dsh-chatgpt-plan

A native DeepSeek Harness plugin that connects a ChatGPT plan account through OpenAI's Sign in with ChatGPT plan-sharing preview. Inference uses the official Responses API directly; Harness keeps ownership of tools and permissions.

Package	dsh-chatgpt-plan (this source repository is dsh-codex-oauth)
Version	0.1.9-sidebar-height.1 — experimental
Harness compatibility	>=0.2.0-rc.2 <0.3.0-0
License	MIT

Disclaimer

• Usage limits are enforced by OpenAI, not by this plugin. ChatGPT plan usage through third-party applications is governed by the Sign in with ChatGPT preview and by limits that OpenAI sets and can change.
• The permitted usage you configure in ChatGPT does not remove app-specific limits. Even with usage set to 100% for this application, OpenAI can apply limits that are specific to third-party applications or to a short rolling window. A turn can therefore fail with a quota error while ChatGPT still reports remaining weekly capacity.
• The plugin never changes the billing path. Quota and rate-limit failures are surfaced as they arrive: no API-key fallback, no model substitution and no inferred reset time. See known issues for a reported instance of this behavior.
• This is experimental software built on a preview feature. Model metadata, limits and availability can change, degrade or stop working without notice, and no compatibility or availability guarantee is provided.

Known issues

• An app-specific limit can remain at its original value after the weekly usage resets (external, open). A connection created through the Sign in with ChatGPT preview can keep failing with subscription_sharing_usage_limit_exceeded (429, "ChatGPT usage limit for this app") while ChatGPT reports that weekly capacity has reset. It was reported to OpenAI and acknowledged as specific to the Sign in with ChatGPT route, and it does not affect Codex usage. This plugin classifies the code as QUOTA and never infers a reset time; it never changes the billing path either, so the limit cannot be cleared from inside the plugin. See diagnostics for the entry.

How it works

• Sign-in. OAuth 2.0 with PKCE through the browser, using the ChatGPT plan-sharing preview and a loopback callback on 127.0.0.1.
• Credentials. Tokens live in Harness's native managed credential store, which also handles refresh, rotation and revocation. The Client surfaces are token-free.
• Inference. Requests go directly to https://api.openai.com/v1/responses with the granted plan permission. There is no API-key configuration, no Codex transport and no intermediary server.
• Model catalog. Pi's openai-codex chat catalog is the only model metadata source. A profile-local cache and a bundled snapshot keep the last valid list available offline.
• Model identity. Exact Pi names and IDs are preserved and listed newest first. The plugin renames, aliases and substitutes nothing; the selected ID is what is sent to OpenAI.
• Tools and attachments. Native history, local tool declarations and results, images and file references are projected into the request. Harness executes tools and checks permissions.

Requirements

• Node.js 24 (nvm use).
• DeepSeek Harness 0.2.0-rc.2 or a later 0.2 release; development is pinned to the tested 0.2.0-rc.2 baseline.
• A ChatGPT account eligible for the plan-sharing preview.
• Browser and Harness host on the same machine, because the OAuth callback listens on 127.0.0.1.

Install

Build the package and install the same tarball in the target profile through Harness's native plugin manager. Keep the previous working tarball for rollback.

nvm use
npm ci
npm run check
npm run pack:plugin

The commands above emit dsh-chatgpt-plan-<version>.tgz in the repository root.

Windows Desktop

& "$env:LOCALAPPDATA\Programs\DeepSeek Harness\resources\runtime\cli\bin\dsh.cmd" plugin --profile desktop add "C:\path\to\dsh-chatgpt-plan-0.1.9-sidebar-height.1.tgz"

macOS Desktop and local Web

"/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" plugin --profile desktop add /absolute/path/dsh-chatgpt-plan-0.1.9-sidebar-height.1.tgz
dsh plugin --profile web add /absolute/path/dsh-chatgpt-plan-0.1.9-sidebar-height.1.tgz

After installing, run the application's restartHost action and reload the existing UI. Source edits alone never change an installed profile. For a new custom Web profile, initialize it from the shipped Web template first; see the plugin guide. For an agent-driven or unattended install, follow the agent installation runbook.

