lonny-agent 0.1.8 → 0.2.1

package/src/agent/compaction.ts

@@ -6,6 +6,12 @@ import type { LLMMessage } from './llm.js'
  */
 
 const DEFAULT_MAX_TOKENS = 256_000
+
+/**
+ * 获取默认的上下文窗口大小（保持向后兼容）
+ * @deprecated 请使用 config.contextWindow 代替
+ */
+export const DEFAULT_CONTEXT_WINDOW = DEFAULT_MAX_TOKENS
 const COMPACTION_THRESHOLD = 0.75 // compact when usage exceeds 75% of budget
 
 /** Rough token estimation (4 chars ~= 1 token).

package/src/agent/prompt-builder.ts

@@ -67,11 +67,14 @@ ${getToolListForMode(config.mode)}`
   // ── Mode-specific instructions ───────────────────────────────────────────
   const modeInstructions =
     config.mode === 'plan'
-      ? `You are a planning agent. Your sole job is to investigate the codebase and produce an actionable implementation plan plus a todo list. You NEVER edit source files.
+      ? `You are a planning agent. Your ONLY job is to investigate the codebase and produce an actionable implementation plan with a todo list.
+
+You CANNOT edit files. You do NOT have access to edit, bash (write mode), exec, or install_skill.
+Any attempt to call these tools will FAIL — they are simply unavailable in this mode.
 
 RULES (plan-specific):
 1. Read first: Use read/grep/glob tools to gather all context you need before planning.
-2. You NEVER edit source files. You ONLY use read-only tools (read/glob/grep/ls/find/git/search) and bash (read-only commands only). You do NOT have access to edit, install_skill, or exec.
+2. You NEVER edit source files. You ONLY use read-only tools (read/glob/grep/ls/find/git/search) and bash (read-only commands only).
 3. Use \`bash\` for investigation only — NEVER to modify files, install packages, or run write operations.
 4. Your ONLY output is a plan file saved via \`write_plan\`. You CANNOT modify the codebase directly.
 5. You MUST persist the final plan AND todo list to a file in \`.lonny/\` using \`write_plan\`. The \`write_plan\` content MUST include both ## Plan and ## Todo List sections.
@@ -117,7 +120,40 @@ RULES (code-specific):
 7. COST OPTIMIZATION (CRITICAL): Each API call costs money. You have a hard limit of ~5 API calls per task.
 8. TODO LIST MAINTENANCE: After completing a task item, update the corresponding plan file in \`.lonny/\` by checking off the TODO item (change \`- [ ]\` to \`- [x]\`). Use \`read\` to find the plan file, then \`edit\` to update the checkbox.`
 
-  // ── Environment section (dynamic content — put LAST for prefix caching) ──
+  // ── Built-in development methodologies ─────────────────────────────────
+  // Embedded directly from Superpowers — no skill files needed.
+  const methodologySection =
+    config.mode === 'code'
+      ? `
+
+## Development Methodology
+
+### Systematic Debugging
+When investigating a bug or test failure, follow this process:
+1. **Root Cause** — Read errors carefully, reproduce consistently, check recent changes
+2. **Pattern Analysis** — Find working examples, compare with broken code
+3. **Hypothesis** — Form single hypothesis, test minimally (one change at a time)
+4. **Implementation** — Create failing test first, implement minimal fix, verify
+
+NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST.
+If you've tried 3+ fixes without success, question the architecture — don't keep guessing.
+
+### Verification Before Completion
+After ANY change, you MUST verify it actually works before claiming success:
+- Run the relevant tests and check the exit code
+- Only then report the result
+
+You CANNOT claim "tests pass" or "fix works" without having run the verification command in this same turn. Evidence before assertions, always.`
+      : config.mode === 'plan'
+        ? `
+
+## Design-First Planning
+Before writing a plan, explore the user's request thoroughly:
+1. Ask clarifying questions — understand purpose, constraints, success criteria
+2. Explore alternatives — consider 2-3 different approaches with trade-offs
+3. Break into bite-sized tasks (2-5 minutes each) with exact file paths and verification steps
+4. Save the plan via \`write_plan\``
+        : ''
   const envSection = `Environment:
 - Platform: ${platform} ${release} (${arch})
 - Shell: ${shell}
@@ -135,5 +171,5 @@ ${isWindows ? '  - Use `type` instead of `cat`, `dir` instead of `ls`, `where` i
 ${sharedRules}`
   return `${modeInstructions}
 
