dsh-subagent-codex

Some tasks in a DSH session are simply too heavy for the main conversation: competitive research, cross-module refactors, deep-dive learning materials. dsh-subagent-codex fills that gap with one sentence in the main agent handing the job to a locally logged-in OpenAI Codex CLI, which runs an entire pipeline in isolation and hands back a finished result. DSH orchestrates; Codex does the work.

Under the hood, every call spawns a single codex exec process. Codex browses sources, decomposes problems, writes and tests code, then returns one final agent_message as the result. Timeout, cancellation, and output truncation are all supported, and because the plugin does not use --ephemeral, every session is persisted under ~/.codex/sessions/ for later codex resume or codex archive archival. The callback always resolves to a terminal state and never throws an exception back into the main conversation.

This plugin is a great fit if you already run Codex CLI locally and use DSH as a daily assistant: need a boss-ready competitive analysis, a thorough study guide on an unfamiliar topic, a full content plan with supporting material, or an end-to-end auth migration with tests: drop it to subagent_codex and get a one-call delivery. The one caveat: if the task depends on DSH-internal memory, session history, or other DSH tools, stick with the regular subagent instead of routing through Codex.