SLOPSHOPPER

cache-warm

Keeps the prompt cache warm on idle chats with a capped fork ping, and tracks cache hits and misses

newcommandpromptmodeltimer
v0.1.0MITupdated 2026-10-04mustafa89/my-claude-code-mods/cache-warm
A shopper browsing a rack in a slop shop
Preview · a replayed session in a sandbox
claude · ~/work/app · cache-warm
› fix the failing auth test and add an audit log call ⏺ Read(src/auth.ts) ⎿ Read 6 lines ⏺ Update(src/auth.ts) ⎿ Added 2 lines, removed 1 line ⏺ Bash(bun test) ⎿ 3 pass, 1 fail ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM › /cache-warm ⎿ cache-warm: keepalive on · ttl 60m · cold · no request yet · hits 0 · misses 0 · pings 0/4 (session 0) ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts
README

my-claude-code-mods

Six mods for Claude Code: a one-line status bar, image previews for pasted screenshots, boxed cards for tool calls, a keepalive for the prompt cache, a weekly calendar of your sessions, and animated diagrams Claude can draw in the transcript.

Mods are Claude Code plugins built from function hooks. They need Claude Code 2.1.287 or newer. Mods are an early-access feature: the hooks API can change between releases.

ModWhat it does
slick-barReplaces the hint line under the prompt with a one-line bar: model, effort, folder, branch, context gauge, prompt-cache health, rate-limit gauges
image-peekShows a preview of a pasted image above the prompt and under the message that sent it
tool-cardsDraws each tool call as a boxed card: highlighted command, output preview, timing footer
cache-warmKeeps the prompt cache of an idle chat warm with a small capped ping, and counts cache hits and misses
week-calendar/week draws a calendar of the week's sessions with commits, hours per project, and a 3-line report
diagram-modGives Claude a show_diagram tool: animated diagrams that play a story step by step, with packets, gauges and a live log

Install

  1. Clone the repo:
   git clone https://github.com/mustafa89/my-claude-code-mods.git ~/.claude/mods
  1. Tell Claude Code to load the mods. Add this to the env block of ~/.claude/settings.json (paths separated by :; ~ is allowed):
   {
     "env": {
       "CLAUDE_CODE_PLUGIN_DIRS": "~/.claude/mods/slick-bar:~/.claude/mods/image-peek:~/.claude/mods/tool-cards:~/.claude/mods/cache-warm:~/.claude/mods/week-calendar:~/.claude/mods/diagram-mod"
     }
   }

Leave out the mods you do not want. slick-bar lists cache-warm as a dependency for its cache segment, so load the two together.

  1. Start a new Claude Code session. Each mod prints <name>: loaded in the transcript.

To try one mod for a single session only:

claude --plugin-dir ~/.claude/mods/slick-bar

If you use a statusLine command, the bar does not replace it. The bar sits on the hint line under the prompt; your status line stays below it. Remove statusLine from your settings if you want only the bar.

slick-bar

slick-bar

One line under the prompt, after Claude Code's own permission-mode label:

  • Model and effort: ✻ Opus 5.5 1M · medium. The ✻ turns into ◉ while a turn runs.
  • Folder and branch: repository name (plus subfolder), branch, ± when the tree has changes, ↑/↓ for commits ahead of or behind the upstream.
  • Context gauge: how full the context window is, green → amber → red.
  • Rate-limit gauges (subscription plans): 5-hour and 7-day usage, the time until each window resets (↻ 3h12m), and a ┃ tick that marks how much of the window has passed. Usage to the right of the tick means you are using it faster than the window allows; the percentage turns amber when the current pace projects past 100%.
  • Prompt cache (with cache-warm): cache ● 97% 52m ✗1 ⟳2. The dot is green while the cache is warm, amber in the last minutes before it expires, and ○ cold after. Then the share of the last request served from cache, the time until the cache expires, misses (✗) and keepalive pings (⟳).

When the terminal is narrow, segments drop in this order: hint text, 7-day gauge, token count, cache, branch, folder. The model and the context gauge always stay.

The bar needs a Nerd Font for the pill shapes and icons.

image-peek

image-peek

  • While you write a prompt, every pasted image shows above the prompt with its size (Image #3 · 1246×846).
  • After you send it, the image shows under your message in the transcript.

image-peek draws real images through the kitty graphics protocol, so it needs a terminal that supports it, such as kitty or Ghostty. Elsewhere it shows the image's label instead.

