dsh-chatgpt-connect

dsh-chatgpt-connect

Use the ChatGPT subscription you already pay for, inside DeepSeek Harness. If you work in Harness but answer coding questions in another app because your paid models live there, this plugin puts those models in the normal Harness model picker, so one session keeps the conversation, tools and attachments together.

Apache-2.0 · independent community Alpha · not affiliated with, sponsored by, or endorsed by OpenAI, ChatGPT, Codex, DeepSeek, or DeepSeek Harness. · Security and permissions · Uninstall and what it leaves behind

English | 中文

1. What this is and whether it works for you

dsh-chatgpt-connect adds one provider, openai-codex, to Harness. Harness keeps managing tools, permissions, approvals, attachments, sessions, compaction and recovery; this plugin only supplies the model connection and an optional image lane. Installing it does not change your default model or search route, and it does not turn a ChatGPT subscription into an OpenAI Platform API key.

Current state. Published on npm and installable through the alpha channel. It runs; it is not a finished product. See Install.

Four different kinds of limit, stated separately because they need different reactions:

• Not built. Ultra orchestration for Codex Sol and a desktop Computer Use executor do not exist here.
• Built but off by default. Proxy, search takeover, image viewing, image generation, auto-review and the quota tool are all disabled on a fresh install; each is opt-in in the settings panel.
• Needs permission you may not have. Model availability, quota and served tier are decided by OpenAI. A model in the list is not proof your account may use it.
• Present but unverified. The fifteen-host installation matrix is a keyless install and doctor regression. It does not demonstrate a live OAuth flow, real model calls, real image editing, or a real Desktop session.

2. What you need before installing

Requirement	Value
DeepSeek Harness	>=0.1.7-rc.1 <0.3.0. Inside that range the plugin loads. Whether a given host is reported compatible or unverified is a separate, recorded verdict — not an admission gate. The current verified list is in compatibility.json; this README deliberately does not copy it, so a Harness update inside the range needs no README change.
Node.js	^22.19.0 || >=24.0.0
Account	A ChatGPT account whose subscription can reach the model you want, authorized through the browser during setup
Profile	A Harness profile is a named configuration (for example web). The plugin installs into one profile and only appears there.

A moving npm tag is not a compatibility guarantee. Do not upgrade or downgrade Harness just to match a row.

3. Install

Run this where you normally run Harness, replacing web with your profile name:

dsh plugin --profile web add dsh-chatgpt-connect

The plugin loads the next time DSH starts. It does not change your default model or search route; selecting a model is the next step.

The command names the package, not a version, so publishing does not require editing this file. npm resolves it through the package's latest tag. If your registry is a mirror that lags behind, or you need to pin a version for a rollout, use the exact command in INSTALL.md. From a DSH source checkout, prefix the command with pnpm.

Removing it: see Uninstall. Uninstalling does not sign you out or delete your credentials — that is a separate, deliberate action.

4. Quick start: finish one real task

Installation success is not task success. An entry appearing in Settings, or doctor passing, only proves the plugin loaded; it says nothing about whether a model answered. The four steps below end at a real answer.

1. Input — a question you actually want answered, typed in a new Harness session (for example: explain what this repository's release workflow does).
2. Action — open the model picker in the composer, choose an openai-codex model, and send. If no openai-codex entry appears, the account is not authorized yet: open Settings → ANB 工具 → ChatGPT Connect → Authorize and finish the browser approval, then pick the model again.
3. Where the result appears — as the normal assistant answer in that session. Tool approvals, file changes and attachments behave as they do for any other model.
4. How to tell it worked — you got a written answer to your question, with no red error block and no overloaded/401-style diagnostic appended. If instead you get an error, go to Problems and next steps; if the answer arrives but you want to know what the request consumed, see the quota section of the settings panel.

5. Daily use

Main line, end to end. In DSH, open or create a session → pick an openai-codex model in the composer → work normally (ask, attach files, approve tool calls) → read the answer in the session, where images also land if you enabled generation.

Shortest path from the second time on. Open a session → pick the model → ask. There is no restart and nothing to re-authorize; the account selection and every toggle persist. If you always want the same model, your profile's default model covers the picker step, so it collapses to open a session and ask.

Things that do change state, so you are not surprised:

• Account — choosing another account affects subsequent requests, not conversations already started; each request keeps the account it captured. There is no automatic rotation or failover.
• Fast Mode and per-conversation speed are scoped to one conversation.
• Attachments and tool calls are read/written by Harness as usual; this plugin adds no file permissions of its own.

