dsh-peak-pricing-warning

A DeepSeek Harness plugin that shows, right beside the model selector, whether DeepSeek is on peak or off-peak pricing, how long until that changes, and what the session has cost so far.

● Off-peak · peak in 8h 36m · $0.0795
● Peak pricing · off-peak in 2h 30m · $0.1240
● Off-peak weekend · peak in 1d 22h · $0.0310

DeepSeek bills off-peak requests at half the peak rate, so the two facts belong together: when you are being billed, and how much you have spent. Plain inline text — no pill, no border, no background — so it reads as part of the composer row rather than as a widget bolted onto it.

It appears only when the session's selected model is a DeepSeek model. Switch provider and it disappears.

Dot	Meaning
🔴 Red	Peak pricing — full rate
🟡 Yellow	Weekday off-peak — half rate
🟢 Green	Weekend — off-peak all day

Hover for the rate class, the per-bucket token split, and the cost caveats.

The pricing rule

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Every other hour is off-peak, which makes weekends off-peak all day.

Source: DeepSeek Models & Pricing.

Note: older documents and blog posts cite a single 16:30–00:30 UTC off-peak window. DeepSeek has since split peak into the two weekday blocks above. This plugin follows the current published schedule, and tests all five boundary shapes — both peak blocks, the gap between them, both weekend edges, and the return to peak after a weekend.

Cost

The cost figure is priced from the tokenUsage session projection, which already exposes DeepSeek's four disjoint billing buckets over the whole durable session log:

Bucket	Rate vs. uncached input
cacheReadTokens	~50x cheaper — the reason to keep a cache warm
cacheWriteTokens	billed as ordinary input
uncachedInputTokens	full input rate
outputTokens	full output rate (reasoning already included)

Because those counts are provider-reported, the plugin does not count tokens — it only applies a rate table. The same projection backs the shipped usage donut.

Not to be confused with ctx.tokenMeter.measure(). TokenMeasurement is a context-pressure estimate for occupancy display, documented as not a billing input. This plugin reads tokenUsage.

Rates

USD per 1M tokens, from DeepSeek's published pricing:

Model	Cache read	Uncached input	Output
deepseek-flash	$0.006	$0.30	$1.20
deepseek-v4-pro	$0.044	$1.32	$3.96

These are peak rates; off-peak is derived as exactly half. The tests assert that relationship, so a pricing change that breaks it fails loudly instead of silently doubling every figure. Two documented simplifications:

• Cache writes are billed as uncached input — DeepSeek publishes a cache read discount but no separate write premium. One field to change if that stops being true.
• Unknown models fall back to Flash and say so in the hover text rather than inventing a rate.

What it is not

An estimate, not a bill. It prices four buckets at list rates for the currently selected model, and cannot see account credits or negotiated discounts, retries billed on another route, or a mid-session model switch that leaves usage split across two price levels. The hover text says so rather than presenting the number as authoritative. The Harness ships no rate table of its own, so this one is the plugin's and can lag a price change.

Install

Not yet published to npm. npm view @imnotndesh/dsh-peak-pricing-warning returns 404, so any add command naming the bare package will fail at the registry. Install from the git URL or a local checkout below.

Note the CLI package is @deepseek-ai/dsh, not dsh — npx dsh … fails with "could not determine executable to run" because no package by that name exists. The binary it installs is called dsh, so a global install gives you a bare dsh command.

From GitHub

npx @deepseek-ai/dsh plugin add github:Imnotndesh/dsh-peak-pricing-warning

Or, with dsh already installed globally:

dsh plugin add github:Imnotndesh/dsh-peak-pricing-warning

Then restart the Host and reload the web GUI.

From a local checkout

git clone https://github.com/Imnotndesh/dsh-peak-pricing-warning.git
dsh plugin add ./dsh-peak-pricing-warning

dsh plugin forwards to a package manager inside the profile directory, so whichever manager you use must be on PATH: the CLI hardcodes pnpm and reports "pnpm not found on PATH" otherwise. Bun also works if you add the profile dependency by hand — bun add file:/path/to/checkout in $DSH_HOME/profiles/web — then add the package name to dsh.profile.bundles in that profile's package.json. That is exactly what the CLI's own reconcilePlugins pass does after a successful install.

Without installing

The plugin also runs as a dynamic Cordis plugin. plugins/peak-pricing-warning/src/plugin.js exports two functions; pass them to cordis_define as code.host and code.client, then cordis_run.

How it works

Host half owns the clock. It computes the pricing window from the authoritative Host Date and answers one Package-private RPC method (peak-status), returning only scalars. The subtlety is which boundary to report: during off-peak the countdown targets the upcoming peak, whose start is found by scanning backward to the opening edge of the run in progress; during peak it scans forward. Scanning the wrong way collapses the countdown to 0m 00s for the entire off-peak stretch.

Client half owns the pixels, reads the model selection and the usage projection, and ticks a local countdown every second, re-polling the Host once a minute for drift and day rollovers.

Three bugs worth documenting

All are easy to reintroduce, so each is called out in the source.

1. React #185 — infinite render loop. useSyncExternalStore compares snapshots by reference. Returning a freshly built object from getSnapshot, or mutating and returning a new object each tick, re-renders forever. Fixed with one cached status object reused until a displayed field changes, and a primitive "provider|model" string instead of an object.

2. React #310 — "rendered more hooks than during the previous render". useProjection was called after two early return null guards. A hook after a conditional return runs on some renders and not others, so the hook count changed. Every hook must run unconditionally before any return. This is the one to watch when adding to this component.

3. A silent catch turned a crash into a hang. A missing helper once made every Host call throw; the catch logged to console and left the status null, so the badge sat on "Checking pricing…" forever with no error and no retry. The poll is now refcounted and self-healing, re-fetches on subscribe, and renders an explicit Pricing status unavailable instead of an endless placeholder.

Verifying it

node test/peak.test.mjs   # 8 schedule boundaries
node test/cost.test.mjs   # 12 published rates + invariants

No Harness runtime needed. peak.test.mjs covers every boundary shape and asserts the countdown target is always in the future. cost.test.mjs asserts the rate table reproduces every published DeepSeek figure exactly, that peak is exactly 2x off-peak, and that cache reads stay >10x cheaper than uncached input — the premise the cost figure depends on.

Layout

plugins/peak-pricing-warning/src/plugin.js   single source of truth (both halves)
lib/                                          generated halves the dsh.bundle path loads
cordis.patch.yml                             composition row `dsh plugin add` applies
test/                                        boundary + rate tests

lib/ is generated from the plugin source by node build.mjs, so the dynamic-plugin copy and the installed package cannot drift.

Compatibility

Built against DeepSeek Harness 0.1.5-rc.2. Uses one additive UI slot (conversation.input.left), one Package-private RPC method, and no services of its own — it replaces no shipped UI.

License

MIT