Usage

1. Open Settings → Models → ChatGPT Plan and choose Continue with ChatGPT.
2. Complete the browser sign-in and authorize ChatGPT plan usage. The first connection shows a one-time confirmation that eligible requests use your plan.
3. Select a model under the native ChatGPT Plan provider in the conversation model selector.
4. Use Update models now to refresh metadata without an inference request, and View usage in the sidebar to open ChatGPT's usage settings.

While a session is connected, the sidebar and the Models header state Using ChatGPT plan; the sign-in invitation is shown only when the profile is not connected.

Disconnecting revokes the remote grant and erases local tokens. To switch accounts, disconnect and choose Change account.

Model catalog

• Every plugin activation loads a valid profile cache or the bundled snapshot, then checks Pi in the background. Startup never waits for the network.
• Each check is bounded to 4 seconds, 2 MiB and 1,000 records. Remote JSON is untrusted input and is validated before publication.
• A valid response replaces the entire list, including a valid empty list. Errors retain the last valid list or the bundled snapshot and raise a catalog warning that is separate from authentication state.
• The cache is stored per profile at .cache/dsh-chatgpt-plan/pi-catalog-v1.json, independently of OAuth grants.
• available and preferredModelAvailable mean locally selectable, not verified account entitlement. OpenAI enforces access on the actual inference request.
• Discovery never calls the OAuth /v1/models endpoint, and no model is added, aliased or substituted manually. See models and restrictions.

Project layout

Path	Contents
packages/plugin	Plugin source, package documentation and third-party notices.
packages/plugin/src	Host account management, Pi catalog and cache, Responses adapter, RPC service and Client UI.
packages/plugin/docs	Architecture, model restrictions, diagnostics and validation evidence.
scripts	Maintainer catalog snapshot generator and Typert preparation.
test	Protocol, native-service integration and React tree tests using fictional credentials.
docs	Development guide and architecture decision records.

Development

nvm use
npm ci
npm run check          # build Host, generated RPC and Client, then run the automated suite
npm run typecheck      # static types
npm run pack:plugin    # emit the installable tarball in the repository root

npm run check builds before testing. OAuth callback tests bind localhost and need a terminal that permits local listeners; they neither authenticate a real account nor perform live inference.

Maintainers regenerate the bundled catalog snapshot from the same Pi service with npm run catalog:snapshot. Review the metadata diff and rerun the checks before packaging. See the development guide.

Compatibility and validation

Runtime peer dependencies and the informational engines.dsh field declare >=0.2.0-rc.2 <0.3.0-0. The -0 upper bound deliberately excludes every 0.3.0 prerelease. A permissive admission result is not an API compatibility guarantee; validate each new Harness release before relying on it.

This release, 0.1.9-sidebar-height.1, changes only the sidebar presentation: one visible status line, a fixed expanded height matching the native account launcher, ellipsis and the embedded ChatGPT mark. Host behavior, credentials, transport, catalog and inference are unchanged. The automated suite and static types pass on Node 24 (154/154, plus 11/11 in the focused sidebar suite). Installation into the real desktop profile through the bundled CLI succeeded, and the installed Client bundle matches the repository build.

Windows Desktop is the only platform executed for this release; macOS Desktop and Linux/Web have not been executed. The package is platform-neutral — prebuilt JavaScript, pure-JavaScript dependencies, no os/cpu restriction and no native module — but that is a static property, not a runtime result. Live browser rendering, theme contrast, pointer hit testing and keyboard focus in the installed GUI remain pending in the local deployment receipt.

The catalog, native-selector and live gpt-6.1-sol evidence belongs to the 0.1.5-pi-catalog series and was not re-executed for this presentation release. Automated fixtures, catalog HTTP checks, installation and live inference are recorded separately in the validation evidence.

Documentation

Document	Contents
Plugin guide	Installation per platform, rollback and connection behavior.
Architecture	Components, request boundaries, catalog lifecycle and credential races.
Models and restrictions	Catalog semantics, capabilities and preview restrictions.
Diagnostics	Symptom-to-action tables for catalog, installation and authorization.
Validation	Version-scoped build, test and platform evidence.
Development guide	Repository layout, release procedure and compatibility policy.
Agent installation	Runbook for installing, configuring, verifying and rolling back the plugin from a terminal or desktop agent.
ADR-0001	Decision to use the Pi catalog as the single metadata source.

License

MIT. Bundled third-party notices, including the Pi model metadata attribution, are reproduced in THIRD_PARTY_NOTICES.md.