agent-island
/agent-island
Instaluje i otwiera Agent Island, panel pokazujący limity subskrypcji, wydatki API i stan zadań podagentów.
- Autor
- Lichenshuo ↗
- Repozytorium
- AutoSwarmagent-island ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 12
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 2
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: dashboard.cmd, install.ps1, island.cmd
Co robi i kiedy
Instaluje i otwiera Agent Island, panel pokazujący limity subskrypcji, wydatki API i stan zadań podagentów.
Oryginalny opis (EN)
Install, open or query Agent Island, a Windows Dynamic-Island style widget and local dashboard for multi-agent work (Claude Code, Codex, OpenCode, AutoSwarm). Shows remaining Claude / Codex subscription quota (5-hour and weekly), Claude API spend, running / waiting-for-approval / failed sub-agent tasks and delegation history. Use when the user asks to install or open the island / dashboard / 灵动岛 / 仪表盘, check remaining quota / 额度 / 花了多少钱, or review what sub-agents are doing / 派活记录.
Jak działa
- Install (once)
- Open
- Answer questions from its data
- Quota sources
- API balance
Podgląd SKILL.md
---
name: agent-island
description: Install, open or query Agent Island, a Windows Dynamic-Island style widget and local dashboard for multi-agent work (Claude Code, Codex, OpenCode, AutoSwarm). Shows remaining Claude / Codex subscription quota (5-hour and weekly), Claude API spend, running / waiting-for-approval / failed sub-agent tasks and delegation history. Use when the user asks to install or open the island / dashboard / 灵动岛 / 仪表盘, check remaining quota / 额度 / 花了多少钱, or review what sub-agents are doing / 派活记录.
---
# Agent Island
A black pill pinned to the top of the screen (hover to expand, drag to move, double-click to re-center) plus a full dashboard at `http://127.0.0.1:8765/`. Everything is local and read-only: it never writes to agent data, only to its own `data/` folder. Windows only.
`<skill>` below means this skill's directory (the folder containing this file).
## Install (once)
Needs Windows 10/11, Python 3.10+ (or `uv`) and the Edge WebView2 runtime. `install.ps1` checks all of this and stops with a clear message when something fails. Git is optional: it only enables AutoSwarm (`requirements-autoswarm.txt`) and with it the delegation history.
```
powershell -ExecutionPolicy Bypass -File "<skill>\install.ps1" # add -Autostart to start at sign-in
```
Optional `<skill>\config.json` (copy `config.example.json`) lists extra OpenCode databases to count API spend from; by default `~/.local/share/opencode/opencode.db` is used. Schema v1/v2 is detected automatically.
## Open
- Island only: run `<skill>\island.cmd`
- Island + dashboard in the browser: run `<skill>\dashboard.cmd`
Both are safe to repeat: a second start only opens the browser if asked. The island runs in the background with no taskbar button or console; quit from the island's expanded view.
## Answer questions from its data
Fetch `http://127.0.0.1:8765/api/data` (start `island.cmd` first if it is not running) and summarize in the user's language:
- `claude.windows` / `codex.windows`: `used` is percent used, `resets_at` is Unix seconds or null. Remaining = 100 − used; a window past `resets_at` has reset. `captured_at` tells how fresh the reading is; a 5-hour reading older than 5 hours is meaningless. `claude.quota_source`: `oauth` (live, from the terminal Claude Code login), `desktop` (Claude Desktop's own samples) or `statusline`.
- `api.today | week | d30 | month`: Claude API spend in USD from OpenCode (`week` and `d30` are rolling 7 and 30 days, `month` is the calendar month); `api.sources` breaks it down per database.
- `calls`: AutoSwarm delegations, newest first. Skip `probe: true` (installation self-checks). `state` is `running`, `waiting` (blocked on an approval, see `approvals`), `completed`, `failed`, `lost` (runner died) or `cancelled`.
- `summary`: counts for today, running, waiting and failed.
- `calls[].usage`: `scope: "call"` means `tokens` / `cost` are what THIS round used (growth of the conversation's counter while it ran); `scope: "session"` means only the whole conversation's total was available. The same agent is reused for many rounds, so never read a call's session total as that call's cost.
- `delegation`: `totals` (agents, calls, completed, failed, running, tokens, cost, today, `by_provider`) and `agents[]` grouped per sub-agent (name, rounds, completed/failed, tokens, last state/task). Use it to answer "how many sub-agents / how many tokens have I dispatched".
- Anomalies: a call counts as an anomaly when `state` is failed / lost / cancelled, it was not re-dispatched afterwards (`superseded`) and the user did not clear it (`dismissed`). `summary.failed` and `delegation.*.failed` count only those; `delegation.*.failed_total` is the history. Clearing only hides records (the list is `data/island-dismissed.json`, shared by island and dashboard; AutoSwarm's own records are never touched): `POST /api/dismiss` with `{"ids":[...]}`, `{"all_bad":true}`, `{"restore":[...]}` or `{"restore_all":true}`.
