SLOPSHOPPER

task-band

Framed band above the prompt: running subagents, workflows, background shells and monitors, each with its own gauge; ctrl+x tab then arrows to page

newbandguardprompttimeragents
A shopper browsing a rack in a slop shop
README

Claude Code mods: field notes from day one (Windows · Thai)

One long session of building Claude Code mods on Windows 11, written up honestly: what shipped, what I tried and threw away, and why. Most mod repos only show the finished screenshot. This one also keeps the failures, because they say more about the mod API than the wins do.

Status: experimental, day one. Built on Claude Code 2.1.291 and still loading on 2.1.292. Tested on one machine: Windows 11, Windows Terminal 1.24, Cascadia Code, a cream background (#F2F0DA), 134 columns, fullscreen mode. The mod API is new and may change under you. Issues welcome.

PieceStatusOne-line lesson
fuel-barin daily useThe engine gives you context, auto-compact threshold and rate limits for free after every turn. No polling, no tokens.
thai-modein daily useYou can re-render tool rows, groups, spinner and turn duration. You cannot touch the prompt box.
task-bandnew, day twoThe engine tells a mod when an agent ends right away, but shell/monitor/workflow ends only reach it at the next tool call while Claude is busy. And there is no progress % for anything.
paste-peeknew, day threeThe mod API can draw images only through the kitty protocol, which Windows Terminal lacks. A tiny native window beside the terminal does the job, and Claude Code already saves each pasted image to disk before you send.
Matrix introin daily useA sequential intro always leaves a blank gap. Run it in parallel and stop on a file Claude writes before its first frame.
Safe updaterin daily useAuto-update on Windows can leave claude.exe as a 500-byte stub. Stage, verify, then rename-and-swap.
Right-side pane + widgetsdroppedThe dock frame belongs to the engine, and Thai text breaks in the Windows Terminal grid with every font I tried.
project-band above the promptdroppedEngine notifications render between your band and the prompt, and you cannot intercept them.
matrix-rain as a moddroppedsession.start fires after the first frame. Too late for a loading screen.
Rounded prompt boxnot possibleThe prompt input is not a render site.
Context Router, prismantisdeclinedAlready built in / clashes with thai-mode and adds tokens per prompt.

fuel-bar: context + quota footer

fuel-bar

Two lines under the prompt:

  1. T77 │ Opus 5.5 · medium │ ━━━━━━──── 53.7% 519.3k/1M │ 5h 26.0% (1h57m) · 7D 47.0% (3d) turn · model · effort · gauge to the auto-compact point (not to the raw window) · 5-hour and 7-day quota with time to reset.
  2. PROJECT · ✦ 24.2° night rain 39% · worked 4:32 hr, plus the latest mod toast on the right for 8 seconds.

Why I built it: I was using ccusage statusline, and on a 1M-context model it showed context above 100% (147% on my screen). The engine already knows the right numbers.

How it works:

  • session.measure fires after every turn with context.tokens, context.window and rateLimits[] (five_hour, seven_day, each with percentUsed and resetsAt). session.usage({ breakdown: 'summary' }) adds autoCompactThreshold. Zero tokens, no process spawned.
  • turn.step carries the current effort.
  • It draws into the PromptHint render site. If you return two lines, the engine keeps its own mode pill (⏵⏵ auto mode on) on the first line and indents the second by the pill's width. I tried drawing my own pill. The engine's pill stays anyway, so I gave up and kept theirs.
  • On narrow terminals it drops, in order: the hint text, the reset times, the 5h/7D block, then wraps to two lines. The token count always stays.
  • Weather: ipwho.is for a rough location, then api.open-meteo.com. No keys, cached 30 min. This sends your IP to ipwho.is. Delete loadWeather if you don't want that.
  • ui.toast can be intercepted (return { value: undefined }), so mod toasts land at the end of line 2 instead of popping over the transcript.
  • dependencies: ["thai-mode"] in plugin.json lets fuel-bar read thai-mode's state and translate the hint itself. Two mods never fight over the same render site.

Colors are hand-picked for a light cream background. On a dark theme you'll want to change the C palette in register.tsx.

thai-mode: Claude Code in Thai

thai-mode

Rewrites, with a built-in dictionary (no model calls):

  • tool rows (ToolUse): Read → อ่านไฟล์, Bash → รันคำสั่ง, and MCP tools as service · Thai verb
  • results (ToolResult): "read 120 lines (40–159 of 300)", "added 4 · removed 1 lines", with the diff still drawn by the engine's Code component
  • collapsed groups (ToolGroup): กำลังเขียน 1 ไฟล์ · แก้ 1 ไฟล์ · รัน 1 คำสั่ง…
  • the spinner word by real phase (requesting / thinking / responding / tool-input / tool-use)
  • turn duration and the ctrl+b hint

/thai toggles it, and the choice persists in $.store.