It also needs macOS: it uses sips to make a smaller copy of each image.

Terminal multiplexers: inside a multiplexer that passes kitty graphics through but identifies itself differently (for example herdr, which reports libghostty), Claude Code turns graphics off. To turn them on, add this to the same env block:

"CLAUDE_CODE_FORCE_TERMINAL_IMAGES": "1"

Only do this if your terminal really supports kitty graphics. Otherwise images draw as empty boxes.

tool-cards

tool-cards

Every tool call becomes a card:

  • Header: tool name and status: ✓ done, ◌ running, ✗ error, ⊘ interrupted. The command's description follows.
  • Bash: the command with syntax colours, then the first 5 lines of output (errors in red).
  • Edit and Write: a split diff, old on the left and new on the right, with line numbers, red and green rows, and the changed words highlighted. A +N -M header with a bar shows the size of the change. Narrow terminals stack old over new. A new file from Write keeps Claude Code's own view.
  • MCP tools: the arguments, then the output folded behind ▾ output · N lines. JSON is indented; errors show in red.
  • Other tools: one line saying what the call acted on, for example src/cart/total.js:10-30 for a Read.
  • Footer: how long the call took, how many words it printed, and the Bash timeout if one was set.
  • Diagrams from diagram-mod: a rounded card headed ◇ diagram, as wide as the diagram, with the animation inside and ▴ collapse / ▾ expand.

Claude Code normally folds runs of reads and searches into one line (Searched for 2 patterns). tool-cards unfolds them, so each call gets its own card.

Expanding output

  • Click ▾ expand on a card to see all of its output or diff (up to 400 lines). Clicks reach the card in fullscreen mode.
  • /cards full expands every card; /cards compact goes back to the preview (5 output lines, 16 diff rows, MCP output folded); /cards switches between the two.

Ctrl+o does not expand a card. Claude Code does not tell mods when the Ctrl+o view is open.

cache-warm

cache-warm

Claude Code caches the conversation on the API side. A request that reads the cache pays a small part of the input price; a request after the cache expired writes the whole conversation again. cache-warm sends one small request just before the cache expires, so an idle chat stays warm.

How it works

  • Timer: a 15-second timer runs inside the Claude Code session. There is no cron job; the timer stops when the session closes or the computer sleeps.
  • Ping: when the chat is idle and the cache is 2 minutes from expiry, the mod calls $.model.fork. The fork sends the session's last request again with a one-line prompt. The API reads the conversation from the cache, and that read starts the cache lifetime again. The fork runs no tools and does not show in the transcript.
  • Hits and misses: after each turn, the mod reads the token counts of the turn's first request. If less than half of the input came from the cache, the turn counts as a miss.

Why it is worth it

Prompt caching prices, as a multiple of the normal input price:

RequestPrice
Cache write, 1-hour cache2×
Cache write, 5-minute cache1.25×
Cache read0.1× (Opus 5.5: 0.05×, Fable 5.1: 0.025×)

Example: a 400k-token chat on Opus 5.5 with the 1-hour cache.

  • The cache expires: your next prompt writes 400k tokens at 2×, the cost of 800k input tokens.
  • The cache is warm: your next prompt reads 400k tokens at 0.05×, the cost of 20k input tokens.
  • One ping: the same read, 20k.

A ping costs 1/40 of a cold restart. Four pings cost 80k, one tenth of a restart, even if you never come back to the chat. The keepalive pays for itself when there is a better than 1-in-40 chance that you come back within the next hour. At the general read price of 0.1×, the break-even is 20 pings.

Limits on the pings

  • Cap: at most maxPings pings (default 4) per idle stretch. A prompt you type resets the count.
  • Rate limit: no pings while the 5-hour usage window is 80% full or more.
  • Expired cache: no ping after the cache expired, for example after the computer slept. A ping then would pay a full cache write.
  • Missed ping: when a ping does not read from the cache, the mod stops pinging until your next turn.
  • Failed ping: an API error does not count as a refresh. The next tick tries again.

Settings and command

SettingValuesDefault
ttl1h, 5m1h
maxPingsa number4

Change them in /config. Claude Code sessions use the 1-hour cache on most plans. Use 5m for the API default, or when your account runs in usage overage.

/cache-warm status shows the cache state, /cache-warm off stops the pings, /cache-warm on starts them again.

week-calendar

week-calendar

The screenshot shows a real week with every title, project and commit replaced.

/week builds a calendar of this week's Claude Code and Codex sessions, opens it in your browser, and prints a 3-line report:

Shipped: rate limiting, CI runner migration, deploy runbook
Most time: api-service, 22h 27m
Next: Prototype caching layer, Upgrade database driver

/week last shows last week; /week -2 goes two weeks back.

What the calendar shows

  • Sessions as blocks, one colour per project. A session splits into blocks at gaps longer than 30 minutes.
  • Commits you made during each block, matched by your git config user.email.
  • Sessions that ended without a commit: a dashed outline and an amber dot. The sidebar lists them, longest first.
  • Active time: hours per project and per day. Sessions that run in parallel count once.
  • Details: click a block to see its time, model, commits, and the first message you typed.

How it works

  • bin/build.mjs reads ~/.claude/projects/**/*.jsonl and ~/.codex/sessions/**/*.jsonl, runs git log in each project, and writes one HTML file to ~/.calendar/week-<monday>.html. Node only, no network.
  • The hook sends the week's commit subjects to Sonnet once to write the "Shipped" line.
  • The HTML file holds all its data inline and loads nothing from the network.

The calendar file holds the first message of each session as plain text. It stays on your machine.

Settings

~/.calendar/config.json is created on the first run:

SettingDefaultWhat it does
dayStartHour6Work before this hour counts toward the previous day
gapMinutes30A gap longer than this starts a new block
minNoCommitMinutes15Shorter blocks are not flagged as ending without a commit
weekStartmonmon or sun
themedarkdark or light
accent, paletteColours for today, the busiest day, and the projects

It needs Node and git. It opens the file with open on macOS or xdg-open on Linux.

diagram-mod

diagram-mod

Ask Claude for a diagram or a visual explanation and it calls the mod's show_diagram tool. The diagram draws in the tool's row in the transcript:

  • Boxes: a label, detail lines and a status, with dashed borders in the colour of their role. A status has a tone: a spinner while running, ✓ ok, ✗ error, ! warning.
  • Edges: dashed arrows with optional labels. The layout runs top to bottom and is computed from the edges; Claude never gives coordinates. Edges that point back up, or skip a row, run along the right margin.
  • Packets: braille "comets" with a fading tail, moving a quarter cell at a time.
  • Gauges: bars (meters) that fill in eighths of a cell, and a moving activity wave (spark).
  • Side panels: a node with side: "left" or "right" becomes a tall panel beside the tree, with rows a step can highlight.
  • Frame: a legend under the title, a caption under the diagram, a log table (time, actor, message, status) and a footer of counters.

Stories

Give the spec steps and the diagram plays them in a loop. Each step names the edges carrying traffic (only those light up), changes node statuses, meters and panel highlights, adds log rows, and moves the counters. Meters ease to their new values and counters count up across the step. Without steps, packets loop on the edges you name.

examples/agent-tree.json is the story in the GIF; examples/dispatcher.json is a simple diagram without steps. The bundled skill tells Claude when to use the tool and how to keep a diagram readable: one idea, about 12 nodes at most, labels of 1-3 words, 3-6 steps.

How it animates

  • About 60 frames a second, repainting one cell grid in place; frames that would not change are skipped.
  • Only the newest diagram moves; older ones keep a still frame.
  • The animation waits while Claude is still writing its reply, so the streaming text stays fast, and pauses while the row is scrolled out of view or folded.
  • A terminal narrower than the diagram gets a still text version instead.

Settings

SettingDefaultWhat it does
displayinlineinline: the diagram animates in the transcript row. pane: it opens in a side pane, closed with q or Esc (Claude Code only seats a pane it did not ask for from 144 columns)

Development

Each mod is a folder:

<mod>/
  .claude-plugin/plugin.json   name, version, description, types
  hooks/hooks.json             { "modules": ["./register.tsx"] }
  hooks/register.tsx           the hooks
  types/index.d.ts             the state the mod keeps
  tests/*.test.tsx             tests

Check and test a mod:

claude plugin validate ./slick-bar
claude plugin test ./slick-bar

Claude Code writes the API types into <mod>/.claude-plugin/types/ the first time it loads a mod; those files are not committed. A session started with --plugin-dir reloads a mod when you save its files.

Limits

These come from the mods API, not from the mods:

  • Permission mode: mods cannot read it, so the bar sits after Claude Code's own mode label.
  • Space above the prompt: the empty row between the transcript and the prompt belongs to Claude Code.
  • Ctrl+o: tool rows do not report the expanded view (see Expanding output).
  • Cache lifetime: mods cannot read which cache lifetime the session uses, so cache-warm takes it from the ttl setting.
  • Tool names: Claude Code lists a tool a mod registers as mcp__<mod>__<tool>, so diagram-mod's tool is mcp__diagram-mod__show_diagram though no MCP server is involved.

License

MIT

Source 2 files
hooks/register.tsx 176 lines
1import { atom, read, update } from 'claude-code'
2import type { EngineInterface, ModelUsage, Register } from 'claude-code'
3
4import type { Health } from '../types'
5
6const TTL_MS: Record<string, number> = { '1h': 3_600_000, '5m': 300_000 }
7
8const health = atom({ plugin: 'cache-warm', key: 'health' } as const, {
9  ttlMs: TTL_MS['1h']!,
10  marginMs: 120_000,
11  maxPings: 4,
12  lastAt: 0,
13  isWarm: false,
14  isEnabled: true,
15  lastRatio: null,
16  hits: 0,
17  misses: 0,
18  pings: 0,
19  pingsTotal: 0,
20  note: '',
21} satisfies Health)
22
23const TICK_MS = 15_000
24// Below this share of the input served from cache, a request counts as a miss.
25const HIT_RATIO = 0.5
26// Requests this small carry no prefix worth caching.
27const MIN_TOKENS = 4_000
28// No pings while the 5h rate-limit window is this full.
29const RATE_LIMIT_PCT = 80
30const PING = 'Keepalive ping, not a task. Reply with the single character: .'
31
32type $ = EngineInterface
33
34function ratio(u: ModelUsage) {
35  const total = u.input_tokens + u.cache_read_input_tokens + u.cache_creation_input_tokens
36  return total >= MIN_TOKENS ? u.cache_read_input_tokens / total : null
37}
38
39// Module variables reset on a reload, which only happens between turns.
40let isBusy = false
41let isPinging = false
42
43async function ping($: $) {
44  isPinging = true
45  try {
46    // The TTL counts from the start of the request that reads the entry.
47    const sentAt = await $.clock.now()
48    const r = await $.model.fork({ prompt: PING })
49    if (!r.isAnswered && r.reason === 'nothing-to-fork') {
50      await update($, health, h => ({ ...h, isWarm: false, note: 'nothing to keep warm' }))
51      return
52    }
53    if (!r.isAnswered) {
54      // The cache was not touched: leave lastAt alone so the next tick retries inside the margin.
55      await update($, health, h => ({ ...h, note: `ping failed: ${r.reason}` }))
56      return
57    }
58    const share = ratio(r.usage)
59    if (share !== null && share < HIT_RATIO) {
60      // The entry had lapsed: this ping paid a full write. Stop until a real turn.
61      await update($, health, h => ({ ...h, isWarm: false, misses: h.misses + 1, note: 'ping missed' }))
62      return
63    }
64    await update($, health, h => ({
65      ...h,
66      lastAt: sentAt,
67      pings: h.pings + 1,
68      pingsTotal: h.pingsTotal + 1,
69      note: 'pinged',
70    }))
71  } finally {
72    isPinging = false
73  }
74}
75
76async function tick($: $) {
77  if (isBusy || isPinging) return
78  const h = await read($, health)
79  if (!h.isEnabled || !h.isWarm || h.lastAt === 0) return
80  const idle = (await $.clock.now()) - h.lastAt
81  // A late timer (sleep, suspend) never pings a cold cache: that would pay a full write.
82  if (idle >= h.ttlMs) {
83    await update($, health, x => ({ ...x, isWarm: false, note: 'expired' }))
84    return
85  }
86  if (idle < h.ttlMs - h.marginMs) return
87  if (h.pings >= h.maxPings) {
88    if (h.note !== 'ping cap reached') await update($, health, x => ({ ...x, note: 'ping cap reached' }))
89    return
90  }
91  const fiveHour = (await $.session.usage()).rateLimits.find(l => l.kind === 'five_hour')
92  if (fiveHour !== undefined && fiveHour.percentUsed >= RATE_LIMIT_PCT) {
93    if (h.note !== 'rate limit high') await update($, health, x => ({ ...x, note: 'rate limit high' }))
94    return
95  }
96  await ping($)
97}
98
99export const register: Register = (on, options) => {
100  const ttlMs = TTL_MS[String(options.ttl)] ?? TTL_MS['1h']!
101  // 30s on a 5m cache, 2m on a 1h one: a fork of a large prefix needs the slack.
102  const marginMs = Math.min(120_000, ttlMs / 10)
103  const maxPings = typeof options.maxPings === 'number' ? options.maxPings : 4
104
105  on('session.start', async ($, e, next) => {
106    const result = await next(e)
107    await update($, health, h => ({ ...h, ttlMs, marginMs, maxPings }))
108    await $.command.register({
109      name: 'cache-warm',
110      description: 'Prompt cache keepalive: /cache-warm on | off | status',
111    })
112    $.clock.every(TICK_MS, () => void tick($).catch(() => undefined))
113    return result
114  })
115
116  on('command.run', { command: 'cache-warm' }, async ($, e) => {
117    const arg = e.args.trim()
118    if (arg === 'on' || arg === 'off') await update($, health, h => ({ ...h, isEnabled: arg === 'on' }))
119    const h = await read($, health)
120    const idle = h.lastAt > 0 ? Math.round(((await $.clock.now()) - h.lastAt) / 1000) : null
121    const parts = [
122      `keepalive ${h.isEnabled ? 'on' : 'off'}`,
123      `ttl ${h.ttlMs / 60_000}m`,
124      h.isWarm ? 'warm' : 'cold',
125      idle !== null ? `idle ${idle}s` : 'no request yet',
126      `hits ${h.hits} · misses ${h.misses}`,
127      `pings ${h.pings}/${h.maxPings} (session ${h.pingsTotal})`,
128      h.note,
129    ]
130    return { text: parts.filter(Boolean).join(' · ') }
131  })
132
133  on('prompt.submit', async ($, e, next) => {
134    isBusy = true
135    await update($, health, h => ({ ...h, pings: 0 }))
136    return next(e)
137  })
138
139  // Every main-thread request refreshes the entry; the first of a turn says whether it survived the gap.
140  on('turn.step', async function* ($, e, next) {
141    // Turns start without a typed prompt too (task notices, queued commands).
142    if (e.agentId === undefined && !isPinging) isBusy = true
143    const sentAt = await $.clock.now()
144    const result = yield* next(e)
145    if (e.agentId !== undefined || isPinging || result.usage === null) return result
146    const share = ratio(result.usage)
147    await update($, health, h => {
148      // The session's first request has nothing to hit.
149      const counts = e.index === 0 && share !== null && h.lastAt > 0
150      return {
151        ...h,
152        lastAt: sentAt,
153        isWarm: true,
154        lastRatio: share ?? h.lastRatio,
155        hits: h.hits + (counts && share >= HIT_RATIO ? 1 : 0),
156        misses: h.misses + (counts && share < HIT_RATIO ? 1 : 0),
157        note: '',
158      }
159    })
160    return result
161  })
162
163  on('turn.complete', async ($, e, next) => {
164    const result = await next(e)
165    if (e.agentId === undefined) isBusy = false
166    return result
167  })
168
169  // A compacted transcript is a new prefix: the next request writes it afresh.
170  on('session.compact', async ($, e, next) => {
171    const result = await next(e)
172    await update($, health, h => ({ ...h, isWarm: false, note: 'compacted' }))
173    return result
174  })
175}
176
types/index.d.ts 29 lines
1export type Health = {
2  /** Cache lifetime and how long before expiry a ping goes out, in ms. */
3  ttlMs: number
4  marginMs: number
5  /** Pings per idle stretch before the cache is left to expire. */
6  maxPings: number
7  /** When the last request that read or wrote the cache finished; 0 before the first. */
8  lastAt: number
9  isWarm: boolean
10  isEnabled: boolean
11  /** Share of the last main-thread request's input the cache served, 0..1. */
12  lastRatio: number | null
13  hits: number
14  misses: number
15  /** Pings in the current idle stretch, and over the session. */
16  pings: number
17  pingsTotal: number
18  /** Why pinging stopped or what the last ping did. */
19  note: string
20}
21
22declare module 'claude-code' {
23  interface PluginState {
24    'cache-warm': {
25      health: Health
26    }
27  }
28}
29