WODE25500/dsh-codex

OpenAI Codex CLI integration for DeepSeek Harness (dsh) — exposes your locally installed Codex CLI to the dsh agent as native tools: one-shot tasks, repository code reviews and session resumes, all via non-interactive codex exec with a read-only sandbox by default. Codex talks to whatever provider is configured in ~/.codex/config.toml (on this machine: DeepSeek — no OpenAI login required) and shares the same provider quota as this session.

Independent community project, not official. The CLI itself comes from OpenAI Codex (Apache-2.0); this plugin only wires codex exec into dsh's tool/skill system, following the same adapter style as dsh-skillopt (Schemastery config + defineTool + shell executor + SKILL.md + bundle patch layer).

Features

• 4 native dsh tools: codex_status / codex_exec / codex_review / codex_resume
• Bundled SKILL.md teaching the agent when and how to use them
• Schemastery config (model / sandbox / cd / profile / timeoutMs …)
• Bundle patch layer (cordis.patch.yml) — drop into any profile
• Default read-only sandbox; writing files requires an explicit workspace-write / danger-full-access escalation
• Sessions are persisted by default, so codex_resume works; outputFile echoes the agent's final message back into the tool result
• codex_status prints a masked provider summary — never any credentials

Prerequisites

• DeepSeek Harness (dsh) installed
• Codex CLI:

npm install -g @openai/codex
codex --version

• A provider configured in ~/.codex/config.toml (e.g. DeepSeek: model_provider, base_url, model) or codex login. The plugin never reads or displays credentials.

Install

As a bundle in a profile

Add dsh-codex to the profile's bundles, or in the profile cordis.patch.yml:

- insert:
    - id: codex
      name: './src/index.js'
      config:
        model: deepseek-v4-flash   # empty = use ~/.codex/config.toml
        sandbox: read-only         # read-only | workspace-write | danger-full-access
        cd: /path/to/workspace

Local patch for development

pnpm dsh web --patch ./dsh-codex/cordis.patch.yml

Then ask the agent: "use codex_status to check the Codex setup".

Usage

Tool	Behavior
codex_status	CLI version, config path, masked provider summary
codex_exec	One-shot task: self-contained prompt + directory + sandbox level
codex_review	Code review of the repo / a diff / a commit (read-only by nature)
codex_resume	Resume a session by id or with last=true

Typical flow:

codex_status
codex_exec prompt="Review F:\8.15.6\src\app.ts and report bugs and optimizations" cd="F:\8.15.6"
codex_review uncommitted=true
codex_resume last=true prompt="Finish the remaining work"

Config (Schemastery)

Key	Default	Description
codexPath	codex	Path to the codex CLI
model	—	Default model override (empty = config.toml)
sandbox	read-only	read-only|workspace-write|danger-full-access
cd	—	Default working root
profile	—	Config profile to layer
skipGitRepoCheck	true	Always pass --skip-git-repo-check
ephemeral	false	Persist sessions (needed for resume)
json	false	JSONL event output
approveForMe	false	Automatic approval review (needs a writable sandbox)
color	never	auto|always|never
timeoutMs	600000	Per-call timeout (default 10 min)

Data boundaries & safety

• Default read-only: codex may inspect but not modify files; escalation is explicit per call.
• codex_status masks the provider config — nothing from the [auth] section is ever printed.
• Every codex_exec / codex_review sends your prompt and repository excerpts to the configured provider and spends its quota — validate on a small sample before large batches.
• Session files are written to disk (~/.codex/sessions) by default; use ephemeral=true when you don't want traces.

Verification (minimal smoke)

codex --version
codex_status                                  # no model spend
codex_exec prompt="Reply with exactly: OK"    # tiny spend

Directory layout

dsh-codex/
  src/index.js            # plugin entry: 4 tools + Schemastery config
  skills/codex/SKILL.md
  docs/                   # Chinese & English docs
  cordis.patch.yml        # bundle patch layer

License

MIT. The CLI itself is copyright OpenAI Codex (Apache-2.0).