-${envSection}${rulesSection}${skillsSection}`
+${envSection}${rulesSection}${methodologySection}${skillsSection}`
 }

package/src/agent/providers/google.ts

@@ -68,7 +68,7 @@ export class GoogleProvider implements LLMProvider {
         if (m.content) {
           parts.push({ text: m.content })
         }
-        if (m.tool_calls) {
+        if (m.tool_calls && m.tool_calls.length > 0) {
           for (const tc of m.tool_calls) {
             parts.push({
               functionCall: {

package/src/agent/providers/openai.ts

@@ -91,7 +91,7 @@ export class OpenAIProvider implements LLMProvider {
           content: m.content || '',
         }
       }
-      if (m.role === 'assistant' && m.tool_calls) {
+      if (m.role === 'assistant' && m.tool_calls && m.tool_calls.length > 0) {
         return {
           role: 'assistant',
           content: m.content,

package/src/agent/session.ts

@@ -388,6 +388,8 @@ export class Session {
     const out = this.output
     printUserMessage(userPrompt, out)
     this.messages.push({ role: 'user', content: userPrompt })
+    // Save immediately after user message so page refresh doesn't lose it
+    this.save()
 
     // Reset per-turn counters
     this.turnInputTokens = 0
@@ -568,6 +570,12 @@ export class Session {
           }
           if (chunk.finish_reason === 'stop' || chunk.finish_reason === 'end_turn') {
             if (toolCalls.length === 0) {
+              const finalAssistantMsg: LLMMessage = {
+                role: 'assistant',
+                content: fullResponse || null,
+                reasoning_content: reasoningContent,
+              }
+              this.messages.push(finalAssistantMsg)
               printTokenStats(
                 this.turnInputTokens,
                 this.turnOutputTokens,
@@ -610,6 +618,12 @@ export class Session {
 
       if (toolCalls.length === 0) {
         if (fullResponse) {
+          const finalAssistantMsg: LLMMessage = {
+            role: 'assistant',
+            content: fullResponse,
+            reasoning_content: reasoningContent,
+          }
+          this.messages.push(finalAssistantMsg)
           printTokenStats(
             this.turnInputTokens,
             this.turnOutputTokens,
@@ -639,6 +653,7 @@ export class Session {
         reasoning_content: reasoningContent,
       }
       this.messages.push(assistantMsg)
+      this.save()
 
       // ── User confirmation for write-type tool calls ──
       if (!this.config.autoApprove && this.output?.confirmTool && toolCalls.length > 0) {
@@ -691,13 +706,15 @@ export class Session {
         this.messages.push(resultMsg)
       }
 
+      this.save()
+
       // End the current turn before next iteration — frontend creates a new message
       bus.emit(EventChannels.TURN_END, { iterations, toolCallCount: toolCalls.length })
 
