context-budget

Reviewed at commit 9519cda

Measures a Claude Code session's fixed startup context payload per item at a pinned binary version, including per-tool attribution of the built-in tool pools…

0 upvotes

Source: github.com/melodic-software/claude-code-plugins/tree/main/plugins/context-budget
Made by: Melodic Software
Pinned commit: 9519cdae8db7003c57e27e2fe123d4634ac1f742
What we checked: Shows a live breakdown of what fills your context window in a pane or above the prompt, and asks you first whenever Claude edits a Claude Code settings file. Its audit skill starts short headless Claude Code sessions to measure what each built-in tool costs at startup and keeps a ledger in the plugin's data folder; project settings are changed only through its fix option after you approve.

Review details

Static review of the source at the pinned commit; not run. The register.tsx module hooks session.start, session.end and prompt.submit only to register the /context-budget command (it does not read prompt text), handles command.run for /context-budget live [pane|band|off], refreshes on session.measure through $.session.usage({ breakdown: 'summary' }) only while the pane or band is shown, and draws through ui.render (Pane and AbovePrompt) with $.ui.open/close/panes/toast and plugin-state atoms; nothing is written to the transcript. Shell hooks: a SessionStart command runs the bundled prerequisites.sh or prerequisites.ps1 (PowerShell with -ExecutionPolicy Bypass on its own script) in node-notice mode, which only checks that node is on PATH, shows a notice if not and creates a once-per-session latch directory in the temp folder; a PreToolUse hook on Write|Edit|NotebookEdit runs settings-write-ask.mjs, which reads the target path, runs git rev-parse to find the main checkout root, and returns permissionDecision ask (never allow) for user, project, local or managed settings files; it fails open on errors and has a userConfig kill switch. The audit skill (no pre-approved tools, so Claude asks before each Bash call) runs skills/audit/scripts/measure.mjs, which spawns the claude binary on PATH (or --binary) as claude -p /context, once per denied tool with --disallowedTools and without permission bypass, or uses the Agent SDK query() with maxTurns 1 when @anthropic-ai/claude-agent-sdk resolves from the plugin data folder or, as a fallback, from the current project's node_modules; reads the claude binary to grep for setting names; and writes snapshot, attribution, ledger and report JSON/Markdown under ${CLAUDE_PLUGIN_DATA}/audit/<state-key>/ (state-key.sh derives the key from the git remote and root with path-injection guards). The fix path edits project .claude/settings(.local).json one approved lever at a time and only prints user-global edits. The SDK npm install is printed for the user, never run; the setup skill pre-approves only Bash(node --version*) and Bash(claude --version*). Claude may fetch the cited docs pages on code.claude.com when re-verifying levers. No network calls in code. Read fully: manifests, hooks.json, register.tsx, parts.tsx, types, both prerequisites scripts, settings-write-ask.mjs, measure.mjs, state-key.sh, levers.test.sh and the audit, check and setup SKILL.md files. By targeted search: measure.test.sh (developer test, not loaded at runtime), the reference JSON/Markdown and evals, and lib/prerequisites.mjs from the clone (only reached by prerequisites.sh check mode, which no hook or skill uses; it probes with spawnSync and writes a marker file, no network). Not reviewed: the Agent SDK and the Claude Code binary it measures.
What it uses: Reads your prompts: no

Sees tool calls: yes

Blocks or changes tool calls: yes

Reads or writes files: yes

Runs programs on your machine: yes

Calls the model (uses tokens): no

Sends data over the network: no

Stores data: yes

Shows something on screen: yes
Hooked events: Prompt

Tool call

Permission

UI

How to install

This mod is not in the official Claude directory, so you load it straight from its source on GitHub. modsforClaude.com is an unofficial community site: we do not host, repackage or install anything. Read the source and the review on this page first.

In the terminal

Clone the repository and start Claude Code with the mod. It stays loaded for that session.

git clone https://github.com/melodic-software/claude-code-plugins
claude --plugin-dir ./claude-code-plugins/plugins/context-budget

In the desktop app

Clone the repository into a folder you keep:

git clone https://github.com/melodic-software/claude-code-plugins ~/claude-mods/claude-code-plugins

Add that folder to CLAUDE_CODE_PLUGIN_DIRS in the env block of ~/.claude/settings.json, then start a new session in the Code tab. On Windows, with your own user name:

"env": {
  "CLAUDE_CODE_PLUGIN_DIRS": "C:\\Users\\you\\claude-mods\\claude-code-plugins\\plugins\\context-budget"
}

On macOS and Linux:

"env": {
  "CLAUDE_CODE_PLUGIN_DIRS": "~/claude-mods/claude-code-plugins/plugins/context-budget"
}

Already loading other mods this way? Add the folder to the same value, separated by ; on Windows or : on macOS and Linux.

Not working, or want to change its settings? How to load a Claude Code mod walks through every step.

Measures a Claude Code session's fixed startup context payload per item at a pinned binary version, including per-tool attribution of the built-in tool pools that /context reports only as totals, derived by A/B deny differencing. Falls back from an SDK meter to a headless /context parser, then to an error rather than a wrong number. Keeps a before/after ledger per lever. Shows a live context breakdown in a pane or above the prompt. Report-only by default; fix applies one approved project trim.

Made this mod? Add a badge to your README

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/context-budget.svg)](https://modsforclaude.com/mods/context-budget/)

← All mods