openrune 1.0.1 → 1.1.0

package/README.ko.md

@@ -19,6 +19,10 @@
   <img src="https://img.shields.io/badge/license-MIT-green" alt="license" />
 </p>
 
+<p align="center">
+  <img src="screenshot-chatting-ui.png" width="800" alt="Rune 데스크톱 UI — 4개 에이전트 협업" />
+</p>
+
 ---
 
 ## 사전 준비
@@ -67,6 +71,7 @@ Rune은 다른 접근을 합니다: **에이전트가 파일입니다.**
 | **스케줄링** | 수동 실행만 가능 | Cron, 파일 변경, git-commit 트리거 |
 | **권한** | 세션에서 상속 | 에이전트별 제어 (`fileWrite`, `bash`, `allowPaths`) |
 | **실행** | 대화형 | 헤드리스, 파이프라인, CI/CD 지원 |
+| **자기 수정** | 기본 제공 없음 | `rune loop` — 자동 리뷰-수정 반복 |
 
 Rune 에이전트는 세션, 머신, 팀을 넘어 살아남습니다. 한 번 만들면 영원히 실행.
 
@@ -210,6 +215,41 @@ rune pipe architect.rune coder.rune "Build a REST API with Express" --auto
 
 architect가 설계 → coder가 구현 (파일 작성, 의존성 설치).
 
+### 자기 수정 루프
+
+에이전트가 자동으로 자신의 작업을 리뷰하고 수정합니다:
+
+```bash
+rune loop coder.rune reviewer.rune "Build a REST API with Express" --until "no critical issues" --max-iterations 3 --auto
+```
+
+```
+🔁 Starting self-correction loop (max 3 iterations)
+   Stop condition: "no critical issues"
+
+  ━━━ Iteration 1/3 ━━━
+
+  ▶ [doer] coder — API 구현
+  ✓ coder done
+
+  ▶ [reviewer] reviewer — 치명적 이슈 2개 발견
+  ✓ reviewer done
+
+  ━━━ Iteration 2/3 ━━━
+
+  ▶ [doer] coder — 이슈 수정
+  ✓ coder done
+
+  ▶ [reviewer] reviewer — "no critical issues found"
+  ✓ reviewer done
+
+  ✅ Condition met: "no critical issues"
+
+🔁 Loop completed after 2 iterations
+```
+
+구현자가 구현하고, 리뷰어가 리뷰합니다. 문제가 발견되면 피드백이 자동으로 구현자에게 전달됩니다 — 조건이 충족되거나 최대 반복 횟수에 도달할 때까지.
+
 ### 자동화 트리거
 
 ```bash
@@ -301,6 +341,10 @@ rune open reviewer.rune
 
 또는 Finder에서 `.rune` 파일을 더블 클릭하세요.
 
+<p align="center">
+  <img src="screenshot-chatting-ui.png" width="800" alt="Rune 데스크톱 UI" />
+</p>
+
 - **실시간 활동** — 도구 호출, 결과, 권한 요청을 실시간으로 확인.
 - **내장 터미널** — Claude Code 출력과 사용자 명령어를 나란히.
 - **우클릭으로 생성** — macOS Quick Action으로 Finder에서 에이전트 생성.
@@ -316,6 +360,7 @@ rune open reviewer.rune
 | `rune new <name> [--role "..."]` | 에이전트 생성 |
 | `rune run <file> "prompt" [--auto] [--output json]` | 헤드리스 실행 |
 | `rune pipe <a> <b> [...] "prompt" [--auto]` | 에이전트 체이닝 |
+| `rune loop <doer> <reviewer> "prompt" [--until "..."] [--max-iterations N] [--auto]` | 자기 수정 루프 |
 | `rune watch <file> --on <event> --prompt "..."` | 자동화 트리거 |
 | `rune open <file>` | 데스크톱 UI |
 | `rune list` | 현재 디렉토리의 에이전트 목록 |

package/README.md

@@ -19,6 +19,10 @@
   <img src="https://img.shields.io/badge/license-MIT-green" alt="license" />
 </p>
 
+<p align="center">
+  <img src="screenshot-chatting-ui.png" width="800" alt="Rune Desktop UI — 4 agents collaborating" />
+</p>
+
 ---
 
 ## Prerequisites
@@ -67,6 +71,7 @@ Rune takes a different approach: **agents are files.**
 | **Scheduling** | Manual execution only | Cron, file-change, and git-commit triggers |
 | **Permissions** | Inherited from session | Per-agent controls (`fileWrite`, `bash`, `allowPaths`) |
 | **Execution** | Interactive | Headless, pipelines, CI/CD-ready |
+| **Self-correction** | Not built-in | `rune loop` — automatic review-fix cycles |
 
 Rune agents survive across sessions, machines, and teams. Build once, run forever.
 
@@ -210,6 +215,41 @@ rune pipe architect.rune coder.rune "Build a REST API with Express" --auto
 
 architect designs → coder implements (writes files, installs deps).
 
+### Self-correction loop
+
+Agents review and fix their own work automatically:
+
+```bash
+rune loop coder.rune reviewer.rune "Build a REST API with Express" --until "no critical issues" --max-iterations 3 --auto
+```
+
+```
+🔁 Starting self-correction loop (max 3 iterations)
+   Stop condition: "no critical issues"
+
+  ━━━ Iteration 1/3 ━━━
+
+  ▶ [doer] coder — implements the API
+  ✓ coder done
+
+  ▶ [reviewer] reviewer — finds 2 critical issues
+  ✓ reviewer done
+
+  ━━━ Iteration 2/3 ━━━
+
+  ▶ [doer] coder — fixes the issues
+  ✓ coder done
+
+  ▶ [reviewer] reviewer — "no critical issues found"
+  ✓ reviewer done
+
+  ✅ Condition met: "no critical issues"
+
+🔁 Loop completed after 2 iterations
+```
+
+The doer implements, the reviewer reviews. If issues are found, feedback goes back to the doer automatically — until the condition is met or max iterations are reached.
+
 ### Automated triggers
 
