dsh-copilot-quota

GitHub Copilot quota indicator for the DeepSeek Harness web GUI.

DeepSeek Harness has no Copilot quota surface of its own: the dsh-oauth-copilot plugin stores the grant and the account-enabled model list, and the harness only reacts once a limit is already exhausted (a QUOTA / RATE_LIMIT failure notice). This plugin fills the gap with a live badge in the composer row, next to the send button.

Copilot 200/200      ← green dot, click to refresh

What it shows

Element	Meaning
Dot	Worst remaining percentage across the metered buckets: green, amber at ≤ 40 %, red at ≤ 15 %
Number	The chat bucket as remaining/total (falls back to the first metered bucket)
Hover	Login, plan and SKU, every bucket, the quota reset date and the time of the last refresh
Click	Force a refresh, bypassing the server cache

The badge is hidden entirely when no Copilot grant exists, so it never claims a meter that is not there. Buckets that report 0 of 0 are dropped for the same reason — a free plan has no AI-credit pool.

Requirements

• DeepSeek Harness >= 0.1.0-rc.6 < 0.3.0, profile web (verified on 0.2.0-rc.2).
• A signed-in Copilot grant:

  npx -y dsh-oauth-copilot login
  

  Sign-in is a human-only terminal operation by design; see that package's README.
• github.com only — GitHub Enterprise is not supported, matching dsh-oauth-copilot.

Install

From GitHub (no clone or build step):

dsh plugin --profile web add github:vowa-antilamer/dsh-copilot-quota

From a local checkout:

dsh plugin --profile web add /path/to/dsh-copilot-quota

A live profile recomposes itself (the server half starts serving immediately and the client bundle starts being served); reload the browser page to load the client module. Restart dsh web only if the badge does not appear after a reload.

Remove:

dsh plugin --profile web remove dsh-copilot-quota

Configuration

Optional entry config (profile patch or the plugin row in the GUI):

- id: copilot-quota
  config:
    cacheMs: 300000   # GitHub response cache, ms (0 disables caching)

How it works

Half	File	Responsibility
Server	index.js	Reads the llm-pi-ai/github-copilot record through ctx.credentials, calls GitHub, serves GET /copilot-quota/api
Client	client.js	Slot conversation.input.right; polls the route every 2 minutes and on window focus

The server half asks the same client endpoint VS Code uses:

GET https://api.github.com/copilot_internal/user
Authorization: Bearer <token>

Both response shapes are supported: quota_snapshots (current accounts) and monthly_quotas + limited_user_quotas (older ones). The answer is cached server-side (5 minutes by default) and the last good reading is kept on screen if a refresh fails.

Tokens never reach the browser. The grant stays in the harness credential store; only derived numbers cross the route. The badge talks to the same origin that served the page (/copilot-quota/api).

Verify without the GUI

curl http://127.0.0.1:3080/copilot-quota/api

Limitations

• copilot_internal/user is not a public GitHub API. If GitHub changes the response shape, the badge reports the error in its tooltip instead of breaking the UI.
• GitHub Enterprise is unsupported (the grant whitelist in dsh-oauth-copilot accepts only the official proxy.individual.githubcopilot.com endpoint).
• The badge reflects the signed-in account, not the model currently selected: it is the account's quota, visible with any model.
• Quota is refetched on a timer, so the number lags real usage by up to the cache window.

License

MIT — see LICENSE.

Русская версия: README.ru.md.