WSL043/dsh-codex-subscription

DSH Codex Subscription

中文 · English · 日本語 · Español · Français · Deutsch · Русский · Português

Use your ChatGPT / Codex subscription directly in DeepSeek Harness

Sign in with ChatGPT Plus / Pro to use GPT-6 and other Codex models, quota, web search, and image generation inside DSH. No OpenAI API key or Codex CLI.

Highlights · Install · Images & sketching · Phone remote control · All features · User guide · Troubleshooting · Agent

Highlights

What you want to do	What the plugin provides
Chat and search	Sign in with your subscription; pick account-provided Codex models, reasoning and speed levels; use subscription web search.
Manage accounts and quota	Switch accounts, view short-window and weekly quota, Credits and resets; see usage history and optional quota alerts in the composer.
Create and edit images	Generate images, mark regions for editing, or draw manually and with the Agent on an editable sketch board. Examples →
Continue from your phone	Pair the ChatGPT mobile app to control this computer's DSH. DeepSeek and other DSH models work too. Beta, off by default →
Use connected apps	Let DSH models use apps connected in ChatGPT, including GitHub, Gmail, Drive and Calendar; write approval follows the current DSH session permissions.

Models and quota depend on what your account returns. Complete feature list →

Let DeepSeek or any other DSH model use the apps you connected in ChatGPT: GitHub, Gmail, Google Drive, Calendar and more. Models search first, then call, and write approval follows the mode you select and the current DSH session permissions. Off by default; turn it on under Web & tools.

Web search and webpage reading sources, ChatGPT connectors and official Codex independent subtasks are on Settings → Codex subscription → Web & tools, with one master switch at the top: when off, conversations on DeepSeek and other non-Codex models no longer use subscription web search (DSH search is used instead), image generation or ChatGPT connectors; Codex subscription models are not affected. On by default.

Prepare DSH

Verified against DSH 0.2.2-rc.1, with 0.2.1-rc.3 and 0.2.1-rc.2 still supported; the minimum supported version is 0.1.7-rc.2. Please upgrade older DSH releases first.

The official DSH Desktop 0.2.2-rc.2 was verified on-device.

DSH 0.2.2-rc.1 CLI/Web requires Node.js 24.9 or newer; the official desktop app uses its bundled runtime.

Also supports the preview host 0.2.2-alpha.2. See the compatibility notes for scope and validation limits.

This plugin supports the latest DeepSeek Harness release recorded in its package metadata and requires a ChatGPT account that currently has Codex access.

• Recommended: the official desktop app. DeepSeek now ships a DeepSeek Harness desktop app (macOS on Apple Silicon and 64-bit Windows; preview, open source). It needs no Node.js setup: install it, then add this plugin from the Plugins page as described below. The desktop app, the web app and the CLI share one DSH plugin system; this plugin's end-to-end acceptance is mainly run on the web app, so please open an issue if the desktop app misbehaves.
• Need Linux, or prefer a community portable build? Use DSH-Portable, a community portable desktop distribution for Windows, macOS, and Linux.
• Prefer the official route? Follow the DeepSeek Harness run guide.

Install

Install from the Plugins page (recommended)

1. Open Plugins → Add plugin in DSH.

2. Paste this package name, with its version, into the Package name or address field:

   dsh-codex-subscription@3.2.15
   

3. Click Install and wait for completion. Follow the page instructions; save your work before restarting if requested.

4. Open Settings → Codex, sign in to ChatGPT, then select a Codex model in your conversation.

Why the version is included: DSH's plugin page only installs versions that are at least 24 hours old, so the bare package name can install an older release during the first day after a new one, and that older release may not support your DSH. With the version you get exactly this release; a day after release the bare dsh-codex-subscription also resolves to the latest stable release. Other version numbers are on the releases page.

dsh plugin --profile web add dsh-codex-subscription@3.2.15

Follow the restart instructions, then sign in under Settings → Codex. Both the Plugins page and the terminal use DSH's installation management.

After signing in and selecting a Codex model in Web, install the same plugin in the Headless profile:

dsh plugin --profile headless add dsh-codex-subscription@3.2.15
dsh --profile headless "Reply with only the word: ok"

Interface demo

Connect your ChatGPT account, choose a Codex model in DSH, and check your subscription quota from the same settings page.

