feat(lite-execute): intelligent task grouping with parallel/sequential execution

Major improvements:
- Implement dependency-aware task grouping algorithm
  * Same file modifications → sequential batch
  * Different files + no dependencies → parallel batch
  * Keyword-based dependency inference (use/integrate/call/import)
  * Batch limits: Agent 7 tasks, CLI 4 tasks

- Add parallel execution strategy
  * All parallel batches launch concurrently (single Claude message)
  * Sequential batches execute in order
  * No concurrency limit for CLI tools
  * Execution flow: parallel-first → sequential-after

- Enhance code review with task.json acceptance criteria
  * Unified review template for all tools (Gemini/Qwen/Codex/Agent)
  * task.json in CONTEXT field for CLI tools
  * Explicit acceptance criteria verification

- Clarify execution requirements
  * Remove ambiguous strategy instructions
  * Add explicit "MUST complete ALL tasks" requirement
  * Prevent CLI tools from partial execution misunderstanding

- Simplify documentation (~260 lines reduced)
  * Remove redundant examples and explanations
  * Keep core algorithm logic with inline comments
  * Eliminate duplicate descriptions across sections

Breaking changes: None (backward compatible)
This commit is contained in:
catlog22
2025-11-22 21:05:43 +08:00
parent 530c348e95
commit 4f533f6fd5

