engineering / skill

Cache Tax

A Claude Code mod that sends keep-warm pings during breaks and shows an estimated rewrite cost before a message sent after the cache has gone cold.

Repository owner karanb192

Not rated yet
0 comments

Like 0 Save

Published 11 Oct 2026Listing updated 12 Oct 2026

At a glance

Purpose: A Claude Code mod that sends keep-warm pings during breaks and shows an estimated rewrite cost before a message sent after the cache has gone cold.
Runs in: Local agent with shell
Operating systems: Not stated
Needs: Runtime: Not stated; Needs internet: Not stated; Needs accounts or API keys beyond your agent: Not stated
Files and data: Reads files: Not stated; Writes files: Not stated; Sends data to other services: No
Tool type: Skill
Delivery: From GitHub
Version: Latest on the main branch
Access: Code on GitHub
Compatibility: The README states a minimum of Claude Code 2.1.287 and reports access validation on 2.1.289. The README says the terminal and Desktop Code tab show a row above the prompt while keepwarm is armed. The README says Ghostty and kitty can draw the image, while other terminals show [>]. The README says included subscription usage defaults to a one-hour cache, while API-key, usage-credit and cloud-provider sessions default to five minutes.
External costs: The repository is MIT-licensed. According to the README, each keep-warm ping costs tokens, including uncapped output, and a message sent after the cache goes cold pays a cache write. The README describes the dollar figures as API-equivalent costs rather than extra subscription charges, and calls them a yardstick rather than the subscription bill.
Permissions: The mod acts inside Claude Code with the access Claude Code gives it. The README describes use within Claude Code sessions billed through included subscription usage, an API key, usage credits or a cloud provider. The README shows plugin installation and says API-key, usage-credit and cloud-provider sessions need promptCacheTtl set to "1h". The README says the mod can see typed prompts, request timing and answer token counts; its threat model describes inspecting a prompt's leading character for a slash and passing prompt text onward without storing or logging it. The README says keepwarm uses a timer to send tool-less API requests over the session transcript within an armed window. The README says detected cold writes automatically arm a window, and /keepwarm always enables windows at each session start. The README describes persistent deadline, ping-period, always and guard-mode settings. The README says the default guard can drop an ordinary message once before allowing a resend.
Usage terms: MIT License, as stated in the repository. See its LICENSE file for the exact terms.

Code on GitHub

karanb192/cache-tax

MIT

TypeScript163 stars24 forksLast pushed 4 Oct 2026 UTC

Keep Claude Code’s prompt cache warm during breaks and show the estimated cost before a cold send.

Topics: claude-code claude-code-plugin claude-mods function-hooks prompt-caching

Public GitHub data fetched 13 Oct 2026 UTC · 0 days old

View on GitHub ↗

What it does

Cache Tax is a Claude Code mod; the repository's file list contains no folder with SKILL.md. The README says the mod sends keep-warm requests during breaks and shows an estimated rewrite cost before a cold send. The README gives /keepwarm 90m as a break command and says pings cost tokens.

The README says /keepwarm arms a bounded window, six hours by default. The README says a timer sends a tool-less fork over the session transcript after 50 idle minutes and stops when a ping reads zero tokens or writes reach 10% of reads. The README says the default guard drops an ordinary message once when its clock estimates a cold cache and context is at least 50,000 tokens; users can resend or start from a note with /clear. The README says a detected cold write automatically arms at least three hours of keepwarm.

The README lists /keepwarm status, /keepwarm off and /cache-tax for state and estimates. The README says an armed window adds a warm, cold or unknown row above the prompt. The README describes displayed dollar figures as API-equivalent estimates rather than extra subscription charges and the refusal token count as an upper bound.

Who it is for

Claude Code users who step away from long sessions and want the prompt cache refreshed or its rewrite cost shown.

Example tasks

Run /keepwarm 90m before a break.
Run /cache-tax to see the cache estimate, the warming window and this session's cold-write tally.
Run /keepwarm status to see the warming window, the next ping and the last receipt.
Run /keepwarm off to stop, forget the window and turn the always setting off.
Resend a dropped message to continue, or run /clear and start from a note.

Getting started

The README shows two shell commands: claude plugin marketplace add karanb192/claude-code-mods and claude plugin install cache-tax@claude-code-mods. Then restart Claude Code or run /reload-plugins in an open session. In a warm session, run /keepwarm to arm six hours, /keepwarm 90m for a shorter window or /keepwarm off to stop. For API-key, usage-credit or cloud-provider sessions, the README says to set promptCacheTtl to "1h". The README has the exact commands.

Limitations

According to the README, it needs Claude Code 2.1.287 or later, a one-hour cache and Claude Code left running. It says mods are on by default, but local settings, safe mode or organization policy can prevent them from loading. It says included subscription usage defaults to a one-hour cache, while API-key, usage-credit and cloud-provider sessions default to five minutes, so those need promptCacheTtl set to "1h", and that a 50-minute timer cannot protect a five-minute main cache. It says pings cost tokens, including uncapped output. It also says a ping that reads warm shows the cache was warm only then, that prefix changes can invalidate the cache regardless of time, and that its keep-warm return was observed in one run in one configuration. It says the guard drops your ordinary message once when the cache is cold, and you resend it. On a subscription, the README says it found no documentation on how a cache read counts against the 5-hour and weekly limits. It says to keep one guard if you also install its separate settings-hook version. The repository can change at any time; check its README before you install.

Data practices

The README says each keep-warm fork is an API request over the session's transcript with a fixed one-line prompt. The README's threat model describes checking a prompt's leading character for a slash and passing prompt text onward without storing or logging it. The README says its store holds session-specific keepwarm deadlines and ping periods plus cross-session always and guard-mode settings. The README notes that a session which arms keepwarm and does not return can leave two stored keys. The README says the cold-write tally is held in memory for the session. Provider-side handling depends on your Claude Code configuration and model provider.