       // Check if compaction is needed
-      if (shouldCompact(this.messages)) {
+      if (shouldCompact(this.messages, this.config.contextWindow)) {
         const before = this.messages.length
-        const result = compact(this.messages)
+        const result = compact(this.messages, this.config.contextWindow)
         if (result.compressed) {
           this.messages = result.messages
           bus.emit(EventChannels.COMPACTION_TRIGGERED, { before, after: result.newCount })

package/src/config/__tests__/context-window.test.ts

@@ -0,0 +1,94 @@
+import { describe, expect, test } from 'vitest'
+import { getContextWindowForModel } from '../index.js'
+
+describe('getContextWindowForModel', () => {
+  // 精确匹配测试
+  test('exact match with known model names', () => {
+    expect(getContextWindowForModel('doubao-seed-2.0-code')).toBe(256_000)
+    expect(getContextWindowForModel('deepseek-v4-pro')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen3.6-plus')).toBe(1_000_000)
+    expect(getContextWindowForModel('claude-sonnet-4.6')).toBe(1_000_000)
+    expect(getContextWindowForModel('gpt-5.5')).toBe(1_000_000)
+  })
+
+  // 大小写不敏感测试
+  test('case insensitive matching', () => {
+    expect(getContextWindowForModel('DOUBAO-SEED-2.0-CODE')).toBe(256_000)
+    expect(getContextWindowForModel('DeepSeek-V4-Pro')).toBe(1_000_000)
+    expect(getContextWindowForModel('QWEN3.6-PLUS')).toBe(1_000_000)
+  })
+
+  // 模糊匹配 - 含空格的变体
+  test('fuzzy match with space variants', () => {
+    expect(getContextWindowForModel('doubao seed 2.0 code')).toBe(256_000)
+    expect(getContextWindowForModel('deepseek v4 pro')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen 3.6 plus')).toBe(1_000_000)
+  })
+
+  // 模糊匹配 - 下划线与连字符变体
+  test('fuzzy match with underscore and hyphen variants', () => {
+    expect(getContextWindowForModel('doubao_seed_2.0_code')).toBe(256_000)
+    expect(getContextWindowForModel('deepseek_v4_pro')).toBe(1_000_000)
+  })
+
+  // 模糊匹配 - 含前缀和后缀的模型名
+  test('fuzzy match with prefixed and suffixed model names', () => {
+    expect(getContextWindowForModel('my-doubao-seed-2.0-code')).toBe(256_000)
+    expect(getContextWindowForModel('deepseek-v4-pro-variant')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen3.6-plus-prod')).toBe(1_000_000)
+  })
+
+  // 模糊匹配 - 关键词匹配（不完整的模型名）
+  test('fuzzy match with keyword matching (partial model names)', () => {
+    // 百万上下文模型
+    expect(getContextWindowForModel('deepseek-v4')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen3.6')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen3.7')).toBe(1_000_000)
+    expect(getContextWindowForModel('claude opus 4.7')).toBe(1_000_000)
+    expect(getContextWindowForModel('gemini 2.5')).toBe(1_000_000)
+
+    // 256K 上下文模型
+    expect(getContextWindowForModel('doubao-seed-2.0')).toBe(256_000)
+    expect(getContextWindowForModel('kimi-k2.5')).toBe(262_144) // 实际值
+    expect(getContextWindowForModel('kimi-k2.6')).toBe(262_144) // 实际值
+
+    // 200K 上下文模型
+    expect(getContextWindowForModel('minimax')).toBe(200_000)
+    expect(getContextWindowForModel('glm-5')).toBe(200_000)
+    expect(getContextWindowForModel('glm-5.1')).toBe(200_000)
+    expect(getContextWindowForModel('claude-haiku')).toBe(200_000)
+
+    // 160K 上下文模型
+    expect(getContextWindowForModel('deepseek-v3')).toBe(160_000)
+    expect(getContextWindowForModel('deepseek-v3.2')).toBe(160_000)
+
+    // 128K 上下文模型
+    expect(getContextWindowForModel('gpt-4o')).toBe(128_000)
+    expect(getContextWindowForModel('gpt-4')).toBe(128_000)
+    expect(getContextWindowForModel('llama-3')).toBe(128_000)
+  })
+
+  // 未知模型名 - 兜底默认值
+  test('unknown model names return default 128K', () => {
+    expect(getContextWindowForModel('unknown-model')).toBe(128_000)
+    expect(getContextWindowForModel('random-model-name')).toBe(128_000)
+  })
+
+  // 用户配置中常用的模型名
+  test('common model names from user configs', () => {
+    // 豆包模型
+    expect(getContextWindowForModel('Doubao-Seed-2.0-Code')).toBe(256_000)
+
+    // DeepSeek 模型
+    expect(getContextWindowForModel('deepseek-chat')).toBe(128_000)
+    expect(getContextWindowForModel('deepseek-reasoner')).toBe(128_000)
+
+    // Kimi 模型
+    expect(getContextWindowForModel('kimi-k2.5')).toBe(262_144) // 实际值
+    expect(getContextWindowForModel('kimi-k2.6')).toBe(262_144) // 实际值
+
+    // 通义千问
+    expect(getContextWindowForModel('qwen3.6-plus')).toBe(1_000_000)
+    expect(getContextWindowForModel('qwen3.7-max')).toBe(1_000_000)
+  })
+})

