dsh-statistics

A DeepSeek Harness bundle that reads every conversation log in the current profile and reports what those conversations actually cost. It adds a Statistics panel to the sidebar, beside Plugins and Schedules.

The panel	Sessions

What the panel shows

• Total tokens with the input / output / cache-read split, conversation count, turns, steps, and tool calls.
• Estimated cost, priced per conversation and per day from a configurable rate card, plus a budget card with the spend of the current period, a progress bar, and a projection to the end of the period.
• Token usage timeline — one stacked column per day, switchable between all tokens and input + output only, with a selectable day detail line.
• Models — token volume and cost per provider/model, four rows in view and the rest scrolling, flagging models that fell back to the catch-all rate.
• Conversations — sortable cards for every conversation with its tokens, estimated cost, and attributed CPU time, newest activity first by default, four rows in view and the rest scrolling.
• CPU activity — the host process utilization history (last few minutes), system utilization, load average, core count, and memory, next to live per-conversation CPU attribution.

Ranges: last 7, last 30, last 90 days, or everything (the default). The header reports the data source and the number of logs scanned.

Where the numbers come from

Token usage is folded from committed assistant/message and compaction/summary events of every stored conversation, read through sessionQuery (live-preferred) or sessionPersistence. Nothing is sampled or estimated on the token side: the counts are the provider's own usage records, attributed to the model named on each message.

Cost is an estimate: tokens are multiplied by the rate card below. The budget card compares that estimate with budgetUsd for the selected period.

A conversation's last active time is the last turn, step, message, tool call, or compaction event — bookkeeping events (session/end-seed, permission/preset, sandbox/mode, workspace/changes) are written whenever a Session is opened again and must not move that date.

CPU is measured in this process: process.cpuUsage() deltas and os.cpus() idle deltas sampled every cpuIntervalMs. Each interval's CPU time is split across the conversations that were running or that produced an event within the last 10 seconds; time with nothing active is reported as idle. Attribution starts when the plugin loads, so the per-conversation CPU column shows — for conversations it never saw running.

Configuration

Set these on the statistics row of the bundle patch (~/.dsh/profiles/desktop/cordis.patch.yml) to override the defaults.

Key	Default	Meaning
rates	see cordis.patch.yml	Rate rules matched against provider/model and model, * wildcards allowed. Values are USD per 1,000,000 tokens.
fallbackRate	0.27 / 1.10 / 0.07	Applied to models no rule matches.
budgetUsd	20	Budget for the current period.
budgetPeriod	month	day, week, month, or all.
cpuIntervalMs	2000	CPU sampling interval.
cpuHistorySize	120	Samples retained for the sparkline.
refreshMs	4000	Panel polling interval.
statsTtlMs	15000	Minimum age before the log scan is repeated.
includeSubagents	true	Count subagent conversations.
maxSessions	500	Upper bound on logs read per scan.

HTTP surface

One route, inside Connection's authentication fence:

GET /api/statistics/snapshot

It returns the whole panel payload — totals, budget, day series, models slot, conversation rows with per-day buckets, and the CPU sample window — so the browser never folds session events itself.

Files

File	Role
index.js	Host route, session source, scan cache, CPU sampler wiring.
lib/fold.js	Pure event folding, pricing, and aggregation.
lib/cpu.js	CPU sampler with injectable probes.
client.js	Sidebar icon and panel, plain JavaScript for the module loader.
cordis.patch.yml	Inserts the statistics row with the default config.
locale/{en,zh}.json	Display metadata read by Plugin Manager.