limitcast 每週更新（10月30日）：已推出內容及一週實際使用數據

Nothing new reached npm this week after 0.17.3 on 28 October. Here is what changed on main, and one week of my own usage numbers with the caveats that come with them.

My total token usage fell 68% week over week, but limitcast can only vouch for an estimated 11.9M Claude context tokens not re-sent, about 1.3% of one 5-hour window. The two figures measure different things, and neither proves limitcast caused the drop.

What shipped

0.17.3 (28 October, npm and GitHub): the Grok status line prints one line per tool, so Claude and Grok windows no longer share one unlabeled row. Release notes.
On main, not in npm 0.17.3: trim --share keeps configuration names and paths private when no measured startup transcript is available, and honors totals-only output for Codex and Grok. Sharing cannot be combined with keep-list edits, --apply or --undo. trim --share --out FILE now prints "Wrote shared summary." instead of the file path. Until the next release, use results --share for voluntary check-ins and read the output before posting.
Also on main: a Discussions space with tester guidance and issue templates, and a README with a CLI showcase built from synthetic demo sessions.

One week of my own usage

This is one person on one machine, real usage, not synthetic data and not a user result. It contains no prompts, project names or account IDs.

Usage recap: Claude Bar's weekly recap shows 28.8M tokens (Codex 16.5M, Claude 12.3M) across 108 sessions and 86 h 25 min, down 68% from last week. That is total usage, which depends on how much work I did as well as how I worked. Claude Bar counts tokens its own way, so its figures are not comparable with limitcast's totals.
limitcast, same week (limitcast results --days 7): an estimated 11.9M Claude context tokens not re-sent, across 7,132 calls in 40 sessions.

What limitcast counts as saved

limitcast reports one saved-token figure: context tokens not re-sent, estimated from characters divided by 4 against my first recorded starting context. Most of them would have been cached input, so 11.9M is about 1.3% of one 5-hour window. It is an estimate. It can include project or configuration changes, it does not prove limitcast caused the reduction, and it is not a lower bill.

Codex and Grok context savings are unavailable, and holds and model changes are not converted into saved tokens. There is no larger total to add up.

Protection and forecasts

2 usage limits reached, both on 24 October. The 5-hour window was already at 100% in the first reading limitcast saw. For the weekly window, limitcast warned 40 minutes ahead.
Applied events: 8 tight-window model changes (subagents ran on Sonnet instead of my main model) and 2 cold-cache holds. The guard did not need to stop a subagent. These are events, not token savings.

Starting context

My Claude Code sessions start at 109.4k tokens before I type. About 7.2k (7%) went unused for 7 days: 7 MCP servers and 68 skills. Claude Code's own prompt and tools are 68.5k of the total.

The part I control (hooks, MCP servers, skills, agents, instruction files) grew from about 39.1k to about 40.9k tokens (+1.8k) since 24 October. The 11.9M estimate covers my whole starting context against the first recorded one, so it can reflect changes beyond that part.

Measure your own numbers

npm install -g limitcast
limitcast onboard
limitcast results --days 7
limitcast trim

To share a summary, run limitcast results --share, read every line, and remove anything you do not want public. Nothing uploads automatically. The weekly updates and a one-week tester check-in are in Discussions.

Originally published at limitcast.dev. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.