View File

@@ -185,276 +185,107 @@ Execution Complete
previousExecutionResults = []
```
### Step 2: Intelligent Task Grouping & Execution Planning
### Step 2: Task Grouping & Batch Creation
**Operations**:
- Analyze task dependencies and file targets
- Group tasks based on context similarity and dependencies
- Support parallel execution for independent tasks
- Respect task limits per execution (Agent: 7 tasks, CLI: 4 tasks)
**Task Analysis & Grouping**:
**Dependency Analysis & Grouping Algorithm**:
```javascript
// Extract file path from task for dependency analysis
function getTaskFile(task) {
return task.file || task.title.match(/in\s+([^\s:]+)/)?.[1] || null
}
// Infer dependencies from task descriptions and file paths
// Infer dependencies: same file → sequential, keywords (use/integrate) → sequential
function inferDependencies(tasks) {
return tasks.map((task, index) => {
const dependencies = []
const taskFile = getTaskFile(task)
const keywords = task.description?.toLowerCase() || task.title.toLowerCase()
return tasks.map((task, i) => {
const deps = []
const file = task.file || task.title.match(/in\s+([^\s:]+)/)?.[1]
const keywords = (task.description || task.title).toLowerCase()
// Check previous tasks for dependencies
for (let i = 0; i < index; i++) {
const prevTask = tasks[i]
const prevFile = getTaskFile(prevTask)
// Same file modification → sequential dependency
if (taskFile && prevFile === taskFile) {
dependencies.push(i)
continue
}
// Keyword-based dependency detection
if (keywords.includes('use') || keywords.includes('integrate') ||
keywords.includes('call') || keywords.includes('import')) {
const prevTitle = prevTask.title.toLowerCase()
if (keywords.includes(prevTitle.split(' ')[0])) {
dependencies.push(i)
}
}
for (let j = 0; j < i; j++) {
const prevFile = tasks[j].file || tasks[j].title.match(/in\s+([^\s:]+)/)?.[1]
if (file && prevFile === file) deps.push(j) // Same file
else if (/use|integrate|call|import/.test(keywords)) deps.push(j) // Keyword dependency
}
return { ...task, taskIndex: index, dependencies }
return { ...task, taskIndex: i, dependencies: deps }
})
}
// Group tasks into execution batches
// Group into batches: independent → parallel [P1,P2...], dependent → sequential [S1,S2...]
function createExecutionCalls(tasks, executionMethod) {
const tasksWithDeps = inferDependencies(tasks)
const maxTasksPerCall = executionMethod === "Codex" ? 4 : 7
const maxBatch = executionMethod === "Codex" ? 4 : 7
const calls = []
const processed = new Set()
// Phase 1: Group independent tasks for parallel execution
// Parallel: independent tasks, different files, max batch size
const parallelGroups = []
tasksWithDeps.forEach(task => {
if (task.dependencies.length === 0 && !processed.has(task.taskIndex)) {
const group = [task]
processed.add(task.taskIndex)
// Find other independent tasks with different files for parallel batch
const taskFile = getTaskFile(task)
tasksWithDeps.forEach(other => {
if (other.dependencies.length === 0 &&
!processed.has(other.taskIndex) &&
group.length < maxTasksPerCall) {
const otherFile = getTaskFile(other)
// Only group if different files (avoid conflicts)
if (!taskFile || !otherFile || taskFile !== otherFile) {
group.push(other)
processed.add(other.taskIndex)
}
tasksWithDeps.forEach(t => {
if (t.dependencies.length === 0 && !processed.has(t.taskIndex)) {
const group = [t]
processed.add(t.taskIndex)
tasksWithDeps.forEach(o => {
if (!o.dependencies.length && !processed.has(o.taskIndex) &&
group.length < maxBatch && t.file !== o.file) {
group.push(o)
processed.add(o.taskIndex)
}
})
parallelGroups.push(group)
}
})
// Phase 2: Group dependent tasks sequentially
const dependentTasks = tasksWithDeps.filter(t => !processed.has(t.taskIndex))
const sequentialBatches = []
while (dependentTasks.length > 0) {
const batch = []
const batchIndices = new Set()
for (const task of dependentTasks) {
// Can add to batch if dependencies are already processed
const depsProcessed = task.dependencies.every(dep =>
processed.has(dep) || batchIndices.has(dep)
)
if (depsProcessed && batch.length < maxTasksPerCall) {
batch.push(task)
batchIndices.add(task.taskIndex)
processed.add(task.taskIndex)
}
}
if (batch.length === 0) break // Prevent infinite loop
sequentialBatches.push(batch)
dependentTasks.splice(0, dependentTasks.length,
...dependentTasks.filter(t => !processed.has(t.taskIndex))
// Sequential: dependent tasks, batch when deps satisfied
const remaining = tasksWithDeps.filter(t => !processed.has(t.taskIndex))
while (remaining.length > 0) {
const batch = remaining.filter((t, i) =>
i < maxBatch && t.dependencies.every(d => processed.has(d))
)
if (!batch.length) break
batch.forEach(t => processed.add(t.taskIndex))
calls.push({ executionType: "sequential", groupId: `S${calls.length + 1}`, tasks: batch })
remaining.splice(0, remaining.length, ...remaining.filter(t => !processed.has(t.taskIndex)))
}
// Combine parallel and sequential batches
parallelGroups.forEach((group, i) => {
calls.push({
method: executionMethod === "Codex" ? "Codex" : "Agent",
executionType: "parallel",
groupId: `P${i + 1}`,
taskSummary: group.map(t => t.title).join(' | '),
tasks: group
})
})
