Subagent Driven Development Codex

---
name: subagent-driven-development-codex
description: Execute an implementation plan by dispatching Codex-backed implementer and reviewer agents while Claude keeps spec-compliance control. Codex-backed variant.
disable-model-invocation: true
---
<!--
Upstream source: obra/superpowers skills/subagent-driven-development/SKILL.md
Last synced: 2026-04-03
Divergence: Codex-backed implementer/reviewer agents replace Claude subagents; adapter-managed resume flow
-->

# Subagent-Driven Development

Execute plan by calling `codex_implement` per task via the MCP tool, with two-stage review after each: Claude-side spec compliance first, then `codex_review` for code quality.

**Core principle:** Codex implementer per task + two-stage review (spec then quality) = high quality, fast iteration

<HARD-GATE>
You MUST dispatch codex_implement for every plan task (Step 3).
You cannot implement code yourself, even for simple tasks.
Spec review requires seeing EXTERNAL implementation. No dispatch = no task completion.
</HARD-GATE>

## Checklist

You MUST create a task for each plan task and complete them in order.

**Setup:**

1. **Read plan** — read the plan file once, extract all tasks with full text and context
2. **Create TodoWrite** — one entry per plan task

**Per task:**

3. **Capture task base** — before dispatching the implementer, record the current revision:

```bash
TASK_BASE_SHA="$(git rev-parse HEAD)"
```

Substitute the captured TASK_BASE_SHA value as the structured review base for this task, so the reviewer inspects only the task's delta rather than all prior tasks.

4. **Dispatch implementer** — call the `codex_implement` MCP tool.
   Pass the task description as the `prompt`:

```json
{
  "tool": "codex_implement",
  "arguments": {
    "taskId": "task-17",
    "prompt": "Implement Task 3 from docs/superpowers/plans/2026-04-03-agent-forwarding.md. Keep the work scoped to agents/codex-implementer.md and tests/prompt-contracts/execution-workflows.test.mjs.",
    "workspaceRoot": "/absolute/path/to/your/repo"
  }
}
```

5. **Handle status** — see Status Handling below
6. **Spec compliance review** — Claude reads the actual code and verifies against spec using `spec-review-template.md`; if issues remain, resume the same implementer thread with `codex_resume`:

```json
{
  "tool": "codex_resume",
  "arguments": {
    "taskId": "task-17",
    "sessionId": "019d4f82-58b8-72d3-9212-2e3d3fc69bcb",
    "prompt": "Fix these issues:\n- The agent still lacks tools: Bash.\n- The prompt-file override is ignored in the TDD path.",
    "workspaceRoot": "/absolute/path/to/your/repo"
  }
}
```

7. **Code quality review** — call `codex_review` only after spec compliance passes:

```json
{
  "tool": "codex_review",
  "arguments": {
    "taskId": "task-17-review",
    "reviewStyle": "structured",
    "scope": { "kind": "base", "base": "<captured TASK_BASE_SHA value>" },
    "prompt": "Review the implementation of Task 4 from docs/superpowers/plans/2026-04-03-agent-forwarding.md. Focus on regressions, incorrect routing, and missing test coverage.",
    "workspaceRoot": "/absolute/path/to/your/repo"
  }
}
```

8. **Mark complete** — only after both gates pass

**After all tasks:**

9. **Final review** — call `codex_review` for the entire implementation with a new task ID and the final base SHA.
10. **Finish branch** — wrap up the development branch

## Status Handling

Handle the implementer's reported status:

| Status | Action |
|---|---|
| **DONE** | Proceed to spec compliance review |
| **DONE_WITH_CONCERNS** | Read concerns; if correctness/scope — address before review; if observations — note and proceed |
| **NEEDS_CONTEXT** | Provide the missing context and re-dispatch |
| **BLOCKED** | Assess: context problem → re-dispatch with more context; too hard → escalate approach; too large → break down; plan wrong → escalate to human |

**Never** ignore an escalation or force the same agent to retry without changes.

## Red Flags

**Never:**

- Skip reviews (spec compliance OR code quality)
- Proceed with unfixed issues
- Dispatch multiple implementer agents in parallel (conflicts)
- Make the Codex agent read the plan file (provide full text instead)
- **Start code quality review before spec compliance passes** (wrong order)
- Move to next task while either review has open issues

**If reviewer finds issues:** resume the implementer thread to fix, reviewer reviews again, repeat until approved.

## Integration

- **superpowers-cc-to-codex:writing-plans-codex** — creates the plan this skill executes
- **superpowers-cc-to-codex:requesting-code-review-codex** — code review via the `codex_review` MCP tool