Things I learned:

  • In this version even a single Edit gets folded into a ToolGroup, so your nice single-row rendering mostly shows up only in ctrl+o.
  • Command output and diffs stay as they are. Translating those would be lying about what ran.

task-band: background work above the prompt

task-band

A framed band above the prompt (AbovePrompt) titled "งานเบื้องหลัง" (background work). Each running subagent, Workflow, background shell and Monitor gets one chip with its own gauge, 3 per page:

  • Workflow: agents finished / agents started (1/2) plus elapsed time.
  • Single agent: an estimated ~62% once this agent type has finished at least 3 times (average kept in $.store), otherwise a sweeping line plus elapsed time.
  • Shell / Monitor: sweeping line plus elapsed time. Nothing reports real progress.
  • Running chips come first, finished ones (✓ / ✗) after them, and finished chips stay until the turn ends. Within each group chips keep the order you started them in.
  • When a task starts or ends, the band jumps to its page for 3 seconds. Ctrl+X then Tab focuses the band, ← / → page, Esc goes back to the prompt.

What the API gives you (build 2.1.292):

  • agent.spawn answers with an agentId. Its end is turn.complete carrying that agentId, and it arrives immediately. Workflow agents come through agent.spawn too, with e.workflow.runId.
  • Background Bash/PowerShell results carry backgroundTaskId. Monitor and Workflow results carry taskId (Workflow also runId and workflowName).
  • Their end arrives as prompt.submit with origin.kind === 'task-notification', with <task-id> and <status> in the text. When Claude is idle that is instant. While a turn runs it waits for the next tool-call boundary (I measured 4.3 s late).
  • Results of tool calls sent in parallel come back in any order. To keep chips in the order you started them, take the order key when the hook is entered, before await next(e).

Things that bit me:

  • Box takes borderStyle but has no border title. Laying a position: "absolute" title over the border line shifted the whole screen sideways and left it garbled until a resize forced a full redraw. The frame is now three plain Text rows sized to bodyColumns (the engine draws [-] at the far right of the band).
  • Windows Terminal gives Thai above/below marks zero cells, the same as the engine does. Don't pad widths to "fix" the mis-spaced look. It only misaligns the frame.
  • Engine notices still render between the band and the prompt (the reason project-band was dropped). Here it matters less, because the band is only there while something runs.

paste-peek: see the image you just pasted (Windows)

paste-peek

Windows only. Experimental, day one. Claude Code shows a pasted image as [Image #1] and nothing else, so you can't tell whether you attached the right screenshot until Claude answers. paste-peek opens a small strip in the bottom-right corner of your Windows Terminal window with a thumbnail of every [Image #N] in your draft:

  • Click a thumbnail to enlarge it, click again to shrink. The strip never takes focus, so you keep typing.
  • Delete [Image #2] from the draft and #2 leaves the strip. Send the message and the strip closes.
  • Works for Alt+V pastes and for files dragged onto the terminal.
  • Stays with its own terminal window: it follows moves, hides when you switch to another app or minimize, and comes back when you return.

Why a separate window: the mod API has an Image element, but the engine draws it only through the kitty graphics protocol, which Windows Terminal doesn't speak, so you just get the alt text. I also tried a Raster of half-block characters (▀, two pixels per cell). It works everywhere, but a 6-row thumbnail is about 21×12 pixels, which is too blurry to tell screenshots apart.

How it works:

  • The mod reads the draft ($.prompt.read()) every 250 ms and writes the list of [Image #N] numbers to a small state file in %TEMP%.
  • On the first image it starts peek.exe (C#, WinForms, about 475 lines in native/peek.cs). The exe watches the state file with a FileSystemWatcher and closes itself when the file goes empty, which the mod does on submit.
  • Where the pictures come from: Claude Code writes every attached image to %TEMP%\claude\<project-slug>\<session-id>\images\<N>.<ext> as soon as you paste or drop it, before you send. This is undocumented and may change in a future version. Claude Code skips this when transcript saving is off (for example a session started from inside another Claude session, which shows a "Transcript saving is off" notice), and the strip then says "no preview". My first build read the clipboard instead, and a dragged file then showed whatever image was last copied.
  • The exe ties itself to the Windows Terminal window that was in front when you pasted, and listens for EVENT_SYSTEM_FOREGROUND and location changes. No polling.

Measured on my laptop: about 24 MB private memory and 0% CPU while the strip is open, nothing at all while there are no images, and the strip appears in about 0.2 s.

No binary in this repo. On first use the mod compiles native/peek.cs with the C# compiler that ships with Windows (%WINDIR%\Microsoft.NET\Framework64\v4.0.30319\csc.exe, .NET Framework 4.x) into %LOCALAPPDATA%\paste-peek\. That takes about a second, once.

Options (/config): label language (en / th), theme (dark / light), and the distance from the window's bottom and right edges. The default bottom distance clears a two-line footer. Raise it if the strip covers your prompt.

