Non-destructive setup
Preview first and create only missing files. Existing AGENTS, README, and uncommitted changes stay intact.
One persistent main Task owns the goal, constraints, key decisions, and final acceptance. Temporary Subagents take only bounded work that can be verified independently.
The point is not to split work into more pieces. It is to give every delegation a clear input, objective evidence, and an explicit return point—then let the main Task integrate the result and write the true state back into the project.
Choose Skip, Light, or Orchestrated for a Codex project, then safely create durable context, architecture, acceptance, and handoff rules without overwriting existing files.
Complete local Marketplace · 37,000 bytes · Plugin source and installation guide included
When a project gets complicated, the real problem is usually not too few agents. It is missing boundaries for context, responsibility, and acceptance.
Preview first and create only missing files. Existing AGENTS, README, and uncommitted changes stay intact.
Only write facts verified from the project. Keep unknowns explicitly marked as 待确认 (To confirm).
The main Task owns the outcome. Subagents receive only bounded, independently verifiable work.
Publishing, paid requests, external messages, permissions, and deletion remain explicit user decisions.
Start with the simplest profile. Upgrade to federated only when cross-repository, cross-deployment, or cross-host boundaries already exist.
For one-off, small, self-contained work. If the project needs no durable context, staying simple is the right choice.
For a single owner, one repository, and short feedback loops. It records the essential product boundaries and project state.
For multiple modules, milestones, or real cross-repository boundaries, with exploration, implementation, review, and integration clearly separated.
The plugin initializes a collaboration structure. It does not automatically create Subagents, Worktrees, separate Tasks, or Automations. Enable parallel work only when the boundaries justify it.
Subagents handle read-only exploration, testing, log analysis, and bounded review.
Worktrees isolate writers only when several people or agents edit the same repository concurrently.
Persistent Tasks are reserved for cross-project, cross-host, or long-lived independent ownership.
project/ ├── AGENTS.md ├── HANDOFF.md ├── README.md ├── .codex/agents/ │ ├── project-explorer.toml │ └── project-reviewer.toml ├── docs/ │ ├── PRODUCT.md │ ├── ARCHITECTURE.md │ ├── ACCEPTANCE.md │ ├── ROADMAP.md │ ├── WORKFLOW.md │ └── DECISIONS.md ├── work/ └── outputs/
Move from “code was generated” to “observable outcomes were verified.” Update HANDOFF after material state changes, not as a running diary.
Read first, then choose the structure.
Turn done into observable outcomes.
Make changes, check them, keep evidence.
Find issues without silently rewriting.
Write the true state back to HANDOFF.
You will get the codex-project-setup-marketplace directory; it is not a double-click installer.
codex plugin marketplace add /absolute/path/to/codex-project-setup-marketplace
codex plugin add codex-project-setup@codex-project-setup-local
Only a new Task started after installation will load $setup-project-collaboration.
Use $setup-project-collaboration to choose and initialize the right Codex collaboration structure for this project.
Inspect the project read-only first. Do not overwrite existing files.
The bundle includes the local Marketplace manifest, plugin source, and INSTALL.md. Installation still requires two Codex commands.
Download Marketplace bundleDownload plugin source ZIP only →0a9e48c5d62f9927ba8766a2e615ac4403afbe7192420b84813798a2a7501aa6No. It only creates documentation, directories, and project-level role configurations. Real Subagents, Tasks, Worktrees, or Automations are enabled only when the work calls for them.
No. The initializer creates only missing files and reports same-name content that needs manual integration. There is no force-overwrite mode.
More coordination layers increase token use, conflicts, and management cost. Federated is worthwhile only for real cross-repository, cross-deployment, cross-host, or long-lived ownership boundaries.