feat(codex): convert team-planex skill to Codex-native format

Add Codex skill package with spawn_agent/wait/send_input/close_agent
patterns, replacing Claude Task/TeamCreate/SendMessage primitives.
This commit is contained in:
catlog22
2026-02-23 23:19:19 +08:00
parent 07b9d06cad
commit 9e8d6af9f1
3 changed files with 1023 additions and 0 deletions

View File

@@ -0,0 +1,416 @@
---
name: team-planex
description: |
2-member plan-and-execute pipeline with Wave Pipeline for concurrent planning and execution.
Planner decomposes requirements into issues, generates solutions, forms execution queues.
Executor implements solutions via configurable backends (agent/codex/gemini).
agents: 2
phases: 4
---
# Team PlanEx
2 成员边规划边执行团队。通过 Wave Pipeline波次流水线实现 planner 和 executor 并行工作planner 完成一个 wave 的 queue 后orchestrator 立即 spawn executor agent 处理该 wave同时 send_input 让 planner 继续下一 wave。
## Architecture Overview
```
┌──────────────────────────────────────────────┐
│ Orchestrator (this file) │
│ → Parse input → Spawn planner → Spawn exec │
└────────────────┬─────────────────────────────┘
│ Wave Pipeline
┌───────┴───────┐
↓ ↓
┌─────────┐ ┌──────────┐
│ planner │ │ executor │
│ (plan) │ │ (impl) │
└─────────┘ └──────────┘
│ │
issue-plan-agent code-developer
issue-queue-agent (or codex/gemini CLI)
```
## Agent Registry
| Agent | Role File | Responsibility | New/Existing |
|-------|-----------|----------------|--------------|
| `planex-planner` | `.codex/skills/team-planex/agents/planex-planner.md` | 需求拆解 → issue 创建 → 方案设计 → 队列编排 | New (skill-specific) |
| `planex-executor` | `.codex/skills/team-planex/agents/planex-executor.md` | 加载 solution → 代码实现 → 测试 → 提交 | New (skill-specific) |
| `issue-plan-agent` | `~/.codex/agents/issue-plan-agent.md` | ACE exploration + solution generation + binding | Existing |
| `issue-queue-agent` | `~/.codex/agents/issue-queue-agent.md` | Solution ordering + conflict detection | Existing |
| `code-developer` | `~/.codex/agents/code-developer.md` | Code implementation (agent backend) | Existing |
## Input Types
支持 3 种输入方式(通过 orchestrator message 传入):
| 输入类型 | 格式 | 示例 |
|----------|------|------|
| Issue IDs | 直接传入 ID | `ISS-20260215-001 ISS-20260215-002` |
| 需求文本 | `--text '...'` | `--text '实现用户认证模块'` |
| Plan 文件 | `--plan path` | `--plan plan/2026-02-15-auth.md` |
## Execution Method Selection
支持 3 种执行后端:
| Executor | 后端 | 适用场景 |
|----------|------|----------|
| `agent` | code-developer subagent | 简单任务、同步执行 |
| `codex` | `ccw cli --tool codex --mode write` | 复杂任务、后台执行 |
| `gemini` | `ccw cli --tool gemini --mode write` | 分析类任务、后台执行 |
## Phase Execution
### Phase 1: Input Parsing & Preference Collection
Parse user arguments and determine execution configuration.
```javascript
// Parse input from orchestrator message
const args = orchestratorMessage
const issueIds = args.match(/ISS-\d{8}-\d{6}/g) || []
const textMatch = args.match(/--text\s+['"]([^'"]+)['"]/)
const planMatch = args.match(/--plan\s+(\S+)/)
const autoYes = /\b(-y|--yes)\b/.test(args)
const explicitExec = args.match(/--exec[=\s]+(agent|codex|gemini|auto)/i)?.[1]
let executionConfig
if (explicitExec) {
executionConfig = {
executionMethod: explicitExec.charAt(0).toUpperCase() + explicitExec.slice(1),
codeReviewTool: "Skip"
}
} else if (autoYes) {
executionConfig = { executionMethod: "Auto", codeReviewTool: "Skip" }
} else {
// Interactive: ask user for preferences
// (orchestrator handles user interaction directly)
}
```
### Phase 2: Planning (Planner Agent — Deep Interaction)
Spawn planner agent for wave-based planning. Uses send_input for multi-wave progression.
