@yuyaxiong/dsh-usage-report

Token usage and cost dashboard for DeepSeek Harness, as a DSH bundle.

Every model call costs tokens, but a session log only records the conversation. This plugin adds a Settings → Usage page that shows where the money went:

• cost / request / token cards for a chosen range,
• an hourly bar chart for a single day, daily charts for longer ranges,
• per-model panels with request and token charts,
• a per-purpose split that separates conversation from auto review, session titles and compaction,
• a custom date range picker and CSV export,
• a manual price form for models outside the built-in table.

The purpose split

Auto review, session titles and compaction call the model through the same route as the conversation, but they never write an assistant/message into the session log. Their tokens are therefore invisible to any log-based accounting — including this plugin's own charts, before 0.2.0.

The plugin now installs an llm/stream listener (lib/tracker.js). It passes chunks through untouched and, when a call finishes, records the reported usage together with the call's purpose:

purpose	source
conversation	the agent loop; already durable as assistant/message.usage
auto-review	dsh-experimental-auto-review
session-title	dsh-session-title-llm
compaction	dsh-compaction-basic
other	any side call without a purpose

The conversation is not recorded twice: the tracker skips requests that isAgentLoopRequest() marks as loop-built, so the existing assistant/message record stays the single source for conversation cost.

Where the records go

1. Session log (preferred). When @deepseek-ai/dsh-session resolves, each side call becomes a plugin:dsh-usage/llm-call record. It is an ignorable event: it never enters the model-visible surface, and fork and resume carry it along.
2. Local ledger (fallback). When that module does not resolve — a profile is free not to install it — records are appended to $DSH_HOME/storages/usage/ledger.jsonl instead.

A call is written to exactly one place, so nothing is counted twice. When a session already has side-call records, its compaction/summary usage is ignored: the tracker already covered it.

Install

npm install -g @yuyaxiong/dsh-usage-report   # or let `dsh plugin` do it
dsh plugin --profile web add @yuyaxiong/dsh-usage-report

Then list the package in dsh.profile.bundles of your profile's package.json (a plugin-market index does both steps for you).

Restart the host after install. The usage page appears under Settings → Usage.

Release

npm version patch --no-git-tag-version   # patch / minor / major
git commit -am "0.2.1"
git tag v0.2.1
git push && git push --tags

The tag push runs .github/workflows/publish.yml, which checks the tag against package.json and publishes to npm. It needs a repository secret NPM_TOKEN: a granular access token with Bypass two-factor authentication, Read and write (publish and stage), scoped to @yuyaxiong.

Manual publishing still works: ./scripts/publish.sh.

Pricing

Costs are estimated, not billed. The built-in table follows DeepSeek's published prices, including peak/off-peak hours:

per 1M tokens	deepseek-flash off/peak	deepseek-v4-pro off/peak
input (cache hit)	$0.003 / $0.006	$0.022 / $0.044
input (cache miss)	$0.15 / $0.30	$0.66 / $1.32
output	$0.60 / $1.20	$1.98 / $3.96

Override anything in $DSH_HOME/storages/usage/prices.json:

{
  "usdCny": 7.1,
  "tiers": { "flash": { "hitPeak": 0.006, "hitOff": 0.003, "missPeak": 0.3, "missOff": 0.15, "outPeak": 1.2, "outOff": 0.6 } },
  "models": { "my-model": { "missPeak": 0.5, "outPeak": 2 } },
  "providers": { "my-gateway": { "models": { "gw-model": { "missPeak": 0.5, "outPeak": 2 } } } }
}

Resolution order: providers[provider].models[model] → providers[provider].default → models[model] → built-in tier. A model that resolves to none is unpriced: its tokens are counted, its cost is excluded, and the page asks you for a price.

Web search calls are billed separately ($0.75 / 1000 successful requests) and counted on their own.

Known limits

• Records start when the plugin first runs. Calls made before that are gone — the model did not write them anywhere.
• A stream that ends without a usage chunk (network abort, provider error) is not recorded. The purpose split therefore counts completed calls only.
• Session logs record provider and model but not base_url and not which credential paid, so two routes of the same adapter cannot be told apart after the fact. Give each route its own provider if you need them separate.
• The purpose split follows the date range. It ignores the model filter, because the tracker records one purpose per call, not a purpose × model grid.
• nodeLinker: hoisted profiles copy plugin files on install. After editing the source under ~/.dsh/profiles/<name>/node_modules/, run the install again.

License

MIT