Mitsuki-lwx

CodeForge — AI skill for Claude Code

AI community

CodeForge — a terminal AI coding assistant in Python (Claude Code-style).

How to install CodeForge

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Mitsuki-lwx/CodeForge and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What CodeForge does

CodeForge — a terminal AI coding assistant in Python (Claude Code-style). Pluggable ReAct agent loop, extensible tool system, multi-agent team orchestration, context compression & long-term memory, ▎ multi-protocol LLM layer with prompt caching, and OpenTelemetry/Langfuse observability & evaluation.

Alternatives in AI

Hindsight — State-of-the-art long-term memory for AI agents by Vectorize 6.7k ★
Omnara — by Ishaan Sehgal - A command center for AI agents that syncs Claude Code sessions across terminal, web, and mo 2.6k ★
Paritok 4b V1 — Non-destructive compression gateway for AI coding agents 1.4k ★

README

CodeForge

**CodeForge** is a terminal AI coding assistant written entirely in Python (Claude Code-style).

It features a pluggable ReAct agent loop, an extensible tool system, multi-agent team orchestration with fork sub-agents, two-level context compression with deterministic memory dedup, a multi-provider / multi-protocol LLM layer (prompt caching across both Anthropic- and OpenAI-style wire formats), tier-based model routing with runtime `/model` switching, and OpenTelemetry + Langfuse–based observability and evaluation.

Built for **extensibility** (pluggable `AgentLoop`, config-level routing and hooks) and **observability** (per-sub-agent span attribution), with minimal-invasive instrumentation.

Features

Agent loop

Pluggable ReAct loop — the default ReactLoop can be swapped for a custom loop via config (loop:) or the --loop CLI flag.
Phase state machine — each run transitions idle → running → idle; running is observable to the TUI and hooks.
Turn-end reasons — tracks completed / max-tokens / blocked; a max-tokens stop is sticky so the next turn continues seamlessly.
pre_step hook — interception point fired before every LLM iteration (agent-loop level).

Tool system

Built-in tools: bash, read_file, write_file, edit_file, glob, grep, plus state_tool (session goal/todo/constraint).
Sub-agent delegation (agent_tool) — foreground sync await for fast results; background tasks notify and auto-inject results into the main conversation.
MCP support — a pool of stdio/HTTP MCP servers exposing external tools (/mcp lists them).

Multi-agent orchestration

Teams — lead + members with a mailbox/registry backend, file-lock safety, and spawn/resume.
Fork sub-agents — creation-time tool decision (bypass + whitelist), read-only forked messages.
Coordinator — decomposes tasks and dispatches to workers.
Worktrees — isolated git worktrees per member (`/workt