```javascript
// Build planner input context
let plannerInput = ""
if (issueIds.length > 0) plannerInput = `issue_ids: ${JSON.stringify(issueIds)}`
else if (textMatch) plannerInput = `text: ${textMatch[1]}`
else if (planMatch) plannerInput = `plan_file: ${planMatch[1]}`
const planner = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: .codex/skills/team-planex/agents/planex-planner.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json
---
Goal: Decompose requirements into waves of executable solutions
## Input
${plannerInput}
## Execution Config
execution_method: ${executionConfig.executionMethod}
code_review: ${executionConfig.codeReviewTool}
## Deliverables
For EACH wave, output structured wave data:
\`\`\`
WAVE_READY:
wave_number: N
issue_ids: [ISS-xxx, ...]
queue_path: .workflow/issues/queue/execution-queue.json
exec_tasks: [
{ issue_id: "ISS-xxx", solution_id: "SOL-xxx", title: "...", priority: "normal", depends_on: [] },
...
]
\`\`\`
After ALL waves planned, output:
\`\`\`
ALL_PLANNED:
total_waves: N
total_issues: N
\`\`\`
## Quality bar
- Every issue has a bound solution
- Queue respects dependency DAG
- Wave boundaries are logical groupings
`
})
// Wait for Wave 1
const wave1 = wait({ ids: [planner], timeout_ms: 600000 })
if (wave1.timed_out) {
send_input({ id: planner, message: "Please finalize current wave and output WAVE_READY." })
const retry = wait({ ids: [planner], timeout_ms: 120000 })
}
// Parse wave data from planner output
const wave1Data = parseWaveReady(wave1.status[planner].completed)
```
### Phase 3: Wave Pipeline (Planning + Execution Interleaved)
Pipeline: spawn executor for current wave while planner continues next wave.
```javascript
const allAgentIds = [planner]
const executorAgents = []
let waveNum = 1
let allPlanned = false
while (!allPlanned) {
// --- Spawn executor for current wave ---
const waveData = parseWaveReady(currentWaveOutput)
if (waveData && waveData.exec_tasks.length > 0) {
const executor = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: .codex/skills/team-planex/agents/planex-executor.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json
---
Goal: Implement all solutions in Wave ${waveNum}
## Wave ${waveNum} Tasks
${JSON.stringify(waveData.exec_tasks, null, 2)}
## Execution Config
execution_method: ${executionConfig.executionMethod}
code_review: ${executionConfig.codeReviewTool}
## Deliverables
For each task, output:
\`\`\`
IMPL_COMPLETE:
issue_id: ISS-xxx
status: success|failed
test_result: pass|fail
commit: <hash or N/A>
\`\`\`
After all wave tasks done:
\`\`\`
WAVE_DONE:
wave_number: ${waveNum}
completed: N
failed: N
\`\`\`
## Quality bar
- All existing tests pass after each implementation
- Code follows project conventions
- One commit per solution
`
})
allAgentIds.push(executor)
executorAgents.push({ id: executor, wave: waveNum })
}
// --- Tell planner to continue next wave ---
if (!allPlanned) {
send_input({ id: planner, message: `Wave ${waveNum} dispatched to executor. Continue to Wave ${waveNum + 1}.` })
// Wait for both: planner (next wave) + current executor
const activeIds = [planner]
if (executorAgents.length > 0) {
activeIds.push(executorAgents[executorAgents.length - 1].id)
}
const results = wait({ ids: activeIds, timeout_ms: 600000 })
// Check planner output
const plannerOutput = results.status[planner]?.completed || ""
if (plannerOutput.includes("ALL_PLANNED")) {
allPlanned = true
} else if (plannerOutput.includes("WAVE_READY")) {
waveNum++
currentWaveOutput = plannerOutput
}
}
}
// Wait for remaining executor agents
const pendingExecutors = executorAgents
.map(e => e.id)
.filter(id => !completedIds.includes(id))
if (pendingExecutors.length > 0) {
const finalResults = wait({ ids: pendingExecutors, timeout_ms: 900000 })
// Handle timeout
if (finalResults.timed_out) {
const pending = pendingExecutors.filter(id => !finalResults.status[id]?.completed)
pending.forEach(id => {
