Module token_usage 

Token/cost accounting for the usage subject.

pricing is the LiteLLM-sourced per-model USD table used to turn JSONL token counts into a notional cost. This root holds the project-folding that maps a ~/.claude/projects/<slug> directory name back to the repo it belongs to.

Re-exports

use std::collections::BTreeMap; 🔒

use std::collections::HashSet; 🔒

use std::io::BufRead; 🔒

use std::path::Path; 🔒

use std::path::PathBuf; 🔒

use std::time::SystemTime; 🔒

use forge_primitives::token_usage::UsageReport; 🔒

use forge_primitives::token_usage::UsageRow; 🔒

use forge_primitives::token_usage::WindowUsage; 🔒

use serde::Deserialize; 🔒

use serde::Deserialize; 🔒

use serde::Serialize; 🔒

use serde::Serialize; 🔒

use time::format_description::well_known::Rfc3339; 🔒

use time::Date; 🔒

use time::Duration; 🔒

use time::Month; 🔒

use time::OffsetDateTime; 🔒

use time_tz::OffsetDateTimeExt; 🔒

use time_tz::Tz; 🔒

use self::pricing::PricingTable; 🔒

Modules

pricing

LiteLLM-sourced per-model pricing: the parser plus the proxy-aware fetch that populates the runtime cache.

Structs

CacheCreation 🔒

FileUsageSummary

One session file’s deduped usage, keyed model -> day -> counts. mtime + size drive the incremental cache: an unchanged file is reused rather than re-parsed. folded_project is the repo the file belongs to, folded from its ~/.claude/projects/<slug> dir name.

Record 🔒

Minimal view of a transcript record: only the fields usage accounting reads. Unknown fields (content, tool blocks, …) are ignored by serde.

RecordMessage 🔒

RecordUsage 🔒

TokenCounts

The five-way token split accumulated for one (model, day) bucket. Cache-write is kept split by TTL tier so pricing can apply the 1h / 5m rates independently.

WindowAcc 🔒

Per-window accumulator keyed project -> model -> counts. Keeping the model split under each project lets the by-project rows sum cost across a project’s differently-priced models.

Constants

WORKTREE_MARKER 🔒

Encoded form of /.claude/worktrees/ in a project slug: Claude Code maps both / and . to -, so a worktree path folds to <parent>--claude-worktrees-<name>.

Functions

basename_fallback 🔒

The trailing --separated component of a slug, used when no richer rule applies.

calendar_day 🔒

The YYYY-MM-DD calendar day in tz for an rfc3339 timestamp, or None when it doesn’t parse. DST-correct: the offset is resolved at the instant, so a late-evening UTC time lands on the next local day.

cost_of 🔒

Notional cost of counts at model’s price; 0 for an unpriced or <synthetic> model.

encoded_projects_prefix 🔒

The slug prefix for paths under <projects_root>: the root path with / and . mapped to -, plus the trailing separator’s -.

fold_project 🔒

Fold a ~/.claude/projects/<slug> directory name to the display project it belongs to, plus whether that label is final (see FileUsageSummary::project_resolved).

fold_project_in 🔒

Testable core of fold_project. projects_prefix is the encoded <home>/Projects/ string to strip; an empty prefix disables the repo-resolution rule (no known home). projects_root is the real directory the candidate repo names are stat’d against.

home_dir 🔒

make_row 🔒

parse_date 🔒

Parse a YYYY-MM-DD day key into a Date.

parse_file

Parse one session file into its per-file usage summary, bucketing each record by its calendar day in tz. None when the file can’t be stat’d or opened; a malformed line is skipped.

resolve_project_name 🔒

Resolve the repo name from a slug remainder (the encoded <name>/<subpath...> after the Projects/ prefix). Picks the longest leading run of --joined tokens that is an existing directory under projects_root; when nothing resolves (the repo is not checked out) the whole run is the best-effort label.

roll_up

Roll per-file summaries up into the four windows the /usage overlay drew. Windows are calendar periods relative to now, in the same local timezone the day buckets already use: today, the current week (from Monday), the current month (from the 1st), and all-time. Each window carries both groupings and a total row, priced via pricing (an unpriced model contributes 0 cost).

sort_by_cost_desc 🔒

usage_files

Every real session file under projects_dir/<slug>/. Syncthing conflict copies (*.sync-conflict-*.jsonl) are skipped: they are stale duplicates of a real session and would double-count usage that per-file message-id dedup can’t catch across files.