thatgfsj-code 1.0.2 → 2.2.9

package/install.sh

@@ -0,0 +1,113 @@
+#!/bin/bash
+# Thatgfsj Code Installer for macOS/Linux
+# Usage: curl -sL https://raw.githubusercontent.com/Thatgfsj/thatgfsj-code/main/install.sh | bash
+
+set -e
+
+# Colors
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+CYAN='\033[0;36m'
+NC='\033[0m' # No Color
+
+echo ""
+echo -e "${CYAN}  Thatgfsj Code 安装向导${NC}"
+echo -e "${CYAN}  =======================${NC}"
+echo ""
+
+# ============== Step 1: Check Node.js ==============
+echo -e "${YELLOW}[*] 检查 Node.js...${NC}"
+
+if command -v node &> /dev/null; then
+    NODE_VERSION=$(node --version)
+    MAJOR_VERSION=$(echo $NODE_VERSION | cut -d'v' -f2 | cut -d'.' -f1)
+    if [ "$MAJOR_VERSION" -ge 18 ]; then
+        echo -e "${GREEN}[✓] Node.js $NODE_VERSION 已安装${NC}"
+    else
+        echo -e "${RED}[✗] Node.js 版本过低，需要 v18+${NC}"
+        echo "    请访问 https://nodejs.org 升级"
+        exit 1
+    fi
+else
+    echo -e "${YELLOW}[*] 未检测到 Node.js，开始安装...${NC}"
+    
+    # Try Homebrew (macOS)
+    if command -v brew &> /dev/null; then
+        echo -e "${YELLOW}[*] 使用 Homebrew 安装...${NC}"
+        brew install node
+    # Try apt (Ubuntu/Debian)
+    elif command -v apt-get &> /dev/null; then
+        echo -e "${YELLOW}[*] 使用 apt 安装...${NC}"
+        curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
+        sudo apt-get install -y nodejs
+    # Try yum (CentOS/RHEL)
+    elif command -v yum &> /dev/null; then
+        echo -e "${YELLOW}[*] 使用 yum 安装...${NC}"
+        curl -fsSL https://rpm.nodesource.com/setup_20.x | sudo bash -
+        sudo yum install -y nodejs
+    else
+        echo -e "${RED}[✗] 未找到包管理器${NC}"
+        echo "    请手动安装: https://nodejs.org"
+        exit 1
+    fi
+fi
+
+# ============== Step 2: Clone/Update ==============
+echo -e "${YELLOW}[*] 准备安装 Thatgfsj Code...${NC}"
+
+INSTALL_DIR="$HOME/thatgfsj-code"
+
+if [ -d "$INSTALL_DIR" ]; then
+    echo -e "${YELLOW}[*] 检测到已有安装，正在更新...${NC}"
+    cd "$INSTALL_DIR"
+    git pull origin main 2>/dev/null || {
+        echo -e "${YELLOW}[*] 更新失败，重新克隆...${NC}"
+        rm -rf "$INSTALL_DIR"
+    }
+fi
+
+if [ ! -d "$INSTALL_DIR" ]; then
+    echo -e "${YELLOW}[*] 克隆仓库...${NC}"
+    git clone https://github.com/Thatgfsj/thatgfsj-code.git "$INSTALL_DIR"
+fi
+
+cd "$INSTALL_DIR"
+
+if [ ! -f "package.json" ]; then
+    echo -e "${RED}[✗] 安装目录无效${NC}"
+    exit 1
+fi
+
+echo -e "${GREEN}[✓] 代码准备完成: $INSTALL_DIR${NC}"
+
+# ============== Step 3: Install Dependencies ==============
+echo -e "${YELLOW}[*] 安装依赖...${NC}"
+
+npm install
+npm run build
+
+echo -e "${GREEN}[✓] 依赖安装完成${NC}"
+
+# ============== Step 4: Link Command ==============
+echo -e "${YELLOW}[*] 设置命令...${NC}"
+
+npm link
+
+echo -e "${GREEN}[✓] 命令 'gfcode' 已可用${NC}"
+
+# ============== Done ==============
+echo ""
+echo -e "${CYAN}  ======================================${NC}"
+echo -e "${GREEN}  安装完成!${NC}"
+echo -e "${CYAN}  ======================================${NC}"
+echo ""
+echo -e "  ${WHITE}使用方法:${NC}"
+echo -e "    gfcode init          - 重新配置"
+echo -e "    gfcode               - 启动交互模式"
+echo -e "    gfcode '你的问题'    - 直接提问"
+echo -e "    gfcode explain '代码' - 解释代码"
+echo -e "    gfcode debug '代码'   - 调试代码"
+echo ""
+echo -e "  文档: ${CYAN}https://github.com/Thatgfsj/thatgfsj-code${NC}"
+echo ""

