dsh-token-audit

Find out where your DeepSeek Harness tokens went — and what to change first.

dsh-token-audit reads your local DSH session logs, prices every model call with the official DeepSeek peak/idle rates, and prints a ranked savings plan. It is a one-shot audit, not another dashboard: no plugin required, no config, no upload.

npx dsh-token-audit

DSH Token Audit
  /home/you/.dsh/sessions

  Window      2026-09-22 → 2026-10-31  (22 active days)
  Sessions    345  ·  59,823 model calls
  Total       CNY 778.78  ·  CNY 35.40/day  ·  CNY 1,061.98/month projected

  Cost split
    cache read      CNY 390.83  [##########........] 50.2%
    output          CNY 297.36  [########..........] 38.2%
    cache miss       CNY 90.59  [##................] 11.6%
    peak hours      CNY 281.49  [#######...........] 36.1%

  Context distribution
    0–50.0k          140 calls  [........................]
    ...
    500.0k–750.0k   2143 calls  [#######.................]

  Savings plan  (gross; combined overlaps, behavior unchanged)
    context cap 150k   save   CNY 210.20  27.0%
    context cap 250k   save   CNY 118.90  15.3%
    output      -20%   save    CNY 59.47   7.6%
    off-peak     40%   save    CNY 56.30   7.2%

  What to do first
    • 64% of calls ran above 200k context. Lower the compaction trigger ...
    • 36% of spend landed in peak hours (09-12, 14-18 Beijing) ...

A synthetic example report is in docs/demo.html; open it in a browser to see the full layout.

Why this exists

DSH is good at keeping a long conversation alive; that also makes it easy to spend far more than you expect. The official DSH surface shows context pressure and token counts, but not which calls cost the most, which hours were expensive, or what to change. This tool answers those three questions from the logs you already have.

It is deliberately not a dashboard:

• One shot. Run it monthly, or after an expensive week.
• Exact accounting. Uses provider-reported usage from assistant/message events, including cache reads, cache misses, and output.
• Official prices. DeepSeek deepseek-flash and deepseek-v4-pro, peak and idle, in CNY per million tokens (as of 2026-10-31).
• Actionable. Ranks context caps, output discipline, off-peak scheduling, and tool-output bloat by the money they would save.

Install

# zero-install
npx dsh-token-audit

# or globally
npm i -g dsh-token-audit
dsh-token-audit --days 7

Node.js 20+. If zstdcat is on PATH it is used to decode logs quickly; otherwise a pure-JS decoder is bundled. On Windows, the pure-JS path is used.

Install as a DSH plugin

The npm package is also a DSH bundle. It adds one read-only agent tool, token_audit:

dsh plugin --profile web add dsh-token-audit

Then ask the agent: "Audit my token spend for the last 30 days and tell me what to change." The tool returns the same report. It never sends session data anywhere.

Usage

dsh-token-audit                              # last 30 days, terminal
dsh-token-audit --days 7                     # last week
dsh-token-audit --since 2026-10-01 --until 2026-10-31   # calendar month (Beijing dates)
dsh-token-audit --all                        # every session on disk
dsh-token-audit --cwd Miner                  # one workspace
dsh-token-audit --session 4160cf07           # drill into one expensive session
dsh-token-audit --format markdown --out report.md
dsh-token-audit --format html --out report.html --open
dsh-token-audit --format json > audit.json

Option	Meaning
--sessions <dir>	Session root. Default $DSH_HOME/sessions or ~/.dsh/sessions.
--days <n>	Only sessions active in the last N days. Default 30.
--since <date>	Only calls on/after this Beijing date (YYYY-MM-DD); implies --all.
--until <date>	Only calls on/before this Beijing date (YYYY-MM-DD); implies --all.
--all	Ignore the time window.
--cwd <text>	Only sessions whose working directory contains the text.
--session <id>	Only sessions whose id contains the text; drill into one expensive session.
--format <fmt>	terminal (default), markdown, html, json.
--out <file>	Write the report to a file. Format is inferred from the extension.
--open	Open an HTML report in the default browser.
--redact	Hide session titles, paths, commands, and file names; safe to share.
--top <n>	Rows in the terminal session table. Default 8.
--no-color	Disable ANSI colors.

