- Create codex version of unified-execute-with-file command
- Supports universal execution of planning/brainstorm/analysis output
- Coordinates multi-agents with smart dependency management
- Features parallel/sequential execution modes
- Unified event logging as single source of truth (execution-events.md)
- Agent context passing through previous execution history
- Knowledge chain: each agent reads full history of prior executions
Codex-specific adaptations:
- Use $VARIABLE format for argument substitution
- Simplified header configuration (description + argument-hint)
- Plan format agnostic parsing (IMPL_PLAN.md, synthesis.json, conclusions.json, debug recommendations)
- Multi-wave execution orchestration
- Dynamic artifact location handling
Execution flow:
1. Parse and validate plan from $PLAN_PATH
2. Extract and normalize tasks with dependencies
3. Create execution session (.workflow/.execution/{sessionId}/)
4. Group tasks into execution waves (topological sort)
5. Execute waves sequentially, tasks within wave execute in parallel
6. Unified event logging: execution-events.md (SINGLE SOURCE OF TRUTH)
7. Each agent reads previous executions for context
8. Final statistics and completion report
20 KiB
description, argument-hint
| description | argument-hint |
|---|---|
| Universal execution engine consuming planning/brainstorm/analysis output. Coordinates multi-agents, manages dependencies, and tracks execution with unified progress logging. | PLAN_PATH="<path>" [EXECUTION_MODE="sequential|parallel"] [AUTO_CONFIRM="yes|no"] [EXECUTION_CONTEXT="<focus area>"] |
Codex Unified-Execute-With-File Prompt
Overview
Universal execution engine that consumes any planning/brainstorm/analysis output and executes it with minimal progress tracking. Coordinates multiple agents (code-developer, test-fix-agent, doc-generator, cli-execution-agent), handles dependencies intelligently, and maintains unified execution timeline.
Core workflow: Load Plan → Parse Tasks → Validate Dependencies → Execute Waves → Track Progress → Report Results
Key features:
- Plan Format Agnostic: Consumes IMPL_PLAN.md, brainstorm synthesis.json, analysis conclusions.json, debug resolutions
- execution-events.md: Single source of truth - unified execution log with full agent history
- Multi-Agent Orchestration: Parallel execution where possible, sequential where needed
- Incremental Execution: Resume from failure point, no re-execution of completed tasks
- Dependency Management: Automatic topological sort and execution wave grouping
- Knowledge Chain: Each agent reads all previous execution history in context
Target Execution Plan
Plan Source: $PLAN_PATH
EXECUTION_MODE: Strategy (sequential|parallel)AUTO_CONFIRM: Skip confirmations (yes|no)EXECUTION_CONTEXT: Focus area/module (optional)
Execution Process
Session Detection:
├─ Check if execution session exists
├─ If exists → Resume mode
└─ If not → New session mode
Phase 1: Plan Loading & Validation
├─ Detect and parse plan file (multiple formats supported)
├─ Extract and normalize tasks
├─ Validate dependencies (detect cycles)
├─ Create execution session folder
├─ Initialize execution.md and execution-events.md
└─ Pre-execution validation
Phase 2: Execution Orchestration
├─ Topological sort for execution order
├─ Group tasks into execution waves (parallel-safe groups)
├─ Execute waves sequentially (tasks within wave execute in parallel)
├─ Monitor completion and capture artifacts
├─ Update progress in execution.md and execution-events.md
└─ Handle failures with retry/skip/abort logic
Phase 3: Progress Tracking & Unified Event Logging
├─ execution-events.md: Append-only unified log (SINGLE SOURCE OF TRUTH)
├─ Each agent reads all previous events at start
├─ Agent executes task with full context from previous agents
├─ Agent appends execution event (success/failure) with artifacts and notes
└─ Next agent reads complete history → knowledge chain