Tested end to end in a fresh session: first-use compile, dark theme, English labels, a pasted image showing in the strip. Not handled yet: more than one terminal tab in the same window (the strip follows the window, not the tab), and terminals other than Windows Terminal.

Matrix boot intro

matrix-intro

Katakana rain in ink-on-cream with the folder name decoding in the middle, shown while Claude Code loads (about 4.7 s on this laptop). It is not a mod. It is a tiny C# exe started by the launcher.

The path there was three failures:

  1. Intro first, then claude. However long the intro runs, you still get a blank or log-only screen between the intro ending and Claude's first frame.
  2. Intro as a mod. session.start fires after Claude has drawn its first screen. The rain lands on top of Claude, and in fullscreen mode Ctrl+L does not clear it.
  3. Intro in PowerShell. pwsh takes ~0.9 s to start and ~1.5 s to set up, so the first second is blank anyway.

What works:

  • claude-launcher.cmd does start "" /b matrix-intro.exe and then call claude at the same moment. The rain shows within ~0.7 s.
  • The exe polls numStartups in <CLAUDE_CONFIG_DIR>\.claude.json every 120 ms. Claude bumps it 0.4–0.7 s before drawing its first frame, which is the "ready" signal. Hard cap: 20 s.
  • Claude's fullscreen mode uses the alternate screen, so the rain on the main screen is hidden behind it. After Claude exits, the launcher runs cls, but only if the intro actually ran (it drops a marker file).
  • The intro is skipped for -p, --version, --help and subcommands. CLAUDE_MATRIX_INTRO=0 disables it.

Batch trap: claude installed by npm is claude.cmd. If you write claude %* instead of call claude %* in a .cmd, nothing after that line ever runs.

Half-width katakana take exactly one cell in Windows Terminal with Cascadia Code, so the columns line up.

Build it with the C# compiler that ships with Windows:

C:\Windows\Microsoft.NET\Framework64\v4.0.30319\csc.exe /nologo /optimize+ /codepage:65001 windows\matrix-intro.cs

Safe updater for the npm install

On this machine, Claude Code's own auto-update repeatedly left node_modules\@anthropic-ai\claude-code\bin\claude.exe as a ~500-byte stub ("This version of claude.exe is not compatible…"), or failed with "claude.exe in use" while another session was open. Its warning also showed up in the middle of the screen:

engine notice

windows/claude-update-safe.ps1:

  1. Takes a global mutex, so two launches don't race.
  2. Compares the installed version with npm view @anthropic-ai/claude-code version.
  3. Installs the new version into a staging folder, picks the real binary (≥10 MB), and checks --version.
  4. Renames the live exe to claude.exe.old-<timestamp> (Windows allows renaming a running exe, not overwriting it), then moves the new one in. Open sessions keep running the old file. New sessions get the new one.
  5. Cleans up old .old-* files on the next run, and logs to %LOCALAPPDATA%\claude-update-safe\claude-update.log.

Set "env": { "DISABLE_AUTOUPDATER": "1" } in settings.json so the built-in updater stops fighting it. matrix-intro.exe launches the script hidden at every start if it sits in the same folder.

This only applies to the global npm install. The native installer manages its own versioned folders.

Tried and dropped

Right-side pane with widgets (project name, a 3-tab reminder list, weather, other open sessions). Inspired by paneline and flightdeck. It worked technically (Pane render site with fullscreen: true), but:

  • The dock frame and its colors belong to the engine and follow the Claude Code theme, not yours.
  • Thai text in Windows Terminal is mis-spaced: combining vowels and tone marks get their own cell (เ ปิด). I tried Noto Sans Thai, Leelawadee, Tahoma, Tlwg Mono and Tlwg Typo. None fixed it. This is a WT grid limit, not a font problem.
  • While the dock is open, viewport.columns in PromptHint is the transcript width, not the full window.
  • Below 144 columns the pane doesn't open by itself. On a 134-column window you have to open it once with /pane.

project-band above the prompt (AbovePrompt, one line: project · weather · time worked). It looked good until Claude Code showed its own notice (Auto-update failed …). Engine notices render between AbovePrompt and the prompt, and they are standing warnings, not ui.toast, so a mod can't catch them. I merged the band into fuel-bar's second line instead.

matrix-rain as a mod. See the intro section: session.start is too late.

Rounded prompt box. I mocked it in a browser first. That was a mistake, because the prompt input is not one of the render sites, so a mod can't change it. A custom theme can recolor the border (promptBorder) but not reshape it.

Considered and declined

  • Context Router (load different context per folder). Claude Code already does this: it loads CLAUDE.md from parent folders, and @path imports work.
  • prismantis. Its default theme is dark. It also re-renders ToolUse, ToolGroup and TurnDuration (the same sites as thai-mode), and its diagramHints adds about 190 tokens to every prompt.

A warning about mod screenshots

The HUD in one popular mod's demo video (LIFE / LEVEL / XP bars across the top) is rendered with Revideo in its demo/ folder. The README says as much: "Nothing in it is captured". It isn't something the mod draws in your terminal. Before you promise someone a feature you saw in a GIF, grep the mod's hooks/ folder.

Testing on the real terminal (Windows)

Mockups in a browser lie about fonts, colors and cell widths. What I used instead:

  • Launch a fresh window with wt -w new <file.cmd>. Start-Process with arguments quotes the whole command line and fails with 0x80070002.
  • Find the window by class CASCADIA_HOSTING_WINDOW_CLASS. Capture it with PrintWindow(hwnd, dc, 2) after SetProcessDPIAware(). That works even when the window is behind another one. CopyFromScreen only sees what's visible.
  • Type with SendInput + KEYEVENTF_UNICODE. SendKeys follows the active keyboard layout: with Thai active, /exit arrives as /59.

Install

claude plugin marketplace add <path-to-clone>\mods
claude plugin install fuel-bar@field-notes-mods --scope user
claude plugin install thai-mode@field-notes-mods --scope user
claude plugin install task-band@field-notes-mods --scope user
claude plugin install paste-peek@field-notes-mods --scope user

Plugins are read from that folder in place: edit a file, then run /reload-plugins. claude plugin test <folder> runs the tests (fuel-bar 7, thai-mode 5, task-band 13, paste-peek 4).

For the intro: compile windows/matrix-intro.cs, then put matrix-intro.exe, claude-update-safe.ps1 and claude-launcher.cmd in one folder on your PATH.


ภาษาไทย (สรุป)

บันทึกการทำ mod ให้ Claude Code บน Windows ในหนึ่งวัน ทั้งชิ้นที่ใช้จริงและชิ้นที่ลองแล้วทิ้ง

  • fuel-bar: แถบ 2 บรรทัดใต้ช่องพิมพ์ แสดง turn, โมเดล, effort, เกจ context (นับถึงจุด auto-compact), โควตา 5 ชม./7 วัน, ชื่อโปรเจกต์, อากาศ และเวลาที่ทำงานมาแล้ว ข้อมูลมาจาก engine หลังจบแต่ละ turn จึงไม่กิน token
  • thai-mode: แปลแถวเครื่องมือ, แถวสรุปที่พับไว้, spinner และเวลาที่ใช้ต่อ turn เป็นภาษาไทยด้วยพจนานุกรมในตัว /thai ใช้สลับเปิด/ปิด
  • task-band: กรอบ "งานเบื้องหลัง" เหนือช่องพิมพ์ แสดง agent, Workflow, คำสั่งเบื้องหลัง และ Monitor ทีละ 3 งาน แต่ละงานมีเกจของตัวเอง งานที่รันอยู่ขึ้นก่อน กด Ctrl+X แล้ว Tab เพื่อเลื่อนหน้าด้วยลูกศร
  • paste-peek (Windows): กรอบรูปเล็กลอยที่มุมขวาล่างของ Windows Terminal แสดงรูปทุก [Image #N] ที่วางหรือลากมาไว้ในข้อความ ก่อนกดส่ง คลิกเพื่อขยาย ลบ [Image #N] แล้วรูปหายจากกรอบ กดส่งแล้วกรอบปิดเอง ซ่อนเมื่อสลับไปหน้าต่างอื่น ใช้ RAM ประมาณ 24 MB เฉพาะตอนมีรูปแนบ ไม่แนบไฟล์ exe ไว้ใน repo แต่จะคอมไพล์จากโค้ดต้นฉบับตอนใช้ครั้งแรก
  • Matrix intro: แสดงฝนตัวอักษรระหว่างรอ Claude โหลด ต้องรันขนานกับ Claude และหยุดเมื่อ numStartups เพิ่ม ถ้ารันเรียงกันจะมีจอว่างเสมอ
  • ตัวอัปเดตปลอดภัย: แก้ปัญหา auto-update ทำให้ claude.exe เหลือไฟล์ 500 ไบต์ ใช้วิธีติดตั้งแยกไว้ก่อน ตรวจว่าใช้ได้ แล้วค่อยเปลี่ยนชื่อสลับไฟล์
  • ที่ลองแล้วไม่เวิร์ค:
  • แผงด้านขวา: กรอบเป็นของ engine และสระ/วรรณยุกต์ไทยเพี้ยนใน Windows Terminal ทุกฟอนต์ที่ลอง
  • แถบเหนือช่องพิมพ์: แจ้งเตือนของ Claude Code ขึ้นมาคั่นกลาง และ mod ดักไม่ได้
  • ฝน Matrix แบบ mod: session.start มาช้ากว่าจอแรก
  • กรอบช่องพิมพ์มุมมน: API ไม่เปิดให้แก้

ถ้าใช้ Windows Terminal กับภาษาไทย ให้เตรียมใจไว้ว่าสระบนและวรรณยุกต์จะกินช่องของตัวเอง ฟอนต์แก้ไม่ได้