6. Settings panel

There is exactly one entry. The panel title is ChatGPT Connect, which does not match the package name dsh-chatgpt-connect — that is the same thing. Path: Settings → ANB 工具 → ChatGPT Connect (Chinese UI: 设置 → ANB 工具 → ChatGPT连接). Everything below lives inside that one card, including the More settings section; there is no separate entry under Models.

Section	What it is for
Accounts	Add up to 16 accounts, pick the active one, sign out.
Authorize / status	Start the browser OAuth flow and see whether the current account still works.
More settings	Every optional capability: proxy, search, image viewing, image generation, auto-review, model catalogue.
Diagnostics	Run doctor and the capability probes from the UI.

Every capability toggle with its effect and default:

Control	Effect when on	Default
enableProxy	Routes this plugin's HTTP(S) traffic — including compressed OAuth and quota responses — through a credential-free proxy. Unrelated Fetch calls keep Harness's transport, and a failed proxy request never retries directly.	off
enableSearch	Makes Codex the search route for the whole profile; turning it off restores the previous route.	off
enableImageTool	Adds view_image for local files and validated public HTTP(S) images.	off
enableImageGeneration	Generate images from a prompt, or edit selected conversation images with ordered references. Also bridges image handles to non-Codex routes.	off
enableAutoReview	Sends bounded approval context to chatgpt.com for a second opinion; it changes who reviews, not what Harness permits.	off
enableQuotaTool	Registers the read-only codex_connect_quota tool. Reload the plugin after changing it.	off

Proxy. Off by default — this plugin connects directly. To route only this plugin's traffic through an HTTP(S) proxy: More settings → Proxy configuration method, then either Detect proxy (scans common local proxy ports and environment variables without changing the saved connection) or Configure manually and enter a proxy origin with no credentials, path, query or fragment. Press Test proxy — activation requires a successful test of that exact address, and the result reports the HTTP status it reached. Only then can you save, and the panel states what will change ("After saving, ChatGPT Connect will use

"). **Disable proxy** returns to a direct connection. A failed test, or a saved proxy that can no longer reach the endpoint, leaves you on the direct path — this plugin never retries a failed proxy silently.

Use the image generation capability included with your current GPT subscription. Successful images appear below the answer in the conversation; generated originals are stored separately from attachment previews, so disabling the capability or uninstalling the plugin does not delete them.

Luna Reserve has been removed. Legacy enableReserveFallback settings are ignored; no Reserve negotiation or automatic model switching remains. Select an ordinary model explicitly for an old gpt-reserve conversation; if Harness cannot change that conversation's model, keep it and start a new one. Old return files are not read, migrated or deleted.

Saving. Change a control, then press Save changes at the bottom of the card. A conflict or a failed save keeps your draft. Volatile capabilities (proxy, auto-review) apply to new requests after saving; the quota tool needs a plugin reload.

Two common changes, step by step.

Add a second account: Settings → ANB 工具 → ChatGPT Connect → Accounts → Add account → complete the browser approval → the new account appears in the list → select it; later requests use it immediately.

Turn off auto-review after trying it: Settings → ANB 工具 → ChatGPT Connect → More settings → clear Codex Auto-review (Approve for me) → Save changes. Approvals go back to you on the next request; nothing already approved is re-run.

When the panel is not visible. It needs the ANB 工具 shell and a served configuration form. From the command line, dsh plugin --profile web exec dsh-chatgpt-connect doctor --json reports the loaded state; the raw configuration lives in the profile's plugin configuration, which you can edit by hand and reload.

7. Problems and next steps

Use the literal text you see; each row tells you what it means and what to do next.

What you see	What it means	Next
No openai-codex entry in the model picker	No account authorized, or the plugin is not loaded in this profile	Settings → ANB 工具 → ChatGPT Connect → Authorize; then re-check you started the same profile
Answer fails with overloaded	A service-side condition on that request. It is not proof your account is blocked	Retry once; if it persists, compare a request that worked. The bounded diagnostic metadata is attached to the error
Answer fails with 401 / re-authorization prompt	The refresh credential was revoked or expired	Authorize again; existing accounts are preserved unless you sign out explicitly
A listed model fails while another works	Model entitlement is decided by OpenAI, not by the list	Try a model you have used before; treat the catalogue as advertising, not permission
Settings shows the panel but every save conflicts	Another window holds a newer draft	Reload the page, re-apply, save once
doctor passes but requests still fail	doctor is local metadata; it makes no request	Do the quick start task instead — that is the only real check

If none of these match, collect the version (dsh --version and the plugin version from the panel), the exact error text, and a minimal reproduction, then use Feedback. Remove credentials, tokens, account identifiers and full session archives before posting.

