audit-usage

Consolidated trail

!java "${CLAUDE_PROJECT_DIR:-.}/.claude/hooks/ArchHook.java" audit summary 2>&1

Everything below reads. This skill never writes into .claude/audit-usage/ and never edits project code — the trail is written by ArchHook.java audit, at Stop, and by nothing else.

Answer in English, the language of the reports themselves, so the consolidated view and the per-run report read as one document. Reports and ledger rows written by an older hook are in Portuguese (✅ sucesso); quote them as they are, never translate a number or a status out of them. Reports written before decision 0134 also carry a cost in USD; quote it only from that report, never as a figure of the trail.

When the block came back empty

The trail is off: .claude/audit-usage/ does not exist, and audit summary prints nothing at all. That is the switch, not a failure. Say so, and give the two commands that turn it on — creating the directory is enough, the hook is already wired in .claude/settings.json:

mkdir -p .claude/audit-usage
printf '.claude/audit-usage/.state/\n' >> .gitignore

Then stop. Do not create the directory yourself: switching a versioned audit trail on is the user's decision, not a side effect of asking what it holds.

If the block says closed runs: 0, no run has finished yet. Any report listed under ⏳ No ledger line is from a run in progress — report it as such and say the ledger line is written only when the run closes.

What the block already did

ArchHook.java audit summary reads history.jsonl (one line per top-level run) and nodes.jsonl (one line per skill or agent that run chained) and does the arithmetic: totals, spend per piece without double counting, failure rate, and the reports worth opening. The aggregation is Java on purpose — the ledgers grow without bound, and injecting them raw would cost tokens on every read and leave the sums to you.

Never recompute a number the block printed, and never read the .jsonl files to re-derive one. Your job is rendering and judgment, not arithmetic.

What the block's vocabulary means:

Piece — a skill (📘) or an agent (🤖) of this project. Plugin skills and runtime agents are not recorded.
Origin — user typed /<skill>; model invoked the piece on its own, with no run open.
Model — the model the run billed on, or a comma-separated list when a subagent ran on another one. — means the run was recorded before the column existed, not that the model is unknown. It is the first thing to read before comparing two runs of the same pipeline: the trail counts tokens only, and a token weighs more on a larger model — on the bill and on the plan — so the same token count is not the same spend.
root · nested · preloaded — the piece opened its own run; it was chained inside another run; or it was loaded through an agent's skills: frontmatter (no tokens of its own — they are its agent's).
own tokens — a piece's own spend. A run's total minus its chained pieces is the root's own; that is why the per-piece bars add up to the totals instead of exceeding them.
Calls — tool calls counted from the transcripts' tool_use blocks, attributed to pieces like the tokens. — means the run was recorded before the column existed.
Peak — the largest single request a piece sent (input + cache read + cache write): how much context one turn carried. An absolute number: never turn it into a percentage of a context window you would have to state from memory.
Plan windows — how many runs of a piece fit a plan's 5-hour and weekly windows, as if nothing else ran: the Pro budget in .claude/audit-usage/plan-limits.json × the plan's multiplier ÷ the piece's mean billable tokens, capped by the window's hours ÷ its mean active duration. Model-blind, like every number of the trail. budget or time says which limit binds. Max 5x and 20x are multiples of Pro on 5 hours, the only multiplier Anthropic publishes; Max weekly has no column. A root row counts the whole run, what typing that command spends; a (chained, own) row only that piece's share.

Procedure

1 · Route on the argument

Argument	Do this
empty	§ 2 — consolidated view
`last` / `último`	§ 3 for the `most recent` report the block names
a skill or agent name (`new-feature`, `java-spring-boot-developer`, …)	§ 2 reduced to that piece's line, then § 3 for its newest report: `ls -1t .claude/audit-usage/*--<name>.md` when it ran as a root; otherwise the newest `run` of its line in `grep '"skill":"<name>"' .claude/audit-usage/nodes.jsonl` — a chained piece's report is its parent's
any other text	treat it as a fragment of a report filename (`ls -1t .claude/audit-usage`); § 3 for the single match. Two or more matches: list them and ask which

