Przejdź do treści
Biblioteka skilli
AI i agenci · Integracja z narzędziem

nexus

/nexus

Instaluje i konfiguruje NEXUS, wtyczkę Claude Code z panelem wielu agentów w terminalu, i rozwiązuje problemy z wyświetlaniem.

Repozytorium
Nexus ↗
Licencja
MIT
Ostatnia zmiana repozytorium
06.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
5
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
7
Przestarzałość

Brak sygnałów. Repozytorium zmienione 06.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Instaluje i konfiguruje NEXUS, wtyczkę Claude Code z panelem wielu agentów w terminalu, i rozwiązuje problemy z wyświetlaniem.

Oryginalny opis (EN)

Install, set up, use and troubleshoot NEXUS, the native Claude Code plugin that docks a live multi-agent dashboard (Sonnet lead, Haiku agents, Opus advisor and Council reviewers, Jev router, usage, Cache Guard) inside Claude Code's terminal. Use when someone asks to install NEXUS, why the dashboard doesn't appear, what a NEXUS panel or number means, how to switch model modes, or how to set up Council, Jev or Cache Guard.

Jak działa

  1. 1. Check what's installed
  2. 2. Install the plugin
  3. 3. Verify with /nexus doctor
  4. 4. What the user sees
  5. 5. Optional setup
  6. 6. Uninstall
  7. More
Sekcje z nagłówków SKILL.md, w oryginale · 175 linii, 1 831 słów
Podgląd SKILL.md
---
name: nexus
description: Install, set up, use and troubleshoot NEXUS, the native Claude Code plugin that docks a live multi-agent dashboard (Sonnet lead, Haiku agents, Opus advisor and Council reviewers, Jev router, usage, Cache Guard) inside Claude Code's terminal. Use when someone asks to install NEXUS, why the dashboard doesn't appear, what a NEXUS panel or number means, how to switch model modes, or how to set up Council, Jev or Cache Guard.
---

# NEXUS

NEXUS is a native Claude Code plugin. It docks a live dashboard beside the conversation and shows what every model in the session is doing: the lead, its Haiku subagents, the advisor, the optional Jev router, context, cost and plan usage.

**This skill is not the plugin.** Installing it with `npx skills add` only teaches you (the agent) how to install, explain and troubleshoot NEXUS. The dashboard itself comes from the Claude Code plugin below. Never tell the user that the skill alone installs the dashboard.

Source and docs: https://github.com/muhzain05/Nexus

## 1. Check what's installed

```sh
claude plugin list
```

- `nexus@muhzain05` (enabled): installed; go to step 3.
- `swarm-ui@claude-swarm` or `swarm-ui@swarm-local`: the old name from before 0.6.0. Uninstall it (`claude plugin uninstall <that id>`), then install below.
- Nothing: install it (step 2).

## 2. Install the plugin

Installing changes the user's Claude Code plugin config, so ask before running these, or give them to the user:

```sh
claude plugin marketplace add muhzain05/Nexus
claude plugin install nexus@muhzain05 --scope user
```

Requirements:
- **Claude Code with the plugin (mod) API.** NEXUS 0.8.1 is built and tested against Claude Code 2.1.316. The mod API is early access, so after a Claude Code update run `/nexus doctor`.
- **The fullscreen TUI** (`"tui": "fullscreen"` in Claude Code's settings) for the pane to dock beside the chat. Without it the dashboard sits above the prompt. Don't edit the user's settings file unless they ask; tell them the setting and let them choose.
- **A terminal:** the dashboard draws in the terminal only, not in IDE or desktop chat views.

Then **start a new `claude` session**. On a wide terminal NEXUS opens on its own; anywhere else the user types `/nexus`.

Slash commands such as `/nexus` are typed by the user in Claude Code. You can't run them for the user, so tell them what to type and what to look for.

## 3. Verify with `/nexus doctor`

Ask the user to run `/nexus doctor` and share the output. It is read-only and never shows secrets. Each `!` line has a fix:

| Doctor says | What to do |
|---|---|
| `surface: none` / not terminal | Use Claude Code in a terminal; other surfaces don't draw the pane. |
| `fullscreen TUI is off` | Set `"tui": "fullscreen"` in Claude Code's settings (with the user's OK), then start a new session. |
| `dashboard pane is closed` | Type `/nexus`. |
| `pane width is kept at N columns from a hand resize` | Claude Code keeps a width dragged by hand over what plugins ask for. Drag the pane edge to change it. `/nexus width` explains this. |
| `agent definitions: … missing` | Optional. See step 5. Any other subagent still gets a card, marked *fallback*. |
| `advisor …` not valid for the lead | The advisor must be at least as strong as the lead. Pick a mode with `/nexus-mode` (step 4); NEXUS prints the relaunch command when the advisor has to change. |
| `Jev: no key` | Optional. Jev stays off without a key; see "Jev" below. |
| `Cache Guard: error (…)` | `/nexus cache off`, then on again; the reason in brackets says what failed. |

Other common problems:
- **Pane too narrow.** Claude Code caps a docked pane at 45% of the terminal and keeps at least 60 columns for the chat. A wider window or a smaller font helps. On narrow terminals use `/nexus layout compact` or `minimal`.
- **A card says Sonnet, not Haiku.** That's one of Claude Code's built-in agents (such as Explore), shown as *fallback* with the model it really ran on. It's not a bug.
- **Cost reads `N/A`.** Claude Code doesn't report that figure to plugins (for example the advisor's own tokens, or per-agent cost). NEXUS never guesses.