License

MIT

Source 3 files
hooks/register.tsx 229 lines
1import { atom, read, update } from 'claude-code'
2import type { EngineInterface, Register } from 'claude-code'
3
4import type { BandJump, BandTask, TaskStatus } from '../types'
5import { average, C, cellWidth, chipSegs, clampPage, clearFinished, insertOrdered, navLabel, pageOf, paginate, parseNotifications, pushSample, segWidth, statusOf, viewOrder, frameTop, frameBottom, rowPad } from './logic'
6
7const tasks = atom({ plugin: 'task-band', key: 'tasks' } as const, [] as BandTask[])
8const page = atom({ plugin: 'task-band', key: 'page' } as const, 0)
9const jump = atom({ plugin: 'task-band', key: 'jump' } as const, null as BandJump | null)
10const tick = atom({ plugin: 'task-band', key: 'tick' } as const, 0)
11
12const JUMP_MS = 3000
13const FRAME = 4
14let pageCount = 1
15let seq = 0
16type Durations = Record<string, number[]>
17
18async function now($: EngineInterface): Promise<number> {
19  return $.clock.now()
20}
21
22async function entryOrder($: EngineInterface): Promise<number> {
23  const s = ++seq % 1000
24  return (await now($)) + s / 1000
25}
26
27async function flash($: EngineInterface, id: string) {
28  const until = (await now($)) + JUMP_MS
29  await update($, jump, () => ({ id, until }))
30  $.clock.after(JUMP_MS + 50, async () => {
31    const j = await read($, jump)
32    if (j && j.until <= (await now($))) await update($, jump, () => null)
33  })
34}
35
36async function addTask($: EngineInterface, t: BandTask) {
37  await update($, tasks, list => insertOrdered(list ?? [], t))
38  await flash($, t.id)
39}
40
41async function finish($: EngineInterface, match: (t: BandTask) => boolean, status: TaskStatus) {
42  const at = await now($)
43  let hit: BandTask | undefined
44  await update($, tasks, list =>
45    (list ?? []).map(t => {
46      if (t.status !== 'running' || !match(t)) return t
47      hit = { ...t, status, endedAt: at }
48      return hit
49    }),
50  )
51  if (!hit) return
52  await flash($, hit.id)
53  if (hit.kind === 'agent' && status === 'done' && hit.agentType) {
54    const all = ((await $.store.get('durations')) as Durations | undefined) ?? {}
55    all[hit.agentType] = pushSample(all[hit.agentType], at - hit.startedAt)
56    await $.store.set('durations', all)
57  }
58}
59
60async function estimateFor($: EngineInterface, agentType: string): Promise<number | undefined> {
61  const all = (await $.store.get('durations')) as Durations | undefined
62  return average(all?.[agentType])
63}
64
65async function onTick($: EngineInterface) {
66  const list = await read($, tasks)
67  if (list && list.some(t => t.status === 'running')) await update($, tick, n => (n + 1) % 1_000_000)
68}
69
70export const register: Register = on => {
71  on('session.start', async ($, e, next) => {
72    $.clock.every(1000, () => onTick($))
73    return next(e)
74  })
75
76  on('agent.spawn', async ($, e, next) => {
77    const order = await entryOrder($)
78    const r = await next(e)
79    const agentId = (r as { agentId?: string }).agentId
80    if (!agentId) return r
81    const at = await now($)
82    if (e.workflow) {
83      const runId = e.workflow.runId
84      const list = (await read($, tasks)) ?? []
85      if (!list.some(t => t.runId === runId)) {
86        await addTask($, { id: `run:${runId}`, kind: 'wf', label: 'workflow', startedAt: at, order, status: 'running', runId, members: [], finished: [] })
87      }
88      await update($, tasks, l => (l ?? []).map(t => (t.runId === runId ? { ...t, members: [...(t.members ?? []), agentId] } : t)))
89      return r
90    }
91    const agentType = e.subagentType || 'general-purpose'
92    await addTask($, { id: agentId, kind: 'agent', label: e.description || agentType, agentType, startedAt: at, order, status: 'running', estMs: await estimateFor($, agentType) })
93    return r
94  })
95
96  on('tool.call', async ($, e, next) => {
97    const bg = e.tool === 'Bash' || e.tool === 'PowerShell' || e.tool === 'Monitor' || e.tool === 'Workflow'
98    const order = bg && !e.agentId ? await entryOrder($) : undefined
99    const r = await next(e)
100    if (e.agentId || !('result' in r) || !r.result) return r
101    const res = r.result as Record<string, unknown>
102    const at = await now($)
103    if (e.tool === 'Bash' || e.tool === 'PowerShell') {
104      const bgId = res.backgroundTaskId
105      if (typeof bgId === 'string') {