 ```bash
@@ -301,6 +341,10 @@ rune open reviewer.rune
 
 Or double-click any `.rune` file in Finder.
 
+<p align="center">
+  <img src="screenshot-chatting-ui.png" width="800" alt="Rune Desktop UI" />
+</p>
+
 - **Real-time activity** — See tool calls, results, and permission requests as they happen.
 - **Built-in terminal** — Claude Code output and your own commands, side by side.
 - **Right-click to create** — macOS Quick Action for creating agents from Finder.
@@ -316,6 +360,7 @@ Or double-click any `.rune` file in Finder.
 | `rune new <name> [--role "..."]` | Create agent |
 | `rune run <file> "prompt" [--auto] [--output json]` | Run headlessly |
 | `rune pipe <a> <b> [...] "prompt" [--auto]` | Chain agents |
+| `rune loop <doer> <reviewer> "prompt" [--until "..."] [--max-iterations N] [--auto]` | Self-correction loop |
 | `rune watch <file> --on <event> --prompt "..."` | Automated triggers |
 | `rune open <file>` | Desktop UI |
 | `rune list` | List agents in current directory |

package/bin/rune.js

@@ -27,6 +27,7 @@ switch (command) {
   case 'open':    return openRune(args[0])
   case 'run':     return runRune(args[0], args.slice(1))
   case 'pipe':    return pipeRunes(args)
+  case 'loop':    return loopRunes(args)
   case 'watch':   return watchRune(args[0], args.slice(1))
   case 'list':    return listRunes()
   case 'uninstall': return uninstall()
@@ -1156,6 +1157,236 @@ async function pipeRunes(args) {
   }
 }
 
