mirror of
https://github.com/catlog22/Claude-Code-Workflow.git
synced 2026-02-28 09:23:08 +08:00
- Created a new orchestrator template for Codex skill design, detailing structure and execution phases. - Introduced the executor role with responsibilities for task execution, including routing to backends and handling implementation. - Added the planner role for requirement breakdown, issue creation, and task dispatching, ensuring a structured planning process.
10 KiB
10 KiB
Command Pattern Template
Pre-built Codex command patterns for common agent interaction scenarios.
Purpose
| Phase | Usage |
|---|---|
| Phase 0 | Read to understand available command patterns |
| Phase 2 | Select appropriate patterns for orchestrator |
| Phase 3 | Apply patterns to agent definitions |
Pattern 1: Explore (Parallel Fan-out)
Use When: Multi-angle codebase exploration needed.
// ==================== Explore Pattern ====================
// Step 1: Define exploration angles
const angles = ["architecture", "dependencies", "patterns", "testing"]
// Step 2: Create parallel exploration agents
const agents = angles.map(angle =>
spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/cli-explore-agent.md (MUST read first)
2. Read: .workflow/project-tech.json
---
Goal: Execute ${angle} exploration for ${task_description}
Scope:
- Include: All source files relevant to ${angle}
- Exclude: node_modules, dist, build artifacts
Context:
- Task: ${task_description}
- Angle: ${angle}
Deliverables:
- Structured findings following output template
- File:line references for key discoveries
- Open questions for unclear areas
Quality bar:
- At least 3 relevant files identified
- Findings backed by concrete evidence
`
})
)
// Step 3: Batch wait
const results = wait({ ids: agents, timeout_ms: 600000 })
// Step 4: Aggregate
const findings = agents.map((id, i) => ({
angle: angles[i],
result: results.status[id].completed
}))
// Step 5: Cleanup
agents.forEach(id => close_agent({ id }))
Pattern 2: Analyze (Multi-Perspective)
Use When: Code analysis from multiple dimensions needed.
// ==================== Analyze Pattern ====================
const perspectives = [
{ name: "security", focus: "OWASP Top 10, injection, auth bypass" },
{ name: "performance", focus: "O(n²), memory leaks, blocking I/O" },
{ name: "maintainability", focus: "complexity, coupling, duplication" }
]
const agents = perspectives.map(p =>
spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/cli-explore-agent.md (MUST read first)
---
Goal: Analyze ${targetModule} from ${p.name} perspective
Focus: ${p.focus}
Scope:
- Include: ${targetPaths}
- Exclude: Test files, generated code
Deliverables:
- Severity-classified findings (Critical/High/Medium/Low)
- File:line references for each finding
- Remediation recommendations
Quality bar:
- Every finding must have evidence (code reference)
- Remediation must be actionable
`
})
)
const results = wait({ ids: agents, timeout_ms: 600000 })
// Merge findings by severity
const merged = {
critical: [], high: [], medium: [], low: []
}
agents.forEach((id, i) => {
const parsed = parseFindings(results.status[id].completed)
Object.keys(merged).forEach(sev => merged[sev].push(...(parsed[sev] || [])))
})
agents.forEach(id => close_agent({ id }))
Pattern 3: Implement (Sequential Delegation)
Use When: Code implementation following a plan.
// ==================== Implement Pattern ====================
const implementAgent = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/code-developer.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json
---
Goal: Implement ${featureDescription}
Scope:
- Include: ${targetPaths}
- Exclude: Unrelated modules
- Constraints: No breaking changes, follow existing patterns
Context:
- Plan: ${planContent}
- Dependencies: ${dependencies}
- Existing patterns: ${patterns}
Deliverables:
- Working implementation following plan
- Updated/new test files
- Summary of changes with file:line references
Quality bar:
- All existing tests pass
- New code follows project conventions
- No TypeScript errors
- Backward compatible
`
})
const result = wait({ ids: [implementAgent], timeout_ms: 900000 })
// Check for open questions (might need clarification)
if (result.status[implementAgent].completed.includes('CLARIFICATION_NEEDED')) {
// Handle clarification via send_input
const answers = getUserAnswers(result)
send_input({ id: implementAgent, message: `## ANSWERS\n${answers}\n\n## CONTINUE\nProceed with implementation.` })
const final = wait({ ids: [implementAgent], timeout_ms: 900000 })
}
close_agent({ id: implementAgent })
Pattern 4: Validate (Test-Fix Cycle)
Use When: Running tests and fixing failures iteratively.