106        const label = (e.description as string | undefined) || String(e.command).split('\n')[0] || e.tool
107        await addTask($, { id: bgId, taskId: bgId, kind: 'sh', label, startedAt: at, order, status: 'running' })
108      }
109    } else if (e.tool === 'Monitor') {
110      const id = res.taskId
111      if (typeof id === 'string') await addTask($, { id, taskId: id, kind: 'mon', label: String(e.description || 'monitor'), startedAt: at, order, status: 'running' })
112    } else if (e.tool === 'Workflow') {
113      const id = res.taskId
114      const runId = typeof res.runId === 'string' ? res.runId : undefined
115      const label = typeof res.workflowName === 'string' ? res.workflowName : 'workflow'
116      if (typeof id !== 'string') return r
117      const list = (await read($, tasks)) ?? []
118      if (runId && list.some(t => t.runId === runId)) {
119        await update($, tasks, l => (l ?? []).map(t => (t.runId === runId ? { ...t, taskId: id, label } : t)))
120      } else {
121        await addTask($, { id, taskId: id, kind: 'wf', label, startedAt: at, order, status: 'running', runId, members: [], finished: [] })
122      }
123    }
124    return r
125  })
126
127  on('turn.complete', async ($, e, next) => {
128    const r = await next(e)
129    const agentId = e.agentId
130    if (!agentId) {
131      await update($, tasks, l => clearFinished(l ?? []))
132      return r
133    }
134    const list = (await read($, tasks)) ?? []
135    const wf = list.find(t => t.status === 'running' && t.members?.includes(agentId) && !t.finished?.includes(agentId))
136    if (wf) {
137      await update($, tasks, l => (l ?? []).map(t => (t.id === wf.id ? { ...t, finished: [...(t.finished ?? []), agentId] } : t)))
138      return r
139    }
140    if (list.some(t => t.id === agentId && t.status === 'running')) {
141      const info = (await $.agent.list().catch(() => [])).find(a => a.id === agentId)
142      if (info && (info.status === 'idle' || info.status === 'waiting')) return r
143      await finish($, t => t.id === agentId, statusOf(info?.status))
144    }
145    return r
146  })
147
148  on('prompt.submit', async ($, e, next) => {
149    if (e.origin?.kind === 'task-notification') {
150      for (const n of parseNotifications(e.text)) {
151        const status = statusOf(n.status)
152        await finish($, t => t.taskId === n.id || t.id === n.id, status)
153      }
154    }
155    return next(e)
156  })
157
158  on('ui.focus', async ($, e, next) => {
159    if (e.component !== 'AbovePrompt' || e.plugin !== 'task-band' || e.origin.kind !== 'person') return next(e)
160    if (e.element === 'prev' || e.element === 'next') {
161      const step = e.element === 'next' ? 1 : -1
162      await update($, jump, () => null)
163      await update($, page, p => clampPage(p + step, pageCount))
164      return {}
165    }
166    return next(e)
167  })
168
169  on('ui.scroll', async ($, e, next) => {
170    if (e.component !== 'AbovePrompt' || e.by === 0) return next(e)
171    const list = (await read($, tasks)) ?? []
172    if (list.length === 0) return next(e)
173    await update($, jump, () => null)
174    await update($, page, p => clampPage(p + (e.by > 0 ? 1 : -1), pageCount))
175    return {}
176  })
177
178  on('ui.render', { component: 'AbovePrompt' }, async ($, e, next) => {
179    const list = viewOrder((await read($, tasks)) ?? [])
180    if (e.props.hasSurvey || list.length === 0) return next(e)
181    const [t, p, j, at] = await Promise.all([read($, tick), read($, page), read($, jump), now($)])
182    const chips = list.map(task => chipSegs(task, at, t))
183    const pages = paginate(chips.map(segWidth), e.props.bodyColumns - 1 - FRAME)
184    pageCount = pages.length
185    const jumpIndex = j && j.until > at ? list.findIndex(x => x.id === j.id) : -1
186    const shown = jumpIndex >= 0 ? pageOf(pages, jumpIndex) : clampPage(p, pages.length)
187    const [from, to] = pages[shown] ?? [0, 0]
188    const { Box, Button, Text } = $.ui.resolve(e)
189    const many = pages.length > 1
190
191    const W = e.props.bodyColumns - 1
192    const nav = many ? `‹ ${navLabel(from + 1, to, list.length)} ›  ` : ''
193    const visible = chips.slice(from, to)
194    const used = cellWidth(nav) + visible.reduce((n, segs, i) => n + segWidth(segs) + (i > 0 ? 3 : 0), 0)
195
196    return (
197      <Box flexDirection="column">
198        <Text color={C.faint}>{frameTop(W)}</Text>
199        <Box flexDirection="row">
200          <Text color={C.faint}>│ </Text>
201          {many ? <Button key="prev" label="‹" plain dimColor onPress={() => undefined} /> : null}
202          {many ? <Text> </Text> : null}
203          {many ? <Button key="pos" label={navLabel(from + 1, to, list.length)} plain autoFocus onPress={() => undefined} /> : null}
204          {many ? <Text> </Text> : null}
205          {many ? <Button key="next" label="›" plain dimColor onPress={() => undefined} /> : null}
206          {many ? <Text>  </Text> : null}
207          {visible.map((segs, i) => {
208            const task = list[from + i]
209            const isJump = jumpIndex === from + i
210            return (
211              <Box key={task?.id ?? `c${i}`} flexDirection="row">
212                {i > 0 ? <Text color={C.faint}> │ </Text> : null}
213                {segs.map((s, k) => (
214                  <Text key={`s${k}`} color={s.color} backgroundColor={isJump ? C.hl : undefined} bold={s.bold}>
215                    {s.text}
216                  </Text>
217                ))}
218              </Box>
219            )
220          })}
221          <Text>{rowPad(W, used)}</Text>
222          <Text color={C.faint}> │</Text>
223        </Box>
224        <Text color={C.faint}>{frameBottom(W)}</Text>
225      </Box>
226    )
227  })
228}
229
hooks/logic.ts 187 lines
1import type { BandTask, TaskStatus } from '../types'
2
3export const C = {
4  fg: '#2d2a24',
5  dim: '#6b675e',
6  faint: '#b9b5a3',
7  green: '#3f8a3a',
8  red: '#b8322b',
9  hl: '#e6e2bf',
10}
11
12export type Seg = { text: string; color?: string; bg?: string; bold?: boolean }
13
14export const GAUGE = 8
15export const LABEL_MAX = 18
16export const MIN_SAMPLES = 3
17export const EST_CAP = 95
18const SPIN = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏']
19const SEP = ' │ '
20const THAI_ZERO = /[ัิ-ฺ็-๎]/u
21
22export function cellWidth(s: string): number {
23  let n = 0
24  for (const ch of s) if (!THAI_ZERO.test(ch)) n++
25  return n
26}
27
28export function fitCells(s: string, max: number): string {
29  if (cellWidth(s) <= max) return s
30  let out = ''
31  for (const ch of s) {
32    if (cellWidth(out + ch) > max - 1) break
33    out += ch
34  }
35  return `${out}…`
36}
37
38export function clock(ms: number): string {
39  const s = Math.max(0, Math.floor(ms / 1000))
40  const h = Math.floor(s / 3600)
41  const m = Math.floor((s % 3600) / 60)
42  const sec = String(s % 60).padStart(2, '0')
43  return h > 0 ? `${h}:${String(m).padStart(2, '0')}:${sec}` : `${m}:${sec}`
44}
45
46export function average(samples: number[] | undefined): number | undefined {
47  if (!samples || samples.length < MIN_SAMPLES) return undefined
48  return samples.reduce((a, b) => a + b, 0) / samples.length
49}
50
51export function pushSample(samples: number[] | undefined, ms: number): number[] {
52  return [...(samples ?? []), ms].slice(-10)
53}
54
55export function estimate(elapsed: number, estMs: number | undefined): number | undefined {
56  if (!estMs || estMs <= 0) return undefined
57  return Math.min(EST_CAP, Math.round((elapsed / estMs) * 100))
58}
59
60function bar(fill: number, color: string): Seg[] {
61  const n = Math.max(0, Math.min(GAUGE, Math.round(fill * GAUGE)))
62  return [{ text: '━'.repeat(n), color }, { text: '─'.repeat(GAUGE - n), color: C.faint }]
63}
64
65function sweep(tick: number): Seg[] {
66  const pos = tick % (GAUGE - 1)
67  return [
68    { text: '─'.repeat(pos), color: C.faint },
69    { text: '━━', color: C.fg },
70    { text: '─'.repeat(GAUGE - 2 - pos), color: C.faint },
71  ]
72}
73
74export function chipSegs(t: BandTask, now: number, tick: number): Seg[] {
75  const elapsed = (t.endedAt ?? now) - t.startedAt
76  const head: Seg[] =
77    t.status === 'done'
78      ? [{ text: '✓ ', color: C.green }]
79      : t.status === 'fail'
80        ? [{ text: '✗ ', color: C.red }]
81        : [{ text: `${SPIN[tick % SPIN.length]} `, color: C.fg }]
82  head.push({ text: `${t.kind} `, color: C.dim }, { text: `${fitCells(t.label, LABEL_MAX)} `, color: C.fg })
83
84  const total = t.members?.length ?? 0
85  const done = t.finished?.length ?? 0
86  const count = t.kind === 'wf' && total > 0 ? `${t.status === 'done' ? total : done}/${total}` : undefined
87
88  if (t.status === 'done') {
89    return [...head, ...bar(1, C.green), { text: ` ${count ?? '100%'}`, color: C.green }]
