kasuboski/pi-extensions/pi-usage

pi-usage

pi writes every session as JSONL. Each assistant message carries a usage block — the single source of truth for token accounting. This skill parses those blocks and sums them into a report.

Where the logs live

~/.pi/agent/sessions/<dir-encoded-cwd>/*.jsonl

Each directory is a project, named with its cwd path encoded (/ → -). Files are named <UTC-timestamp>_<session-uuid>.jsonl. One directory per project; one file per session.

The usage schema

Every assistant message contains:

"provider":"<provider>","model":"<model>","usage":{"input":N,"output":N,"cacheRead":N,"cacheWrite":N,...}

input is fresh prompt tokens, cacheRead is cached prompt tokens served from the cache, cacheWrite is tokens written to the cache (often 0 — many providers never report it), output is generated tokens. Sum these per provider/model.

Steps

1. Parse and aggregate. Run this from ~/.pi/agent/sessions — it extracts

every usage block across all sessions and sums per provider/model:

```bash cd ~/.pi/agent/sessions && \ rg --no-filename -o \ '"provider":"([^"])","model":"([^"])","usage":\{"input":([0-9]+),"output":([0-9]+),"cacheRead":([0-9]+),"cacheWrite":([0-9]+)' \ --glob '*.jsonl' -r '$1/$2|$3|$4|$5|$6' \

END { for (k in in_) printf "%s|%d|%d|%d|%d|%d\n", k, n_[k], in_[k], cr_[k], cw_[k], out_[k] }' \

```

Fields out: provider/model | calls | input | cacheRead | cacheWrite | output.

Completion criterion: every model that logged usage appears once, with non-negative totals; a TOTAL row sums all columns.

1. Present the table. Print the breakdown with a TOTAL row. Columns: calls,

input, cache-read, cache-write, output. Add a TOTAL computed across all models.

Completion criterion: the table has all five token columns plus a TOTAL row whose column sums match the per-model rows.

1. Add a short findings note. Flag the dominant model, whether caching is

in use (compare cacheRead vs input), and any all-zero models (errors or no usage returned — common for *-free and errored providers).

Completion criterion: findings state which provider/model dominates, whether cacheRead exceeds input, and list any all-zero models.

Regrouping (per-project, per-date)

The parse is fixed; only the grouping key changes. Re-run step 1 but swap the key:

• Per-project — group by directory instead of provider/model. Count usage

calls per session directory: for d in ~/.pi/agent/sessions/*/; do rg -c '"usage":\{"input"' "$d"; done | sort -rn.

• Per-date — session filenames start with the UTC timestamp; group the

parse output by filename date prefix.

Caveats

• cacheWrite is frequently 0 across all models — many providers never return

it. A zero column is expected, not a bug.

• Some providers log calls with all-zero usage (errored or no usage returned).

Report them as a row but don't let their zeros skew findings.

• Only providers with prompt caching enabled show meaningful cacheRead;

local (ollama) and most *-free models show zero caching.