Explorer
Runs bounded, read-heavy investigations before planning.
A role-based workflow for Codex and Claude Code
Separate planning, implementation, and review. Challenge the premise before refining the plan. Keep every task independently checked and every publication under your control.
Work path
4focused roles
2explicit approval gates
1task in progress
0automatic merges
Deliberate separation
Each agent gets a narrow responsibility and the access needed to perform it—no more.
Runs bounded, read-heavy investigations before planning.
Turns intent and repository evidence into a minimal, decision-complete plan.
Implements exactly one approved, reviewable task at a time.
Challenges premises, then checks plans, task deltas, and cumulative risk.
Each role has a purpose-matched model and reasoning-effort assignment. See the README for runtime-specific details.
Built-in safeguards
The reviewer must pass both the plan premise and its specification.
No later task starts until the current change is independently correct, validated, reviewed, and committed.
A distinct final review checks cross-task contracts only when it adds coverage.
Ready when you are
Claude Code installs the skill and all four roles in one step, with no manual file copying — Codex needs one extra manual step for its custom agents. Read-only access isn't equivalent either: Codex sandboxes writes at the OS level, while Claude Code's read-only roles simply aren't granted Bash or MCP tools.
Codex
git clone https://github.com/jameswei/plan-build-review.git \
~/.codex/skills/plan-build-review
mkdir -p ~/.codex/agents
cp ~/.codex/skills/plan-build-review/assets/agents/*.toml \
~/.codex/agents/
Use $plan-build-review to deliver this feature or milestone.
Claude Code
/plugin marketplace add jameswei/plan-build-review
/plugin install plan-build-review@plan-build-review
/reload-plugins
Use /plan-build-review:plan-build-review to deliver this feature or milestone.