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thatgfsj-code",
-  "version": "1.0.2",
+  "version": "2.2.9",
   "description": "Thatgfsj Code - AI Coding Assistant",
   "main": "dist/cmd/index.js",
   "type": "module",
@@ -12,7 +12,10 @@
     "build": "tsc",
     "start": "node dist/cmd/index.js",
     "dev": "tsc && node dist/cmd/index.js",
-    "link": "npm link"
+    "link": "npm link",
+    "test": "node tests/smoke-thinking.mjs && node tests/smoke-tool.mjs && node tests/smoke-pollution.mjs && node tests/edge-thinking.mjs && node tests/edge-tool.mjs && node tests/edge-pollution.mjs && node tests/edge-session.mjs && node tests/edge-cli.mjs",
+    "test:single": "node tests/",
+    "prepublishOnly": "npm run build && npm test"
   },
   "keywords": [
     "ai",

package/src/app/index.ts

@@ -10,6 +10,7 @@ import { ToolRegistry } from '../tools/index.js';
 import { HookManager } from '../hooks/index.js';
 import { SystemPromptBuilder } from '../prompts/index.js';
 import { SkillRegistry } from '../skills/index.js';
+import { compressThinking } from '../utils/thinking.js';
 import type { ChatMessage, ChatResponse } from '../types.js';
 
 export class App {
@@ -20,6 +21,12 @@ export class App {
   hooks: HookManager;
   prompts: SystemPromptBuilder;
   skills: SkillRegistry;
+  /**
+   * v2.2.5 (product 0.4.2): toggle <think> block compression. Default
+   * true. Toggled by `--show-thinking` on the CLI or `/thinking on|off`
+   * in the REPL.
+   */
+  showThinking: boolean = false;
 
   private constructor(
     config: ConfigManager,
@@ -83,6 +90,14 @@ export class App {
 
   /**
    * Run a single prompt (non-interactive mode)
+   *
+   * v2.2.4 (port from v2.1.0): persistence of the assistant message
+   * uses addMessageSafe, which drops the message if it contains
+   * pollution markers like "[已中断]".
+   *
+   * v2.2.5 (product 0.4.2): persistence also strips <think> blocks
+   * (and similar reasoning delimiters) when showThinking is false,
+   * so the conversation log stays compact.
    */
   async runPrompt(prompt: string): Promise<string> {
     this.session.addMessage('user', prompt);
@@ -94,15 +109,16 @@ export class App {
         fullResponse += chunk;
       }
     } catch (err) {
-      // Re-throw after saving partial response
-      if (fullResponse) {
-        this.session.addMessage('assistant', fullResponse);
-      }
+      // Re-throw without persisting partial response. Persisting
+      // truncated output here was the source of the [已中断] loop in
+      // v2.2.3.
       throw err;
     }
 
     console.log();
-    this.session.addMessage('assistant', fullResponse);
+    // v2.2.5: compress <think> blocks before persisting.
+    const toPersist = compressThinking(fullResponse, this.showThinking);
+    this.session.addMessageSafe('assistant', toPersist);
     return fullResponse;
   }
 }

package/src/cmd/index.tsx

@@ -2,18 +2,60 @@
 
 /**
  * Thatgfsj Code - CLI Entry Point
+ *
+ * v2.2.4 (product 0.4.1): three concrete fixes for the bugs visible
+ * in the 2.2.3 (product 0.4.0) session log:
+ *
+ *   1. Encoding: chcp 65001 used to be wrapped in a silent try/catch,
+ *      and stdout never had its default encoding set. On Windows this
+ *      caused Chinese characters from tool output to render as
+ *      mojibake (ļ ʹ ϵͳ Ƿ ...).
+ *
+ *   2. [已中断] hallucination loop: 1.0.4's SessionManager had no
+ *      anti-pollution filter. When the user pressed Ctrl+C mid-
+ *      stream, the truncated assistant message (containing the
+ *      model's literal "[已中断]" marker) got persisted to history.
+ *      On the next turn, the model saw that and kept echoing it.
+ *      v2.1.0 had the right fix (_wasAborted flag + filter) — we
+ *      port it back here.
+ *
+ *   3. Tool output visual: 1.0.4 printed `@@TOOL@@{...}` markers
+ *      inline with text, no separator, no truncation. Long tool
+ *      output (registry dumps, dir listings) bled into the next
+ *      AI message. Add separator + truncation + color separation.
  */
 