Open Settings → Codex → Start and select Browser sign-in. Once connected, this page shows your saved accounts and the quota reported by ChatGPT. Expand Usage overview or Where the allowance went for more detail.

Account and quota values are examples.

Open Settings → Codex → Models to set a default model and reasoning effort. You can also choose a model for one conversation from the selector in the composer. Enter a prompt and send it to use your subscription in DSH.

Settings tab	What it controls
Start	Sign-in and accounts, a switch for every feature, subscription quota, quota display and alerts
Models	Default model, default reasoning effort and model list, answer style; the folded Advanced group has context length, retry, connection mode, cloud compaction and the review model
Web & tools	The master switch for other models, search source, ChatGPT connectors, official Codex subtasks
Create	Image generation, editing, @image in the composer, image viewer, sketch board and agent drawing
Phone & more	Phone remote control, experimental voice calls, local cache, support diagnostics

Voice conversation is off by default. Enable it under Phone & more, then start a call from the composer. If a native window, including Linux, does not support WebRTC, DSH offers to continue the same conversation in a browser. Voice samples remain available, and your existing subscription is used without an API key.

Model-aware context follows your account's model directory; refreshing preserves your unsaved draft. Settings show the raw window; conversation budgets reserve headroom using the official percentage (5% when omitted), then DSH manages output reservations and compaction. No manual subtraction is needed. Quota groups use backend-provided values, including separate buckets such as Spark.

Subscription failures remain visible: no silent fallback to another paid route.

Images and sketching

Generate images with your subscription, view and download the originals, or mark a region and add notes to continue editing. Image generation and editing are controlled under Settings → Codex → Create.

Settings shown directly as cropped screenshots from the actual DSH interface, following the light or dark theme.

Sketch board (Beta)

The sketch board and Agent drawing are Beta and off by default; turn them on in the Create page. Draw by hand from the sketch button in the composer, or send @sketch with a request to have the Agent draw.

The board supports layers, native curves, brushes, gradients, blur, blend modes, clipping, undo/redo and shortcuts. Save multiple drafts and export PNG or editable drafts. With a reference, the Agent can compare images, measure colours and positions, and undo. Finished drawings can be replayed and recorded as video.

The agents drew these examples on the sketch board using native editable strokes, without calling image generation. All three used the same reference and drawing-tool baseline. These are individual runs with different effort levels and durations.

After a connection interruption, the 6.1 drawing was restored from its autosave and original successful calls. The model then checked the preview and saved it, with no manual drawing added. This was not an uninterrupted successful run.

The Agent first drew an editable sketch, then used that sketch as the reference for subscription image generation. Sol and Astra below name the conversation models directing the workflow, not the image-generation backend.

Clocktower cutaway

Blue-green landscape

Image editing and draft storage →

Phone remote control (Beta)

Pair the Codex section of the ChatGPT mobile app with the DSH on your computer. Start or continue conversations, use tools and handle approvals from your phone, with DeepSeek or any other DSH model.

Off by default. Enable it under Settings → Codex → Phone & more → Phone remote control and pair by QR or manual code. Your computer must stay on, online and running DSH.

This is a third-party compatibility layer using OpenAI's remote connection service. Official Codex remote control exists, but DSH host compatibility is not explicitly guaranteed by OpenAI and may change with service updates. You can remove paired devices or disable it at any time.

• follow, start, continue, interrupt and archive conversations; search conversations and keywords inside one;
• switch model, reasoning level and permission (workspace write / full access);
• approvals and questions pop up on the phone; if it drops, they pop up again when you reopen it, otherwise they go back to the computer;
• replies appear as they are written; commands, file edits with diffs, the task list, web searches, connector calls and generated images use the phone's native views;
• send photos, @ search files on the computer, / pick DSH skills, view changes with read-only git, and start chats in a real folder picker;
• DSH workspaces appear as projects; fork a conversation, see usage limits, lifetime usage and context usage; where DSH supports /compact, compact the context from the phone;
• model errors show on the phone; paired phones are listed in settings and can be removed.

All features

Model access and quota depend on your account.

Choose the model for the task, then adjust reasoning for how much analysis you need and speed for how quickly a supported request runs. Response verbosity changes the requested level of detail. In Web & tools, use search to find pages and webpage reading to open a URL you already have.

