dsh-balance-meter

English | 中文

DeepSeek account balance and session-cost readout for the DeepSeek Harness (DSH) Web GUI.

• Live account balance (queries the official Get User Balance endpoint)
• Current session estimated spend (token usage x official pricing)
• Per-model pricing: reads the model actually driving each session from its request header (flash vs pro), so the cost tracks the model you used instead of a fixed default
• Auto-fetches the official pricing page every 6h, so price changes and the 2026-08-17 peak/off-peak pricing rollout never require a plugin update
• Peak-hour band (Beijing 09:00-12:00 / 14:00-18:00) applied automatically once the peak pricing goes live

Features

The composer dock shows a chip with the account total balance and the current session's estimated cost:

Balance CNY 4.16 · This session CNY 2.57

Clicking the chip reveals the per-currency balance breakdown (granted + top-up) and the per-bucket cost breakdown (input / cache read / output). Clicking while an error is shown forces an immediate refresh.

Requirements

• DeepSeek Harness 0.1.0-rc.6 or newer (web profile)
• A DeepSeek API key stored through the DSH credentials seam (DEEPSEEK_API_KEY — the web Models page writes it)

Installation

From a git URL (no npm account needed):

dsh plugin --profile web add https://github.com/Ghost011118/dsh-balance-meter

Or from a local checkout:

git clone https://github.com/Ghost011118/dsh-balance-meter.git
dsh plugin --profile web add link:$(pwd)/dsh-balance-meter

Restart dsh web, then refresh the page. The balance chip appears in the composer dock next to the conversation stats line.

Configuration

The plugin is zero-config by default (uses DEEPSEEK_API_KEY and the official pricing page). Optional composition settings:

- insert:
    - id: balance
      name: 'dsh-balance-meter'
      config:
        model: auto         # 'auto' (default) | 'flash' | 'pro'
        pricingRefreshHours: 6

Key	Type	Default	Meaning
model	'auto' | 'flash' | 'pro'	auto	auto detects each session's model from its request header (flash/pro); flash/pro force that preset regardless of auto-detection
pricingRefreshHours	number	6	Hours between automatic official-pricing refreshes
apiKeyEnv	string	DEEPSEEK_API_KEY	Credential ref storing the DeepSeek API key
baseUrl	string	https://api.deepseek.com	API base URL (gateway/compat override)
refreshIntervalSeconds	number	30	Minimum seconds between balance queries

How the cost is estimated

The plugin reads DSH's durable tokenUsage projection (the same accounting the built-in stats line uses) and converts the four buckets — uncached input, cache read, cache write, output — to money using prices parsed from the official pricing page. Cache-write tokens are not billed separately by DeepSeek and default to 0.

For the price set, in auto mode (the default) it uses the model actually driving the session: each session's request header records the provider/model of the most recent request, and the plugin maps that id (deepseek-v4-flash → flash, deepseek-v4-pro → pro) to the matching per-million prices. A session is therefore priced at whatever model produced its usage, not a hard-coded flash. When no header exists yet or the model id is unrecognized, auto falls back to flash. Setting model: flash or model: pro explicitly forces that preset and ignores auto-detection, so you can pin the estimate to one model when you want to.

Before the 2026-08-17 peak-pricing rollout the current single prices stay authoritative; after it, the peak/off-peak band for the current Beijing hour is applied. If the pricing page cannot be fetched, built-in presets (flash: 0.02 / 1 / 2 CNY per 1M) are used. Explicit cost.* overrides in the composition config take precedence over any preset. The cost JSON also reports pricingKey and model so the chip can show which model was priced.

Troubleshooting

"no API key for provider route `deepseek-official`"

The host reads your key from the DSH credentials store — the file <harness home>/.credentials.yaml (default ~/.dsh/.credentials.yaml), the same store the web Models page writes. This plugin's balance query and the LLM route resolve through that same seam.

• If this error appears, make sure the document contains DEEPSEEK_API_KEY: sk-... (a strict mapping of reference to non-empty string). Editing it while DSH runs is fine — the provider hot-reloads and re-reads the file.
• When a credentials seam is mounted, both the LLM route and this plugin read the key only from the credential store; a plain export DEEPSEEK_API_KEY is ignored in that case. Exporting still helps for the plugin's own fallback when no seam is present.
• Prefer running dsh web through a single supervised instance (e.g. dsh-autostart) instead of launching several ad-hoc npx dsh web processes that can race on the same port and settle different credential snapshots. If you encounter this right after killing a manual instance, confirm the other (still-supervised) instance read the key — the balance chip recovering to a live total means the key resolved.
• The error is transient-friendly: the balance chip auto-recovers, because an error state is never cached as fresh — the next poll re-queries the provider.

Balance stuck on "unavailable" and only updates on click

An error/unavailable snapshot used to be served from the cache until it aged out, so a transient failure could hold the chip on "unavailable" until you clicked to force a refresh. Now an erroneous view is never reused as a fresh cache: every poll re-queries the provider, so the chip recovers on its own as soon as the underlying condition clears (balance reachable, network back, key stored).

License

BSD-3-Clause. Copyright (c) 2026, Ghost011118.