 if (process.platform === 'win32') {
+  // Switch Windows console to UTF-8 BEFORE any other code runs.
+  // v2.2.3 wrapped this in `try { } catch {}` which silently swallowed
+  // failures (e.g. when running under a non-interactive shell where
+  // chcp is meaningless). Now we still try, but we also fall back to
+  // setting the codepage via the parent process's stdio handles.
   try {
-    require('child_process').execSync('chcp 65001', { stdio: 'ignore', windowsHide: true });
-  } catch {}
+    require('child_process').execSync('chcp 65001 >NUL', { stdio: 'ignore', windowsHide: true });
+  } catch {
+    // Non-Windows shell, or chcp unavailable (e.g. git-bash on CI) —
+    // that's fine, Node defaults to UTF-8 in that environment.
+  }
+}
+
+// Force stdout/stderr to UTF-8 regardless of platform. Without this,
+// Node will emit GBK-encoded bytes for non-ASCII characters on
+// Windows even after `chcp 65001`, because the underlying file
+// descriptors still report a non-UTF-8 code page.
+if (process.stdout.setDefaultEncoding) {
+  process.stdout.setDefaultEncoding('utf8');
+}
+if (process.stderr.setDefaultEncoding) {
+  process.stderr.setDefaultEncoding('utf8');
 }
 
 import { program } from 'commander';
 import chalk from 'chalk';
 import { App } from '../app/index.js';
 import { WelcomeScreen } from '../tui/welcome.js';
+import { compressThinking, summarizeThinking, splitThinking } from '../utils/thinking.js';
+import { PRODUCT_VERSION } from '../version.js';
 
 process.on('uncaughtException', (error) => {
   console.error(chalk.red('\n  Error:'), error.message);
@@ -28,11 +70,12 @@ process.on('unhandledRejection', (reason) => {
 program
   .name('gfcode')
   .description('Thatgfsj Code - AI Coding Assistant')
-  .version('1.0.2')
+  .version(PRODUCT_VERSION)
   .argument('[prompt]', 'Task to execute (omit to start interactive mode)')
   .option('-m, --model <model>', 'Specify model')
   .option('-i, --interactive', 'Force interactive mode')
-  .action(async (prompt: string | undefined, options: { model?: string; interactive?: boolean }) => {
+  .option('--show-thinking', 'Show full <think>...</think> reasoning blocks (default: compress to one-line summary)')
+  .action(async (prompt: string | undefined, options: { model?: string; interactive?: boolean; showThinking?: boolean }) => {
     try {
       const app = await App.create();
 
@@ -78,16 +121,36 @@ program
 
         app.session.addMessage('user', prompt);
         let fullResponse = '';
+        // v2.2.4: track whether the stream was aborted so we don't
+        // persist a truncated assistant message (which is what caused
+        // the [已中断] hallucination loop in v2.2.3).
+        const abortCtrl = new AbortController();
+        const onSigInt = () => { abortCtrl.abort(); };
+        process.once('SIGINT', onSigInt);
+        // v2.2.5 (product 0.4.2): compress <think> blocks at render time
+        // AND before persisting to history. Without this, every assistant
+        // message balloons history with reasoning the user has already
+        // seen compressed, and context windows blow up fast.
+        const showThinking = !!options.showThinking;
 
         try {
-          process.stdout.write(chalk.gray('  Thinking...\r'));
+          process.stdout.write(chalk.gray('  Thinking...'));
           const stream = app.streamResponse();
 
+          // v2.2.5: stream-time rendering stays simple — we just buffer
+          // the full response. Post-process at the end (compress
+          // thinking blocks once). This avoids a state machine for
+          // "are we inside <think> right now" mid-stream which would
+          // be flaky if the model splits the tag across chunks.
           for await (const chunk of stream) {
-            // Clear thinking line
-            process.stdout.write('\r' + ' '.repeat(40) + '\r');
+            if (abortCtrl.signal.aborted) break;
+            // Clear the entire "Thinking..." line — must use enough
+            // spaces to overwrite any longer thinking indicator.
+            process.stdout.write('\r' + ' '.repeat(80) + '\r');
 
