convez.ai
AgenteAutomaçãoAvançado
Martin-Hausleitner

Martin's Agent Manager

Martin-Hausleitner/martins-agent-manager

Orquestrador research-first para frota de agentes com relatórios coloridos

Ver no GitHub
InicialPluginScore · de 0 a 100

Orquestrador minimalista para gerenciar uma frota de agentes de codificação (Claude Code, Antigravity, Codex) com observabilidade em tempo real. Implementa relatórios tabulares coloridos, monitoramento de drift e delegação disciplinada de trabalho sem raciocínio centralizado.

Instalação

$npx convez-ai Martin-Hausleitner/martins-agent-manager

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Coordenar múltiplos agentes de IA em tarefas de desenvolvimento complexas
  • Monitorar progresso e saúde de pipelines de trabalho descentralizados
  • Implementar padrões de orquestração com observabilidade e babysitting automático

Veredito

Implementação sólida de padrão de orquestração para múltiplos agentes com foco em observabilidade legível e delegação disciplinada. Bem estruturado para cenários de produção com frota de modelos heterogêneos, mas requer adaptação ao contexto específico do usuário.

README

A tradução para o português deste README ainda está na fila.

🧭 Martin's Agent Manager

A tiny manager / orchestrator discipline for running a fleet of coding agents (Claude Code, Antigravity, Codex) as watchable terminal sub-sessions.

The manager is a thin orchestrator: it never does the work itself and never thinks/plans itself — it delegates all non-trivial reasoning to worker lanes, researches before planning, babysits on a timer, and reports back in one fixed, colour-coded format that's easy to scan (built for a dyslexic operator: calm, line-based, emoji-keyed, no noisy ASCII boxes).

This repo is the public, anonymised version of the convention. It contains no hostnames, secrets, or private project names — just the reusable pattern.


🤖 Engine legend (icons + colours)

Every worker sub-session is named with its engine's icon + colour, so you instantly see who is running.

  • 🦀 Claude Code — 🟧 orange — priority HIGH 🔺 — 🦀 cc·frontend
  • 🟦 Antigravity / agy — blue (square icon) — light tasks 🔹, limit irrelevant — 🟦 agy·mock
  • 🟣 Codex — lila (circle) — mid priority 🔸 — 🟣 codex·tests

Watch/guard sessions stay neutral (grey).

Subagent = Antigravity by default (esp. light tasks like IDR). Pick model first: Flash = light; Gemini 3.1 Pro High = harder. Important → CC; light → agy. Empty context → CC + Sonnet. agy limit irrelevant. If unsure which engine, ask the operator.


📋 Report format — Variant A: sectioned report

The manager reports in STATUS SECTIONS — no Markdown table (renders buggy in CMUX) and no ASCII/box frames. Each section has a header emoji, slim indented lines, progress bar, and a confidence word. Questions are restated at the END with answers in [square brackets].

✅ ERLEDIGT
  <engine-icon> **Thema**       ██████████ 100%
🔄 LÄUFT
  <engine-icon> **Thema**       ██░░░░░░░░  20%   🔎 prüfen
      🧠 IDR 🔁2  🕐 1h ago   🔬 DR ⚡ running
👀 BITTE DRÜBERSCHAUEN
  <engine-icon> **Thema**       ███░░░░░░░  30%   🔎 unsicher
🚨 ALARM            (critical only)
  <engine-icon> **Thema** — what to do right now

❓ FRAGEN
  Q1  <question>
      [1] … ⭐   [2] …   [3] …
  • Sections (in order): ✅ ERLEDIGT · 🔄 LÄUFT · 👀 BITTE DRÜBERSCHAUEN · 🚨 ALARM (only when truly critical, outranks all) · ❓ FRAGEN
  • Progress: bar + percent ██████░░░░ 60% (10-block Unicode bars)
  • Confidence word (after bar): sicher · prüfen · unsicher
  • Bold (**…**): important lane names / themes and key numbers so the operator can scan fast.
  • Colour: chat reports use emoji + bold (Markdown); heartbeat TUI uses ANSI colours.
  • Research indicator line (indented line under the lane, when relevant): 🔬 = Deep Research · 🧠 = IDR (interactive deep research) · 🔁 N = iteration count · 🕐 <when> = last run · ⚡ = running RIGHT NOW Example: 🧠 IDR 🔁3 🕐 2h ago 🔬 DR ⚡ running
  • Questions restated at the END in ❓ FRAGEN; answers in [square brackets] with [1] … ⭐ on the recommended option. Running numbers never restart across questions. Operator replies with just the number(s), e.g. 1 3.
  • Engine icons: 🦀 cc-orange · 🟦 agy-blue · 🟣 codex-purple. Priority optional. Drop nonsensical "continue?" rows.
  • Short 📋 Summary: after a ─── divider at the end.