+// ── loop (self-correction) ──────────────────────
+
+async function loopRunes(args) {
+  // Parse: rune loop coder.rune reviewer.rune "prompt" [--until "condition"] [--max-iterations N] [--auto]
+  const runeFiles = []
+  let prompt = ''
+  let untilCondition = ''
+  let maxIterations = 5
+  let autoMode = false
+
+  for (let i = 0; i < args.length; i++) {
+    if (args[i] === '--until' && args[i + 1]) {
+      untilCondition = args[++i]
+    } else if (args[i] === '--max-iterations' && args[i + 1]) {
+      maxIterations = parseInt(args[++i], 10)
+    } else if (args[i] === '--auto') {
+      autoMode = true
+    } else if (args[i].endsWith('.rune')) {
+      runeFiles.push(args[i])
+    } else if (!prompt) {
+      prompt = args[i]
+    }
+  }
+
+  if (runeFiles.length < 2 || !prompt) {
+    console.log('Usage: rune loop <doer.rune> <reviewer.rune> "prompt" [--until "condition"] [--max-iterations N] [--auto]')
+    console.log('')
+    console.log('Options:')
+    console.log('  --until "..."          Stop when the reviewer\'s output contains this text')
+    console.log('  --max-iterations N     Maximum number of loop iterations (default: 5)')
+    console.log('  --auto                 Allow agents to write files and run commands')
+    console.log('')
+    console.log('Example:')
+    console.log('  rune loop coder.rune reviewer.rune "Build a REST API" --until "no critical issues" --max-iterations 3 --auto')
+    console.log('')
+    console.log('The first agent implements, the last agent reviews.')
+    console.log('If the reviewer finds issues, feedback is sent back to the first agent automatically.')
+    process.exit(1)
+  }
+
+  // Validate files
+  for (const file of runeFiles) {
+    const filePath = path.resolve(process.cwd(), file)
+    if (!fs.existsSync(filePath)) {
+      console.error(`  ❌ File not found: ${filePath}`)
+      process.exit(1)
+    }
+  }
+
+  const doerFile = runeFiles[0]
+  const reviewerFile = runeFiles[runeFiles.length - 1]
+  const doerPath = path.resolve(process.cwd(), doerFile)
+  const reviewerPath = path.resolve(process.cwd(), reviewerFile)
+
+  let currentPrompt = prompt
+  let iteration = 0
+  let converged = false
+
+  console.log(`\n🔁 Starting self-correction loop (max ${maxIterations} iterations)`)
+  if (untilCondition) console.log(`   Stop condition: "${untilCondition}"`)
+  console.log('')
+
+  while (iteration < maxIterations && !converged) {
+    iteration++
+    console.log(`  ━━━ Iteration ${iteration}/${maxIterations} ━━━\n`)
+
+    // Step 1: Doer implements
+    const doerRune = JSON.parse(fs.readFileSync(doerPath, 'utf-8'))
+    const doerFolder = path.dirname(doerPath)
+    const doerSystem = []
+    if (doerRune.role) doerSystem.push(`Your role: ${doerRune.role}`)
+    if (doerRune.memory && doerRune.memory.length > 0) {
+      doerSystem.push('Saved memory:')
+      doerRune.memory.forEach((m, j) => doerSystem.push(`${j + 1}. ${m}`))
+    }
+
+    const doerContext = iteration > 1
+      ? `You are in iteration ${iteration} of a self-correction loop. The reviewer found issues with your previous work:\n\n${currentPrompt}\n\nFix the issues and improve your implementation.`
+      : currentPrompt
+
+    console.log(`  ▶ [doer] ${doerRune.name} (${doerRune.role || 'assistant'})`)
+
+    const doerOutput = await runAgent(doerRune.name, doerFolder, doerSystem, doerContext, autoMode)
+
+    doerRune.history = doerRune.history || []
+    doerRune.history.push({ role: 'user', text: doerContext, ts: Date.now() })
+    doerRune.history.push({ role: 'assistant', text: doerOutput, ts: Date.now() })
+    fs.writeFileSync(doerPath, JSON.stringify(doerRune, null, 2))
+
+    console.log(`  ✓ ${doerRune.name} done\n`)
+
+    // Step 2: Reviewer reviews
+    const reviewerRune = JSON.parse(fs.readFileSync(reviewerPath, 'utf-8'))
+    const reviewerFolder = path.dirname(reviewerPath)
+    const reviewerSystem = []
+    if (reviewerRune.role) reviewerSystem.push(`Your role: ${reviewerRune.role}`)
+    if (reviewerRune.memory && reviewerRune.memory.length > 0) {
+      reviewerSystem.push('Saved memory:')
+      reviewerRune.memory.forEach((m, j) => reviewerSystem.push(`${j + 1}. ${m}`))
+    }
+
+    const reviewerContext = `You are the reviewer in iteration ${iteration} of a self-correction loop. Review the work done by ${doerRune.name}:\n\n${doerOutput}\n\nIf there are issues, describe them clearly so the implementer can fix them. If the work is satisfactory, say so clearly.`
+
+    console.log(`  ▶ [reviewer] ${reviewerRune.name} (${reviewerRune.role || 'assistant'})`)
+
