- 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
Codex Agent Patterns
Core Codex subagent API patterns reference for skill generation.
Purpose
| Phase | Usage |
|---|---|
| Phase 0 | Read to understand available Codex patterns |
| Phase 2 | Reference when generating orchestrator patterns |
| Phase 3 | Reference when designing agent interactions |
1. API Reference
1.1 spawn_agent
Creates a new subagent with independent context.
const agentId = spawn_agent({
message: "task message", // Required: task assignment
agent_type: "type" // Optional: preset baseline
})
// Returns: agent_id (string)
Key Facts:
- Each agent has isolated context (no shared state)
agent_typeselects preset behavior baseline- Role definition must be loaded via MANDATORY FIRST STEPS
- Returns immediately — use
wait()for results
1.2 wait
Retrieves results from one or more agents.
const result = wait({
ids: [agentId1, agentId2], // Required: agent IDs to wait for
timeout_ms: 300000 // Optional: max wait time (ms)
})
// Returns: { timed_out: boolean, status: { [id]: { completed: string } } }
Key Facts:
- Primary result retrieval method (NOT close_agent)
- Supports batch wait for multiple agents
timed_out: truemeans some agents haven't finished — can re-wait- Can be called multiple times on same agent
1.3 send_input
Continues interaction with an active agent.
send_input({
id: agentId, // Required: target agent
message: "follow-up", // Required: continuation message
interrupt: false // Optional: interrupt current processing
})
Key Facts:
- Agent must NOT be closed
- Preserves full conversation context
- Use for: clarification answers, phase transitions, iterative refinement
interrupt: true— use with caution (stops current processing)
1.4 close_agent
Permanently terminates an agent.
close_agent({ id: agentId })
Key Facts:
- Irreversible — no further wait/send_input possible
- Do NOT use to retrieve results (use wait instead)
- Delay until certain no more interaction needed
- Call for ALL agents at end of workflow (cleanup)
2. Interaction Patterns
2.1 Standard (Single Agent, Single Task)
spawn_agent → wait → close_agent
Use When: Simple, one-shot tasks with clear deliverables.
const agent = spawn_agent({ message: taskPrompt })
const result = wait({ ids: [agent], timeout_ms: 300000 })
close_agent({ id: agent })
2.2 Parallel Fan-out (Multiple Independent Agents)
spawn_agent × N → batch wait({ ids: [...] }) → close_agent × N
Use When: Multiple independent tasks that can run concurrently.
const agents = tasks.map(t => spawn_agent({ message: buildPrompt(t) }))
const results = wait({ ids: agents, timeout_ms: 600000 })
// Aggregate results
const merged = agents.map(id => results.status[id].completed)
// Cleanup all
agents.forEach(id => close_agent({ id }))
Split Strategies:
| Strategy | Description | Example |
|---|---|---|
| By responsibility | Each agent has different role | Research / Plan / Test |
| By module | Each agent handles different code area | auth / api / database |
| By perspective | Each agent analyzes from different angle | security / performance / maintainability |
2.3 Deep Interaction (Multi-round with send_input)
spawn_agent → wait (round 1) → send_input → wait (round 2) → ... → close_agent
Use When: Tasks needing iterative refinement or multi-phase execution within single agent context.
const agent = spawn_agent({ message: initialPrompt })
// Round 1
const r1 = wait({ ids: [agent], timeout_ms: 300000 })
// Round 2 (refine based on r1)
send_input({ id: agent, message: refinementPrompt })
const r2 = wait({ ids: [agent], timeout_ms: 300000 })
// Round 3 (finalize)
send_input({ id: agent, message: finalizationPrompt })
const r3 = wait({ ids: [agent], timeout_ms: 300000 })
close_agent({ id: agent })
2.4 Two-Phase (Clarify → Execute)
spawn_agent → wait (questions) → send_input (answers) → wait (solution) → close_agent
Use When: Complex tasks where requirements need clarification before execution.
const agent = spawn_agent({
message: `
## TASK ASSIGNMENT
...
### Phase A: Exploration & Clarification
Output findings + Open Questions (CLARIFICATION_NEEDED format)
### Phase B: Full Solution (after receiving answers)
Output complete deliverable
`
})
// Phase A
const exploration = wait({ ids: [agent], timeout_ms: 600000 })
if (exploration.status[agent].completed.includes('CLARIFICATION_NEEDED')) {
// Collect answers
const answers = getUserAnswers(exploration)
// Phase B
send_input({
id: agent,
message: `## CLARIFICATION ANSWERS\n${answers}\n\n## PROCEED\nGenerate full solution.`
})
const solution = wait({ ids: [agent], timeout_ms: 900000 })
}
close_agent({ id: agent })
2.5 Pipeline (Sequential Agent Chain)
spawn(A) → wait(A) → close(A) → spawn(B, with A's output) → wait(B) → close(B)
Use When: Tasks where each stage depends on the previous stage's output.