// ==================== Validate Pattern ====================
const validateAgent = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/test-fix-agent.md (MUST read first)
---
Goal: Validate ${component} — run tests, fix failures, iterate
Scope:
- Include: ${testPaths}
- Exclude: Unrelated test suites
Context:
- Recent changes: ${changedFiles}
- Test framework: ${testFramework}
Deliverables:
- All tests passing (or documented blocked tests)
- Fix summary with file:line references
- Coverage report
Quality bar:
- Pass rate >= 95%
- No new test regressions
- Max 5 fix iterations
`
})
const round1 = wait({ ids: [validateAgent], timeout_ms: 600000 })
// Check if more iterations needed
let iteration = 1
while (
iteration < 5 &&
round1.status[validateAgent].completed.includes('TESTS_FAILING')
) {
send_input({
id: validateAgent,
message: `## ITERATION ${iteration + 1}\nContinue fixing remaining failures. Focus on:\n${remainingFailures}`
})
const roundN = wait({ ids: [validateAgent], timeout_ms: 300000 })
iteration++
}
close_agent({ id: validateAgent })
Pattern 5: Review (Multi-Dimensional)
Use When: Code review from multiple dimensions.
// ==================== Review Pattern ====================
const dimensions = [
{ name: "correctness", agent: "cli-explore-agent" },
{ name: "security", agent: "cli-explore-agent" },
{ name: "performance", agent: "cli-explore-agent" },
{ name: "style", agent: "cli-explore-agent" }
]
const agents = dimensions.map(d =>
spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/${d.agent}.md (MUST read first)
---
Goal: Review ${targetCode} for ${d.name}
Scope: ${changedFiles}
Deliverables: Findings with severity, file:line, remediation
`
})
)
const results = wait({ ids: agents, timeout_ms: 600000 })
// Aggregate review findings
const review = {
approved: true,
findings: [],
blockers: []
}
agents.forEach((id, i) => {
const parsed = parseReview(results.status[id].completed)
review.findings.push(...parsed.findings)
if (parsed.blockers.length > 0) {
review.approved = false
review.blockers.push(...parsed.blockers)
}
})
agents.forEach(id => close_agent({ id }))
Pattern 6: Deep Interact (Merged Explore + Plan)
Use When: Exploration and planning are tightly coupled.
// ==================== Deep Interact Pattern ====================
const agent = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/cli-explore-agent.md (MUST read first)
2. **Also read**: ~/.codex/agents/cli-lite-planning-agent.md (dual role)
---
### Phase A: Exploration
Goal: Explore codebase for ${task_description}
Output: Structured findings + CLARIFICATION_NEEDED questions (if any)
### Phase B: Planning (activated after clarification)
Goal: Generate implementation plan based on exploration + answers
Output: Structured plan following plan schema
Deliverables:
- Phase A: exploration findings (Summary/Findings/Open questions)
- Phase B: implementation plan (after receiving clarification answers)
`
})
// Phase A: Exploration
const exploration = wait({ ids: [agent], timeout_ms: 600000 })
if (exploration.status[agent].completed.includes('CLARIFICATION_NEEDED')) {
const answers = getUserAnswers(exploration)
// Phase B: Planning (same agent, preserved context)
send_input({
id: agent,
message: `
## CLARIFICATION ANSWERS
${answers}
## PROCEED TO PHASE B
Generate implementation plan based on your exploration findings and these answers.
`
})
const plan = wait({ ids: [agent], timeout_ms: 900000 })
}
close_agent({ id: agent })
Pattern 7: Two-Phase (Clarify → Execute)
Use When: Task requires explicit clarification before execution.
// ==================== Two-Phase Pattern ====================
// Phase 1: Clarification
const agent = spawn_agent({
message: `
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/${agentType}.md (MUST read first)
---
### PHASE: CLARIFICATION ONLY
Goal: Understand ${task_description} and identify unclear points
Output ONLY:
1. Your understanding of the task (2-3 sentences)
2. CLARIFICATION_NEEDED questions (if any)
3. Recommended approach (1-2 sentences)
DO NOT execute any changes yet.
`
})
const clarification = wait({ ids: [agent], timeout_ms: 300000 })
// Collect user confirmation/answers
const userResponse = processUserInput(clarification)
// Phase 2: Execution
send_input({
id: agent,
message: `
## USER CONFIRMATION
${userResponse}
## PROCEED TO EXECUTION
Now execute the task with full implementation.
Output: Complete deliverable following structured output template.
`
})
const execution = wait({ ids: [agent], timeout_ms: 900000 })
close_agent({ id: agent })
Pattern Selection Guide
| Scenario | Recommended Pattern | Reason |
|---|---|---|
| Explore codebase from N angles | Pattern 1: Explore | Parallel fan-out, independent angles |
| Analyze code quality | Pattern 2: Analyze | Multi-perspective, severity classification |
| Implement from plan | Pattern 3: Implement | Sequential, plan-driven |
| Run tests + fix | Pattern 4: Validate | Iterative send_input loop |
| Code review | Pattern 5: Review | Multi-dimensional, aggregated verdict |
| Explore then plan | Pattern 6: Deep Interact | Context preservation, merged phases |
| Complex/unclear task | Pattern 7: Two-Phase | Clarify first, reduce rework |
| Simple one-shot task | Standard (no pattern) | spawn → wait → close |