Bottom Info Bar

English | 中文

A drop-in replacement for the stats row under the DeepSeek Harness composer.

It keeps everything that row already showed — turns and steps, LLM time, tool calls, cache hit rate, input/output tokens — and adds what you actually want to see while you work: your real balance (or subscription quota, or this month's bill), the provider and exact model, peak/off-peak pricing with a countdown to the next switch, and what this conversation has cost so far.

Install once, restart once, then it activates on every launch. The billing mode is detected automatically, and every figure is either read from a provider API or explicitly labelled as an estimate.

💬 Want to suggest something, or just talk about AI? Scan the QR code to join our WeChat group DeepThinking. _{The code is generated by WeChat and lasts about 7 days — if it has expired, please open an issue and we will refresh it.}

_{The screenshot uses the Chinese UI; with DSH's language set to English the labels switch automatically. From top to bottom: ChatGPT subscription, DeepSeek balance, OpenCode Go subscription — each shown in full view followed by compact view.}

At a glance

• Three billing modes, auto-detected — balance, subscription quota, or cloud bill. They replace each other and never overlap.
• Real data, honestly labelled — balances, quotas, plans and bills come from official provider APIs. The one figure that cannot (OpenAI has no public balance API) is derived from your spending rate and marked (estimated) in the bar.
• Exactly what DSH shows — the provider and model match DSH's model switcher, and newly published models are picked up automatically.
• Peak / off-peak pricing — both prices plus a countdown to the next switch (weekends count as off-peak all day).
• Honest spend tracking — this conversation (including subagents), today, last 30 days, and all time — persisted to disk, nothing lost on restart.
• Native by design — it replaces the native row instead of duplicating it; click to switch full/compact, and choose which fields and colors appear.

Requirements

• DeepSeek Harness with the web interface (dsh web). This plugin is web-only — the bar is a web UI component.
• pnpm — used by dsh plugin to manage profile packages.
• An API key for whichever provider you use, set in DSH under Settings → Models.

Quick start

dsh plugin --profile web add dsh-bottom-info-bar

Then restart dsh web — plugins are composed when the host process starts, so a page refresh is not enough.

That is the whole setup: configure your provider's API key, restart, done.

git clone https://github.com/songoao25/dsh-bottom-info-bar.git
dsh plugin --profile web add /path/to/dsh-bottom-info-bar/plugin

git clone https://github.com/songoao25/dsh-bottom-info-bar.git
cd dsh-bottom-info-bar
./install.sh                # installs into the "web" profile; --profile <name> to override

What it does not do

The boundaries are part of the design, and they are worth stating plainly:

• It never passes a guess off as a fact. Every figure comes from a provider API — or, in the one case where no API exists (OpenAI has no public balance endpoint), is computed from your spending rate and labelled (estimated) in the bar. When a number is simply unavailable, the bar says so instead of filling the gap.
• It never updates itself. The version reminder only tells you a newer version exists and prepares the command; nothing on your machine changes until you run it.
• It never binds your ChatGPT account. That belongs to the companion plugin dsh-chatgpt-subscription. This bar only reads the local token.
• It never stores conversations. The ledger records tokens, model, provider and cost — no prompts, no messages, no API keys.
• It does not work outside the web interface. There is no terminal or headless equivalent of the bar.

Billing modes

The bar follows the active session in DSH and picks the display from the provider — no manual mode selection.

Balance mode (DeepSeek, Kimi, OpenRouter, StepFun, Xiaomi MiMo, OpenAI reference)

Shows the real balance from the provider's /user/balance (or equivalent) API. It refetches the moment the bar opens, refreshes, or the provider changes, then polls every 60 seconds. The last known snapshot is kept on failure so the row never goes blank.

When the balance drops below ¥20, the amount and a Low label turn red.

Subscription mode (ChatGPT / Codex, OpenCode Go, Zhipu, Xiaomi MiMo Token Plan, Command Code)

Shows quota remaining per window (5-hour / weekly / monthly, where remaining = 100 − used) and a countdown to the next reset. Quota and countdown always come from the same window, so they can never disagree.

• ChatGPT / Codex — decoded locally from ~/.codex/auth.json, showing the real plan tier and expiry date (for example ChatGPT · Plus | Expires 2026-09-16). Zero network calls: the values come straight from OpenAI's own login token, never estimated. Not signed in → Refresh failed with a reauthorization hint. Binding, token refresh and the openai-codex model route belong to the companion plugin dsh-chatgpt-subscription — this bar only reads the token, and never writes it back.
• OpenCode Go — reads quota from opencode.ai/zen/go/v1/usage using OPENCODE_GO_API_KEY (Settings → Models) or the opencode CLI login at ~/.local/share/opencode/auth.json. Missing key → a "not configured" hint, not an error.
• Zhipu — GLM Coding Plan quota via ZAI_CODING_CN_API_KEY (fallback ZAI_API_KEY): plan tier plus the 5-hour and weekly windows (including the credit-based plans introduced on 2026-07-30).
• Xiaomi MiMo Token Plan — monthly Credits quota via XIAOMI_TOKEN_PLAN_CN/SGP/AMS_API_KEY per region (fallback XIAOMI_API_KEY): plan name plus the monthly window.
• Command Code — reads the official CLI quota endpoints using COMMAND_CODE_API_KEY or CMD_API_KEY from Settings → Models, the same environment variables, or the CLI login at ~/.commandcode/auth.json: monthly credits plus 5-hour and weekly windows. Credits are displayed as credits, never as currency; unknown plan IDs do not get an invented monthly percentage.

In compact mode the bar prefers the shortest window (5-hour > weekly > monthly), because it refreshes fastest; if the 5-hour window is unavailable it falls back to weekly, then monthly.

Billing mode (Together, Fireworks, AWS Bedrock, Cloudflare)

Shows this month's real spend from the provider's official billing API, for example Together | This month $12.34 or AWS Bedrock | This month $45.60 · Budget 46%. Cloudflare also shows daily free-quota remaining and a UTC-midnight reset countdown — but only when the API actually reports a free allowance; otherwise it shows real usage alone rather than inventing a quota. Missing keys → a "not configured" hint.

Spend tracking

Every llm/stream request is recorded (usage × unit price) and aggregated four ways: this conversation (including subagents), today, last 30 days, and all time.

Subagents bill to the same provider account, so their records are folded into the current session. The billed price is fixed the moment a response completes, so later price-table updates never rewrite history. Models with no known price keep their token counts but are excluded from money totals, and are never given an invented cost.

Nothing is lost on restart: the journal is synchronously confirmed before the UI counts a new charge, and if that write fails the bar says Spend not saved instead of quietly dropping the amount. Ledger details auto-archive into summaries and stats are computed incrementally, so refreshing stays fast even after months of use.

Updating

Once per full DSH startup the plugin checks npm for a newer version. If one exists, a red Update available label appears.

Click the label and the update command is copied to your clipboard — paste it into a terminal, run it, then restart dsh web. The label briefly confirms with "Update command copied".

Hovering the label explains both paths: ask an agent that has terminal access, or click to copy the command yourself.

The copied command matches how the plugin was installed:

It is never applied automatically. The plugin only tells you a newer version exists; nothing on your machine changes until you run the command yourself.

Supported providers

The bar detects the provider from DSH's current model list — no configuration needed. Set the key up in DSH (Settings → Models) and the bar recognises it. If DSH has not supplied a model yet, the bar waits rather than guessing.

Balance-based

Subscription-based (quota windows)

Cloud-billing (real monthly bill)

Anything else shows a Not supported hint instead of borrowing another provider's numbers.

About the model name DeepSeek-V41-Flash

If you have seen V41-Flash and read it as a typo, here is the whole story:

• It is DeepSeek V4.1 Flash. That is the model actually serving your requests, priced at V4.1 Flash's official peak/off-peak rates.
• DeepSeek-V41-Flash is how DSH's own model catalogue spells it — model id deepseek-flash, display name written without the dot. The bar mirrors DSH's model switcher, so the odd spelling is DSH's, not this plugin's.
• It has no effect on anything you do. Same model, same pricing, same quota. The bar shows whatever DSH shows, so if DSH ever renames it, the bar follows automatically with no update needed.

DeepSeek-V4-Flash is a different, older model id — not the same model.

Interface language

The plugin follows Settings → General → Language in DSH. Switching languages updates the bar without a reload, and there is no separate plugin language selector. Host-side messages use DSH's saved preference; a remote browser's unsaved choice cannot change host-only text, and external provider messages keep their original wording. See localization notes.

Data storage

Everything the plugin records lives in its own directory, isolated from other plugins and from DSH configuration:

~/.dsh/dsh-bottom-info-bar/
├── usage-records.json           # the complete, human-readable ledger
├── usage-records.journal.jsonl  # recovery journal — do not edit by hand
└── usage-records.json.bak       # previous complete snapshot, for recovery

• Permissions — directory 0700, files 0600, hardened automatically at startup: readable only by your user.
• Relocate — set DSH_BOTTOM_INFO_BAR_DATA_DIR to move the whole directory (external drive, synced folder, …).
• Inspect or migrate — open usage-records.json in any editor. To move machines, copy the directory while DSH is closed.
• Contents — one entry per model response (id / ts / model / provider / sessionId / input / cacheRead / cacheWrite / output / currency / cost / status), where status is completed or interrupted. No conversation content, prompts, or API keys are ever stored.
• Retention — there is no silent entry cap; keep the directory in your normal backups if you need history beyond this machine.
• Manage — Settings → Info Bar → Billing data can export CSV/JSON or clear the plugin's records after confirmation. Settings, sign-in information and pricing data are untouched, and uninstalling does not delete your data.

Money is aggregated only in the active provider's currency (CNY for DeepSeek, USD for the OpenAI reference prices) — currencies are never mixed. DeepSeek peak hours are weekdays 09:00–12:00 and 14:00–18:00 Beijing time; weekends are off-peak all day.

Uninstall

cd dsh-bottom-info-bar
./uninstall.sh
# or: dsh plugin --profile web remove dsh-bottom-info-bar

After a restart the native stats row returns with no residue. Your usage records stay in ~/.dsh/dsh-bottom-info-bar/ — export or clear them from Settings → Info Bar → Billing data if you want a clean slate.

ChatGPT binding and token maintenance belong to the separate plugin dsh-chatgpt-subscription; removing this info bar does not touch it.

FAQ

Development

• Source — plugin/src/host.js (host) and plugin/src/client-bundle.js (client)
• Build — cd plugin && npm run build (generates lib/)
• Test — node tests/run-all.mjs (builds first, then runs every suite)
• Release history — CHANGELOG.md
• Contributing — CONTRIBUTING.md; day-to-day workflow and the release process: docs/WORKFLOW.md

Every rule this project learned the hard way is enforced by CI rather than by convention — the test suite, the source guards and the release-chain contract all run on every pull request. See AGENTS.md for the constraints an agent must respect.

License

MIT © 2026 songoao25