pi-deepseek-suite

English | 中文

DeepSeek cost accounting for pi, in one extension:

1. Peak/off-peak re-pricing — every DeepSeek assistant message is charged at the official rate in effect at that message's own UTC timestamp, so pi's session totals, footer, and exports match what DeepSeek actually bills.
2. Local ledger — one JSONL record per priced call, rolled up by day, model, and tier. Plain files you can query with jq.
3. Account balance — the official balance endpoint, with a burn-rate runway and an optional low-balance warning.

No API calls beyond the balance lookup. No telemetry. Nothing about your conversation leaves the machine.

pi install npm:pi-deepseek-suite

Then /reload (or restart pi).

---

Why one extension instead of three

message_end handlers chain: a handler sees the replacement returned by an earlier handler, and the final replacement is what pi persists. If re-pricing and recording live in separate plugins, the ledger's correctness depends on their order in settings.json — put the recorder first and it logs the uncorrected cost forever, silently.

Here the corrected cost is computed once and used by both consumers, so there is no ordering to get wrong.

Why pricing matches on model id, not provider name

Several DeepSeek plugins key everything off provider === "deepseek". Provider names are local configuration: relay and router tools (cc-switch, new-api, one-api, …) generate models.json, pick their own provider name, and keep the key inline. On such a setup a provider-name check silently disables the whole plugin — the footer just never appears, with no error.

This extension matches on model id, and discovers the API key from DEEPSEEK_API_KEY, then auth.json, then any models.json provider whose baseUrl points at api.deepseek.com. So it works whether your provider is called deepseek, cc-switch-deep-seek, or anything else.

Commands

Command	What it shows
/ds [Nd]	Cost report from the ledger: total, calls, tokens, peak vs off-peak split, per-day and per-model breakdown. Defaults to today; /ds 7d, /ds 30d.
/ds-balance	Account balance (CNY-aware), how usable the account is, average daily burn, and runway at that rate.
/ds-tier	Current tier, the exact rates in effect for the current model, and the next switch.

The status line stays compact:

off-peak · $0.0340 today · ¥42.50

How pricing works

Source of truth: https://api-docs.deepseek.com/quick_start/pricing

• Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday. Everything else, weekends included, is off-peak.
• Off-peak is exactly half of peak, so the built-in table stores the off-peak column only and multiplies during peak.
• Cache writes are free for DeepSeek.

Off-peak rates (USD per 1M tokens) shipped in the built-in table:

Model id	input	output	cacheRead	cacheWrite
deepseek-flash	0.15	0.60	0.003	0
deepseek-v4-flash	0.15	0.60	0.003	0
deepseek-v4-flash-vision-exp	0.15	0.60	0.003	0
deepseek-v4-pro	0.66	1.98	0.022	0

deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired ids that DeepSeek still accepts and bills at the current Flash price.

deepseek-v4-pro keeps its own rates. DeepSeek announced a 2026-09-14 retirement that would have re-routed Pro to Flash pricing and then reversed that decision — so no cutover is encoded. A plugin that still assumes the reversal never happened will under-report Pro usage by roughly 3–4× from that date onward.

Updating prices without waiting for a release

Set rates in the config file. It replaces the built-in table entirely, so include every model you need:

{
  "rates": {
    "deepseek-flash": { "input": 0.15, "output": 0.6, "cacheRead": 0.003, "cacheWrite": 0 },
    "some-new-model": { "input": 1, "output": 2, "cacheRead": 0.1, "cacheWrite": 0 }
  }
}

Rates are always the off-peak figures; peak is derived by multiplication.

Configuration

Global: ~/.pi/agent/deepseek-suite.json Project: <cwd>/.pi/deepseek-suite.json (honoured only for a trusted project)

Project values override global ones. Anything malformed is ignored rather than fatal. Both a flat object and a { "deepseekSuite": { ... } } wrapper are accepted.

Key	Default	Meaning
peakMultiplier	2	Multiplier applied during peak hours.
rates	null	Replaces the built-in table. Off-peak USD per 1M tokens.
reprice	true	Write the corrected cost back into the session.
ledger	true	Append one JSONL record per priced message.
status	true	Show the status-line entry.
currency	null	Prefer this currency when the account reports several.
lowBalanceThreshold	null	Warn below this balance. Re-arms after a top-up.
balanceRefreshSeconds	300	Balance poll interval.

Files

~/.pi/agent/deepseek-suite.json                    config (global)
~/.pi/agent/deepseek-suite/ledger/YYYY/MM/DD.jsonl  one line per priced call

A ledger record:

{
  "ts": 1789219121803,
  "provider": "cc-switch-deep-seek",
  "model": "deepseek-flash",
  "tier": "offPeak",
  "tokens": { "input": 134, "output": 2, "cacheRead": 4351, "cacheWrite": 0 },
  "cost": { "input": 0.0000201, "output": 0.0000012, "cacheRead": 0.000013053, "cacheWrite": 0, "total": 0.000034353 },
  "ratesSource": "official"
}

provider is recorded exactly as pi reports it — never rewritten. Pricing keys off model. tier is the tier that was in effect for that call, which is what makes "how much would this have cost off-peak" answerable later.

jq recipes:

# Today's total
jq -s 'map(.cost.total) | add' ~/.pi/agent/deepseek-suite/ledger/2026/09/12.jsonl

# Spend by tier this month
jq -s 'group_by(.tier) | map({tier: .[0].tier, total: map(.cost.total) | add})' \
  ~/.pi/agent/deepseek-suite/ledger/2026/09/*.jsonl

Scope and limitations

• Only assistant messages are priced. Tool-result and compaction usage has no model of its own, so it is left to pi.
• Models outside the table are left completely alone: pi's own cost is preserved rather than replaced with a guess.
• Day boundaries in the ledger use local time; tier boundaries use UTC, because DeepSeek publishes them in UTC.
• The ledger is append-only and never rewritten. Changing rates does not retroactively re-price history.
• A balance lookup is the only network call, and only when a balance surface asks for it (session_start, model_select, the poll timer, or /ds-balance).

Requirements

• pi @earendil-works/pi-coding-agent (any recent version; no hard floor)
• Node 20+ (uses AbortSignal.timeout)

License

MIT