Phase 4: Completion & Summary
├─ Collect execution statistics
├─ Update execution.md with final status
├─ execution-events.md contains complete execution record
└─ Report results and offer follow-up options
Implementation Details
Session Setup & Plan Detection
const getUtc8ISOString = () => new Date(Date.now() + 8 * 60 * 60 * 1000).toISOString()
// Plan detection from $PLAN_PATH
let planPath = "$PLAN_PATH"
// If not provided, auto-detect
if (!planPath || planPath === "") {
const candidates = [
'.workflow/.plan/IMPL_PLAN.md',
'.workflow/plans/IMPL_PLAN.md',
'.workflow/IMPL_PLAN.md',
'.workflow/brainstorm/*/synthesis.json',
'.workflow/analyze/*/conclusions.json'
]
// Find most recent plan
planPath = findMostRecentFile(candidates)
if (!planPath) {
throw new Error("No execution plan found. Provide PLAN_PATH or ensure .workflow/IMPL_PLAN.md exists")
}
}
// Session setup
const executionMode = "$EXECUTION_MODE" || "parallel"
const autoConfirm = "$AUTO_CONFIRM" === "yes"
const executionContext = "$EXECUTION_CONTEXT" || ""
const planContent = Read(planPath)
const plan = parsePlan(planContent, planPath)
const executionId = `EXEC-${plan.slug}-${getUtc8ISOString().substring(0, 10)}-${randomId(4)}`
const executionFolder = `.workflow/.execution/${executionId}`
const executionPath = `${executionFolder}/execution.md`
const eventLogPath = `${executionFolder}/execution-events.md`
bash(`mkdir -p "${executionFolder}"`)
Plan Format Parsers
Support multiple plan sources:
function parsePlan(content, filePath) {
const ext = filePath.split('.').pop()
if (filePath.includes('IMPL_PLAN')) {
return parseImplPlan(content)
} else if (filePath.includes('brainstorm') && filePath.includes('synthesis')) {
return parseSynthesisPlan(content)
} else if (filePath.includes('analyze') && filePath.includes('conclusions')) {
return parseConclusionsPlan(content)
} else if (filePath.includes('debug') && filePath.includes('recommendations')) {
return parseDebugResolutionPlan(content)
} else if (ext === 'json' && content.includes('tasks')) {
return parseTaskJson(content)
}
throw new Error(`Unsupported plan format: ${filePath}`)
}
// IMPL_PLAN.md parser
function parseImplPlan(content) {
return {
type: 'impl-plan',
title: extractSection(content, 'Overview'),
phases: extractPhases(content),
tasks: extractTasks(content),
criticalFiles: extractCriticalFiles(content),
estimatedDuration: extractEstimate(content)
}
}
// Brainstorm synthesis.json parser
function parseSynthesisPlan(content) {
const synthesis = JSON.parse(content)
return {
type: 'brainstorm-synthesis',
title: synthesis.topic,
ideas: synthesis.top_ideas,
tasks: synthesis.top_ideas.map(idea => ({
id: `IDEA-${slugify(idea.title)}`,
type: 'investigation',
title: idea.title,
description: idea.description,
dependencies: [],
agent_type: 'universal-executor',
prompt: `Implement: ${idea.title}\n${idea.description}`,
expected_output: idea.next_steps
}))
}
}
Phase 1: Plan Loading & Validation
Step 1.1: Parse Plan and Extract Tasks
const tasks = plan.tasks || parseTasksFromContent(plan)
// Normalize task structure
const normalizedTasks = tasks.map(task => ({
id: task.id || `TASK-${generateId()}`,
title: task.title || task.content,
description: task.description || task.activeForm,
type: task.type || inferTaskType(task), // 'code', 'test', 'doc', 'analysis', 'integration'
agent_type: task.agent_type || selectBestAgent(task),
dependencies: task.dependencies || [],
// Execution parameters
prompt: task.prompt || task.description,
files_to_modify: task.files_to_modify || [],
expected_output: task.expected_output || [],
// Metadata
priority: task.priority || 'normal',
parallel_safe: task.parallel_safe !== false,
// Status tracking
status: 'pending',
attempts: 0,
max_retries: 2
}))
// Validate and detect issues
const validation = {
cycles: detectDependencyCycles(normalizedTasks),
missing_dependencies: findMissingDependencies(normalizedTasks),
file_conflicts: detectOutputConflicts(normalizedTasks),