+    const reviewerOutput = await runAgent(reviewerRune.name, reviewerFolder, reviewerSystem, reviewerContext, false)
+
+    reviewerRune.history = reviewerRune.history || []
+    reviewerRune.history.push({ role: 'user', text: reviewerContext, ts: Date.now() })
+    reviewerRune.history.push({ role: 'assistant', text: reviewerOutput, ts: Date.now() })
+    fs.writeFileSync(reviewerPath, JSON.stringify(reviewerRune, null, 2))
+
+    console.log(`  ✓ ${reviewerRune.name} done\n`)
+
+    // Check convergence
+    if (untilCondition) {
+      const lower = reviewerOutput.toLowerCase()
+      if (lower.includes(untilCondition.toLowerCase())) {
+        converged = true
+        console.log(`  ✅ Condition met: "${untilCondition}"`)
+      }
+    }
+
+    if (!converged) {
+      currentPrompt = reviewerOutput
+    }
+  }
+
+  if (!converged && iteration >= maxIterations) {
+    console.log(`  ⚠️  Max iterations (${maxIterations}) reached`)
+  }
+
+  console.log(`\n🔁 Loop completed after ${iteration} iteration${iteration > 1 ? 's' : ''}\n`)
+}
+
+async function runAgent(name, folderPath, systemParts, prompt, autoMode) {
+  if (autoMode) {
+    const mcpPath = path.join(folderPath, '.mcp.json')
+    const mcpBackup = path.join(folderPath, '.mcp.json.loop.bak')
+    let mcpHidden = false
+    if (fs.existsSync(mcpPath)) {
+      fs.renameSync(mcpPath, mcpBackup)
+      mcpHidden = true
+    }
+
+    const claudeArgs = ['-p', '--print',
+      '--dangerously-skip-permissions',
+      '--verbose',
+      '--output-format', 'stream-json',
+    ]
+    if (systemParts.length > 0) {
+      claudeArgs.push('--system-prompt', systemParts.join('\n'))
+    }
+    claudeArgs.push(prompt)
+
+    const output = await new Promise((resolve, reject) => {
+      const child = spawn('claude', claudeArgs, {
+        cwd: folderPath,
+        stdio: ['ignore', 'pipe', 'pipe'],
+        env: { ...process.env },
+      })
+
+      let fullOutput = ''
+      let buffer = ''
+      child.stdout.on('data', (data) => {
+        buffer += data.toString()
+        const lines = buffer.split('\n')
+        buffer = lines.pop()
+        for (const line of lines) {
+          if (!line.trim()) continue
+          try {
+            const event = JSON.parse(line)
+            if (event.type === 'assistant' && event.message && event.message.content) {
+              for (const block of event.message.content) {
+                if (block.type === 'tool_use') {
+                  const tool = block.name || 'unknown'
+                  const input = block.input || {}
+                  if (tool === 'Bash') console.log(`    ▶ Bash: ${(input.command || '').slice(0, 120)}`)
+                  else if (tool === 'Write') console.log(`    ▶ Write: ${input.file_path || ''}`)
+                  else if (tool === 'Edit') console.log(`    ▶ Edit: ${input.file_path || ''}`)
+                  else if (tool === 'Read') console.log(`    ▶ Read: ${input.file_path || ''}`)
+                  else console.log(`    ▶ ${tool}`)
+                } else if (block.type === 'text' && block.text && block.text.trim()) {
+                  console.log(`    💬 ${block.text.trim().slice(0, 200)}`)
+                }
+              }
+            }
+            if (event.type === 'result') {
+              fullOutput = event.result || ''
+            }
+          } catch {}
+        }
+      })
+      child.stderr.on('data', (d) => { process.stderr.write(d) })
+      child.on('close', (code) => {
+        if (mcpHidden && fs.existsSync(mcpBackup)) fs.renameSync(mcpBackup, mcpPath)
+        if (code !== 0) reject(new Error(`Agent ${name} exited with code ${code}`))
+        else resolve(fullOutput.trim())
+      })
+    })
+
+    return output
+  } else {
+    const tmpdir = require('os').tmpdir()
+    const claudeArgs = ['-p', '--print', '--add-dir', folderPath]
+    if (systemParts.length > 0) {
+      claudeArgs.push('--system-prompt', systemParts.join('\n') + `\nWorking folder: ${folderPath}`)
+    }
+    claudeArgs.push('--', prompt)
+
+    const output = await new Promise((resolve, reject) => {
+      const child = spawn('claude', claudeArgs, {
+        cwd: tmpdir,
+        stdio: ['ignore', 'pipe', 'pipe'],
+        env: { ...process.env },
+      })
+
+      let stdout = ''
+      child.stdout.on('data', (d) => { stdout += d.toString() })
+      child.stderr.on('data', (d) => { process.stderr.write(d) })
+      child.on('close', (code) => {
+        if (code !== 0) reject(new Error(`Agent ${name} exited with code ${code}`))
+        else resolve(stdout.trim())
+      })
+    })
+
+    return output
+  }
+}
+
 // ── watch (triggers) ────────────────────────────
 
