SPYQWER1/dsh-codex-tools

中文

Image generation and image understanding for the DeepSeek Harness — powered by your ChatGPT subscription. No OPENAI_API_KEY needed.

A DeepSeek Harness plugin that registers three Codex-backed model tools:

Tool	What it does
web_search	Searches the public web and returns a concise summary with source URLs.
image_gen	Generates a bitmap image (illustrations, icons, logos, photos, concept art, UI mockups) through the ChatGPT Codex backend.
image_vision	Describes or answers questions about an image with a multimodal model — giving text-only harness models (e.g. deepseek-v4-flash) plug-in vision.

All three tools reuse the ChatGPT login state (the same OAuth tokens the official Codex CLI uses) and the harness-owned transports in scripts/ — zero npm dependencies, Node built-in https only.

How it works

web_search / image_gen / image_vision (model tool)
        │  harness.registerTool + credentials service
        ▼
index.js (Cordis bundle entry)
        │  shell service, env-only argument passing
        ▼
scripts/codex-common.mjs / codex-imagegen.mjs / codex-vision.mjs / codex-search.mjs   ← harness-owned transports
        │  OAuth refresh (auth.openai.com) + POST chatgpt.com/backend-api/codex/responses
        ▼
        gpt-5.5 (multimodal) → PNG file / text description

Auth precedence (all transports): CODEX_ACCESS_TOKEN / CODEX_REFRESH_TOKEN / CODEX_ACCOUNT_ID environment variables (resolved from the DSH credentials service, keys OPENAI_CODEX_API_KEY / OPENAI_CODEX_REFRESH_TOKEN / optional OPENAI_CODEX_ACCOUNT_ID) → $CODEX_HOME/auth.json → ~/.codex/auth.json (codex login). On HTTP 401 the transport refreshes the access token once and persists it back to the selected auth file (atomic, 0600). This plugin consumes the login state; it does not provide Codex authentication or an LLM provider.

Requirements

• DeepSeek Harness (the plugin is installed as a profile-wide Cordis bundle; tested against current builds)
• Node.js ≥ 22 (for the transports)
• A ChatGPT subscription login state:
  • ~/.codex/auth.json from codex login (recommended), or
  • the OPENAI_CODEX_API_KEY / OPENAI_CODEX_REFRESH_TOKEN credentials in DSH

Install

The package ships as a bundle: installing it into a profile registers the tools in the host registry, so every session of that profile gets them.

# from a git host (no build step, so no pnpm allowBuilds permission needed)
dsh plugin --profile web add github:SPYQWER1/dsh-codex-tools

# or from a tarball (pnpm pack)
dsh plugin --profile web add ./dsh-codex-tools-1.0.0.tgz

# or from npm, once published
dsh plugin --profile web add dsh-codex-tools

Then restart the profile (dsh web / dsh --profile web) — the tools appear in every session. dsh plugin --profile web remove dsh-codex-tools uninstalls. Pin a commit for git installs (github:SPYQWER1/dsh-codex-tools#<sha>) so a later push cannot change what runs.

Bundle install resolves the in-box @deepseek-ai/dsh-tools peer from the harness installation; no extra npm packages are fetched. The bundle entry is index.js, which registers all three tools into the host registry and shares the scripts/ transports via tools.js.

As standalone CLI

The transports are plain Node scripts and work without the harness:

# generate an image (writes the PNG, prints JSON to stdout)
CG_PROMPT="a cute whale icon, flat vector style" CG_OUT=whale.png CG_SIZE=1024x1024 \
  node scripts/codex-imagegen.mjs

# describe an image
VG_IMAGE=whale.png VG_QUESTION="what is this?" node scripts/codex-vision.mjs

# search the public web
CS_QUERY="latest DeepSeek Harness release" CS_FRESHNESS=live node scripts/codex-search.mjs

Usage

Ask the harness in natural language — the model drives the tools itself:

• "Search for the latest DeepSeek Harness release" → web_search
• "生成一个鲸鱼图标" → image_gen
• "看看这个图片讲了什么：output/photo.png" → image_vision

web_search parameters

Param	Type	Notes
query	string (required)	Public-web research question.
maxSources	integer	1–10, default 5.
freshness	string	cached (default) or live for time-sensitive queries.
model	string	ChatGPT backend model, default gpt-5.4-mini.

The result contains summary and sources, where each source has a title, URL, and short snippet.

image_gen parameters

Param	Type	Notes
prompt	string (required)	Detailed description (subject, style, composition, palette, constraints).
out	string	Output path relative to the workspace. Default output/imagegen/<timestamp>.png.
size	string	1024x1024, 1536x1024, 1024x1536, 2048x2048, 2048x1152, or auto (default).
format	string	png (default), jpeg, webp.
model	string	ChatGPT backend model, default gpt-5.5.

image_vision parameters

Param	Type	Notes
image	string (required)	Path to the image (png/jpeg/webp/gif), workspace-relative or absolute.
question	string	Optional focus question; defaults to a full description.
model	string	Default gpt-5.5.

Gallery

image_gen — generate an image from a natural-language request:

image_vision — a text-only model reads an image:

Caveats

• chatgpt.com/backend-api/codex/responses is the same internal endpoint the official Codex CLI uses. It is not a documented public API — OpenAI may change or restrict it at any time.
• Web search and image generation use the metered Codex-usage bucket of your ChatGPT plan.
• Per OpenAI Terms of Use, do not use your ChatGPT subscription to power a public-facing image generation service.
• Transparent-background output is not supported (use a chroma-key background + local removal instead).

License

MIT — see LICENSE. The protocol shape follows the public behavior of the Codex CLI and the chatgpt-imagegen project; the implementation here is original.