warnings: []
}
if (validation.cycles.length > 0) {
throw new Error(`Circular dependencies detected: ${validation.cycles.join(', ')}`)
}
Step 1.2: Create execution.md
const executionMarkdown = `# Execution Progress
**Execution ID**: ${executionId}
**Plan Source**: ${planPath}
**Started**: ${getUtc8ISOString()}
**Mode**: ${executionMode}
**Plan Summary**:
- Title: ${plan.title}
- Total Tasks: ${normalizedTasks.length}
- Phases: ${plan.phases?.length || 'N/A'}
---
## Execution Plan
### Task Overview
| Task ID | Title | Type | Agent | Dependencies | Status |
|---------|-------|------|-------|--------------|--------|
${normalizedTasks.map(t => `| ${t.id} | ${t.title} | ${t.type} | ${t.agent_type} | ${t.dependencies.join(',')} | ${t.status} |`).join('\n')}
### Dependency Graph
\`\`\`
${generateDependencyGraph(normalizedTasks)}
\`\`\`
### Execution Strategy
- **Mode**: ${executionMode}
- **Parallelization**: ${calculateParallel(normalizedTasks)}
- **Estimated Duration**: ${estimateTotalDuration(normalizedTasks)}
---
## Execution Timeline
*Updates as execution progresses*
---
## Current Status
${executionStatus()}
`
Write(executionPath, executionMarkdown)
Step 1.3: Pre-Execution Confirmation
if (!autoConfirm) {
AskUserQuestion({
questions: [{
question: `准备执行 ${normalizedTasks.length} 个任务,模式: ${executionMode}\n\n关键任务:\n${normalizedTasks.slice(0, 3).map(t => `• ${t.id}: ${t.title}`).join('\n')}\n\n继续?`,
header: "Confirmation",
multiSelect: false,
options: [
{ label: "开始执行", description: "按计划执行" },
{ label: "调整参数", description: "修改执行参数" },
{ label: "查看详情", description: "查看完整任务列表" },
{ label: "取消", description: "退出不执行" }
]
}]
})
}
Phase 2: Execution Orchestration
Step 2.1: Determine Execution Order
// Topological sort for execution order
const executionOrder = topologicalSort(normalizedTasks)
// For parallel mode, group tasks into waves
let executionWaves = []
if (executionMode === 'parallel') {
executionWaves = groupIntoWaves(executionOrder, parallelLimit = 3)
} else {
executionWaves = executionOrder.map(task => [task])
}
Step 2.2: Execute Task Waves
let completedCount = 0
let failedCount = 0
const results = {}
for (let waveIndex = 0; waveIndex < executionWaves.length; waveIndex++) {
const wave = executionWaves[waveIndex]
console.log(`\n=== Wave ${waveIndex + 1}/${executionWaves.length} ===`)
console.log(`Tasks: ${wave.map(t => t.id).join(', ')}`)
// Launch tasks in parallel
const taskPromises = wave.map(task => executeTask(task, executionFolder))
// Wait for wave completion
const waveResults = await Promise.allSettled(taskPromises)
// Process results
for (let i = 0; i < waveResults.length; i++) {
const result = waveResults[i]
const task = wave[i]
if (result.status === 'fulfilled') {
results[task.id] = result.value
if (result.value.success) {
completedCount++
task.status = 'completed'
console.log(`✅ ${task.id}: Completed`)
} else if (result.value.retry) {
console.log(`⚠️ ${task.id}: Will retry`)
task.status = 'pending'
} else {
console.log(`❌ ${task.id}: Failed`)
}
} else {
console.log(`❌ ${task.id}: Execution error`)
}
}
// Update execution.md summary
appendExecutionTimeline(executionPath, waveIndex + 1, wave, waveResults)
}
Step 2.3: Execute Individual Task with Unified Event Logging
async function executeTask(task, executionFolder) {
const eventLogPath = `${executionFolder}/execution-events.md`
const startTime = Date.now()
try {
// Read previous execution events for context
let previousEvents = ''
if (fs.existsSync(eventLogPath)) {
previousEvents = Read(eventLogPath)
}
// Select agent based on task type
const agent = selectAgent(task.agent_type)
// Build execution context including previous agent outputs
const executionContext = `
## Previous Agent Executions (for reference)
${previousEvents}
---
## Current Task: ${task.id}
**Title**: ${task.title}
**Agent**: ${agent}
**Time**: ${getUtc8ISOString()}
### Description
${task.description}