-            // Parse tool messages
+            // Parse tool messages — these should NOT be compressed even
+            // when thinking is hidden, because they carry real signal
+            // (tool name + args + result).
             if (chunk.includes('@@TOOL@@')) {
               const parts = chunk.split('\n');
               for (const part of parts) {
@@ -97,26 +160,93 @@ program
                     if (data.action === 'call') {
                       console.log();
                       console.log(chalk.cyan(`  ⚙ ${data.name}: ${formatArgs(data.args)}`));
+                      console.log(chalk.gray('  ' + '─'.repeat(40)));
                     } else if (data.action === 'result') {
                       const output = data.output || data.error || '';
-                      const lines = output.split('\n').slice(0, 10);
-                      for (const line of lines) {
+                      const lines = output.split('\n');
+                      const MAX_TOOL_LINES = 20;
+                      const truncated = lines.length > MAX_TOOL_LINES;
+                      const visible = truncated ? lines.slice(0, MAX_TOOL_LINES) : lines;
+                      for (const line of visible) {
                         console.log(chalk.gray('    │ ') + line);
                       }
+                      if (truncated) {
+                        console.log(chalk.gray(`    │ ... (${lines.length - MAX_TOOL_LINES} more lines truncated)`));
+                      }
+                      console.log(chalk.gray('  ' + '─'.repeat(40)));
                     }
                   } catch {}
                 } else if (part) {
+                  // Mid-stream text chunk — leave it raw. The post-
+                  // process step below will compress any <think>
+                  // blocks before printing them as the final rendered
+                  // output. (We can't reliably strip mid-stream
+                  // because the closing </think> might be in a later
+                  // chunk.)
                   process.stdout.write(chalk.cyan('  │ ') + part);
                 }
               }
             } else {
-              process.stdout.write(chunk);
+              // Plain AI text chunk — buffer only. Same reasoning:
+              // post-process at end avoids mid-stream delimiter
+              // races.
+              fullResponse += chunk;
+              // Show the chunk as it streams, but we'll re-render the
+              // final compressed version below. To avoid double-
+              // printing, only echo when we're NOT going to post-
+              // process (i.e., when showThinking is true and we want
+              // to see everything live).
+              if (showThinking) {
+                process.stdout.write(chalk.cyan('  │ ') + chunk);
+              }
+            }
+          }
+
+          // v2.2.5: post-process. Strip <think> blocks, summarize,
+          // and print the cleaned conclusion if we suppressed it
+          // during streaming.
+          if (!showThinking && !abortCtrl.signal.aborted && fullResponse) {
+            const split = splitThinking(fullResponse);
+            if (split.thinking) {
+              // Clear what we already streamed (we suppressed text
+              // chunks during streaming, so there's nothing to clear
+              // unless showThinking was on; in that case the text
+              // already contains the <think> block live and there's
+              // no point double-printing).
+              const summary = summarizeThinking(split);
+              if (summary) {
+                console.log(chalk.gray(`  ${summary}`));
+              }
+              if (split.conclusion) {
+                console.log(chalk.cyan('  │ ') + split.conclusion);
+              }
+            } else if (fullResponse.trim()) {
+              // No thinking block found — print the conclusion if we
+              // hadn't already (showThinking=false suppresses
+              // streaming text).
+              console.log(chalk.cyan('  │ ') + fullResponse);
             }
-            fullResponse += chunk;
           }
 
           console.log();
-          app.session.addMessage('assistant', fullResponse);
+          // v2.2.4: skip persistence on abort.
+          if (!abortCtrl.signal.aborted) {
+            // v2.2.5: persist only the conclusion (no <think> blocks)
+            // when compression is enabled. This keeps the conversation
+            // log readable AND keeps context-window usage low.
+            const toPersist = showThinking
+              ? fullResponse
+              : compressThinking(fullResponse, false);
+            const accepted = app.session.addMessageSafe('assistant', toPersist);
+            if (!accepted) {
+              console.error(chalk.yellow(
+                '  ⚠️  Dropped assistant message containing [已中断] marker.\n' +
+                '      (prevents hallucination loop — try your question again)'
+              ));
+            }
+          } else {
+            console.log(chalk.yellow('  ⏹  Cancelled (response not saved)'));
+          }
         } catch (error: any) {
           const msg = error.message || String(error);
           if (msg.includes('401') || msg.includes('403') || msg.includes('Unauthorized')) {
@@ -131,6 +261,8 @@ program
           } else {
             console.error(chalk.red(`\n  Error: ${msg}`));
           }
+        } finally {
+          process.removeListener('SIGINT', onSigInt);
         }
       }
     } catch (error: any) {

package/src/session/index.ts

@@ -5,12 +5,65 @@
 import type { ChatMessage } from '../types.js';
 import { ContextCompactor } from './compactor.js';
 
