dsh-llm-github-copilot

English | 中文

GitHub Copilot LLM adapter for DeepSeek Harness.

Sign in with your GitHub account and use every Copilot model — including GPT-4.1, Claude Sonnet, Gemini, and GPT-5 family — directly inside DeepSeek Harness. Vision-capable models accept pasted or dragged images in the chat composer, images from /goal and /plan, and images returned by tools such as read_image and MCP servers.

Requirements

• DeepSeek Harness 0.1.1-rc.2 or newer. The vision path uses native image
  APIs (AttachmentStore.readImageRequest, offloadRequestImagesWithPolicy,
  requestImageHandleText) introduced in 0.1.1-rc.2; earlier releases will
  not run this adapter. Upgrade with npm install -g @deepseek-ai/dsh@latest.
• Node.js ≥ 24.

Install

From npm (recommended)

dsh plugin --profile web add @lujianjun19/dsh-llm-github-copilot

From GitHub

dsh plugin --profile web add github:lujianjun19/dsh-llm-github-copilot

pnpm 10 and newer block git dependency build scripts by default. If installation asks you to approve the package build, add the exact package key it reports to the profile's pnpm-workspace.yaml:

allowBuilds:
  dsh-llm-github-copilot: true

Then repeat the install command.

After installation, register the plugin in the profile's cordis.patch.yml:

- insert:
    - id: llm-github-copilot
      name: '@lujianjun19/dsh-llm-github-copilot'

Restart DSH to activate:

dsh web

Update

The DSH profile uses pnpm internally, and its lockfile pins the exact installed version. Re-running dsh plugin add alone may not pick up a newer release. Use either of the following commands from any directory:

# Using npm (simplest — no extra flags needed)
npm install --prefix ~/.dsh/profiles/web @lujianjun19/dsh-llm-github-copilot@latest

# Using pnpm
pnpm add --dir ~/.dsh/profiles/web @lujianjun19/dsh-llm-github-copilot@latest --no-frozen-lockfile

Then restart DSH:

dsh web

To confirm the installed version, open Settings → GitHub Copilot in the Harness browser UI — the version number is shown at the bottom of the panel.

Sign in

Run the login command inside the Harness chat:

/copilot-login

Follow the instructions — open the verification URL in your browser, enter the displayed code, and authorize the GitHub App. The token is stored automatically once you complete authorization. Run /copilot-status to confirm the connection and see the available models.

To sign out:

/copilot-logout

You can also set the token directly as an environment variable (useful for CI or headless setups):

export GITHUB_COPILOT_OAUTH_TOKEN=<your-github-oauth-token>

Supported token types:

Prefix	Source
gho_	OAuth token (gh auth login)
github_pat_	Fine-grained PAT (requires Copilot permission)
ghu_	GitHub App user token (VS Code client)

ghp_ classic PATs are not accepted by the Copilot API.

Features

Model discovery — available models are fetched live from https://api.githubcopilot.com/models on each login and cached for 5 minutes. No static list to maintain.

Vision support — models that declare supports.vision: true (e.g. gpt-4.1, gpt-4o) accept images from every source Harness produces: pasted or dragged images in the composer, /goal and /plan attachments, and tool-result images (read_image, MCP servers). Images are derived per model route through the Harness attachment service (readImageRequest), tagged with a stable handle, and sent over both wire protocols. When a request exceeds a model's image count or the local inline byte budget, older request images are offloaded first while the current user submission and the latest tool-result batch are protected; a stable placeholder marks any omitted image without altering the durable history. Set imageOverflowPolicy: error to reject over-limit requests instead.

Two wire protocols — the adapter speaks both OpenAI Chat Completions (/chat/completions) and the newer Responses API (/responses). The correct endpoint is chosen automatically per model.

Reasoning control — effort levels (low / medium / high / max) are forwarded to models that declare them (gpt-5.x, Claude thinking budget, Gemini reasoning).

Automatic token refresh — the short-lived Copilot API token is renewed transparently before it expires; no action required.

Settings page — the plugin adds a dedicated GitHub Copilot section to the Harness Web settings UI (open DSH in your browser → click the gear icon → GitHub Copilot). From there you can sign in, view authentication status and the available model list, and sign out — no slash commands required.

Configure

The plugin works with no configuration. To override defaults, edit the profile's cordis.patch.yml:

- id: llm-github-copilot
  config:
    oauthTokenEnv: GITHUB_COPILOT_OAUTH_TOKEN   # env var that holds the GitHub OAuth token
    baseURL: https://api.githubcopilot.com       # override Copilot API host
    defaultContextWindow: 262144
    defaultMaxTokens: 32768
    streamIdleTimeoutMs: 300000
    imageOverflowPolicy: offload-oldest         # offload-oldest | error
    defaultImagePixelBudget: 4194304            # request-image pixel budget (2048×2048)
    maxInlineRequestImageBytes: 20971520        # total Base64 request-image budget (20 MiB)
    inlineImageOffloadByteQuantum: 10485760     # oldest-image removal step (10 MiB)
    models: []   # optional static fallback catalog; leave empty to use live discovery

Static fallback models may declare vision capability explicitly (used only when
live /models discovery fails); capability is never inferred from a model name:

    models:
      - id: custom-vision-model
        inputModalities: [text, image]
        vision:
          maxImageBytes: 3145728
          maxImages: 1
          mediaTypes: [image/jpeg, image/png, image/webp]
          imagePixelBudget: 4194304

Develop

Requires Node.js ≥ 24 and npm.

cd /path/to/dsh-llm-github-copilot
npm run build      # concatenate source fragments → lib/index.js and lib/client.js
npm test           # build + deterministic artifact, i18n, and metadata tests
npm run check      # build + tests + npm pack dry-run
npm run deploy     # test → build → atomic deploy with rollback backup

Install the checkout into a local profile directly:

dsh plugin --profile web add .

After Host changes restart DSH; after Client-only changes a hard refresh (Ctrl+Shift+R) is usually enough.

Rollback

npm run deploy keeps a timestamped backup before each install. To roll back, stop DSH and restore the wanted backup:

cp -r ~/.dsh/plugin-backups/dsh-llm-github-copilot/<timestamp> \
      ~/.dsh/profiles/web/node_modules/@lujianjun19/dsh-llm-github-copilot
dsh web

Troubleshooting

Model selector shows no GitHub Copilot group after restart
The plugin loaded in a previous process that didn't pick up the new package. Restart dsh web. The stored token is preserved; no need to sign in again.

Models appear but Claude is missing
Your egress IP is restricted. Export HTTPS_PROXY pointing to a proxy that exits from an unrestricted region, then restart dsh web.

/copilot-login times out or reports a network error
Transient network issue during device-code polling. Run /copilot-login again to get a fresh code (the old one is invalidated automatically).

configurable provider "github-copilot" is already declared
An older version of this plugin used the route name github-copilot, which conflicts with a DSH built-in. This version uses github-copilot-official. Verify that your cordis.patch.yml uses id: llm-github-copilot and name: '@lujianjun19/dsh-llm-github-copilot'.

Token expired
No action needed. The plugin stores the long-lived GitHub OAuth token and refreshes the short-lived Copilot API token automatically before it expires. Only an explicit sign-out or token revocation requires a new /copilot-login.

License

MIT