The plugin claudetto-sessions shows what a Claude Code session costs. It has two reports and one note. It works without the core plugin, and it imports nothing from another plugin.

The slices are start-cost, usage, and awareness. Each report is a CLI, and a command.run hook shows its output with no model turn.

The command /start-cost runs claudetto-start. It reads a transcript and reports the tokens of the first request, by source: system prompt, tools, skills, agents, and memory files. The sources are an estimate from the characters of the transcript, and the line First request holds the real token count.

Exit status	Meaning
0	The report ran, and each bound that was given holds.
1	A bound did not hold, or --text-max could not run.
2	The transcript has no request with usage, so there is no data.

• --text-max fails with status 1 when the transcript has no prompt_snapshot record. The report then says that the check did not run.
• --text-min only advises. It never changes the exit status.

The command /claudetto-sessions:usage runs claudetto-usage. It reports the model calls, the tokens that each call reads again from the cache, the cost, and the skills that a period did not use. The built-in /usage is another command and runs as usual.

• Each label names one count. A turn is a prompt, and a call is one request to the model. The two counts never share a label.
• A call with a model id that has no price keeps its token counts. The report names the model, and its cost does not count in the total.
• The report names the Claude Code versions that the transcript records, and the version of the plugin.
• At each session start, the plugin keeps its version for the session in $.store, for at most 500 sessions. The command gives this record to the report, and the report counts the sessions of the period by plugin version. A session that started before the record had it counts as not recorded.

When managed settings install the plugin with appendPlugins, $.session.identity() answers. Each report then starts with a line that names the account. A plugin install gets a deny answer, and the reports name no account.

Claude Code shows the rate-limit state, so the plugin does not draw it again. When the seven-day window reaches 80 percent, the session.measure hook makes one note. The main agent gets the note once, after the result of its next tool call. The bound is WEEKLY_WARN_PERCENT in slices/awareness/limits.ts.

Each hook of the plugin catches its errors, writes them to the debug log, and lets the session go on. The plugin writes no stamp into the transcript.

• Add-On Plugins: the list of the add-on plugins.
• Bounds