Verify a change (before / after)

Save a baseline JSON report, make one change, wait a comparable window, then compare:

dsh-token-audit --format json > before.json
# ... change compaction / effort / scheduling for a week ...
dsh-token-audit --format json > after.json
dsh-token-audit compare before.json after.json

The comparison leads with cost per call, not just total cost: a cheaper week with half the work is not a win. It also shows context, peak share, cache/output composition, and a verdict.

Generate a Lean preset

The audit tells you the shape of the problem; the preset applies a first fix:

dsh-token-audit preset --out ./dsh-lean-preset
dsh plugin --profile web add ./dsh-lean-preset

It keeps the standard toolset, moves the compaction trigger from 80% to 45%, shrinks the retained tail, prunes tool results at 4k chars, and adds a cost-discipline persona. Tune with --threshold 0.5 --retain 0.08 --prune 6144. Choose Lean for a new session; existing sessions keep their preset.

What the numbers mean

Every assistant/message in a DSH session log carries provider-reported usage:

• cacheReadTokens — input served from the prompt cache (cheap)
• inputTokens — uncached input (expensive)
• outputTokens — generated tokens, including reasoning (most expensive)

Cost is computed per call, using the call's Beijing timestamp to pick peak or idle pricing:

Model	Cache hit	Cache miss	Output
deepseek-flash idle	0.02	1.00	4.00
deepseek-flash peak	0.04	2.00	8.00
deepseek-v4-pro idle	0.15	4.50	13.50
deepseek-v4-pro peak	0.30	9.00	27.00

Peak = Beijing weekdays 09:00–12:00 and 14:00–18:00, excluding China statutory holidays. Unknown models fall back to Flash rates and are flagged in the report.

Savings estimates are gross, not promises

The Savings plan assumes behavior stays the same. Capping context at 250k means older history is compacted earlier; the model may lose detail and take more steps. Treat each row as an upper bound, then verify on a real task. The report labels the combined row as overlapping for the same reason.

The most reliable levers we have measured on real logs:

1. Context size. Cache-read cost is proportional to context length on every call. A 250k cap is often 15–25% of the bill.
2. Output discipline. On reasoningEffort: max, output can be 35–40% of cost. Use high for routine work.
3. Off-peak scheduling. Idle prices are half. Move batch jobs after 18:00 or before 09:00 Beijing time.
4. Tool output. bash is usually the largest source of context bloat; tail/head/quiet flags and redirecting logs to files are free.

Privacy

Everything runs locally. The tool reads ~/.dsh/sessions/**, computes in memory, and writes only the report you ask for. There is no network call, no telemetry, and no upload. The plugin tool is read-only.

Use --redact before sharing a report with anyone: it removes session titles, working directories, bash commands, and file names while keeping every number.

Limitations

• This is an estimate, not your invoice. Subscription plans, credits, refunds, and provider discounts are not visible in session logs.
• Peak/idle holidays are hardcoded for China 2026. Weekends are always idle.
• Third-party providers are priced at Flash rates until a catalog is added.
• Session logs must be readable (.jsonl or .jsonl.zstd).
• Context caps are a trade-off; quality is not modeled.

Development

git clone <repo>
cd dsh-token-audit
npm install
npm test
node bin/dsh-token-audit.mjs --days 3

The audit engine (src/audit.mjs) is independent of the CLI and can be imported. Pricing lives in src/pricing.mjs; renderers in src/report.mjs.

Need this applied to your setup?

The audit is free. If you want the configuration work done for you — compaction thresholds, pruner settings, model routing, off-peak scheduling, and a before/after verification — see docs/service.md. Fixed scope, fixed price, pay only after the report.

License

MIT. Pricing data belongs to DeepSeek; this project is unofficial.