Feature	What you can do	Availability / setup
Subscription sign-in	Sign in to ChatGPT and use Codex subscription models without an API key	Normal chat needs no Codex CLI
Model catalog sync	Read available models and reasoning levels from your account, with manual refresh; the last successful catalog is stored locally per account and remains available during the background refresh after restart, with a 24-hour lifetime; models marked for retirement show their retirement date next to the model name, without automatically switching your selection	New models appear when your account offers them; the account catalog is the only source, so no models are shown when signed out or when neither a remote read nor a valid cache is available
Model list	Show or hide individual models in the composer picker on the Models page; the picker keeps DSH's own model labels (such as the new-model badge)	Display only; selected models and requests still work
Composer model menu	Auto yields to another model picker and shows the subscription menu when used on its own. DSH standard leaves the slot to the host or another plugin.	Under Models. Select Subscription explicitly to use its menu and speed controls on supported models. Saved choices are preserved.
Input predictions (Beta)	After a completed reply, the empty composer may suggest a follow-up. Tab puts it in the draft for editing.	Off by default under Models. Only supported Codex models with prediction access qualify. Requests may consume usage.
Default model	Choose the model a new conversation opens with on the Models page	Writes the DSH default model setting, the same one a model switch inside a chat writes; unset it to follow the DSH default
Default reasoning effort	Choose a default level for subscription models on the Models page; Model default follows each model's official catalog recommendation	Manual choices take priority. Unsupported levels fall back to the model's valid default; existing manual conversation choices are preserved
Reasoning levels	Choose supported reasoning effort in the composer	Levels vary by model
Ultra reasoning	Lets DSH delegate suitable independent work to native subagents using the model’s supported reasoning effort.	Offered only when the model and Agent preset support it. Separate from the Ultrafast speed tier.
Speed tiers	Switch supported models between Standard and Fast; Ultrafast (Beta) too when the account catalog offers it, with a lightning indicator	Standard by default; Fast and Ultrafast increase usage, and Ultrafast is Pro 500 only today
Response verbosity	Request short, medium, or detailed responses	Follows the model default unless changed
Input image detail	Choose Low, High, Original, or Auto detail for images sent to the chat model	Affects input tokens; Auto by default
Subscription web search	Auto, DSH, and Codex sources	Auto prefers the source matching the session model and tries the other only for missing credentials or an unavailable service. Explicit DSH / Codex choices fix the source and report failures directly. With Other models off, non-Codex chats use native DSH search only. The plugin neither rewrites DSH search settings nor adds a DeepSeek account fallback; native search charges depend on the selected service
Webpage reading source	Choose how the tools read a known URL, independently of the search source.	Web & tools. DSH is the default. Codex requires a compatible host and reports reading failures directly.
Other-models switch	One switch that stops conversations on DeepSeek and other non-Codex models from using subscription web search, image generation and ChatGPT connectors	Settings → Codex subscription → Web & tools; on by default, Codex subscription models are not affected; phone remote control and voice are started by you and are not affected
Search scope	Choose live / cached search, disable search, or restrict domains	Configure under Web & tools, in the search source group
Stream inactivity timeout	Set how long a request can receive no streamed data before it is aborted	10 minutes by default; not an overall response deadline
Context budget	Use standard, extended, or custom budgets per model	Limited by the model catalog
GPT-Reserve	Select the reserve model when your account offers it	Experimental; no promise of a separate reserve quota

Start with Start → Subscription quota for the remaining allowance and reset times. A short and a weekly window can apply at the same time. The usage breakdown describes past activity, while Runway estimates what will happen if your recent pace continues. Enable Composer quota to see the remaining allowance without leaving a chat.