- `codex.tokens_today` / `codex.tokens_week`: Codex tokens (cache included) from its session counters.
## Quota sources
- **Claude**: best is a signed-in terminal Claude Code (`claude auth login`); the island then reads live usage with reset times. Otherwise it falls back to Claude Desktop's local usage samples, which are only refreshed occasionally, and says how old they are.
If the live reading stops (`claude.live_error`: `token_expired` = the terminal login token expired, `no_login`, `rate_limited`), the island falls back to the freshest of Claude Desktop's samples / the statusline snapshot and shows why. `server.py --quota-debug` prints every source side by side (no secrets) when the numbers look wrong. The island keeps the token alive by launching a short headless Claude Code process (no prompt, so no quota) when it has under 20 minutes left; Claude Code does the refresh with its own lock, the island only watches the expiry. `server.py --renew-token` runs this once and reports which method worked (`claude auth status`, then a headless session; the winner is remembered in `data/claude-renew.json`). If it fails, `claude.renew.ok` is false and the user must run `claude auth login`. A used-up weekly window means the 5-hour one is shown as blocked too. Plans without a 5-hour window (e.g. Codex Pro) are detected from the data: no window of 5 hours or less in `windows` means only the weekly one is shown.
- **Codex**: asked live every 3 minutes from `codex app-server` (`account/rateLimits/read`, Codex handles its own login); if that fails (`codex.live_error`) it falls back to the latest `rate_limits` in `~/.codex/sessions`, which only changes while Codex runs. Set `"codex_exe"` in `config.json` if `codex` is not on PATH.
- **API spend**: summed from per-message costs in the OpenCode databases.
## API balance
`api.providers[]` splits OpenCode's recorded spend per provider (`anthropic` = Claude, `openai` = OpenAI / Codex, `google`, ...), each with its own `billing`; the top-level `api.*` numbers are Claude's. Neither Anthropic nor OpenAI offers an API for a prepaid balance, so every provider's remaining balance is an estimate from the last calibration (`--provider openai` for the others) unless Claude's `anthropic_admin_key` is set. `api.billing` holds Claude's balance. `source: "admin"` means exact numbers from Anthropic's Cost API (`anthropic_admin_key` in `config.json`). `source: "estimate"` means: last balance the user reported − local spend since then × `factor`. Local OpenCode costs undercount real billing (and miss other callers on the same key), so `factor` is learned automatically at each calibration (real spend ÷ local spend, smoothed). `stale: true` means the estimate is old or mostly extrapolated: say so and ask for a fresh number instead of presenting it as exact.
When the user tells you the balance the Console shows (e.g. "API 余额现在 5 美元,本月 128"), record it:
```
"<skill>\.venv\Scripts\python.exe" "<skill>\app\server.py" --calibrate 5.00 --month 128
```
`--provider <id>` picks another provider (stored under `provider_credit` in `config.json`). `--month` (this month's spend) and `--total` (all credit ever topped up) are optional. It prints the learned factor; a running island picks it up within 20 s. The user can also do this themselves: click the API tile on the island, or "校准余额" on the dashboard's API card.
Pliki w paczce (12)
- .gitattributes 1 KB
- .gitignore 1 KB
- CLAUDE.md 1 KB
- LICENSE 1 KB
- README.md 14 KB
- SKILL.md 7 KB
- config.example.json 1 KB
- dashboard.cmd 1 KB
- install.ps1 5 KB
- island.cmd 1 KB
- requirements-autoswarm.txt 1 KB
- requirements.txt 1 KB
Wersje
W bazie nie ma innych skilli o nazwie agent-island.
Migawka bazy z 13.11.2026 · wersja paczki 2d571260a5cf (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100leave-me-something
Wstrzymuje zautomatyzowaną pracę AI, gdy limit subskrypcji osiągnie ustawiony procent okna 5-godzinnego lub tygodniowego, przez haki Claude Code i Codex. - 100external-agents
Deleguje pracę innym agentom CLI, takim jak Codex, Gemini czy qodercli, zbiera drugą opinię i sprawdza limity. - 100quota
Pokazuje stan limitów Claude: pięciogodzinną sesję i zużycie tygodniowe z lokalnej pamięci podręcznej. - 100tokenmeter
Otwiera lokalny panel Tokenmeter, który pokazuje zużycie tokenów, koszty, okna limitów i chybienia cache dla Claude Code i Codex. - 90harness-observatory
Buduje lokalny panel obserwacji sesji agentów z logów Claude Code i Codex oraz proponuje poprawki CLAUDE.md i AGENTS.md. - 100dev-swarm-install-ai-code-agent
Instaluje narzędzia CLI agentów kodujących AI: claude-code, gemini-cli, codex i github copilot-cli.