Cacheback

DeepSeek's prefix cache has been paying you back. Cacheback prints the receipt.

Cacheback is a DeepSeek Harness plugin that turns the session logs already on your disk into a line-by-line cache receipt. It replays every model call, prices it at DeepSeek's published peak/off-peak rates, and answers one question:

How much did DeepSeek's prefix cache actually save me?

On the machine this was built on: 1.57B cached tokens, a 99.66% cache hit rate, and ≈¥2,386 saved — the same prompts would have cost about 30× more at the cache-miss price. Project names are redacted in the screenshot; your receipt uses your own workspace names.

English | 中文

The 30-second version

A cache hit on deepseek-flash costs $0.003 per 1M tokens off-peak. A cache miss costs $0.15 — 50× more. DSH already records both numbers for every call, but nothing puts them on one page. Cacheback does:

Cacheback · DeepSeek cache receipt
2026-08-14 → 2026-11-03 · 11 sessions · 5 projects · 4,226 calls

DeepSeek prefix caching saved you ≈¥2,385.88 ($331.37).
Without cache $342.78 → you paid $11.41 · 30× cheaper · 99.66% hit rate

Spend: cache read $6.76 · cache miss $1.29 · output $3.35
Tokens: 1.57B hit · 5.3M miss · 3.64M out

Projects
  Project A            $185.55  ·  2,186 calls  ·  99.66%
  Project B             $70.53  ·  1,041 calls  ·  99.69%
  Project C             $68.86  ·    780 calls  ·  99.67%

What to improve
  · Peak premium: 2,119 calls ran at peak. Off-peak is half price.
  · 4 cold starts cost $0.14. Resume a session instead of starting a new one.
  · Output is 29.4% of the bill. Shorter answers cut the biggest line.
  · Keep the tool prefix stable: plugin changes rewrite the cached prefix.

Cache score 99.7 · S

Then make it shareable:

/cache receipt   # self-contained HTML receipt, print to PDF
/cache card      # 1200×630 SVG card for X / GitHub / Slack
/cache 30d       # last 30 days only
/cache redact    # hide project names and paths before sharing

Install

dsh plugin --profile web add github:yunhenk/dsh-cacheback

Or install from a local checkout while developing:

dsh plugin --profile web add link:/absolute/path/to/dsh-cacheback

The plugin has zero runtime dependencies, makes zero network requests, and reads only the session logs DSH already writes. A restart (or HMR reload) makes /cache available.

Or install from Discover. Add this index source under Plugins → Discover → Index sources:

https://raw.githubusercontent.com/yunhenk/dsh-cacheback/53d96752f562e5e5280e07733dcb2ee9d7869986/market/index.json

What the receipt measures

Figure	Meaning
Saved	What the same calls would have cost at the cache-miss price, minus what you actually paid.
Cheaper	No-cache cost ÷ actual cost. On a warm session this is usually 20–50×.
Cache hit rate	Cache-read tokens ÷ (cache-read + cache-miss) tokens.
Peak premium	Extra dollars paid because calls landed in DeepSeek's peak window (09:00–12:00 / 14:00–18:00 Beijing weekdays). Off-peak is half price.
Cold starts	Calls that began with an empty cache and paid the full miss price for the prefix.
Spend split	Cache read vs cache miss vs output. Output is often the surprise.
Cache score	Hit-rate score with a letter grade, plus the four tips with the largest impact.

There is no sampling and no estimation of token counts: every number comes from the provider usage DSH persisted on each assistant/message. Pricing uses the official table (see docs/method.md), with the Chinese public-holiday calendar applied to the peak window.

Privacy

• Local only. Cacheback reads ~/.dsh/sessions and writes receipts to ~/.dsh/cacheback. It never opens a socket.
• No telemetry. No analytics, no “anonymous usage”, no remote font or image.
• Share safely. The SVG card contains aggregates only — no project names, session titles, or paths. /cache redact also replaces project names in the HTML receipt with Project A/B/C….
• Delete anytime. Remove ~/.dsh/cacheback and the plugin's profile entry. Session logs belong to DSH.

How it works

1. Discover the newest session*.jsonl(.zstd) under every ~/.dsh/sessions/<project>/<session>/.
2. Decode the append-only log. DSH writes one Zstandard frame per batch; Node's decoder stops after the first, so Cacheback splits on the frame magic and decodes every frame independently (torn trailing frames are skipped).
3. Replay each assistant/message: provider usage (inputTokens, cacheReadTokens, cacheWriteTokens, outputTokens), model id, and timestamp. Seeded/forked history is skipped so calls are never billed twice.
4. Price each call at the rate for its model and its Beijing-time window. Cache writes are billed as misses.
5. Aggregate by project, model, and day; compute the counterfactual, the score, and actionable tips.
6. Render the chat summary, a self-contained HTML receipt, or a 1200×630 SVG card.

Parsed sessions are cached in ~/.dsh/cacheback/index.json, keyed by file size and mtime, so a repeat run only re-reads sessions that changed.

Accuracy and limits

• Savings are a counterfactual. “Saved” assumes the same prompt would otherwise have been sent without a cache hit. That is the standard way to quantify prefix-cache value; it is not a refund.
• CNY is approximate. DeepSeek bills in USD; the receipt shows ≈¥ at a configurable rate (default 7.2).
• Peak calendar is 2026. The official 2026 State Council holiday list is included. Other years fall back to weekdays only; override with the plugin config if you need a different calendar.
• Only DeepSeek calls are priced. Calls from another provider are counted and reported as excluded, because their cache pricing is different.
• Logs must still exist. If you delete session history, its receipt line disappears with it.
• Not a bill. Provider dashboards and invoices remain the source of truth; this is a local, auditable receipt.

Configuration

The plugin accepts optional config through the profile patch:

- id: cacheback
  config:
    usdToCny: 7.2
    lang: zh          # zh | en, default for the chat summary
    open: false       # open generated receipts in the browser by default

Key	Default	Meaning
usdToCny	7.2	Display conversion for the approximate CNY figure.
lang	zh	Default language for the chat summary; the HTML receipt and card are bilingual.
open	false	Open generated files in the default browser. /cache receipt opens explicitly.

Development

git clone https://github.com/yunhenk/dsh-cacheback
cd dsh-cacheback
node --test test/*.test.js

The test suite covers the pricing calendar, cache-write accounting, multi-frame Zstandard decoding, seeded-history deduplication, redaction/escaping, command parsing, and the HTML/SVG renderers. No model calls, no network.

See docs/method.md for the exact formulas and event schema.

For the DSH team

If you run DSH support, growth, or community: Cacheback is a local, privacy-preserving answer to “is the cache working, and why is my bill high?” It needs no telemetry, no server, and no account. Point a user at /cache; the share card tells an accurate, positive cache-value story with numbers from their own machine.

License

MIT. Built for the DSH plugin ecosystem.