2 · Consolidated view

Present the block, in its order, keeping its tables and the text bar chart intact — bars are proportional to billable tokens, not to duration, which is wall-clock and includes time the user spent thinking. Then add what the block can't:

One line of reading above the tables — where the tokens went, in words.
Where it went — from the Where the pieces spent block, the piece whose calls are dominated by one tool, or whose peak stands out from the rest. Name the tool and the number; that is what points at the step to trim.
Health — name the piece with the worst failure ratio. A piece that fails repeatedly is a bad spec, not bad luck.
What to read next — the single most relevant report, and the exact /audit-usage <fragment> that opens it.

If the plan windows say not configured, say so and point at .claude/audit-usage/plan-limits.json — {"pro": {"five_hour_tokens": <n>, "weekly_tokens": <n>}}, the billable tokens the person saw a Pro window hold (the audit's total when /usage reached 100%). Never suggest a budget: Anthropic publishes none, and one written from memory looks as authoritative as a measured one. Never convert anything to a currency either: the trail prices nothing, and a rate written from memory is wrong the day after it changes.

3 · Single run

Read the .md file. It is already a finished, icon-rich report — do not re-render it and do not paste it back in full. Summarize in at most six lines: what was run, the outcome, its billable tokens, the longest step, where it spent (from 🔎 Where the run spent: the dominant tool, the costliest turn, the peak context), and anything that deserves attention (⏳ in progress, compaction incidents, permissions added, repeated tool failures — quote the error's first line the report already redacted, never more). Then give the path so the user can open the whole thing.

Point out ⏳ in progress when you see it: it means the run never closed — the session was killed, or it is still open right now.

What this skill will not do

Never edits or deletes anything under .claude/audit-usage/. Pruning old reports is the user's call; if they ask, show the command and let them run it.
Never prices the trail. The numbers come from audit summary, in tokens; a cost computed from them needs a rate, and a rate written from memory is wrong the day after it changes while looking exactly as authoritative as a correct one. Asked for money, point at the runtime's own /cost for the session (decision 0134).
Never reads .state/. Those are the hook's open append-only logs; the rendered .md is the readable form.
Never quotes a redacted value. The reports blank credential-shaped text on the way in. If something still looks like a live secret in a report, say so and point at .claude/schemas/extensions.json's audit.redact block — a pattern is missing there.

Why this is a skill, and why the hook ignores it

The hook already answers "what happened in this run", and audit summary does the sums across runs. Nothing turns that into "which piece is eating the budget, and which report should I open" — judgment over data, routed by an argument, which is a procedure, not a norm and not enforcement.

audit-usage is one of the skills the trail does not record — the observers: its class, observer in skill_classes of .claude/schemas/extensions.json, declares audited: false, and arch-doctor sits in the same class. Without that, reading the trail would append a report about reading the trail, and every later read would be mostly reads.

Invoking it still closes whatever run is open, and that is the useful half: within a single session a report stays stamped ⏳ in progress until something ends the run, so asking for the report is what finalizes it.

arch-adopt is never recorded either — fixed in its own class override. Any other piece can be switched per project in .claude/audit-usage/audited.json ({"skills": {"<name>": false}, "agents": {}}), the one file of .claude/ the project owns and /arch-adopt never overwrites; doctor's Audit overrides line validates it.

Runs on sonnet with effort: low: audit summary already did the sums, and what is left is ranking and rendering (@.claude/decisions/0080-procedural-skills-on-sonnet.md).

Contract

Class: observer — the territory is skill_classes.observer in @.claude/schemas/extensions.json, and it is empty: this skill writes no file. The reports under .claude/audit-usage/ are written by the audit hook, not by whoever reads them. ArchHook.java guard enforces it.
Reads — the output of ArchHook.java audit summary; .claude/audit-usage/*.md; .claude/audit-usage/nodes.jsonl only to find the report of a chained piece
Writes — nothing
Handoff — none. Terminal skill: it reports and stops