🚀 dsh-subagent-codex

> Let agents in DeepSeek Harness hand real work off to OpenAI Codex.

English | 中文

---

What it is, in one line

A DSH plugin that adds a subagent_codex tool to your Harness sessions. You — or your main agent — hand it a task, and it runs that task in your local Codex CLI, completely isolated from the main conversation, then brings the result back. Codex runs with its own model, its own toolchain, its own working directory.

In short: DSH orchestrates, Codex executes.

When to use it

A few real-world examples (the kind of heavier work Codex excels at):

• Workplace: competitive research — "Have codex run a competitor analysis: compare pricing, features, and pros/cons of 3 mainstream note-taking apps, then recommend one for my use case, formatted as a report I can send to my boss" — Codex researches online, compares across dimensions, and delivers a ready-to-use document.
• Learning: master an unfamiliar topic — "Have codex help me truly understand blockchain: break the question into 5 progressive sub-questions (what it is, how it works, what problem it solves, how it differs from traditional systems, what the controversies are), cross-check each one against multiple sources, explain with everyday analogies, diagrams, and a glossary, and deliver a complete study document with an FAQ" — This is not "look it up and write an article": Codex approaches it like research — systematically breaking down, verifying, and organizing — and delivers a deep, well-structured learning package in one go.
• Creation: content planning — "Have codex plan a viral Xiaohongshu post: 3 topic directions, each with a title, a hook, and a body outline, plus 5 supporting sources" — one call delivers a full creation plan, from topic to material.
• Development: cross-module refactoring — "Have codex migrate user auth from JWT to OAuth2: update the auth middleware, add database migration scripts, write unit and integration tests, run everything green, then summarize the changes as a commit message" — a complete dev pipeline: edit code, add tests, run them, and summarize; you just review the whole change set.

When NOT to use it: tasks that must stay inside DSH (calling DSH memory, session history, or other DSH tools). Leave those to a regular subagent.

Quick start

1. Prerequisites

• Codex CLI installed and logged in (~/.codex/auth.json exists).
• DeepSeek Harness installed.

2. Install

# via the dsh plugin command (the dsh.bundle declaration auto-mounts the plugin)
dsh plugin --profile web add dsh-subagent-codex
# or from a local tarball
dsh plugin --profile web add file:/path/to/dsh-subagent-codex-0.1.1.tgz

3. Restart and talk

Restart DSH and open a new session, then just say:

Have codex research the latest papers on LLM inference acceleration and summarize them into a comparative review
Have the codex subagent do X

The tool is named subagent_codex; each invocation starts one independent Codex CLI task.

Configuration

The plugin registers two rows in the config tree:

Supported config keys on the provider row:

- id: subagent-codex
  name: 'dsh-subagent-codex'
  config:
    command: codex            # codex CLI executable (PATH name or absolute path)
    cwd: /path/to/workdir     # optional; child working dir (defaults to parent session workspace)
    model: o3                 # optional; model override (codex exec -m)
    sandbox: workspace-write  # optional: read-only | workspace-write | danger-full-access
    timeoutMs: 600000         # per-task timeout in ms
    maxOutputChars: 40000     # output cap returned to the delegating agent

How it works (technical details)

• The plugin implements the DSH SubagentProvider seam (@deepseek-ai/dsh-subagent, out-of-process contract) and registers a provider named codex.
• Being an external process, the provider advertises no start capabilities (NO_START_CAPABILITIES): an external CLI cannot honor outputSchema / maxDepth / toolFilter / persona.
• Each delegation spawns one codex exec --json --skip-git-repo-check <prompt>, parses the JSONL event stream, and takes the last agent_message text as the final output.
• Cancellation (AbortSignal → SIGKILL), timeout, and output truncation are supported. Results always resolve (completed / aborted / error all settle to a terminal state), so failures never blow up the main conversation.

Traceability: every run leaves a full record

The plugin deliberately does not pass --ephemeral, so every Codex invocation persists its complete session to disk:

~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-<timestamp>-<session-id>.jsonl

That file holds the full conversation (input, Codex's intermediate steps, final output). You can:

• recover the session with codex resume and keep talking to it;
• archive it long-term with codex archive <session-id> (moves it to ~/.codex/archived_sessions/).

FAQ

Q: Why don't I see codex subagent runs in the DSH sidebar? A: By design. The DSH sidebar subagent tree only lists DSH-internal subagents (those backed by a DSH session). Codex is an out-of-process provider — it never creates a DSH session at runtime — so the GUI does not show it. To watch a run, look at ~/.codex/sessions/ as described above.

Q: Do I need the Codex desktop app? A: No. The plugin calls the codex CLI directly. The desktop app is irrelevant.

Q: Does this consume my OpenAI quota? A: Yes. Every invocation runs Codex with the account your ~/.codex is logged in as, consuming Codex token quota.

License

MIT