package/src/config/index.ts

@@ -8,6 +8,7 @@ export interface Config {
   mode: 'code' | 'plan' | 'ask'
   provider: 'openai' | 'anthropic' | 'google' | 'ollama'
   model: string
+  contextWindow: number
   cwd: string
   autoApprove: boolean
   thinking?: boolean
@@ -19,6 +20,243 @@ export interface Config {
   tavilyApiKey?: string
 }
 
+/**
+ * 模型上下文窗口大小映射表（基于 2026 年最新数据）
+ * 单位：tokens
+ */
+const MODEL_CONTEXT_WINDOWS: Record<string, number> = {
+  // Doubao 系列
+  'doubao-seed-2.0-code': 256_000,
+  'doubao-seed-2.0-pro': 256_000,
+  'doubao-seed-2.0-lite': 256_000,
+  'doubao-seed-2.0-mini': 256_000,
+  'doubao-seed-2.0-code-preview-260215': 256_000,
+  'doubao-seed-1.8': 256_000,
+  'doubao-seed-1.6': 256_000,
+
+  // DeepSeek 系列
+  'deepseek-v4-pro': 1_000_000,
+  'deepseek-v4-flash': 256_000,
+  'deepseek-v3.2': 160_000,
+  'deepseek-r1': 128_000,
+  'deepseek-chat': 128_000,
+  'deepseek-reasoner': 128_000,
+
+  // Kimi 系列
+  'kimi-k2.5': 262_144,
+  'kimi-k2.6': 262_144,
+  'kimi-k2-thinking': 256_000,
+
+  // MiniMax 系列
+  'minimax-m2.5': 200_000,
+  'minimax-m2.7': 200_000,
+  'minimax-m2': 205_000,
+
+  // GLM 系列
+  'glm-5': 200_000,
+  'glm-5.1': 200_000,
+  'glm-4.7-flashx': 200_000,
+  'glm-4-flash': 128_000,
+  'glm-4': 128_000,
+
+  // Qwen 系列
+  'qwen3.6-plus': 1_000_000,
+  'qwen3.7-max': 1_000_000,
+  'qwen3-max': 256_000,
+  'qwen3.5-plus': 1_000_000,
+  'qwen-flash': 1_000_000,
+
+  // Claude 系列
+  'claude-opus-4.7': 1_000_000,
+  'claude-opus-4.6': 1_000_000,
+  'claude-sonnet-4.6': 1_000_000,
+  'claude-haiku-4.5': 200_000,
+
+  // GPT 系列
+  'gpt-5.5': 1_000_000,
+  'gpt-5.4': 1_000_000,
+  'gpt-5.1': 1_000_000,
+  'gpt-5.2-codex': 512_000,
+  'gpt-4.1': 1_000_000,
+  'gpt-4.1-mini': 1_000_000,
+  'gpt-4o': 128_000,
+  'gpt-4': 128_000,
+
+  // Gemini 系列
+  'gemini-2.5-pro': 1_000_000,
+  'gemini-2.5-flash': 1_000_000,
+  'gemini-3.1-pro': 1_000_000,
+  'gemini-3.1-flash': 1_000_000,
+
+  // Llama 系列
+  'llama-4-maverick': 1_000_000,
+  'llama-4-scout': 10_000_000,
+  'llama-3.1': 128_000,
+  'llama-3': 128_000,
+
+  // Mistral 系列
+  'mistral-small-3.1': 128_000,
+  'mistral-small-3.2': 128_000,
+}
+
+/**
+ * 模糊匹配规则
+ */
+interface FuzzyMatchRule {
+  keywords: string[]
+  windowSize: number
+}
+
+/**
+ * 模糊匹配规则列表（从高优先级到低优先级）
+ */
+const fuzzyMatchRules: FuzzyMatchRule[] = [
+  // 首先匹配 128K 模型（防止 gpt-4 匹配到 gpt-5 规则）
+  {
+    keywords: [
+      'gpt-4o',
+      'gpt 4o',
+      'gpt-4',
+      'gpt 4',
+      'llama-3',
+      'llama 3',
+      'mistral-small',
+      'mistral small',
+      'glm-4-flash',
+      'glm 4 flash',
+      'glm-4',
+    ],
+    windowSize: 128_000,
+  },
+  // 百万上下文模型 - 更高优先级（更精确的关键词）
+  {
+    keywords: [
+      'deepseek-v4',
+      'deepseek v4',
+      'qwen3.6',
+      'qwen 3.6',
+      'qwen3.7',
+      'qwen 3.7',
