LemCAE/dsh-balance

dsh-balance

A Host + Web Client composition plugin for deepseek-harness (dsh): queries the DeepSeek Open Platform account balance and estimates the current session’s spend. Installable via dsh plugin add.

Features

• Balance: queries the official GET https://api.deepseek.com/user/balance using the harness’s own DEEPSEEK_API_KEY credential (the key travels only over a bounded node subprocess’s stdin — never in command lines, logs, or UI).
• Session spend estimate: folds the provider-reported token usage from the session log (uncached input / cache-hit input / output) and prices each step with the official peak/off-peak table (peak 9:00–12:00 and 14:00–18:00 Beijing time). Estimate only — the official bill is authoritative.
• Top-bar chip (session header): 余额 ¥x | 会话 ≈¥y, click to refresh; hover (500 ms) shows a detail tooltip below the button, horizontally centered and viewport-clamped.
• Settings page (设置 → DeepSeek 余额): balance rows, an auto-refresh on/off switch, refresh-interval selector (15 s … 5 min or custom), UI-language selector (auto follows the host UI / 中文 / English), and an editable price table (off-peak / peak per model). Changes persist in the settings document across restarts.
• Model tool: deepseek_balance returns balance + the calling session’s estimated spend.
• Pause-aware refresh: after 2 refresh cycles without a new user or assistant message the auto-refresh pauses (5-minute detection cadence); it resumes when a new conversation appears. Auto-refresh can also be switched off manually (Settings switch or /dsh-balance auto-refresh off); while off, no queries are issued.
• Self-contained: no host-repository changes are required to deploy (communication rides the built-in commands Remote namespace).

Installation

Standard (bundle install, recommended):

dsh plugin --profile web add @lemcae/dsh-balance

The installer adds the package to the web profile’s dependencies and bundle list; after restarting dsh web, the loader applies the in-package cordis.patch.yml automatically. Verify:

• Open any session → the 余额 ¥x | 会话 ≈¥y chip appears in the header with a hover detail tooltip;
• 设置 → DeepSeek 余额 shows the full card (balance, interval, language, price table);
• Ask the model to call the deepseek_balance tool.

Manual install (same mechanism, bypassing the installer): edit $DSH_HOME/profiles/web/package.json — add "@lemcae/dsh-balance": "<latest version>" (as on npm) to dependencies and "@lemcae/dsh-balance" to the dsh.profile.bundles array — then run pnpm install in that directory and restart.

Peer dependencies are the official @deepseek-ai/* packages (^0.1.0-rc.5 line, covering rc.5 and rc.6; @deepseek-ai/cordis ^4.0.1) plus react, provided by the host.

Usage

• Chip: shows 余额 ¥x | 会话 ≈¥y; click to refresh; hover for details (breakdown, model, 更新于, refresh cadence / pause note).
• Settings: 设置 → DeepSeek 余额 — balance rows, 自动刷新间隔 (15 s … 5 min or custom), 自动刷新开关, 界面语言 (auto / 中文 / English), price-table editor (保存 persists), peak-hours hint.
• Command (also usable from the command palette): /dsh-balance [refresh | interval <毫秒> | prices <JSON> | language <auto|zh-CN|en> | auto-refresh <on|off>].
• Tool: deepseek_balance (no arguments).

Configuration

Settings namespace dsh-balance:

Prices are picked by Beijing hour: peak for 9:00–12:00 and 14:00–18:00, offPeak otherwise.

Known Limitations and Deferred Work

• Estimate vs. bill: the spend is computed from the local session log and may differ from the official bill (provider-side caching policy, unlogged requests, model renames, price changes). Edit the price table in Settings to keep it current.
• Unknown models are priced with the default entry (v4-flash rates).
• Subagents have their own session ids and are not included.
• Compaction: a compacted session resets event seqs; the incremental fold may keep pre-compaction totals (acceptable for an estimate).
• Session-log noise: every auto-refresh runs a slash command, appending command/run + command/done events; pause-downshifting reduces this.
• Pause recovery latency: while paused, the client probes once per PAUSED_REFRESH_MS (5 min); a new conversation resumes the active cadence at the next probe, so recovery can lag by up to that interval.
• Balance is cached 10 s; the tool and chip may share one API call per cycle.

Model Experience

Request context and condition

What the model sees

The tool schema deepseek_balance (zero parameters) plus its description, which states it queries the official balance endpoint with the harness credential and returns the session spend estimate.

Token effect

Zero direct tokens; the tool result is data-dependent (payload with balance, consumption, prices, idle state).

KV Cache effect

Prefix-stable: tool name, description, and schema are constant; the result varies per call, which does not invalidate prefix reuse.