Feature	What you can do	Availability / setup
Multiple accounts	Add, switch, remove, or sign out of all accounts	Settings → Codex → Start
Quota details	View short / weekly windows, reset times, and separate quota groups	Only backend-provided windows and groups are shown
Where the allowance went	In the quota card, see each week's allowance split by model, surface, source and trigger	From the official usage report: approximate, may lag a day or two; hidden when the account has no report
Usage overview	In the quota card, see lifetime tokens, busiest day, streaks, most used reasoning level and a half-year daily heatmap	From the official usage statistics; hidden when the account has none
Credits and reset cards	View extra Credits, spending caps, and available reset cards	Shown when returned for your account
Composer quota	See remaining quota as a percentage or progress bar while chatting	Off by default; choose your display mode
Runway estimate	Hover over or click the composer quota to see when recent usage would exhaust it	Beta; optional estimate, not a guarantee
Quota alerts	Use fixed, early, or custom short / long-window thresholds	20% remaining by default; can be disabled
Redeem reset cards	Use an available reset card after confirmation	Confirmation and waiting safeguards; server decides the result
Retry after reset	Wait for a confirmed short-window reset before continuing	Off by default; waiting can be stopped

Use generation to create an image and a reference or comment pin to request a specific edit. Closing the viewer saves your pins. Add to composer puts them in a draft you can review and send. Export PNG for sharing or a native draft to keep layers and strokes editable.

Feature	What you can do	Availability / setup
Image generation	Generate through your subscription and adjust available request model / quality options; ask for a transparent-background PNG when needed	On by default, can be disabled
Reference image editing	Continue refining an image with references	On by default, can be disabled
Image viewing and download	Completed images appear after the answer without expanding tool activity; zoom, pan, fit to window, and download verified originals	Built-in viewer; older images may only have previews
Region-based editing	Mark regions, add notes, and continue editing with location references	Fills the composer for you to review and send
Manual canvas	Choose an aspect ratio and draw with pen, pencil, marker, and erasers	Beta; canvas off by default
Shapes and text	Draw lines, arrows, shapes, native curves, and editable text / objects	Available in the sketch canvas
Gradients, blur, blend and clip	Fill with linear / radial gradients; add blur (glow, shadow, haze), blend modes (multiply, screen, overlay, …) and shape clipping to a stroke; layers can have a blend mode too	All stay editable properties and survive moving, scaling, drafts and export; the agent gets the same through SVG gradients, feGaussianBlur, mix-blend-mode and clip-path
Layers and image import	Work in layers and bring images into the canvas	Images are managed as layers
Smoothing and navigation	Smooth complete strokes after release; pan, zoom, undo, and redo	Shortcuts can be customized or disabled
Multiple drafts	Save and reopen up to 20 local drafts	Stored in the current browser
Import and export	Export PNG, or native drafts that preserve editable strokes	Drafts can be imported again to keep editing
Agent drawing	Ask the agent to draw with @sketch, follow progress, and stop it; the agent can draw whole stages in SVG, imported as editable native strokes; with a reference it can compare side by side, read a difference heat map, zoom into a region, measure colours and edge positions, and undo, and every batch reports how the error changed	Beta; off by default, requires the canvas. The tools only measure and show; they never draw for the model
Replay and video	The Replay button redraws the process stroke by stroke (lines along their path, fills outlined first and then filled top to bottom), with speed presets and a scrubber, and records it as a WebM video	The video is previewed in the board and only saved when you click Save video; nothing downloads automatically
Result preview for the model	Return a canvas image after drawing for subsequent review	Beta; off by default, adds image input usage

Use phone control to reach a running DSH from the ChatGPT mobile app. Use a voice call to speak in a conversation, or a subtask to delegate part of the work. For a failure, open Phone & more → Support diagnostics and check which prerequisite or recent operation the report covers.