send_input({ id, message: "Please finalize current task and output results." })
})
wait({ ids: pending, timeout_ms: 120000 })
}
}
```
### Phase 4: Result Aggregation & Cleanup
```javascript
// Collect results from all executors
const pipelineResults = {
waves: [],
totalCompleted: 0,
totalFailed: 0
}
executorAgents.forEach(({ id, wave }) => {
const output = results.status[id]?.completed || ""
const waveDone = parseWaveDone(output)
pipelineResults.waves.push({
wave,
completed: waveDone?.completed || 0,
failed: waveDone?.failed || 0
})
pipelineResults.totalCompleted += waveDone?.completed || 0
pipelineResults.totalFailed += waveDone?.failed || 0
})
// Output final summary
console.log(`
## PlanEx Pipeline Complete
### Summary
- Total Waves: ${waveNum}
- Total Completed: ${pipelineResults.totalCompleted}
- Total Failed: ${pipelineResults.totalFailed}
### Wave Details
${pipelineResults.waves.map(w =>
`- Wave ${w.wave}: ${w.completed} completed, ${w.failed} failed`
).join('\n')}
`)
// Cleanup ALL agents
allAgentIds.forEach(id => {
try { close_agent({ id }) } catch { /* already closed */ }
})
```
## Coordination Protocol
### File-Based Communication
Since Codex agents have isolated contexts, use file-based coordination:
| File | Purpose | Writer | Reader |
|------|---------|--------|--------|
| `.workflow/.team/PEX-{slug}-{date}/wave-{N}.json` | Wave plan data | planner | orchestrator |
| `.workflow/.team/PEX-{slug}-{date}/exec-{issueId}.json` | Execution result | executor | orchestrator |
| `.workflow/.team/PEX-{slug}-{date}/pipeline-log.ndjson` | Event log | both | orchestrator |
| `.workflow/issues/queue/execution-queue.json` | Execution queue | planner (via issue-queue-agent) | executor |
### Wave Data Format
```json
{
"wave_number": 1,
"issue_ids": ["ISS-20260215-001", "ISS-20260215-002"],
"queue_path": ".workflow/issues/queue/execution-queue.json",
"exec_tasks": [
{
"issue_id": "ISS-20260215-001",
"solution_id": "SOL-001",
"title": "Implement auth module",
"priority": "high",
"depends_on": []
}
]
}
```
### Execution Result Format
```json
{
"issue_id": "ISS-20260215-001",
"status": "success",
"executor": "agent",
"test_result": "pass",
"commit": "abc123",
"files_changed": ["src/auth/login.ts", "src/auth/login.test.ts"]
}
```
## Lifecycle Management
### Timeout Handling
| Timeout Scenario | Action |
|-----------------|--------|
| Planner wave timeout | send_input to urge convergence, retry wait |
| Executor impl timeout | send_input to finalize, record partial result |
| All agents timeout | Log error, abort with partial state |
### Cleanup Protocol
```javascript
// Track all agents created during execution
const allAgentIds = []
// ... (agents added during phase execution) ...
// Final cleanup (end of orchestrator or on error)
allAgentIds.forEach(id => {
try { close_agent({ id }) } catch { /* already closed */ }
})
```
## Error Handling
| Scenario | Resolution |
|----------|------------|
| Planner wave failure | Retry once via send_input, then abort pipeline |
| Executor impl failure | Record failure, continue with next wave tasks |
| No issues created from text | Report to user, abort |
| Solution generation failure | Skip issue, continue with remaining |
| Queue formation failure | Create exec tasks without DAG ordering |
| Pipeline stall (no progress) | Timeout handling → urge convergence → abort |
| Missing role file | Log error, use inline fallback instructions |
## Helper Functions
```javascript
function parseWaveReady(output) {
const match = output.match(/WAVE_READY:\s*\n([\s\S]*?)(?=\n```|$)/)
if (!match) return null
// Parse structured wave data
return JSON.parse(match[1])
}
function parseWaveDone(output) {
const match = output.match(/WAVE_DONE:\s*\n([\s\S]*?)(?=\n```|$)/)
if (!match) return null
return JSON.parse(match[1])
}
function resolveExecutor(method, taskCount) {
if (method.toLowerCase() === 'auto') {
return taskCount <= 3 ? 'agent' : 'codex'
}
return method.toLowerCase()
}
```