This project is designed to be conservative by default, but several behaviors are intentionally sharp because the goal is deep Claude/Codex debugging and local power-user workflows.
The default debug profile logs raw MCP traffic, tool arguments/results, prompt text, progress events, and Codex stdin/stdout/stderr traffic. Treat these logs and debug bundles as sensitive project data.
Use quieter logging for normal work:
export CODEX_SUBAGENTS_LOG_PROFILE=productionDisable logging entirely:
export CODEX_SUBAGENTS_LOG_LEVEL=silentLegacy async one-shot jobs are process-local. They keep Claude responsive during
long one-shot work, but they do not survive MCP process restart. These legacy
tools are hidden unless CODEX_SUBAGENTS_ENABLE_LEGACY_TOOLS=1 is set.
Use codex_task with background: true for long-running work, then keep the
returned session_id and call codex_followup with mode: "wait" or
mode: "steer".
codex_followup with mode: "steer" delivers live steering only when the
session is running through Codex app-server.
If app-server is unavailable and the session falls back to codex exec, steering
degrades to a high-priority queued follow-up turn. It cannot alter an already
running codex exec process in place.
The default sandbox is read-only. Full local access requires this per-call flag:
{
"full_access": true
}That maps to Codex's unrestricted local mode. It can write files, mutate git state, use DNS/network, and run package installs. Use it only when the user explicitly asks for that capability.
There is also an advanced workspace-write sandbox. It can write inside the
configured workspace but is not full local access and does not enable DNS/network
or arbitrary filesystem writes. The patcher role uses workspace-write by
default; use other roles or omit subagent_type for strictly read-only work.
Nested Codex subagents are supported, including Spark. Keep nested work scoped,
set explicit advanced.subagent_tasks, and keep
advanced.subagent_runtime.max_depth at 1 unless recursive delegation is
deliberately needed.
The plugin has a skill, tool descriptions, and codex://usage, but Claude can
still make better choices when the user names the intended shape:
- "ask one Codex agent"
- "run three Codex agents in parallel"
- "start Codex in the background"
- "steer the running Codex session"