sequentialBatches.forEach((batch, i) => {
calls.push({
method: executionMethod === "Codex" ? "Codex" : "Agent",
executionType: "sequential",
groupId: `S${i + 1}`,
taskSummary: batch.map(t => t.title).join(' → '),
tasks: batch
})
})
return calls
// Combine results
return [
...parallelGroups.map((g, i) => ({
method: executionMethod, executionType: "parallel", groupId: `P${i+1}`,
taskSummary: g.map(t => t.title).join(' | '), tasks: g
})),
...calls.map(c => ({ ...c, method: executionMethod, taskSummary: c.tasks.map(t => t.title).join(' ') }))
]
}
// Create execution calls with IDs and execution method
executionCalls = createExecutionCalls(planObject.tasks, executionMethod).map((call, index) => ({
...call,
id: `[${call.groupId}]`
}))
executionCalls = createExecutionCalls(planObject.tasks, executionMethod).map(c => ({ ...c, id: `[${c.groupId}]` }))
// Create TodoWrite list with execution type indicators
TodoWrite({
todos: executionCalls.map(call => {
const typeIcon = call.executionType === "parallel" ? "⚡" : "→"
const typeLabel = call.executionType === "parallel" ? "Parallel" : "Sequential"
return {
content: `${typeIcon} ${call.id} [${typeLabel}] (${call.tasks.length} tasks: ${call.taskSummary})`,
status: "pending",
activeForm: `Executing ${call.id} - ${typeLabel} batch (${call.tasks.length} tasks)`
}
})
todos: executionCalls.map(c => ({
content: `${c.executionType === "parallel" ? "⚡" : "→"} ${c.id} (${c.tasks.length} tasks)`,
status: "pending",
activeForm: `Executing ${c.id}`
}))
})
```
**Grouping Strategy Examples**:
```javascript
// Example 1: Independent tasks, different files → Parallel groups
Tasks: [
{ title: "Create auth.ts", file: "src/auth.ts" },
{ title: "Create utils.ts", file: "src/utils.ts" },
{ title: "Create types.ts", file: "src/types.ts" }
]
Result: [P1] Parallel (3 tasks)
// Example 2: Same file modifications → Sequential batch
Tasks: [
{ title: "Add function in auth.ts", file: "src/auth.ts" },
{ title: "Update function in auth.ts", file: "src/auth.ts" }
]
Result: [S1] Sequential (2 tasks)
// Example 3: Mixed dependencies → Multiple batches
Tasks: [
{ title: "Create base.ts", file: "src/base.ts" }, // No deps
{ title: "Create helper.ts", file: "src/helper.ts" }, // No deps
{ title: "Use base in main.ts", file: "src/main.ts" }, // Depends on base.ts
{ title: "Use helper in app.ts", file: "src/app.ts" } // Depends on helper.ts
]
Result:
[P1] Parallel (2 tasks: base.ts | helper.ts)
[P2] Parallel (2 tasks: main.ts | app.ts)
// Example 4: Exceeds batch limit → Multiple batches
Agent (7 tasks max): 10 tasks [P1: 7 tasks] + [P2: 3 tasks]
Codex (4 tasks max): 10 tasks [P1: 4 tasks] + [P2: 4 tasks] + [P3: 2 tasks]
```
**TodoWrite Display Examples**:
```
Parallel execution (independent files):
[ ] ⚡ [P1] [Parallel] (3 tasks: Create auth.ts | Create utils.ts | Create types.ts)
Sequential execution (same file or dependencies):
[ ] → [S1] [Sequential] (2 tasks: Add function in auth.ts → Update function in auth.ts)
Mixed execution (multiple batches):
[ ] ⚡ [P1] [Parallel] (4 tasks: Create base.ts | Create helper.ts | Create config.ts | Create types.ts)
[ ] ⚡ [P2] [Parallel] (3 tasks: Use base in main.ts | Use helper in app.ts | Use config in index.ts)
[ ] → [S1] [Sequential] (2 tasks: Integrate components → Add tests)
```
### Step 3: Launch Execution
**Execution Strategy**:
- **Parallel batches**: Launch all parallel batches simultaneously (CLI has no concurrency limit)
- **Sequential batches**: Execute in order, waiting for previous completion
- **Mixed workflow**: Process parallel groups first, then sequential groups
**Execution Loop**:
**Execution Flow**: Parallel batches concurrently → Sequential batches in order
```javascript
// Phase 1: Launch all parallel batches concurrently
const parallelCalls = executionCalls.filter(c => c.executionType === "parallel")
const sequentialCalls = executionCalls.filter(c => c.executionType === "sequential")
const parallel = executionCalls.filter(c => c.executionType === "parallel")
const sequential = executionCalls.filter(c => c.executionType === "sequential")
if (parallelCalls.length > 0) {
// Mark all parallel calls as in_progress
TodoWrite({
todos: executionCalls.map(call => ({
content: `${call.executionType === "parallel" ? "⚡" : "→"} ${call.id} [${call.executionType}] (${call.tasks.length} tasks)`,
status: call.executionType === "parallel" ? "in_progress" : "pending",
activeForm: `Executing ${call.id}`
}))
})
// Launch all parallel batches using single message with multiple tool calls
// Use Task tool or Bash tool to launch each parallel execution
// Collect results as they complete
parallelResults = await Promise.all(parallelCalls.map(call => executeBatch(call)))
// Phase 1: Launch all parallel batches (single message with multiple tool calls)
if (parallel.length > 0) {
TodoWrite({ todos: executionCalls.map(c => ({ status: c.executionType === "parallel" ? "in_progress" : "pending" })) })
parallelResults = await Promise.all(parallel.map(c => executeBatch(c)))
previousExecutionResults.push(...parallelResults)
// Mark parallel calls as completed
TodoWrite({
todos: executionCalls.map(call => ({
