何内容（干净的 0 token 停止），同样进行拦截。
2. 状态机自动续写 (addContinueStep)：
   • 在 agent_end 触发时，自动通过 deliverAs: "followUp" 向当前会话投递针对性的续写指令。
   • 具有严格的防死循环保护（默认上限 2 次），达到上限自动报警终止。
   • TUI 完整可见：消息气泡、底栏状态进度、Toast 通知三重视觉感知。

安装与使用

方式 1：通过 npm 安装（推荐）

pi install npm:pi-gemini-empty-continuation

方式 2：通过 Git 仓库安装

pi install git:github.com/sususu98/pi-gemini-empty-continuation

方式 3：作为本地全局扩展安装

直接下载并放入全局扩展目录：

mkdir -p ~/.pi/agent/extensions
curl -sL https://raw.githubusercontent.com/sususu98/pi-gemini-empty-continuation/main/extensions/index.ts \
  -o ~/.pi/agent/extensions/gemini-empty-continuation.ts

在 Pi 对话框中输入 /reload 即可立即热重载生效。

交互指令

• /gemini-continuation：查看当前拦截状态及连续重试统计。
• /gemini-continuation on / off：快速开启或关闭自动续写。
• /gemini-continuation 3：调整最大连续重试次数（默认为 2 次）。

---

Links

• Pi Package Catalog: https://pi.dev/packages/pi-gemini-empty-continuation
• npm Registry: https://www.npmjs.com/package/pi-gemini-empty-continuation
• GitHub Repository: https://github.com/sususu98/pi-gemini-empty-continuation

License

MIT © 2026 sususu98