Feature	What you can do	Availability / setup
Phone remote control	Drive the DSH on your computer from the Codex section of the ChatGPT mobile app: start, continue and archive conversations, switch model and permissions, answer approvals and questions; DeepSeek and any other DSH model work	Beta; off by default; Settings → Codex → Phone & more → Phone remote control, pair by QR or manual code
Voice conversation (Beta)	Start a spoken conversation from the composer. A browser can host the call when the native window lacks WebRTC.	Off by default under Phone & more. Grant microphone access when prompted.
Codex independent subtasks	Reuse subscription sign-in and workspace permissions; select allowed models and reasoning levels	Beta, optional, most people do not need it: by default DSH runs its own subagents (they can use subscription models too); install the optional component (including a dedicated Codex CLI) and turn on model selection under Plugins → Subagents only if you want subtasks run by the official Codex agent. It uses its own data directory and your subscription sign-in, never the Codex login on this computer
Subtask component management	Install, disable, or uninstall the official Codex subtask component from settings	Managed by supported DSH hosts; no extra login
Cloud context compaction	Continue long conversations using Codex compaction while retaining DSH history and native compaction	Beta / experimental; off by default, enabled requests use SSE
WebSocket transport	Experimentally reuse connections and context transfers, with SSE fallback on connection failure	Beta / experimental; SSE by default, no speed guarantee
Codex reviewer for Auto review	DSH Auto review judges tool calls with Codex's own review model; DSH keeps its review policy and allow/deny flow	Beta; off by default, applies when your account catalog offers the model
ChatGPT connectors	Let DeepSeek or any DSH model use the apps you connected in ChatGPT: GitHub, Gmail, Google Drive, Calendar and more; models search first, then call, and write approval follows the chosen mode and current DSH session permissions	Off by default, Settings → Codex → Web & tools → ChatGPT connectors; connect the apps in ChatGPT first
Support diagnostics	Inspect prerequisites and recent operation results, and copy a short report (about 1 KB; the full one can be copied separately) without credentials or conversation text	Settings → Codex → Phone & more; unverified does not mean broken
Forecast cache cleanup	Inspect and clear local quota forecast history	Keeps sign-in, drafts, originals, and shared dependencies

Reference for agents and contributors

DSH owns the conversation, tool approvals and workspace permissions. The subscription plugin supplies the Codex route and adds the settings and tools described above. When investigating a missing feature, check the DSH version, account capabilities, selected model and setting before tracing the relevant tool or request.

Detailed user guide · Sketch tool contract · Setting defaults · Development and documentation index

Reuse cache when changing effort

On by default under Settings → Codex → Models → Advanced. It can help supported models reuse their cache when you change reasoning effort. Ordinary prompt caching remains available with this setting off.

How it works and when it is paused

Update and uninstall

Find this plugin on the DSH Plugins page and use its update or uninstall action. Follow any restart instructions. Uninstalling this plugin does not remove other plugins.

dsh plugin --profile web update dsh-codex-subscription@3.2.15

Run only when you want to uninstall:

dsh plugin --profile web remove dsh-codex-subscription

Troubleshooting

They are unrelated and can be installed together. The official Codex / Claude Code bundles let DSH run Codex or Claude Code style hooks configs (scripts that run before and after tool calls). This plugin signs in with your ChatGPT subscription and provides Codex models, quota, search, images and phone remote control.

Use Plugins → Add plugin in DSH and paste the package name. No terminal setup is needed.

Install from the DSH you actually use. For terminal installation, run commands in the intended DSH environment and check its profile. Do not delete profiles or change your system PATH to force an install.

Generate Support diagnostics under Settings → Codex → Phone & more, paste it into the bug report form, and include the steps that caused the problem.

Diagnostics exclude credentials, account identifiers, raw responses, and full logs. Never attach sign-in URLs, authorization codes, or browser callback addresses.

For Codex independent subtasks, install Codex Subagent from DSH Plugins, then choose Prepare CLI in the subscription settings. Setup, removal and the full guide →

Project and support

• ChatGPT / Codex subscription only, done in depth: quota and forecasting, images and sketches, subtasks, diagnostics. Other subscriptions such as Claude or Grok are out of scope.
• Fail explicitly: requests never silently fall back to another paid route, and quota windows the account does not return are never invented.
• Tracks official DSH releases: every release is accepted end to end (install, start, remove, reinstall) against official DSH channels such as latest and next; verified versions are recorded in the compatibility notes.
• Models, quota and features depend on what your account actually returns; the ChatGPT backend and DSH can change independently.

This community project is not affiliated with or endorsed by DeepSeek or OpenAI.

• Use the bug report form for project feedback.
• Use the feature request form for focused product suggestions.
• Focused fixes and compatibility improvements are welcome; see CONTRIBUTING.md.
• Help maintain and develop the project: take ongoing responsibility for a module, follow up on related issues or review code. Open an issue to discuss.
• For DSH plugin discussion, visit DeepSeek Harness Discussions.
• Read SECURITY.md before reporting sensitive issues.

If this project is useful, the Star button helps more DSH users find it.

中文 · English · 日本語 · Español · Français · Deutsch · Русский · Português · MIT