// Stage 1: Research
const researcher = spawn_agent({ message: researchPrompt })
const research = wait({ ids: [researcher] })
close_agent({ id: researcher })
// Stage 2: Plan (uses research output)
const planner = spawn_agent({
message: `${planPrompt}\n\n## RESEARCH CONTEXT\n${research.status[researcher].completed}`
})
const plan = wait({ ids: [planner] })
close_agent({ id: planner })
// Stage 3: Execute (uses plan output)
const executor = spawn_agent({
message: `${executePrompt}\n\n## PLAN\n${plan.status[planner].completed}`
})
const execution = wait({ ids: [executor] })
close_agent({ id: executor })
2.6 Merged Exploration (Explore + Clarify + Plan in Single Agent)
spawn(dual-role) → wait(explore) → send_input(clarify) → wait(plan) → close
Use When: Exploration and planning are tightly coupled and benefit from shared context.
Advantages over Pipeline:
- 60-80% fewer agent creations
- No context loss between phases
- Higher result consistency
const agent = spawn_agent({
message: `
## DUAL ROLE ASSIGNMENT
### Role A: Explorer
Explore codebase, identify patterns, generate questions
### Role B: Planner (activated after clarification)
Generate implementation plan based on exploration + answers
### Phase 1: Explore
Output: Findings + CLARIFICATION_NEEDED questions
### Phase 2: Plan (triggered by send_input)
Output: plan.json
`
})
const explore = wait({ ids: [agent] })
// ... handle clarification ...
send_input({ id: agent, message: answers })
const plan = wait({ ids: [agent] })
close_agent({ id: agent })
3. Message Design
3.1 TASK ASSIGNMENT Structure
## TASK ASSIGNMENT
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/{agent-type}.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json
---
Goal: One-sentence objective
Scope:
- Include: allowed operations
- Exclude: forbidden operations
- Directory: target paths
- Dependencies: dependency constraints
Context:
- Key paths: relevant file paths
- Current state: system status
- Constraints: must-follow rules
Deliverables:
- Output structured following template
Quality bar:
- Criterion 1
- Criterion 2
3.2 Structured Output Template
Summary:
- One-sentence completion status
Findings:
- Finding 1: description
- Finding 2: description
Proposed changes:
- File: path/to/file
- Change: modification detail
- Risk: impact assessment
Tests:
- Test cases needed
- Commands to run
Open questions:
1. Unresolved question 1
2. Unresolved question 2
3.3 Clarification Format
CLARIFICATION_NEEDED:
Q1: [question] | Options: [A, B, C] | Recommended: [A]
Q2: [question] | Options: [A, B] | Recommended: [B]
4. Error Handling
4.1 Timeout
const result = wait({ ids: [agent], timeout_ms: 30000 })
if (result.timed_out) {
// Option 1: Continue waiting
const retry = wait({ ids: [agent], timeout_ms: 60000 })
// Option 2: Urge convergence
send_input({ id: agent, message: "Please wrap up and output current findings." })
const urged = wait({ ids: [agent], timeout_ms: 30000 })
// Option 3: Abort
close_agent({ id: agent })
}
4.2 Agent Recovery (post close_agent)
// Cannot recover closed agent — must recreate
const newAgent = spawn_agent({
message: `${originalPrompt}\n\n## PREVIOUS ATTEMPT OUTPUT\n${previousOutput}`
})
4.3 Partial Results (parallel fan-out)
const results = wait({ ids: agents, timeout_ms: 300000 })
const completed = agents.filter(id => results.status[id]?.completed)
const pending = agents.filter(id => !results.status[id]?.completed)
if (completed.length >= Math.ceil(agents.length * 0.7)) {
// 70%+ complete — proceed with partial results
pending.forEach(id => close_agent({ id }))
}
5. Role Loading
5.1 Path Reference Pattern (Recommended)
spawn_agent({
message: `
### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/${agentType}.md (MUST read first)
...
`
})
Why: Keeps message lean, agent loads its own role context.
5.2 Role Mapping
| Agent Type | Role File |
|---|---|
| cli-explore-agent | ~/.codex/agents/cli-explore-agent.md |
| cli-lite-planning-agent | ~/.codex/agents/cli-lite-planning-agent.md |
| code-developer | ~/.codex/agents/code-developer.md |
| context-search-agent | ~/.codex/agents/context-search-agent.md |
| debug-explore-agent | ~/.codex/agents/debug-explore-agent.md |
| doc-generator | ~/.codex/agents/doc-generator.md |
| action-planning-agent | ~/.codex/agents/action-planning-agent.md |
| test-fix-agent | ~/.codex/agents/test-fix-agent.md |
| universal-executor | ~/.codex/agents/universal-executor.md |
| tdd-developer | ~/.codex/agents/tdd-developer.md |
| ui-design-agent | ~/.codex/agents/ui-design-agent.md |
6. Design Principles
- Delay close_agent: Only close when certain no more interaction needed
- Batch wait over sequential: Use
wait({ ids: [...] })for parallel agents - Merge phases when context-dependent: Use send_input over new agents
- Structured output always: Enforce uniform output template
- Minimal message size: Pass role file paths, not inline content
- Explicit lifecycle: Every spawn must have a close (balanced)
- Timeout handling: Always specify timeout_ms, always handle timed_out