8. Update, roll back, uninstall

Action	How	What happens to your data
Update	Install a newer exact version with the same dsh plugin … add command	Settings and accounts are kept. Read the changelog first; a version bump does not migrate anything for you
Roll back	Install the previous exact version the same way	Settings are kept. A newer setting that the older version does not know is ignored, not deleted
Uninstall	dsh plugin --profile web remove dsh-chatgpt-connect	The plugin code goes; credentials and generated images stay
Remove credentials	Settings → ANB 工具 → ChatGPT Connect → Sign out all accounts, or run logout before uninstalling	Deletes the OAuth state at $DSH_HOME/.openai-codex-auth.json
Remove generated originals	Delete the image directory yourself; the plugin stores originals separately from attachment previews	Irreversible

Uninstall and what it leaves behind

dsh plugin --profile web remove dsh-chatgpt-connect

Uninstalling is not a sign-out: the OAuth state file and any generated images remain on disk, because removing a plugin should not silently destroy data you may still want. If you do want it gone, sign out first (or delete $DSH_HOME/.openai-codex-auth.json) and delete the image directory. The plugin never reads or modifies ~/.codex/auth.json; that file belongs to a different tool.

9. Data, permissions and privacy

• Reads — your ChatGPT OAuth credentials for the authorized account; the model catalogue; and, only when you enable them, local images you point view_image at and the approval context needed for auto-review.
• Sends off the machine — requests go to OpenAI at chatgpt.com / api.openai.com, using the account's own credentials. Nothing is sent to the project's maintainers, and there is no telemetry endpoint. If you enable a proxy, that proxy is an egress you chose and configure.
• Writes to disk — OAuth state at $DSH_HOME/.openai-codex-auth.json; generated image originals in the session's image directory; nothing else outside Harness's own session files.
• Who can read it — the OS account running Harness. The plugin adds no sharing, no server-side store and no multi-user access. Treat the auth file as a password.
• Default switches — every optional capability is off on a fresh install (see the settings table). The one thing that is on by default is the network path to OpenAI once you authorize, which is the point of the plugin.
• Credentials in logs — the plugin's diagnostics report exit codes, allowlisted error codes and version strings. It does not print tokens, authorization URLs or account identifiers. Do not paste those into an issue yourself.

10. Compatibility: required, verified, unconfirmed

• Required — a Harness inside >=0.1.7-rc.1 <0.3.0, on Node ^22.19.0 || >=24.0.0.
• Verified — this exact release installed and reported doctor cleanly on each of the fifteen hosts listed in What you need, in the CI installation matrix for release 0.3.1-alpha.2.
• Not confirmed — live OAuth against a real account, real model responses, real image generation and editing, a real Desktop session, and migrating tasks across a Harness upgrade. Treat those as untested for this release.

11. Documentation and development

docs/ is not shipped in the npm package, so those links point at GitHub and stay current there.

• Installation and upgrades · Versioning policy · Migration from dsh-codex
• Configuration, diagnostics and recovery
• Architecture and security details
• Auto-review behaviour
• Chinese README

Building from source is a separate path, only for working on the plugin:

pnpm install --frozen-lockfile
pnpm run check
pnpm run test:browser
pnpm run check:dsh-install

lib/ is a committed build product; rebuild it on Linux, because the same sources produce different bytes on Windows. See the release runbook.

12. Feedback

Open an issue at https://github.com/yiyunet/dsh-chatgpt-connect/issues. The minimum useful material is: the plugin version, the Harness version, the exact error text, and the smallest reproduction you can manage. For a suspected credential or authorization problem, follow the security policy instead of opening a public issue. Remove tokens, account identifiers and full session archives from anything you post.

13. Licence and provenance

Copyright 2026 Frank Song for the original Codex Connect work; Copyright 2026 yiyunet for this fork's modifications. Distributed under Apache-2.0 — full text in LICENSE, notices in NOTICE. This project contains software derived from Yan-Zero/dsh-codex and is forked from franksong2702/dsh-codex-connect; all upstream copyrights are retained. It is not affiliated with, sponsored by, or endorsed by OpenAI, ChatGPT, Codex, DeepSeek, or DeepSeek Harness.

The published artifact carries a SLSA provenance attestation naming the commit and workflow that built it.