@estebanforge/pi-token-cost-ledger

Token & cost ledger for Pi. Captures every assistant message to a per-day JSONL ledger and exposes a /token-usage command with by-model and by-period breakdowns showing both real USD and API-equivalent USD.

Why two dollar figures

On a flat coding plan the marginal cost/token is $0 — useless for comparison. So alongside the provider's real cost, every report shows api-equiv: what the same tokens would cost pay-as-you-go on GLM's public API. That's the apples-to-apples axis vs Anthropic/OpenAI.

Install

pi install npm:@estebanforge/pi-token-cost-ledger

Usage

/token-usage                   opens a range menu (bare, in TUI)
/token-usage today
/token-usage day [YYYY-MM-DD]      today, or a specific day
/token-usage week [N]              current ISO week; N = weeks ago
/token-usage days [N]              rolling N-day window incl. today (default 30)
/token-usage month [YYYY-MM]       current month, or a specific one
/token-usage year [YYYY]           current year, or a specific one
/token-usage all                   full history
/token-usage model <name>          one model across all history, by month
/token-usage chart [period]        render a usage dashboard (SVG + optional PNG)

Run /token-usage with no argument (in the TUI) to pick a range from a menu: Today · Last 7 days · This month · Last 30 days · This year · Last 365 days · All. Arrow keys to move, Enter to select, Esc to cancel. In headless/RPC mode, bare /token-usage still defaults to today. Each menu item maps to the typed form shown above, so the menu doubles as a cheat sheet.

Each report prints:

• Total — real $ · api-equiv $ · total tokens · calls
• By model — per-model breakdown
• By period — day or month subtotals (when the range spans more than one)

Charts

/token-usage chart [period] renders a usage dashboard as a self-contained SVG to ~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/charts/usage-<start>_<end>.svg, then prints the path. A PNG is written alongside when rsvg-convert or inkscape is on PATH — no npm dependencies are added either way (SVG is pure string templating; PNG relies on a system binary you may already have).

/token-usage chart              last 30 days (default)
/token-usage chart today
/token-usage chart days 7
/token-usage chart month
/token-usage chart all

The dashboard adapts the single-vendor "usage details" convention to pi's multi-provider reality:

• KPI row — Total token consumption · top model consumption · top provider consumption (so the cross-provider split is visible at a glance).
• By model — one line per model, colored by usage rank; legend reads provider:model. The busiest model always takes the amber line.
• Total (all models combined) — a second chart with a single filled line summing every model, for the overall trend.

The period accepts anything the text command does (today, day [D], week [N], days [N], month [YYYY-MM], year [YYYY], all). SVG needs nothing; PNG is emitted only when a converter is found on PATH. These are optional system dependencies — the extension adds no npm packages for rendering.

Linux (Debian / Ubuntu):

sudo apt install librsvg2-bin      # rsvg-convert — fastest, headless-native
# or
sudo apt install inkscape

Fedora / RHEL:

sudo dnf install librsvg2          # provides rsvg-convert
# or
sudo dnf install inkscape

Arch:

sudo pacman -S librsvg             # provides rsvg-convert
# or
sudo pacman -S inkscape

macOS (Homebrew):

brew install librsvg               # provides rsvg-convert (recommended)
# or
brew install --cask inkscape

rsvg-convert (from librsvg) is preferred — small, fast, and built for headless conversion. inkscape works but is heavier. If neither is present, the SVG is still written and the notify notes how to enable PNG.

How it works

• Capture. A message_end hook appends {ts, provider, model, tokens, cost} to ~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/YYYY/MM/DD.jsonl. Format is unchanged from the legacy ledger, so existing files are valid with zero migration.

• Pricing. api-equiv is computed at query time from current prices (reprice semantics — history always reflects the latest price file). A default prices.json ships with the extension; ~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/prices.json overrides it if present. See Updating prices below.

