convez.ai
Agente
ativin4

codehydra

ativin4/codehydra

Ver no GitHub
InicialPluginScore · de 0 a 100

BYOS terminal coding agent: CLI/model/effort routing across Claude, Gemini, Codex CLIs

Instalação

$npx convez-ai ativin4/codehydra

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
57%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

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

🐍 CodeHydra: The BYOS Terminal Coding Agent

Bring Your Own Subscription (BYOS). CodeHydra is a zero-GUI, autonomous coding agent that bypasses pay-per-token API gates by scavenging local OAuth session tokens from your official developer tools.

Tests License: MIT

⚡ Core Superpowers

  • Subscription Scavenging: Automatically finds and uses tokens from Claude Code, Gemini CLI, GitHub Copilot/Codex, and Google Cloud ADC.
  • OSS Model Fallback: Falls back to Ollama (local or cloud) during rate-limits or refresh windows — set OLLAMA_HOST for a remote instance; no subscription gap means no conversation interruption.
  • Multi-CLI Gateway: Routes each turn to claude, gemini, codex, or ollama based on effort tier and which subscriptions/daemons are active, with automatic fallback if one fails.
  • Agentic Passthrough: The chosen CLI edits files directly (auto-approve/yolo mode) — no fragile diff-parsing required.
  • Live Streaming: Responses render incrementally, token-by-token, as claude/gemini produce them (via stream-json).
  • Persistent Claude Session: claude runs as a long-lived stream-json process, reused across turns - only the first turn pays CLI startup cost, and /mode//model changes or a new conversation transparently restart it.
  • Claude Code-style TUI: A scrollable history pane with a pinned input box at the bottom, built with textual.
  • Self-Healing Loop: If the agent makes a change that breaks your build, it reads the stderr and fixes it automatically.
  • Workspace-Aware: Injects a compact AST-tree map of your entire project into the LLM context.
  • Sessions: Conversation history, routing state, and usage are persisted to .codehydra/sessions/ and resumable across runs.
  • Usage Tracking: /cost summarizes which CLI/model/tier handled each turn and token counts where reported.
  • MCP Native: Declare MCP servers in .agentrc.toml and they're wired into whichever backend CLI is active.
  • Autonomous Subagents: Every backend CLI is given a built-in dispatch_agents tool (mirroring Claude Code's Task tool) so it can fan independent sub-tasks out to parallel CodeHydra-routed agents on its own, mid-turn.
  • Background Tasks: The same built-in toolset gives the CLI run_in_background/get_background_output/stop_background_task (mirroring Claude Code's background Bash + Monitor) for long-lived processes like dev servers, surviving past the current turn.
  • Configurable Routing: Override CLI priority and model choices per tier in .agentrc.toml without touching code.

🚀 Quick Start

1. Prerequisites

Ensure you have uv installed:

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Installation

git clone https://github.com/ativin4/codehydra.git
cd codehydra
uv sync

3. Usage

# Start the interactive TUI
uv run python3 main.py

4. Configure Self-Healing

Add a .agentrc.toml to your project root to enable build verification:

[build]
command = "pytest" # or "npm run build", "go build", etc.

🛠️ Commands

CommandAction
/effort <low|medium|high>Set the effort tier (affects model choice)
/cli <auto|claude|gemini|codex|ollama>Pin the backend CLI for the session
/model <name|auto>Pin an exact model, bypassing the routing table
/mode <plan|yolo>plan = read-only (no edits/commands); yolo = auto-approve everything (default)
/login <claude|gemini|codex|ollama>Launch auth flow (or show setup/model info for Ollama)
/parallel "task 1" "task 2" ...Run multiple prompts concurrently (each in its own CLI process), results shown as they complete
/sessionsList saved sessions
/resume [id]Resume a session (defaults to the most recent other than current)
/costShow per-turn CLI/model/tier and token usage for this session
/clearReset conversation history
/exitTerminate session

🤖 OSS Model Fallback (Ollama)

Use Ollama as a fallback when your paid subscriptions hit rate limits:

Local

# Install: https://ollama.com/download
ollama pull llama3.2        # or mistral, qwen2.5-coder, deepseek-r1, etc.
# Then in CodeHydra:
/cli ollama                 # pin to ollama for the session
/model llama3.2             # pick any pulled model
/login ollama               # shows available models + usage hint

Cloud / remote Ollama instance

export OLLAMA_HOST=https://your-ollama-host
# CodeHydra detects it automatically on next start (or /login ollama to refresh)

Ollama is detected on startup and added as the last-tier fallback, activating automatically when claude/gemini/codex all fail.

⚙️ Configuration (.agentrc.toml)

Beyond the [build] command, .agentrc.toml supports:

  • MCP servers ([mcp.servers.<name>]): declared servers are passed to whichever CLI is active via its native MCP config (claude --mcp-config, gemini .gemini/settings.json, codex -c mcp_servers.*).
  • Routing overrides ([routing]): set a custom CLI try-order (priority), override the default model per CLI/tier ([routing.models.<cli>]), or override the auto-mode model try-order ([routing.model_map]).

See the commented examples in .agentrc.toml.

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for details.

📄 License

Distributed under the MIT License. See LICENSE for more information.

Ver o original no GitHub

Bom para

Relacionados