 function watchRune(file, restArgs) {
@@ -1425,6 +1656,10 @@ Usage:
     --log <file.json>       Save structured log (tool calls, cost, duration)
   rune pipe <a.rune> <b.rune> ... "prompt"   Chain agents in a pipeline
     --output json|text      Output format (default: text)
+  rune loop <doer.rune> <reviewer.rune> "prompt"   Self-correction loop
+    --until "condition"     Stop when reviewer output contains this text
+    --max-iterations N      Max iterations (default: 5)
+    --auto                  Allow agents to write files and run commands
   rune watch <file.rune>    Set up automated triggers
     --on <event>            Event: file-change, git-commit, git-push, cron
     --prompt "..."          Prompt to send when triggered
@@ -1438,6 +1673,7 @@ Examples:
   rune new reviewer --role "Code reviewer, security focused"
   rune run reviewer.rune "Review the latest commit"
   rune pipe coder.rune reviewer.rune "Implement a login page"
+  rune loop coder.rune reviewer.rune "Build a REST API" --until "no critical issues" --max-iterations 3 --auto
   rune watch reviewer.rune --on git-commit --prompt "Review this commit"
   rune watch monitor.rune --on cron --interval 5m --prompt "Check server health"
   echo "Fix the bug in auth.ts" | rune run backend.rune

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openrune",
-  "version": "1.0.1",
+  "version": "1.1.0",
   "description": "Persistent AI agents for Claude Code — build once, run forever.",
   "keywords": ["ai", "agent", "claude", "desktop", "electron", "mcp", "claude-code", "toolkit", "automation"],
   "repository": {