• Multi-root. Reads ~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/roots.conf (one path/line, # comments, first = primary write target). Disjoint ledgers (e.g. a construct sandbox) are unioned read-only; no dedup needed. Honors PI_COST_LEDGER / PI_COST_LEDGERS env overrides.

• Cache convention. CACHE_CONV=separate (default) treats cacheRead as a subset of input; CACHE_CONV=included treats input as the full prompt. Matches the old scripts.

• Number format. Controlled by the token-cost-ledger-numbers flag or TOKEN_COST_LEDGER_NUMBERS env:

  • auto (default) — detect from the terminal locale (LC_ALL > LC_NUMERIC > LANG) via Intl.NumberFormat. Respects each user's locale.
  • comma — Latin/European: 1.148,23 (dot thousands, comma decimal)
  • dot — Anglo: 1,148.23 (comma thousands, dot decimal)

  Set with /token-cost-ledger (opens an interactive menu — cycle with Enter/Space, persists on close), /token-cost-ledger comma (one-shot shorthand), pi config set token-cost-ledger-numbers comma, or per-session with TOKEN_COST_LEDGER_NUMBERS=comma pi.

Commands

Command	Description
/token-usage	Open the range menu (Today / Last 7-30-365 days / This month / This year / All).
/token-usage <period>	Query usage directly (today/day/week/days/month/year/all/model). See Usage above.
/token-usage chart [period]	Render a usage dashboard (SVG + optional PNG) to <ledger>/charts/. Default: last 30 days.
/token-cost-ledger refresh	Pull latest costs from models.dev into the override (network).
/token-cost-ledger	Open the interactive options menu (number format / refresh prices).
/token-cost-ledger <auto|comma|dot>	Set number format directly (one-shot shorthand; persists + reloads).

Updating prices

The bundled extensions/prices.json holds canonical first-party rates per 1M tokens (input / cache-read / output), keyed by the exact model string pi logs. Source: models.dev — an open-source database of AI model specs and pricing.

Refresh from models.dev (in pi)

Run /token-cost-ledger refresh (or open /token-cost-ledger and set Refresh prices now → yes) to pull the live models.dev catalog and update the override file at ~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/prices.json. Load precedence means the next /token-usage reflects it immediately — no reload.

The refresh is update-only: it refreshes costs for models already tracked, and preserves meta keys, hand-curated extras (e.g. grok / deepseek / mimo from non-first-party providers), and per-model _tier_note annotations. It does not add new models — those arrive via bundled-file releases. A network failure changes nothing.

Manual edit / regenerate

To regenerate from scratch or add models, extract from the catalog with jq:

curl -sL https://models.dev/catalog.json -o /tmp/models-dev-catalog.json

# Extract canonical first-party pricing (zai, anthropic, minimax, google, openai)
jq -rc '.providers | to_entries[] | .key as $p
  | select($p|test("^(zai|anthropic|minimax|google|openai)$"))
  | .value.models // {} | to_entries[]
  | select(.value.cost != null)
  | {provider:$p, model:.key, cost:.value.cost, name:.value.name}' \
  /tmp/models-dev-catalog.json

Map catalog {input, output, cache_read} → this file's {i, o, c}. Keys MUST match the ledger's exact model string (case-sensitive — e.g. MiniMax-M3, not minimax-m3). Models with context-tier pricing use the default tier (<200K / <512K context); the over-tier rate is noted in _tier_note. After editing, no reload is needed — queries re-read the file each call.

models.dev alternatives: the catalog JSON (catalog.json / models.json / providers.json), the @opencode-ai/models npm SDK (typed snapshot for offline use), or TOML sources on GitHub.

Data layout

~/.pi/extensions-data/estebanforge/pi-token-cost-ledger/
├── roots.conf          # ledger roots (optional; auto-detects host + construct)
├── prices.json         # GLM prices override (optional; bundled default otherwise)
├── 2026/07/22.jsonl    # one record per assistant message, append-only
└── 2026/07/23.jsonl

License

MIT