+/**
+ * v2.2.4 (port from v2.1.0): patterns that, if found in an assistant
+ * message, indicate the message is a truncated/aborted response that
+ * should NOT be persisted into history. Without this filter, the next
+ * turn's LLM sees the marker and starts echoing it back, creating a
+ * self-reinforcing "[已中断]" hallucination loop.
+ *
+ * v2.2.6.1 tightening (smoke-test driven): the design here is a
+ * two-tier check:
+ *   1. STRONG markers — unambiguous pollution, always drop:
+ *      - `[已中断]` bracketed (the model emits this exact form)
+ *      - `[interrupted]` bracketed English equivalent
+ *   2. WEAK markers — only drop if the message is also short (<200
+ *      chars). Long messages mentioning "response truncated" are
+ *      legitimate conversation (e.g. user complaints about past
+ *      behavior), not pollution.
+ *
+ * Earlier drafts matched bare `已中断` substrings and "response
+ * truncated" anywhere in the message — both were too greedy and
+ * dropped legitimate Chinese/English conversation.
+ */
+const POLLUTION_STRONG: RegExp[] = [
+  /\[已中断\]/,
+  /\[interrupted\]/i,
+];
+
+const POLLUTION_WEAK: RegExp[] = [
+  /^\s*[\*#>\-`]*\s*\[已中断\]/m,
+  /^\s*[\*#>\-`]*\s*\[interrupted\]/im,
+  /\bresponse (was )?(truncated|cut off|interrupted)\b/i,
+  /\boutput (was )?(truncated|cut off|interrupted)\b/i,
+];
+
+function looksPolluted(content: string): boolean {
+  if (!content) return false;
+  // Common gates (apply to both tiers):
+  //   - Length: model got cut off mid-stream → pollution is short.
+  //     A long message with [已中断] in the middle is the model
+  //     legitimately referencing the marker, not pollution.
+  //   - Question: ends with '?' → user complaining about past behavior.
+  //   - Temporal: contains "last time" / "earlier" / etc. → past tense.
+  if (content.length >= 200) return false;
+  if (/\?\s*$/.test(content.trim())) return false;
+  if (/\b(last time|earlier|before|previously|yesterday)\b/i.test(content)) return false;
+  // STRONG markers — bracketed truncation markers, unambiguous.
+  if (POLLUTION_STRONG.some(p => p.test(content))) return true;
+  // WEAK markers — bare "response truncated" phrases.
+  if (POLLUTION_WEAK.some(p => p.test(content))) return true;
+  return false;
+}
+
 export class SessionManager {
   private messages: ChatMessage[] = [];
   private sessionId: string;
   private createdAt: Date;
   private compactor: ContextCompactor;
   private maxMessages: number;
+  /** v2.2.4: counter for messages dropped by the pollution filter. */
+  private droppedCount: number = 0;
 
   constructor(maxMessages = 50) {
     this.maxMessages = maxMessages;
@@ -24,6 +77,26 @@ export class SessionManager {
     this.autoCompact();
   }
 
+  /**
+   * v2.2.4 (port from v2.1.0): same as addMessage but returns false
+   * (and skips the push) if the message content matches a known
+   * truncation/abort pollution pattern. Use this for assistant
+   * messages whose stream might have been aborted.
+   */
+  addMessageSafe(role: ChatMessage['role'], content: string, extras?: Partial<ChatMessage>): boolean {
+    if (role === 'assistant' && looksPolluted(content)) {
+      this.droppedCount++;
+      return false;
+    }
+    this.addMessage(role, content, extras);
+    return true;
+  }
+
+  /** v2.2.4: total messages dropped by addMessageSafe since session start. */
+  getDroppedCount(): number {
+    return this.droppedCount;
+  }
+
   getMessages(): ChatMessage[] {
     return [...this.messages];
   }
@@ -62,8 +135,16 @@ export class SessionManager {
 
   truncate(maxMessages?: number): void {
     if (maxMessages) {
+      // v2.2.7 edge fix: preserveRecent must be <= maxMessages - 1,
+      // otherwise the compactor's "always keep recent N" branch
+      // overrides the trim and no messages actually get dropped.
+      // Set preserveRecent to maxMessages-1 so anything beyond the
+      // last maxMessages-1 user/assistant msgs gets summarized.
       this.maxMessages = maxMessages;
-      this.compactor = new ContextCompactor({ maxMessages });
+      this.compactor = new ContextCompactor({
+        maxMessages,
+        preserveRecent: Math.max(1, maxMessages - 1),
+      });
     }
     const { compacted } = this.compactor.compact(this.messages);
     this.messages = compacted;