90  }
91  if (t.status === 'fail') {
92    const fill = total > 0 ? done / total : (estimate(elapsed, t.estMs) ?? 50) / 100
93    return [...head, ...bar(fill, C.red), { text: ' fail', color: C.red }, { text: ` ${clock(elapsed)}`, color: C.dim }]
94  }
95  if (count) {
96    return [...head, ...bar(done / total, C.fg), { text: ` ${count}`, color: C.fg }, { text: ` ${clock(elapsed)}`, color: C.dim }]
97  }
98  const pct = t.kind === 'agent' ? estimate(elapsed, t.estMs) : undefined
99  if (pct !== undefined) {
100    return [...head, ...bar(pct / 100, C.fg), { text: ` ~${pct}%`, color: C.dim }, { text: ` ${clock(elapsed)}`, color: C.dim }]
101  }
102  return [...head, ...sweep(tick), { text: ` ${clock(elapsed)}`, color: C.dim }]
103}
104
105export function segWidth(segs: Seg[]): number {
106  return segs.reduce((n, s) => n + cellWidth(s.text), 0)
107}
108
109export const MAX_PER_PAGE = 3
110
111export function navLabel(first: number, last: number, total: number): string {
112  return first === last ? `${first}/${total}` : `${first}-${last}/${total}`
113}
114
115export function navWidth(total: number): number {
116  return cellWidth(`‹ ${navLabel(total, total, total)} ›  `) + 2
117}
118
119export function paginate(widths: number[], columns: number): Array<[number, number]> {
120  const room = columns - navWidth(widths.length) - 1
121  const pages: Array<[number, number]> = []
122  let i = 0
123  while (i < widths.length) {
124    let used = widths[i] ?? 0
125    let j = i + 1
126    while (j < widths.length && j - i < MAX_PER_PAGE && used + SEP.length + (widths[j] ?? 0) <= room) {
127      used += SEP.length + (widths[j] ?? 0)
128      j++
129    }
130    pages.push([i, j])
131    i = j
132  }
133  return pages
134}
135
136export function pageOf(pages: Array<[number, number]>, index: number): number {
137  const p = pages.findIndex(([a, b]) => index >= a && index < b)
138  return p < 0 ? 0 : p
139}
140
141export function clampPage(page: number, count: number): number {
142  return Math.max(0, Math.min(page, count - 1))
143}
144
145export function parseNotifications(text: string): Array<{ id: string; status: string }> {
146  const out: Array<{ id: string; status: string }> = []
147  const blocks = text.match(/<task-notification>[\s\S]*?<\/task-notification>/g) ?? []
148  for (const b of blocks) {
149    const id = b.match(/<task-id>([^<]+)<\/task-id>/)?.[1]?.trim()
150    const status = b.match(/<status>([^<]+)<\/status>/)?.[1]?.trim()
151    if (id && status) out.push({ id, status })
152  }
153  return out
154}
155
156export function statusOf(word: string | undefined): TaskStatus {
157  if (word === 'failed' || word === 'killed' || word === 'error' || word === 'stopped') return 'fail'
158  return 'done'
159}
160
161export function insertOrdered(list: BandTask[], t: BandTask): BandTask[] {
162  return [...list.filter(x => x.id !== t.id), t].sort((a, b) => (a.order ?? a.startedAt) - (b.order ?? b.startedAt))
163}
164
165export function viewOrder(list: BandTask[]): BandTask[] {
166  return [...list.filter(t => t.status === 'running'), ...list.filter(t => t.status !== 'running')]
167}
168
169export function clearFinished(tasks: BandTask[]): BandTask[] {
170  return tasks.filter(t => t.status === 'running')
171}
172
173export const TITLE = 'งานเบื้องหลัง'
174
175export function frameTop(width: number): string {
176  const head = `┌─ ${TITLE} `
177  return head + '─'.repeat(Math.max(0, width - cellWidth(head) - 1)) + '┐'
178}
179
180export function frameBottom(width: number): string {
181  return '└' + '─'.repeat(Math.max(0, width - 2)) + '┘'
182}
183
184export function rowPad(width: number, used: number): string {
185  return ' '.repeat(Math.max(0, width - 4 - used))
186}
187
types/index.d.ts 32 lines
1export type TaskKind = 'agent' | 'wf' | 'sh' | 'mon'
2export type TaskStatus = 'running' | 'done' | 'fail'
3
4export type BandTask = {
5  id: string
6  kind: TaskKind
7  label: string
8  startedAt: number
9  status: TaskStatus
10  endedAt?: number
11  order?: number
12  taskId?: string
13  agentType?: string
14  estMs?: number
15  runId?: string
16  members?: string[]
17  finished?: string[]
18}
19
20export type BandJump = { id: string; until: number }
21
22declare module 'claude-code' {
23  interface PluginState {
24    'task-band': {
25      tasks: BandTask[]
26      page: number
27      jump: BandJump | null
28      tick: number
29    }
30  }
31}
32