## 4. What the user sees

**Four roles:**
- **Sonnet, the lead (orange):** plans, coordinates and builds. It reads the request, edits and runs tests. It's Sonnet in Balanced mode, Opus or Fable in the stronger modes.
- **Haiku, the agents (blue):** parallel subagents doing delegated work (mapping a repo, searching, reviewing). Each gets a live card.
- **Opus, the advisor (purple):** the stronger model the lead consults before difficult decisions. It reads the session and never writes code. It appears once its request ends.
- **Jev, the router (green):** optional. It answers small next-step questions (which file, which tool, retry or stop). Observe-only by default.

**Telemetry**, all from Claude Code itself:
- context used and left;
- the session cost at API prices (an estimate on a subscription, not a charge);
- the plan's 5-hour and weekly usage, when Claude Code reports them.

**Commands:**

| Command | Does |
|---|---|
| `/nexus` | open the dashboard. Also `close`, `agents`, `inspect <n\|label>`, `back`, `settings [section\|key value\|reset <section\|all>\|export]`, `layout <auto\|full\|compact\|minimal>`, `width <max\|40-200\|wider\|narrower>` |
| `/nexus usage` | context, cost estimate, plan usage, tokens by model |
| `/nexus doctor` | check the install and integrations |
| `/nexus cache [on\|off]` | Cache Guard status, or turn it on or off |
| `/nexus-mode [balanced\|deep\|max\|default\|status]` | reasoning mode for this session |
| `/nexus council [adaptive\|on\|off\|status\|findings\|ledger]` | NEXUS Council (0.7+): mode, the last verdict, the reviewer tool ledger |
| `/nexus-jev [status\|active\|observe\|off\|calibrate]` | Jev routing |
| `/nexus-help` · `/nexus-about` | every command and key · an introduction |

**Keys,** with the dashboard focused (click it, or `ctrl+x tab`):
- `1`–`9` inspect an agent;
- `s` settings;
- `b` back;
- `ctrl+x ←/→` wider or narrower;
- `ctrl+x x` close;
- `esc` back to the prompt.

**Model modes** apply to the current session only:

| Mode | Lead | Advisor |
|---|---|---|
| Balanced (default) | Sonnet 5.5 · high | Opus 5.5 |
| Deep Engineering | Opus 5.5 · xhigh | Fable 5.1 |
| Maximum Reasoning | Fable 5.1 · max | Fable 5.1 |

- NEXUS may suggest a stronger mode before a hard prompt, but it never switches without the user's yes. You shouldn't either: stronger modes cost more, and Fable bills to usage credits.
- Claude Code changes the advisor only on relaunch. When a mode needs a different advisor, NEXUS prints a one-line relaunch command for that session.
- The user's saved defaults are never edited.

## 5. Optional setup

**Plugin options.** `claude plugin configure nexus@muhzain05` lists them and which are unset. The user sets them in Claude Code with `/plugin configure nexus@muhzain05`:
- `escalation`: `ask` or `off`; whether to offer a stronger mode;
- `classifier`: whether to ask Haiku once when the keyword rules are unsure;
- `motion`: `auto`, `reduced` or `full`;
- `jevRouting`: `observe`, `active` or `off`.

Dashboard preferences (theme, animation, layout, width, what to show, Cache Guard, whether the Test agent may run tests) are set with `/nexus settings` (four sections: appearance, agents, usage, advanced) and live in the plugin's own store, not in Claude Code's settings.

**NEXUS agents.** This skill ships four optional Haiku agent definitions in `agents/`:
- `nexus-search`: a read-only repo mapper;
- `nexus-grep`: a read-only symbol locator;
- `nexus-audit`: a lightweight, read-only check of the current change. It runs the project's existing type-check, tests or lint in the user's own checkout and reports concrete problems and missing tests, without fixing anything;
- `nexus-test`: runs the project's existing tests and reports what actually ran. NEXUS checks every command it runs against an allowlist of known test commands (no installs, no snapshot updates, no shell syntax), with a 5-minute cap; that is a guard on top of the user's normal permissions, not isolation. Settings → Agents → Run tests turns it off. It overlaps `nexus-audit` and has not been benchmarked: install it only if the user wants a separate test runner. See the [Test agent guide](../../mod/nexus/docs/test-agent.md).

