| summary | Pi and OMP local token history, source selection, and cost accounting. | ||
|---|---|---|---|
| read_when |
|
Enable Pi in Settings → Providers to show Pi and OMP token history as a separate local source in the menu, Usage & Spend, Overview, and widgets. Pi has no subscription quota or account balance. Costs are API-rate estimates from recorded assistant usage, not a billing statement. No provider login or credential is required to read the transcripts.
Pi uses the official pi.dev logo palette: blue #4D9ABF for the default chart and widget accent, with coral #F09082, blue, and yellow #F1BE58 for confetti. Blue meets the existing light/dark contrast threshold. Saved accent overrides take precedence. The dashboard link opens the current Pi repository.
The scanner prices the openai-codex, anthropic, and amazon-bedrock backends. Bedrock sessions stay in standalone Pi history and use the exact model ID in the cached models.dev Bedrock catalog, preserving regional and global prices. They are never mirrored into native Claude or Codex history. Other backends retain valid timestamped usage as unpriced tokens in standalone Pi history, labeled with their backend and model; they never borrow native-provider prices or enter native Claude/Codex history. Missing or malformed usage keeps standalone history incomplete without invalidating supported partitions. Supported usage with an unknown model retains its recorded tokens and is marked unpriced. Mixed priced and unpriced history preserves the known subtotal without presenting it as a complete cost.
The cacheWrite1h counter (cttl.ephemeral1h on OMP) is the one-hour subset of cacheWrite, priced at twice the input rate; remaining writes use the default cache-write tariff. The first present field wins in this order: cacheWrite1h, cache_write_1h, cttl.ephemeral1h, cttl.ephemeral_1h. Missing one-hour counters retain the default tariff. Invalid counts, including a one-hour count larger than total writes, keep history incomplete; a malformed first-present counter cannot fall through to another spelling. Pi's recorded usage.cost is not used to replace the independent estimate.
Cost collection can refresh the public models.dev pricing catalog. Transcript contents stay local; the catalog request needs no credential. Existing cached prices and bundled rates remain available when a pricing refresh fails.
Default session roots include ~/.pi/agent/sessions and the supported OMP agent/profile stores. Discovery honors PI_CODING_AGENT_DIR, PI_CODING_AGENT_SESSION_DIR, OMP configuration/XDG roots, and OMP_PROFILE (or PI_PROFILE when absent). A named profile limits discovery to that profile. Invalid or unresolved explicit selectors produce incomplete history.
Root canonicalization preserves an explicit directory marker even when the session directory does not exist yet.
Discovering the same roots in a different order does not change the cache scope. Older unsorted scope fingerprints are refreshed once while retaining the same source selection and totals.
Running Pi/OMP processes also contribute their environment, profile, --session-dir, and project settings. Relative paths resolve against that process's working directory. A missing working directory cannot turn an unresolved relative selector into a successful empty scan. Retained roots from explicit command-line or settings selectors survive process exit; settings are revalidated before reuse. Removing a setting from an accessible project drops its former root, while an inaccessible project or broken settings symlink preserves the previous scoped report and its original age.
Live processes with an unreadable environment and no explicit command-line session/profile selector are skipped during cost-root discovery. A single summary diagnostic records the skipped count per discovery pass without logging arguments, paths, or environment values. Default/configured roots and other readable processes still contribute history, including on the first scan. This is best-effort discovery: a custom root visible only in an unreadable process environment cannot be discovered. Explicit selectors that cannot resolve still keep history incomplete; an absolute --session-dir remains usable without an environment or working directory.
Assistant turns are bucketed by their own timestamp in the selected cost time zone. Matching entry IDs within the same session count once across overlapping roots. Distinct turns remain separate. The scanner retains per-message prices and token classes rather than repricing a daily aggregate.
With Pi disabled, unscoped Claude and Codex history can include their supported Pi/OMP backend partitions. With Pi and local cost tracking enabled, the app shows native Claude/Codex history alongside standalone Pi. Combined CLI selections follow the same rule. Account-scoped Codex history always remains native because machine-local Pi history does not establish account ownership. Overview and Usage & Spend use the same source accounting, including during cached hydration and completed Codex catch-up.
codexbar cost --provider pi --format json --pretty
codexbar cost --provider both --format json --pretty
codexbar cost --provider codex --provider-native-onlyFor a combined report, enable Claude, Codex, and Pi in the config, then run codexbar cost --format json --pretty without a provider override. Selecting Pi with Claude/Codex excludes mirrored Pi rows from those native providers. A standalone Claude/Codex selection (including --provider both) retains its existing inclusive behavior unless --provider-native-only is supplied. The same selection rule applies to dashboard and HTTP cost collection.
The cache is ~/Library/Caches/CodexBar/cost-usage/pi-sessions-v9.json on macOS. It records source scope, coverage, and unsupported-history evidence, and is replaced atomically on macOS and Linux. Version 8 is rebuilt once from transcripts because it did not record sufficient scope and completeness evidence for safe reuse. An unavailable source during this upgrade leaves history unavailable until a valid scan can complete; it does not borrow an old cache's timestamp or totals.
Incomplete refreshes can preserve a previously valid report with its original source scope and scan time. Malformed records, truncated tails, inaccessible roots, and unrepresentable aggregate totals cannot advance cache freshness. Cached and debounced reads check the recorded file inventory and metadata before declaring coverage complete, without reparsing transcripts. Failed root transitions never combine old and new datasets. Missing optional Pi history leaves available native history explicitly incomplete and immediately eligible for another refresh. For Claude, an established native spend subtotal stays visible as a partial estimate in Usage & Spend, Overview, and sharing. CLI totals remain available with incomplete-history metadata; repairing the mirror restores complete coverage. A verified empty source remains distinct from an uninspected source.
Files replaced while being read leave the refresh incomplete. A pricing-catalog change requires a complete reparse before new estimates replace the prior report. An explicit refresh reparses Pi history even when file size and modification time are unchanged, including Pi usage shown under Claude or Codex.
The one-hour counter correction invalidates the Pi pricing key once, including estimates mirrored into Claude or Codex reports. Existing Pi cache rows do not retain the one-hour split, so all Pi sessions in the scan window must be reparsed before publishing corrected prices. The cache schema and native Claude/Codex caches are unchanged; subsequent refreshes resume normal cache reuse.
The addition of unsupported-backend token history also invalidates the Pi pricing key once so previously ignored rows are collected. The cache schema stays unchanged.