### Context
- Modified Files: ${task.files_to_modify.join(', ')}
- Expected Output: ${task.expected_output.join(', ')}
### Requirements
${task.requirements || 'Follow the plan'}
### Constraints
${task.constraints || 'No breaking changes'}
`
// Execute based on agent type
let result
if (agent === 'code-developer' || agent === 'tdd-developer') {
result = await Task({
subagent_type: agent,
description: `Execute: ${task.title}`,
prompt: executionContext,
run_in_background: false
})
} else if (agent === 'test-fix-agent') {
result = await Task({
subagent_type: 'test-fix-agent',
description: `Execute Tests: ${task.title}`,
prompt: executionContext,
run_in_background: false
})
} else {
result = await Task({
subagent_type: 'universal-executor',
description: task.title,
prompt: executionContext,
run_in_background: false
})
}
// Capture artifacts
const artifacts = captureArtifacts(task, executionFolder)
// Append to unified execution events log
const eventEntry = `
## Task ${task.id} - COMPLETED ✅
**Timestamp**: ${getUtc8ISOString()}
**Duration**: ${calculateDuration(startTime)}ms
**Agent**: ${agent}
### Execution Summary
${generateSummary(result)}
### Key Outputs
${formatOutputs(result)}
### Generated Artifacts
${artifacts.map(a => `- **${a.type}**: \`${a.path}\` (${a.size})`).join('\n')}
### Notes for Next Agent
${generateNotesForNextAgent(result, task)}
---
`
appendToEventLog(eventLogPath, eventEntry)
return {
success: true,
task_id: task.id,
output: result,
artifacts: artifacts,
duration: calculateDuration(startTime)
}
} catch (error) {
// Append failure event to unified log
const failureEntry = `
## Task ${task.id} - FAILED ❌
**Timestamp**: ${getUtc8ISOString()}
**Duration**: ${calculateDuration(startTime)}ms
**Agent**: ${agent}
**Error**: ${error.message}
### Error Details
\`\`\`
${error.stack}
\`\`\`
### Recovery Notes for Next Attempt
${generateRecoveryNotes(error, task)}
---
`
appendToEventLog(eventLogPath, failureEntry)
// Handle failure: retry, skip, or abort
task.attempts++
if (task.attempts < task.max_retries && autoConfirm) {
console.log(`⚠️ ${task.id}: Failed, retrying (${task.attempts}/${task.max_retries})`)
return { success: false, task_id: task.id, error: error.message, retry: true, duration: calculateDuration(startTime) }
} else if (task.attempts >= task.max_retries && !autoConfirm) {
const decision = AskUserQuestion({
questions: [{
question: `任务失败: ${task.id}\n错误: ${error.message}`,
header: "Decision",
multiSelect: false,
options: [
{ label: "重试", description: "重新执行该任务" },
{ label: "跳过", description: "跳过此任务,继续下一个" },
{ label: "终止", description: "停止整个执行" }
]
}]
})
if (decision === 'retry') {
task.attempts = 0
return { success: false, task_id: task.id, error: error.message, retry: true, duration: calculateDuration(startTime) }
} else if (decision === 'skip') {
task.status = 'skipped'
skipDependentTasks(task.id, normalizedTasks)
} else {
throw new Error('Execution aborted by user')
}
} else {
task.status = 'failed'
skipDependentTasks(task.id, normalizedTasks)
}
return {
success: false,
task_id: task.id,
error: error.message,
duration: calculateDuration(startTime)
}
}
}
function appendToEventLog(logPath, eventEntry) {
if (fs.existsSync(logPath)) {
const currentContent = Read(logPath)
Write(logPath, currentContent + eventEntry)
} else {
Write(logPath, eventEntry)
}
}
Phase 3: Progress Tracking & Event Logging
execution-events.md is the SINGLE SOURCE OF TRUTH:
- Append-only, chronological execution log
- Each task records: timestamp, duration, agent type, execution summary, artifacts, notes for next agent
- Failures include error details and recovery notes
- Format: Human-readable markdown with machine-parseable status indicators (✅/❌/⏳)
Event log format (appended entry):
## Task {id} - {STATUS} {emoji}
**Timestamp**: {time}
**Duration**: {ms}
**Agent**: {type}
### Execution Summary
{What was done}
### Generated Artifacts
- `src/types/auth.ts` (2.3KB)
### Notes for Next Agent
- Key decisions made
- Potential issues
- Ready for: TASK-003
Phase 4: Completion & Summary
After all tasks complete or max failures reached:
const statistics = {
total_tasks: normalizedTasks.length,
completed: normalizedTasks.filter(t => t.status === 'completed').length,
failed: normalizedTasks.filter(t => t.status === 'failed').length,
skipped: normalizedTasks.filter(t => t.status === 'skipped').length,
success_rate: (completedCount / normalizedTasks.length * 100).toFixed(1)
}
// Update execution.md with final status
appendExecutionSummary(executionPath, statistics)
Post-Completion Options (unless auto-confirm):
AskUserQuestion({
questions: [{
question: "执行完成。是否需要后续操作?",
header: "Next Steps",
multiSelect: true,
options: [
{ label: "查看详情", description: "查看完整执行日志" },
{ label: "调试失败项", description: "对失败任务进行调试" },
{ label: "优化执行", description: "分析执行改进建议" },
{ label: "完成", description: "不需要后续操作" }
]
}]
})
Session Folder Structure
.workflow/.execution/{executionId}/
├── execution.md # Execution plan and overall status
└── execution-events.md # SINGLE SOURCE OF TRUTH - all agent executions
# Both human-readable AND machine-parseable
# Generated files go directly to project directories (not into execution folder)
# E.g., TASK-001 generates: src/types/auth.ts (not artifacts/src/types/auth.ts)
# execution-events.md records the actual project paths
Agent Selection Strategy
function selectBestAgent(task) {
if (task.type === 'code' || task.type === 'implementation') {
return task.includes_tests ? 'tdd-developer' : 'code-developer'
} else if (task.type === 'test' || task.type === 'test-fix') {
return 'test-fix-agent'
} else if (task.type === 'doc' || task.type === 'documentation') {
return 'doc-generator'
} else if (task.type === 'analysis' || task.type === 'investigation') {
return 'cli-execution-agent'
} else if (task.type === 'debug') {
return 'debug-explore-agent'
} else {
return 'universal-executor'
}
}
Parallelization Rules
function calculateParallel(tasks) {
// Group tasks into execution waves
// Constraints:
// - Tasks with same file modifications must be sequential
// - Tasks with dependencies must wait
// - Max 3 parallel tasks per wave (resource constraint)
const waves = []
const completed = new Set()
while (completed.size < tasks.length) {
const available = tasks.filter(t =>
!completed.has(t.id) &&
t.dependencies.every(d => completed.has(d))
)
if (available.length === 0) break
// Check for file conflicts
const noConflict = []
const modifiedFiles = new Set()
for (const task of available) {
const conflicts = task.files_to_modify.some(f => modifiedFiles.has(f))
if (!conflicts && noConflict.length < 3) {
noConflict.push(task)
task.files_to_modify.forEach(f => modifiedFiles.add(f))
}
}
if (noConflict.length > 0) {
waves.push(noConflict)
noConflict.forEach(t => completed.add(t.id))
}
}
return waves
}
Error Handling & Recovery
| Situation | Action |
|---|---|
| Task timeout | Mark as timeout, ask user: retry/skip/abort |
| Missing dependency | Auto-skip dependent tasks, log warning |
| File conflict | Detect before execution, ask for resolution |
| Output mismatch | Validate against expected_output, flag for review |
| Agent unavailable | Fallback to universal-executor |
Usage Recommendations
Use this execution engine when:
- Executing any planning document (IMPL_PLAN.md, brainstorm conclusions, analysis recommendations)
- Multiple tasks with dependencies need orchestration
- Want minimal progress tracking without clutter
- Need to handle failures gracefully and resume
- Want to parallelize where possible but ensure correctness
Consumes output from:
/workflow:plan→ IMPL_PLAN.md/workflow:brainstorm-with-file→ synthesis.json → execution/workflow:analyze-with-file→ conclusions.json → execution/workflow:debug-with-file→ recommendations → execution/workflow:lite-plan→ task JSONs → execution
Now execute the unified execution workflow for plan: $PLAN_PATH