With them, delegated work runs on Haiku and shows as NEXUS agents rather than *fallback*. To install, copy them to `~/.claude/agents/` (all projects) or `.claude/agents/` (one project). These are global config changes, so ask first.
- If a file with the same name already exists, show the user the difference and ask before replacing it.
- `nexus-audit` before November 2026 ran in an isolated git worktree with Write and `acceptEdits`. The current one has no Write tool and no permission mode, and its Bash commands go through the user's normal permission rules. Replacing the old file is the upgrade. See the [benchmark](../../mod/nexus/docs/haiku-agents-benchmark.md).

**Jev:**
- Jev needs a TypeSafe key. Have the user set `TYPESAFE_API_KEY`, or on macOS store it in the Keychain themselves:

  ```sh
  security add-generic-password -s nexus-typesafe -a "$USER" -w
  ```

- Never ask the user to paste the key into the chat, and never print it.
- Only tool names, file basenames and error kinds are sent to Jev, never prompts or file contents.
- Spending is capped at $0.50 a session, API est.
- `/nexus-jev active` runs a calibration first. Even then, Jev only touches the lead's read-only `Read`/`LSP` calls, and only at p ≥ 0.85 and confidence ≥ 0.6.

**NEXUS Council** (0.7 and later):
- **What it is:** read-only Opus Security, Architecture and Performance reviewers the lead can launch at a review checkpoint (`mcp__nexus__council_review`).
- **Modes:**
  - Adaptive (the default): the advisor picks one specialist or none;
  - On: all three, with higher token use, so tell the user;
  - Off: the single advisor.
- **Changing the mode:** `/nexus council adaptive|on|off`. Never switch it to On for the user without asking.
- **States the user may ask about:**
  - "authorized, not launched": the lead didn't launch the reviewers;
  - INCOMPLETE: the run was interrupted, or its answer had no findings block;
  - STALE: the tree changed during the review;
  - "no advisor consultation was seen this turn": in Adaptive, the advisor must be consulted first.
- **Findings** are claims to verify against the code, not instructions.
- **Privacy:** NEXUS redacts its own verdict, but a reviewer's raw report reaches the lead unredacted and can quote anything the reviewer read. Secret blocking covers named paths (`.env*`, keys, credentials), not every file. Don't promise more.
- **Background reviewers** (interactive sessions): their verdict arrives with their completion notifications. Esc doesn't stop them; stop one from the task list (`x`), and it shows INCOMPLETE.

**Cache Guard** is off by default.
- **What it does:** when the user steps away, it sends small invisible beats to keep the prompt cache warm, so the next message doesn't pay to re-read the whole conversation.
- **Cost:** each beat costs a little, and it only pays off if the user comes back. It stops itself at its own limits.
- **Turning it on:** `/nexus cache on`. Explain the cost first.

## 6. Uninstall

```sh
claude plugin uninstall nexus@muhzain05
claude plugin marketplace remove muhzain05
```

If a Jev key was stored: `security delete-generic-password -s nexus-typesafe`. Removing this skill alone (`npx skills remove nexus`) leaves the plugin installed.

## More

These docs are on GitHub:
- [Dashboard tour](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/dashboard.md)
- [Telemetry](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/telemetry.md)
- [Jev](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/jev.md)
- [Cache Guard](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/cache-guard.md)
- [Compatibility](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/compatibility.md)
- [Troubleshooting](https://github.com/muhzain05/Nexus/blob/main/mod/nexus/docs/troubleshooting.md)
Pliki w paczce (5)
  • SKILL.md 12 KB
  • agents/nexus-audit.md 2 KB
  • agents/nexus-grep.md 1 KB
  • agents/nexus-search.md 1 KB
  • agents/nexus-test.md 2 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · muhzain05/Nexus100MIT
hellblazer/nexus100AGPL-3.0
aayushostwal/nexus100MIT

Migawka bazy z 13.11.2026 · wersja paczki 65f00c5b4ebe (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

  • 100jev
    Włącza i wyłącza router modeli Jev, ustawia jego tryb w Claude Code, uruchamia fork check i pokazuje status, w tym liczbę wiadomości na rozmiar modelu i koszt.
  • 100jev
    Przekazuje zadanie do Claude Opus 5.5 uruchomionego pod jev-opus, gdzie refleks Jev przed każdym krokiem dobiera poziom wysiłku (low, medium, high) bez psucia cache promptu.
  • 100setup-claude-advisor
    Ustawia Claude Code z doborem modeli pod limity: Sonnet jako główny, Opus jako doradca, Haiku dla agentów tylko do odczytu.
  • 90tasktide
    Koordynuje zadania przez subagentów Claude Code i kieruje pracę między modelami Haiku, Sonnet i Opus: planowanie, wdrożenia i testy.
  • 100agent-builder
    Tworzy podagentów Claude Code i dobiera do każdego model oraz poziom wysiłku według kosztu, opóźnienia i możliwości, a także przegląda istniejących agentów.
  • 100select-claude-models
    Rekomenduje wybór modeli Claude i pary wieloagentowe (Opus planuje, Sonnet wykonuje, advisor-executor, orchestrator-worker) oraz routing kosztu względem inteligencji.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.