+      'qwen3.5',
+      'qwen 3.5',
+      'qwen-flash',
+      'qwen flash',
+    ],
+    windowSize: 1_000_000,
+  },
+  {
+    keywords: [
+      'claude-opus',
+      'claude sonnet',
+      'claude-sonnet',
+      'gpt-5.',
+      'gpt 5',
+      'gpt-4.1',
+      'gpt 4.1',
+      'gemini-2.5',
+      'gemini 2.5',
+      'gemini-3.1',
+      'gemini 3.1',
+      'llama-4',
+      'llama 4',
+    ],
+    windowSize: 1_000_000,
+  },
+  // 256K 上下文模型
+  {
+    keywords: [
+      'doubao-seed-2.0',
+      'doubao seed 2.0',
+      'doubao-seed-1',
+      'doubao seed 1',
+      'kimi-k2',
+      'kimi k2',
+      'kimi-k2.5',
+      'kimi k2.5',
+      'kimi-k2.6',
+      'kimi k2.6',
+      'qwen3',
+      'qwen 3',
+      'qwen3-max',
+      'qwen3 max',
+    ],
+    windowSize: 256_000,
+  },
+  // 200K 上下文模型
+  {
+    keywords: [
+      'minimax',
+      'glm-5',
+      'glm 5',
+      'glm-5.1',
+      'glm 5.1',
+      'glm-4.7',
+      'glm 4.7',
+      'claude-haiku',
+      'claude haiku',
+    ],
+    windowSize: 200_000,
+  },
+  // 160K 上下文模型
+  {
+    keywords: ['deepseek-v3', 'deepseek v3', 'deepseek-v3.2', 'deepseek v3.2'],
+    windowSize: 160_000,
+  },
+  // 最后匹配通用关键词（deepseek）
+  {
+    keywords: ['deepseek'],
+    windowSize: 128_000,
+  },
+]
+
+/**
+ * 根据模型名称获取上下文窗口大小（改进的模糊匹配）
+ * @param modelName 模型名称（不区分大小写）
+ * @returns 上下文窗口大小（tokens）
+ */
+export function getContextWindowForModel(modelName: string): number {
+  // 处理空字符串
+  if (!modelName) {
+    return 128_000
+  }
+
+  const normalizedModelName = modelName.toLowerCase().replace(/[-_]/g, ' ')
+
+  // 1. 精确匹配（原始键名）
+  if (MODEL_CONTEXT_WINDOWS[modelName]) {
+    return MODEL_CONTEXT_WINDOWS[modelName]
+  }
+
+  // 2. 精确匹配（小写键名）
+  if (MODEL_CONTEXT_WINDOWS[modelName.toLowerCase()]) {
+    return MODEL_CONTEXT_WINDOWS[modelName.toLowerCase()]
+  }
+
+  // 3. 前缀匹配（支持部分匹配）
+  for (const [pattern, size] of Object.entries(MODEL_CONTEXT_WINDOWS)) {
+    const normalizedPattern = pattern.toLowerCase().replace(/[-_]/g, ' ')
+    if (
+      normalizedModelName.includes(normalizedPattern) ||
+      normalizedPattern.includes(normalizedModelName)
+    ) {
+      return size
+    }
+  }
+
+  // 4. 模糊匹配（关键词匹配）- 按照规则顺序（优先级）处理
+  for (const rule of fuzzyMatchRules) {
+    for (const keyword of rule.keywords) {
+      const normalizedKeyword = keyword.toLowerCase().replace(/[-_]/g, ' ')
+      if (normalizedModelName.includes(normalizedKeyword)) {
+        return rule.windowSize
+      }
+    }
+  }
+
+  // 5. 兜底默认值
+  return 128_000
+}
+
 interface JsonConfig {
   apiKey?: string
   baseUrl?: string
@@ -69,6 +307,9 @@ export function loadConfig(options?: {
   // Auto-enable cache for DeepSeek models unless explicitly disabled
   const enableCache = jsonConfig.enableCache ?? (isDeepSeekModel(model, baseUrl) || undefined)
 
+  // 根据模型名称获取上下文窗口大小
+  const contextWindow = getContextWindowForModel(model)
+
   return {
     apiKey: jsonConfig.apiKey || '',
     baseUrl,
@@ -79,6 +320,7 @@ export function loadConfig(options?: {
       | 'ollama',
     mode: options?.mode || 'code',
     model,
+    contextWindow,
     cwd: options?.cwd || process.cwd(),
     autoApprove: options?.autoApprove ?? jsonConfig.autoApprove ?? false,
     thinking: jsonConfig.thinking,

package/src/pi-tui/tui.ts

@@ -1351,26 +1351,20 @@ export class TUI extends Container {
     const targetRow = Math.max(0, Math.min(cursorPos.row, totalLines - 1))
     const targetCol = Math.max(0, cursorPos.col)
 
-    // Move cursor from current position to target
-    const rowDelta = targetRow - this.hardwareCursorRow
-    let buffer = ''
-    if (rowDelta > 0) {
-      buffer += `\x1b[${rowDelta}B` // Move down
-    } else if (rowDelta < 0) {
-      buffer += `\x1b[${-rowDelta}A` // Move up
+    // Use absolute cursor positioning (CUP) to avoid scroll artifacts.
+    // Relative movement (\x1b[NB/\x1b[NA) with a stale baseline can
+    // cause the terminal to scroll, inserting blank lines, especially
+    // when the editor overlay cursor is far from the rendered content.
+    let buffer = `\x1b[${targetRow + 1};${targetCol + 1}H`
+
+    if (this.showHardwareCursor) {
+      buffer += '\x1b[?25h'
     }
-    // Move to absolute column (1-indexed)
-    buffer += `\x1b[${targetCol + 1}G`
 
     if (buffer) {
       this.terminal.write(buffer)
     }
 
     this.hardwareCursorRow = targetRow
-    if (this.showHardwareCursor) {
-      this.terminal.showCursor()
-    } else {
-      this.terminal.hideCursor()
-    }
   }
 }