Subagent-Driven Development (Codex Variant)

Execute a plan by dispatching every role — implementer, spec reviewer, quality reviewer — to Codex CLI via the bridge script. The orchestrator (you) stays in this session, plans tasks, ferries context between Codex sessions, and gates progress on a Codex-specific two-stage review.

Core principle: One Codex SESSION_ID per role per task + per-agent planning dir + two-stage review (spec then quality) = high-quality delegation to a second model, all from a single Claude session.

Announce at start: "I'm using the subagent-driven-codex skill — every implementer and reviewer dispatch will go to Codex via the bridge."

Why This Exists

This skill is retained because it uses an external Codex CLI executor/reviewer path. For Claude-native high-parallelism work, prefer Claude Code dynamic workflows instead of rebuilding orchestration in plugin skills.

Use this variant when:

• You want a second model's eyes on every line of generated code without you reviewing each diff manually.
• The work is bounded and verifiable from disk (Codex's sweet spot per the user's collaboration policy).
• You want fresh, isolated executor context per role without spending Claude dynamic-workflow worker budget.
• You want the orchestrator's conversation context kept clean — the bridge runs in the background and only returns the final reply.

Prefer a Claude Code dynamic workflow when the task needs many Claude agents, broad parallel review, large migrations, or cross-checked research. Codex CLI does support MCP if the user has it configured — that is not a reason to avoid Codex.