Continuous-Claude-v3
parcadei · parcadei/Continuous-Claude-v3
This plugin transforms Claude Code into a continuously learning agent by managing context, orchestrating subagents, and maintaining memory across sessions to prevent token waste and context loss.
Context management for Claude Code. Hooks maintain state via ledgers and handoffs. MCP execution without context pollution. Agent orchestration with isolated context windows.
Preview
What you can do with it
- Maintain context and memory for Claude Code across extended development sessions.
- Orchestrate specialized subagents for complex, multi-step tasks.
- Prevent context pollution and token waste during MCP execution in Claude Code.
Safety check
Usable, but read the notes first.
- ●Installer downloads a remote script and pipes it to the shell.
README.md
Automated scan of the repo files. Always review code you run.
From the README
Continuous Claude
A persistent, learning, multi-agent development environment built on Claude Code
Continuous Claude transforms Claude Code into a continuously learning system that maintains context across sessions, orchestrates specialized agents, and eliminates wasting tokens through intelligent code analysis.
Table of Contents
- Why Continuous Claude?
- Design Principles
- How to Talk to Claude
- Quick Start
- Architecture
- Core Systems
- Workflows
- Installation
- Updating
- Configuration
- Contributing
- License
Why Continuous Claude?
Claude Code has a compaction problem: when context fills up, the system compacts your conversation, losing nuanced understanding and decisions made during the session.
Continuous Claude solves this with:
| Problem | Solution |
|---|---|
| Context loss on compaction | YAML handoffs - more token-efficient transfer |
| Starting fresh each session | Memory system recalls + daemon auto-extracts learnings |
| Reading entire files burns tokens | 5-layer code analysis + semantic index |
| Complex tasks need coordination | Meta-skills orchestrate agent workflows |
| Repeating workflows manually | 109 skills with natural language triggers |
The mantra: Compound, don't compact. Extract learnings automatically, then start fresh with full context.
Why "Continuous"? Why "Compounding"?
The name is a pun. Continuous because Claude maintains state across sessions. Compounding because each session makes the system smarter—learnings accumulate like compound interest.
Design Principles
An agent is five things: Prompt + Tools + Context + Memory + Model.
| Component | What We Optimize |
|---|---|
| Prompt | Skills inject relevant context; hooks add system reminders |
| Tools | TLDR reduces tokens; agents parallelize work |
| Context | Not just what Claude knows, but how it's provided |
| Memory | Daemon extracts learnings; recall surfaces them |
| Model | Becomes swappable when the other four are solid |
Anti-Complexity
We resist plugin sprawl. Every MCP, subscription, and tool you add promises improvement but risks breaking context, tools, or prompts through clashes.
Our approach:
- Time, not money — No required paid services. Perplexity and NIA are optional, high-value-per-token.
- Learn, don't accumulate — A system that learns handles edge cases better than one that collects plugins.
- Shift-left validation — Hooks run pyright/ruff after edits, catching errors before tests.
The failure modes of complex systems are structurally invisible until they happen. A learning, context-efficient system doesn't prevent all failures—but it recovers and improves.
How to Talk to Claude
You don't need to memorize slash commands. Just describe what you want naturally.
The Skill Activation System
When you send a message, a hook injects context that tells Claude which skills and agents are relevant. Claude infers from a rule-based system and decides which tools to use.
> "Fix the login bug in auth.py"
🎯 SKILL ACTIVATION CHECK
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ CRITICAL SKILLS (REQUIRED):
→ create_handoff
📚 RECOMMENDED SKILLS:
→ fix
→ debug
🤖 RECOMMENDED AGENTS (token-efficient):
→ debug-agent
→ scout
ACTION: Use Skill tool BEFORE responding
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Priority Levels
| Level | Meaning |
|---|---|
| ⚠️ CRITICAL | Must use (e.g., handoffs before ending session) |
| 📚 RECOMMENDED | Should use (e.g., workflow skills) |
| 💡 SUGGESTED | Consider using (e.g., optimization tools) |
| 📌 OPTIONAL | Nice to have (e.g., documentation helpers) |
Natural Language Examples
| What You Say | What Activates |
|---|---|
| "Fix the broken login" | /fix workflow → debug-agent, scout |
| "Build a user dashboard" | /build workflow → plan-agent, kraken |
| "I want to understand this codebase" | /explore + scout agent |
| "What could go wrong with this plan?" | /premortem |
| "Help me figure out what I need" | /discovery-interview |
| "Done for today" | create_handoff (critical) |
| "Resume where we left off" | resume_handoff |
| "Research auth patterns" | oracle agent + perplexity |
| "Find all usages of this API" | scout agent + ast-grep |
Why This Approach?
| Benefit | How |
|---|---|
| More Discoverable | Don't need to know commands exist |
| Context-Aware | System knows when you're 90% through context |
| Reduces Cognitive Load | Describe intent naturally, get curated suggestions |
| Power User Friendly | Still supports /fix, /build, etc. directly |
Skill vs Workflow vs Agent
| Type | Purpose | Example |
|---|---|---|
| Skill | Single-purpose tool | commit, tldr-code, qlty-check |
| Workflow | Multi-step process | /fix (sleuth → premortem → kraken → commit) |
| Agent | Specialized sub-session | scout (exploration), oracle (research) |
See detailed skill activation docs →
Quick Start
Prerequisites
- Python 3.11+
- uv package manager
- Docker (for PostgreSQL)
- Claude Code CLI
Installation
# Clone
git clone https://github.com/parcadei/Continuous-Claude-v3.git
cd Continuous-Claude-v3/opc
# Run setup wizard (12 steps)
uv run python -m scripts.setup.wizard
Note: