dsh-plugin-price-aware
让你的编程代理了解成本：实时额度、按模型计价、任务预估费用和预算门槛。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:121212165/dsh-plugin-price-aware#50be80f14e7b37b3a849c3f6184804ba1292fb81Plugin Price Aware 以 dsh-plugin-price-aware 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Let your coding agent know what it costs.
dsh meters tokens well and money not at all. There is no price table, no /cost, no balance call, no spend cap anywhere in the harness — the tokenUsage projection counts uncachedInput / output / cacheRead / cacheWrite, and then the bill happens to whoever owns the account.
This plugin closes that gap:
MIT · TypeScript · zero runtime dependencies beyond @deepseek-ai/cordis.
dsh plugin add github:121212165/dsh-plugin-price-aware
or from a local checkout (dsh plugin add ./dsh-plugin-price-aware), or pin a commit:
dsh plugin add github:121212165/dsh-plugin-price-aware#main
Discovery for community plugins is the GitHub topic dsh-plugin. From a local checkout:
dsh plugin add ./dsh-plugin-price-aware
Then restart dsh. /money should print prices. The mount lives in cordis.patch.yml; every field is documented there and in Configuration.
injectIntoPrompt: true adds one dynamic context block per turn — 110 tokens, ¥0.0005 at peak pro rates, and a test fails if it grows past 260:
钱: deepseek-v4-pro 单价(每百万 token, CNY) 输入 4.5/9 · 缓存命中 0.15 · 输出 13.5/27【当前高峰，价×2】
钱: 大批量动作等 1h30m 后（错峰可省一半）
钱: 账户余额 ¥41.20 · 本会话已花 ¥3.84/¥10.00（模式 normal） · 34 轮 · 缓存命中 71%
纪律: 单步预计 ≥ ¥1.50 时，先给 A/B/C 三档让用户选档再开工；不重复读大文件，保持前缀稳定以命中缓存；能用 flash 解决就别用 pro。
If the model is not in the price table, the block says so and forbids itself from inventing a number. An estimate you cannot defend is worse than no estimate.
| Name | Who calls it | What it does |
|---|---|---|
/money | you | unit price, balance, session bill, remaining budget |
/budget [5] | you | read the policy; changing it prints the config edit instead of silently mutating your cap |
/estimate <kind> [turns] [residentK] | you | A/B/C quote for a task shape |
price_status | the model | same facts, callable mid-task before it commits to something expensive |
quote_task | the model | A/B/C quote for a task it is about to start |
An agent turn does not cost "the prompt you typed". It costs the whole history re-sent, times the number of turns:
promptTokens = turns × resident + growth × turns × (turns − 1) / 2
resident (system prompt + tool schemas + history) is what actually burns you: a 14-turn refactor with 20k resident tokens re-sends ~462k prompt tokens, and only the cache-discounted part of that is cheap. The estimator prices that shape, splits it into cached/uncached/output, and applies the model's current regime.
Default error band is ±35–60%. Above reconThresholdMajor (default ¥0.5) the plugin offers to spend a small read-only recon pass — which tightens the band to ±12% and costs under a quarter of the job it is pricing. Below that threshold, paying to estimate is a bad trade and it says so.
A quote is only an upper bound if the model stops talking. The live test found a whole-file generation that filled every cap it was given, 4320 → 6263 tokens, twice. quoteTrust() therefore labels each figure as a bound or a floor based on that model's observed selfFinishedRate, and escalateCap() turns the retry into a priced experiment instead of a fake estimate.
calibrate() learns from your own history: feed it {predicted, actual} pairs and it returns the median correction (median, not mean, so one 5× outlier does not bend future quotes).
price-aware:
mode: normal # economy ¥2 · normal ¥10 · max uncapped · custom below
sessionCapMajor: 10 # whole currency units; only read when mode: custom
warnPercent: 75 # this fraction of the cap triggers the ask
taskAskMajor: 1.5 # one step above this gets an A/B/C quote first
balanceFloorMajor: 3 # stop before draining the account below this
holidays: ['2026-10-01'] # Beijing dates billed off-peak
inputIncludesCache: false
prices:
- id: jiyuan/deepseek-v4-flash # provider-scoped rows beat generic ones
currency: CNY
perMillion: { cacheRead: 0.5, uncachedInput: 3, output: 12 }
Relay and reseller endpoints are first-class: your provider's price is not DeepSeek's price, so a prices row is how you say so. Until you do, the plugin refuses to quote in currency.
src/pricing, ledger, gate and estimate are pure. import { costOf, resolveModel, DEEPSEEK_CATALOG } from 'dsh-plugin-price-aware/pricing' and you have the same money math for your own harness, dashboard or CI budget check.
npm install # dsh types land in node_modules so the wiring typechecks for real
npm run check # typecheck + 63 tests + build
The plugin's apply() compiles against the actual @deepseek-ai/dsh-* declaration files, and lib/index.js imports clean. What is not verified is a live mount inside a running dsh session — that needs an installed harness (this machine's ~/.dsh profile symlinks currently point at a deleted clone).
5 rounds, real money, real gateway (TokenRhythm / 基元律动, ¥18 of credits), driven entirely through this plugin's own code:
| 结果 | |
|---|---|
| Spent | ¥0.406 / ¥18.00 across 22 billed calls |
| Price source | /v1/models publishes CNY rates → catalogFromProviderModels() ingests any self-reporting gateway |
| Official-sheet assumption would have been wrong | relay deepseek-v4-flash-0731 = ¥3/9/0.1 per 1M vs DeepSeek's own ¥1/4/0.02 — 3× under-billed if matched by name |
| Cache | cached_tokens was 0 in every call, including 3 identical 3.7k-token prefixes → this gateway eats no prefix discount; cacheableShare must be 0 here |
| Reasoning | up to 100% of completion_tokens invisible; several round-2 calls billed a full cap and returned 0 characters |
| Cost per useful output | best ¥0.0032/千可见字符, worst ¥9.53/千 — a ~3000× spread the token-price table alone cannot see |
| Deliverable | report/content.js, 10 287 chars, passes vm.Script, covers all 6 F1 sub-requirements |
| Final calibration | median bias ×0.858 with headroom reported against the quote, not the bill |
The headline is that prompt discipline is a cost control: the same model, same cap (6263) and same task filled the cap and truncated in round 4, then self-finished at 2904 tokens for ¥0.028 in round 5 after the system prompt forbade fences and prose. Reproduce with node examples/live-run.ts && node examples/live-run2.ts && node examples/live-run3.ts && node examples/live-run4.ts && node examples/live-run5.ts; every run writes to report/.
asOf in src/pricing/catalog.ts, DeepSeek's page as of 2026-09-21). After 45 days the block admits it may be stale. For gateways that publish rates on /v1/models, use catalogFromProviderModels() instead of the snapshot — that path is the one the live test exercised.peakMinutes changes at the rules layer.dsh is a developer preview with no ABI stability promise. @deepseek-ai/dsh-base is pinned for typechecking; expect churn between alpha tags.perMillion.cacheWrite set.tools/pre-execute; a hard per-request cap would need a model-call veto, which the harness reserves for its own loop.This plugin cannot make the harness refuse a request, and it will not pretend to. What it does is make the price visible early enough that the model can pick a cheaper plan, and you can pick a cheaper tier, before the tokens are gone.
MIT