이 mod 소개
컨텍스트 창 사용량과 턴별 증가량을 보여줘 컨텍스트 정리 시점을 판단하도록 돕습니다.
커뮤니티 작성자가 관리하는 프로젝트입니다. 기능, 버전 요구 사항, 설정 방법은 아래 원본 README를 확인하세요. 이 사이트는 권한 스캔 결과를 제공합니다. 설치 전에 소스와 외부 의존성을 확인하세요.
컨텍스트 창 사용량과 턴별 증가량을 보여줘 컨텍스트 정리 시점을 판단하도록 돕습니다.

컨텍스트 창 사용량과 턴별 증가량을 보여줘 컨텍스트 정리 시점을 판단하도록 돕습니다.
커뮤니티 작성자가 관리하는 프로젝트입니다. 기능, 버전 요구 사항, 설정 방법은 아래 원본 README를 확인하세요. 이 사이트는 권한 스캔 결과를 제공합니다. 설치 전에 소스와 외부 의존성을 확인하세요.
먼저 작성자의 README에서 marketplace와 플러그인 이름을 확인하세요. 저장소 구조에 따라 명령어가 달라질 수 있습니다.
claude plugin marketplace add Arunjay4213/claude-mods claude plugin install context-lens
A live /context.
Claude Code's /context is a snapshot you have to ask for.
context-lens keeps the same figures on screen: a pinned line under the prompt that updates after every turn, and a pane with the window broken down by category.
A pinned status line under the prompt, from the moment the session starts:
ctx 5% · 45.9k/1.0M · +44 last turn · 99+ turns to full
Reading it left to right: how full the context window is, the tokens used against the window size, what the last turn added, and how many more turns fit at the recent average growth.
Any figure the engine has not reported yet is left out rather than guessed, so early in a session the line is shorter.
Before the first answer comes back it reads ctx waiting for the first answer.
/context-lens opens a pane with the same reading in full:
45.9k of 1.0M used · 5% · claude-opus-5
window ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 45.9k 5%
messages █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 22.4k 2%
system tools █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 14.2k 1%
memory files █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 4.9k 0%
system prompt █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 3.1k 0%
autocompact bu… ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 45.0k 5%
free space ████████████████████████████████████████████░░░░ 910k 91%
growth +4.2k last turn, +3.8k average over 5 turns
headroom 12 turns at that rate, 754k left
limit 800k, where auto-compaction runs
counted locally, as an estimate - /context-lens refresh counts exactly
Each bar is sized to the pane's own width, so it fits the dock and the inline placement alike.
Running /context-lens again closes the pane.
/context-lens refresh counts every category with the token-count API, the way /context does, and says so in the counted row.

Everything comes from one call, $.session.usage.
After each finished turn the mod asks for breakdown: "summary", which estimates locally and sends no requests, so the line costs nothing.
/context-lens refresh asks for breakdown: "full", which sends one token-count request per tool and memory file.
That is the only thing in this mod that spends anything, and it only runs when you ask for it.
The growth figures come from a small in-memory list of the token count after each turn. A compaction clears it, since the window it measured no longer exists.
The compaction line prefers the engine's own autoCompactThreshold from the breakdown.
When auto-compaction is off it falls back to the compaction window (rawMaxTokens), and when there is no breakdown at all it falls back to the context window size and says which it used, so the number is never presented as more than it is.
claude plugin marketplace add Arunjay4213/claude-mods
claude plugin install context-lens@claude-mods
Mods are early access, so the module only loads when the function-hooks flag is set.
Add it to ~/.claude/settings.json:
{
"env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" }
}
or set it on the command line for one run:
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude
To run it from a checkout instead:
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude --plugin-dir plugins/context-lens
$.session.usage({ breakdown }).
Claude Code 2.1.272 and later answer it with the categories and the auto-compaction threshold, and the pane draws them.
Claude Code 2.1.270 and earlier answer with only tokens, window and percent, so on those builds the pane draws the window bar alone, says so in place of the category rows, and measures headroom against the whole window.tokens and percent come from the last API response, so a fresh or just-compacted session has no reading until its next answer.