Example

✅ ERLEDIGT
  🦀 **Frontend lane**    ██████████ 100%
  🟦 **Mock lane**        ██████████ 100%

🔄 LÄUFT
  🟣 **Tests lane**       ██████░░░░  60%   🔎 prüfen
      🧠 IDR 🔁1  🕐 30min ago

👀 BITTE DRÜBERSCHAUEN
  🧹 **Dirty worktree**   ████░░░░░░  40%   🔎 unsicher

🚨 ALARM
  💥 **Auth broken** — tokens expired, all lanes blocked — fix immediately

❓ FRAGEN
  Q1  Deploy target?
      [1] managed ⭐   [2] self-host
  Q2  Scope?
      [3] optional module now ⭐   [4] later

───
📋 Summary: 2 done, 1 running (tests 60%), 1 needs review, 1 alarm. Auth must be fixed before deploy.

🔒 Hard rule: manager only

flowchart TD
  OP([Operator]) -->|goal| MGR{{Manager}}
  MGR -->|research first| IDR[Interactive Deep Research]
  IDR -->|grounded plan| MGR
  MGR -->|brief + delegate| L1[🦀 Claude Code lane]
  MGR -->|brief + delegate| L2[🟦 Antigravity lane]
  MGR -->|brief + delegate| L3[🟣 Codex lane]
  L1 & L2 & L3 -->|progress| MGR
  MGR -->|colour-coded lines + summary| OP
  MGR -.->|every 30 min babysit tick| MGR

The manager may: spawn/brief worker lanes, schedule ticks, write memory, synthesise reports, ask clarifying questions.

The manager may not: edit files, run builds/tests/installs, call task-doing APIs, or implement. Everything that does the task goes to a worker lane in a watchable session.

Never think/plan yourself — delegate by default

The manager never reasons out plans, research, or project-logic itself. Any non-trivial thinking is spawned as a worker agent (default: 🟦 Antigravity + interactive deep research) that reports back a concise result, so the manager's context stays clean and cheap. The manager only briefs, schedules, and synthesises.

Default execution target = the remote worker host, never the local machine — unless the task strictly needs local hardware (e.g. an iOS build).


💡 Strategic token-aware orchestration

At every heartbeat the manager makes a strategic orchestration decision based on (a) each lane's priority and (b) each provider's remaining token budget (CodexBar-style: weekly / daily / session for Claude Code / Codex / Antigravity).

When a provider's budget runs low:

  • Do NOT burn expensive tokens on re-questioning the current state — route state-questions to NotebookLM/IDR instead (cheap).
  • Stop or deprioritise low-priority lanes.
  • Allocate the remaining high-value tokens to the highest-priority lanes only.
  • Surface these routing decisions in the report so the operator can see them.

Priority legend: 🔺 high (all CC) · 🔸 mid (Codex) · 🔹 low (agy, limit irrelevant).


🔬 Research-first

Before planning anything new, research first to ground the plan — before specs/architecture/code.

flowchart LR
  A[New idea / plan request] --> B[Deep research pass]
  B --> C{One clarifying question}
  C -->|operator answers| D[Deeper pass + iterate]
  D --> E[Grounded report + Feature Matrix]
  E --> F[Write spec / plan]

Comparisons MUST produce a Feature Matrix — one row per program, columns for the deciding criteria, and a GitHub link per program. Example shape:

ToolLanguageLicenseKey capabilityGitHub
Tool ARustMIT…github.com/org/tool-a
Tool BGoApache-2.0…github.com/org/tool-b

📦 What's here

  • skills/agent-manager/SKILL.md — the reusable skill (drop into ~/.claude/skills/).
  • CLAUDE.md — an anonymised excerpt of the global rules (report format + engine legend + research-first).

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados