内容
# Codex agent configuration
## Repository family
- [Codex Claude Orchestrator](https://github.com/coredo-eu/codex-claude-orchestrator) — local Codex-to-Claude worker transport, ownership policy, and lifecycle controls.
- [Codex agent configuration](https://github.com/coredo-eu/codex-agent-config) — portable Codex guidance, native-agent roles, and configuration template.
- [Claude agent configuration](https://github.com/coredo-eu/claude-agent-config) — portable standalone Claude guidance, agents, permissions, and CodeIndexer hook.
This repository is a portable, public snapshot of an opinionated Codex
agent-working setup. It captures the policy and role definitions that determine
how work is owned, delegated, verified, and handed back. It contains no Codex
runtime, account data, or machine state.
## Operating model
- Codex remains the owner of user intent, architecture decisions, authority,
conflict resolution, independent verification, and the final verdict.
- Work is delegated only when isolation or parallelism lowers total cost or
elapsed time without weakening correctness or safety.
- A worktree has one edit-capable owner at a time. Parallel writers use isolated
roots and return custody explicitly.
- Native agents and Codex-owned Claude workers are bounded executors. They do
not gain commit, push, deploy, credential, destructive, or production
authority from their role definition.
- Verification effort follows consequence and uncertainty rather than a fixed
checklist or task-size label.
- CodeIndexer is an optional discovery and impact-analysis surface. Its index is
derived evidence; authoritative source and observed runtime remain the SSOT.
The full contract is in [`AGENTS.md`](AGENTS.md).
## Repository contents
| Path | Purpose |
| --- | --- |
| [`AGENTS.md`](AGENTS.md) | Global authority, delegation, ownership, safety, and verification policy. |
| [`agents/`](agents) | Native Codex role definitions with model, reasoning, sandbox, and bounded instructions. |
| [`config.example.toml`](config.example.toml) | Sanitized, portable subset of the active Codex configuration. |
## Native agent roles
| Role | Default model | Effort | Sandbox | Intended use |
| --- | --- | --- | --- | --- |
| `source_explorer` | `gpt-5.6-luna` | `medium` | read-only | Direct repository reconstruction and impact evidence. |
| `codeindexer_explorer` | `gpt-5.6-luna` | `medium` | read-only | Indexed semantic, symbol, and dependency discovery. |
| `mech_executor` | `gpt-5.6-terra` | `medium` | workspace-write | One bounded implementation after explicit edit-custody transfer. |
| `test_runner` | `gpt-5.6-luna` | `low` | workspace-write | Tests, builds, linters, and smoke checks without source edits. |
| `reviewer` | `gpt-5.6-terra` | `high` | read-only | Independent correctness and regression review. |
| `security_reviewer` | `gpt-5.6-sol` | `high` | read-only | Focused security, privacy, credential, and authorization review. |
| `scout` | `gpt-5.6-luna` | `medium` | read-only | Local runtime and operational-state observation. |
Model identifiers reflect the source environment. Replace them with models
available in the target Codex installation when necessary.
## Configuration snapshot
The example records these current choices without copying generated state:
- no top-level model or reasoning-effort override: the main session keeps the
model and effort selected by the user;
- up to four native agent threads and one level of child delegation;
- official OpenAI developer-documentation MCP;
- loopback CodeIndexer MCP at `http://127.0.0.1:8978/mcp`;
- document, PDF, spreadsheet, presentation, browser, sites, visualization, and
Codex-Claude orchestrator plugins enabled;
- JS REPL disabled.
The current high-trust `approval_policy` and `sandbox_mode` values are preserved
as comments. They are intentionally not activated by copying the example.
## Installation
Requirements:
- a Codex installation supporting `AGENTS.md` and native agent definitions;
- CodeIndexer only if the indexed explorer or MCP endpoint is used;
- `codex-claude-orchestrator` only if Codex-owned Claude PTY workers are used.
Clone the repository, review differences against the current local files, and
then install only the parts you want:
```bash
git clone https://github.com/coredo-eu/codex-agent-config.git
cd codex-agent-config
mkdir -p ~/.codex/agents
install -m 0644 AGENTS.md ~/.codex/AGENTS.md
install -m 0644 agents/*.toml ~/.codex/agents/
```
Merge `config.example.toml` into `~/.codex/config.toml` manually. Do not replace
generated marketplace, browser, notification, project-trust, or MCP runtime
entries wholesale.
## Validation
The tracked TOML files can be parsed with Python 3.11 or newer:
```bash
python3 - <<'PY'
import pathlib, tomllib
tomllib.loads(pathlib.Path("config.example.toml").read_text())
for path in pathlib.Path("agents").glob("*.toml"):
tomllib.loads(path.read_text())
print("configuration: OK")
PY
```
## Deliberate exclusions
This repository does not contain credentials, licenses, trusted-client
fingerprints, project trust entries, absolute home paths, notifications,
sessions, caches, logs, plugin caches, MCP OAuth state, or generated runtime
files. Those remain owned by each local Codex installation.
MCP 市场