Valor — Colega IA Autônomo
tomcounsell/ai
Colega IA que trabalha autonomamente, entrega código e se auto-recupera
Ver no GitHubUm servidor MCP que transforma Claude em um colega de trabalho autônomo capaz de planejar, codificar, testar e mergear features end-to-end sem supervisão. Integra-se a Telegram, Email e outras plataformas, mantém memória de decisões passadas e recupera-se automaticamente de falhas. Executa um pipeline SDLC completo com critério de modelo (Opus para raciocínio, Sonnet para execução).
Instalação
npx convez-ai tomcounsell/ai1 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
Casos de uso
- Automação de desenvolvimento full-stack sem intervenção constante
- Triagem e execução de tarefas via Telegram ou Email
- Pipelines SDLC autônomos com ciclos de teste, patch e review
Veredito

Valor é uma arquitetura ambiciosa e bem-estruturada para agentes IA realmente autônomos. Indicado para equipes que buscam delegar fluxos de desenvolvimento estruturados; exige infraestrutura dedicada (máquina própria) e calibração cuidadosa de escopos para evitar excesso de autonomia.
README
A tradução para o português deste README ainda está na fila.
Valor
An autonomous AI coworker. Not an assistant, not a chatbot — a colleague that owns its own machine and does real work.
Why Valor?
- Works where you work. Receives tasks via Telegram or Email and replies in the same thread — no context switching, no dashboards to check.
- Ships code end-to-end. Plans a feature, critiques the plan, writes the code, runs tests, opens a PR, and merges it — without being babysat through each step.
- Remembers across sessions. A subconscious memory system injects relevant past decisions and corrections as
<thought>blocks at exactly the right moment. - Fixes itself. A watchdog + self-healing pipeline restarts on crashes, hibernates on API failures, and escalates to you only when automation can't recover.
- Extensible by design. Add a new skill in
.claude/skills/, wire a tool intools/+ MCP, or connect a new comms channel inbridge/— the architecture absorbs extensions without surgery.
What Is This?
Valor wraps agent harnesses (like Claude Code) and bridges them to the comms channels humans actually use (Telegram, Email, LinkedIn, and more). The supervisor assigns work and provides direction. Valor executes autonomously on its own Mac, reaching out when necessary.
How It Works
Three layers:
- Comms layer — bridges to the channels where work actually happens: Telegram, Email, LinkedIn, etc. Messages come in, replies go out, session context survives across conversations.
- Harness layer — wraps agent harnesses like Claude Code, giving Valor tools, memory, skills, and a full SDLC workflow.
- Execution layer — a standalone worker service runs sessions against the configured harness. Sessions come in three role flavors: PM (orchestrates work), Dev (writes code), and Teammate (conversational).
Architecture
See docs/features/bridge-worker-architecture.md for the full design.
📸 TODO: Add demo GIF or screenshot here — see docs/assets/ for placement instructions.
The SDLC Pipeline
Valor ships real features through a structured pipeline — a directed cyclic graph with failure cycles, not a straight line. Each stage is a skill the agent invokes, and each stage runs on the model best suited to its cognitive load: Opus for hard reasoning (Plan, Critique, Review) and Sonnet for plan execution and tool-heavy work (Build, Test, Patch, Docs).
- Cycles are first-class. Critique loops back to Plan when concerns surface. Test→Patch→Test iterates until the fix holds. Review→Patch→Test→Review repeats until findings clear.
- Hard patches resume the builder. When a Patch needs the original builder's accumulated context, PM resumes the Build session's Claude Code transcript instead of starting fresh — the dashed edge in the diagram.
- PM orchestrates, Dev executes. A read-only PM session steers the pipeline stage-by-stage; each stage spawns a full-permission Dev session that runs one skill and reports back.
The pipeline graph is defined in agent/pipeline_graph.py. See .claude/skills/sdlc/SKILL.md for the ground truth on stage definitions and docs/features/pipeline-graph.md for the feature doc.
Subsystems
| Subsystem | Purpose |
|---|---|
| Subconscious memory | Long-term memory with bloom-filter recall, intentional saves, and post-session extraction |
| Reflections | Daily autonomous maintenance pipeline (log review, audits, cleanup, reporting) |
| Autoexperiment | Nightly prompt optimization for observer/summarizer targets |
| Self-healing | Watchdog service with crash tracking and automatic recovery |
| Session steering | Inject guidance into running sessions from any process |
| Worktree isolation | Per-feature git worktrees for parallel work without collisions |
| Dashboard | Web UI showing sessions, health, reflections, and machine state |
Quick Start
# 1. Install dependencies
pip install -e .
# 2. Configure environment (symlink to vault)
ln -sf ~/Desktop/Valor/.env ~/src/ai/.env
# All secrets live in ~/Desktop/Valor/.env (iCloud-synced vault)
# 3. Start the bridge and worker
./scripts/start_bridge.sh
./scripts/valor-service.sh worker-start
Service Management
./scripts/valor-service.sh status # Bridge status
./scripts/valor-service.sh restart # Restart bridge after code changes
./scripts/valor-service.sh worker-status # Worker status
./scripts/valor-service.sh worker-restart # Restart worker
tail -f logs/bridge.log # Stream logs
Repository Layout
ai/
├── agent/ # Session queue, SDK client, output routing
├── bridge/ # Comms bridges (Telegram, etc.) and nudge loop
├── worker/ # Standalone worker service (python -m worker)
├── tools/ # Local Python tools
├── ui/ # Dashboard web UI
├── monitoring/ # Watchdog, crash tracker, health checks
├── .claude/
│ ├── skills/ # SDLC and utility skills
│ ├── commands/ # Slash commands
│ └── agents/ # Subagent definitions
├── config/ # identity.json, persona segments, projects.json
├── scripts/ # Service management, setup, deployment
├── docs/features/ # Feature-level documentation
└── tests/ # Unit, integration, e2e
Development
pytest tests/unit/ -n auto # Fast unit tests in parallel
pytest tests/ # Full suite
python -m ruff format . # Format
See Also
| Resource | Purpose |
|---|---|
| CLAUDE.md | Development principles and working guide |
| docs/features/README.md | Feature index — how things work |
| config/identity.json | Structured identity data |
| config/personas/segments/ | Composable persona segments |
| tests/README.md | Test suite index and contribution guide |