content: `${call.executionType === "parallel" ? "⚡" : "→"} ${call.id} [${call.executionType}] (${call.tasks.length} tasks)`,
status: parallelCalls.includes(call) ? "completed" : "pending",
activeForm: `Executing ${call.id}`
}))
})
TodoWrite({ todos: executionCalls.map(c => ({ status: parallel.includes(c) ? "completed" : "pending" })) })
}
// Phase 2: Execute sequential batches in order
for (const call of sequentialCalls) {
// Update TodoWrite: mark current sequential call in_progress
TodoWrite({
todos: executionCalls.map(c => ({
content: `${c.executionType === "parallel" ? "⚡" : "→"} ${c.id} [${c.executionType}] (${c.tasks.length} tasks)`,
status: c === call ? "in_progress" : (parallelCalls.includes(c) ? "completed" : "pending"),
activeForm: `Executing ${c.id}`
}))
})
// Launch sequential execution with previousExecutionResults context
// Phase 2: Execute sequential batches one by one
for (const call of sequential) {
TodoWrite({ todos: executionCalls.map(c => ({ status: c === call ? "in_progress" : "..." })) })
result = await executeBatch(call)
previousExecutionResults.push(result)
// Update TodoWrite: mark completed
TodoWrite({
todos: executionCalls.map(c => ({
content: `${c.executionType === "parallel" ? "⚡" : "→"} ${c.id} [${c.executionType}] (${c.tasks.length} tasks)`,
status: c.id <= call.id ? "completed" : "pending",
activeForm: `Executing ${c.id}`
}))
})
TodoWrite({ todos: executionCalls.map(c => ({ status: "completed" or "pending" })) })
}
```
**Important Notes**:
- Parallel batches use single Claude message with multiple Bash/Task tool calls
- CLI tools (Codex/Gemini/Qwen) execute in foreground (NOT background)
- Each batch receives `previousExecutionResults` for context continuity
**Option A: Agent Execution**
When to use:
@@ -617,35 +448,9 @@ bash_result = Bash(
**Result Collection**: After completion, analyze output and collect result following `executionResult` structure
### Step 4: Track Execution Progress
### Step 4: Progress Tracking
**Real-time TodoWrite Updates** at batch level:
```javascript
// Parallel batches executing concurrently
TodoWrite({
todos: [
{ content: "⚡ [P1] [Parallel] (4 tasks)", status: "in_progress", activeForm: "Executing P1" },
{ content: "⚡ [P2] [Parallel] (3 tasks)", status: "in_progress", activeForm: "Executing P2" },
{ content: "→ [S1] [Sequential] (2 tasks)", status: "pending", activeForm: "..." }
]
})
// After parallel completion, sequential execution
TodoWrite({
todos: [
{ content: "⚡ [P1] [Parallel] (4 tasks)", status: "completed", activeForm: "..." },
{ content: "⚡ [P2] [Parallel] (3 tasks)", status: "completed", activeForm: "..." },
{ content: "→ [S1] [Sequential] (2 tasks)", status: "in_progress", activeForm: "Executing S1" }
]
})
```
**User Visibility**:
- Parallel batches show simultaneous execution (multiple "in_progress" at once)
- Sequential batches execute one at a time
- Icons distinguish parallel (⚡) from sequential (→)
- Task count shows batch size for each execution group
Progress tracked at batch level (not individual task level). Icons: ⚡ (parallel, concurrent), → (sequential, one-by-one)
### Step 5: Code Review (Optional)
@@ -706,50 +511,9 @@ codex --full-auto exec "[Verify task.json acceptance criteria at ${task.json}]"
## Best Practices
### Execution Intelligence
1. **Intelligent Task Grouping**: Dependency-aware task allocation
- Same file modifications → Sequential batch
- Different files + no dependencies → Parallel batch
- Respects batch limits: Agent (7 tasks), CLI (4 tasks)
- Infers dependencies from file paths and keywords
2. **Parallel Execution**: Maximize throughput for independent tasks
- All parallel batches launch simultaneously
- No concurrency limit for CLI tools
- Reduces total execution time for independent work
- Uses single Claude message with multiple tool calls
3. **Context Continuity**: Each batch receives previous results
- Prevents duplication across batches
- Maintains coherent implementation flow
- Builds on completed work from parallel/sequential batches
4. **Flexible Execution**: Multiple input modes supported
- In-memory: Seamless lite-plan integration
- Prompt: Quick standalone execution
- File: Intelligent format detection
- Enhanced Task JSON (lite-plan export): Full plan extraction
- Plain text: Uses as prompt
### Task Management
1. **Smart Batch Creation**: Automatic dependency analysis
- File path extraction and comparison
- Keyword-based dependency inference
- Context signature grouping (same files = sequential)
- Batch size optimization per execution method
2. **Live Progress Tracking**: Real-time batch-level updates
- Visual distinction: ⚡ (parallel) vs → (sequential)
- Shows concurrent execution for parallel batches
- Clear batch completion tracking
- Task count per batch for progress visibility
3. **Execution Strategies**:
- Parallel-first: All independent batches execute concurrently
- Sequential-after: Dependent batches execute in order
- Context passing: Each batch receives all previous results
**Input Modes**: In-memory (lite-plan), prompt (standalone), file (JSON/text)
**Batch Limits**: Agent 7 tasks, CLI 4 tasks
**Execution**: Parallel batches use single Claude message with multiple tool calls (no concurrency limit)
## Error Handling