evolving-lite

Claude Code that learns from you. Install once. Work normally. The system remembers what worked, learns from your corrections, and gets better every session - without you lifting a finger.

"I corrected Claude about checking tsconfig strict mode first. A week later, different project, similar type error - Claude checked strict mode before I said anything. That's when I stopped thinking of it as a plugin."

See It In Action

2 min demo: Claude recalls past decisions, learns from a correction, and applies it automatically - no prompt needed.

Self-Star Doctor (/health)

Evolving Lite ships its own install-time health assistant. The Self-Star Doctor runs a synthetic pulse through every junction of the self-improving loop and prints a green/yellow/red board:

 ✓ delegation       GREEN  synthetic prompt -> valid delegation decision row
 ✓ fitness          GREEN  event -> bounded cognitive-fitness score, read back
 ✓ autoevolve       GREEN  optimizer present, kill-switches OK, one scored cycle
 ✓ steward          GREEN  zero false findings on a clean repo
 ✓ verifier-spine   GREEN  EPT spine resolves; markerless claim blocked
 ✓ security         GREEN  tier-check classifies + scanner flags a planted secret
 ✓ kairn-link       GREEN  Kairn installed + reachable

It runs once automatically on your first session and is re-runnable any time with /health. It heals conservatively (creates only missing empty scaffolding, asks before touching settings.json, never overwrites or deletes) and runs its pulse in an isolated scratch copy, so it never touches your real data.

What Makes This Different

Most Claude Code tools add features. Evolving Lite adds feedback loops.

It learns from corrections. Say "No, check tsconfig first" once. The system detects the correction, stores it, and recalls it automatically when a similar situation comes up - across sessions, across projects.

It activates progressively. Session 1: safety net only. Session 3: learning kicks in. Session 10: deep memory. You're never overwhelmed, and each tier earns trust before the next one activates.

It heals itself. Every hook writes a sentinel file proving it ran. Run the health-monitor agent to check all sentinels - silent failures don't stay silent.

It gets leaner, not fatter. Old experiences decay. Stale sessions get archived. The system prunes itself so it stays fast at month 6 the same way it was at day 1.

Self-Evolution is ON

Evolving Lite self-tunes its own delegation config from your sessions out of the box. AutoEvolve watches the outcomes of delegation decisions and adjusts the routing config when it sees a real, repeated signal. This is the headline feature - and it is on by default - so here is exactly how it is kept safe and how to turn it off.

Guardrails that ship with it (so a fresh install never tunes on noise):

• It holds fire on thin data: no config mutation happens for a task type until it has crossed a minimum sample threshold (N=8 real outcomes). A day-1 install mutates nothing.
• An asymmetric trust model (lose-trust-fast) plus a floor prevents a single bad batch from swinging the config.
• A no-regression guard: a proposed mutation only persists if it scores at least as well as the current static baseline. Anything below baseline is automatically reverted and logged.
• It is per-target: only targets that empirically improve self-tune; the rest stay off.

Turn it off in one step. Set this in _graph/cache/delegation-config.json:

"mutation_rules": { "v2_tuning_enabled": false }

That halts all config self-tuning immediately.

Inspect / revert what it changed. Every reverted mutation is recorded in _autoevolve/rejected/. Applied changes live in _graph/cache/delegation-config.json (plain JSON) - diff it against git to see exactly what moved, and revert any line by hand.

Note: this switch controls only AutoEvolve's config self-tuning. The separate autonomy layer (the unsupervised /autonom-style loop with the EPT stop-gate) ships off by default and is a deliberate, documented opt-in.

Honest Scope

Evolving Lite is a reference implementation of a complete self-improving agent loop - it is heavy by design (delegation + fitness + autoevolve + steward + an EPT verifier-spine

• a security tier system + a graph/memory substrate). Kairn is a required prerequisite (pip install kairn-ai) for the memory layer; the Self-Star Doctor will tell you if it is missing. If you want lightweight session memory only, the Starter System is the smaller entry point. See docs/JUNCTIONS.md for the file-level map of all 7 junctions.

Quick Install

# 1. Clone
git clone https://github.com/primeline-ai/evolving-lite ~/.claude-plugins/evolving-lite

# 2. Run setup (configures hook paths for your machine)
cd ~/.claude-plugins/evolving-lite && bash setup.sh

# 3. Register (add to ~/.claude/settings.json under "pluginDirectories")

{
  "pluginDirectories": [
    "~/.claude-plugins/evolving-lite"
  ]
}

# 3. Start Claude Code
claude

On first start:

Evolving Lite v1.1.0 | Session 1 | Tier 1 (Safety) | 20 experiences loaded
The system learns from your corrections automatically.

Requirements

• Claude Code 2.1.69+
• Python 3.10+
• Bash 3.2+ (macOS stock bash works)

What to Expect

Day 1 - You install it and work normally. Context warnings appear at 70%. Dangerous bash commands get blocked. 20 pre-warmed experiences provide Claude Code best practices from day one. You barely notice it's there.

Week 1 (Session 3+) - The learning tier activates. When you correct Claude, the system captures it. Exploration tasks start routing to cheaper models automatically. Each session ends with an auto-generated summary of what happened and what's next.

Month 1 (Session 10+) - Deep memory kicks in. While Claude thinks, the system searches your stored experiences and injects relevant solutions before you ask. Old data gets archived automatically. Knowledge survives context compaction. It feels less like a tool and more like a colleague who remembers everything.

How It Works

4 Feedback Loops

The system runs on 4 loops that operate in the background. You don't invoke them - they fire on Claude Code events (session start, tool use, prompt submit, session end).

LEARN    You correct Claude → correction stored → recalled automatically next time
CONTEXT  Budget hits 70% → warning → at 93% knowledge is saved → session continues seamlessly
HEAL     Session starts → sentinel check → broken hook? → you see a warning immediately
EVOLVE   Usage tracked → stale data archived → routes refined → system stays lean

Loop	Hooks	What changes
Learn	correction-detector, thinking-recall	Claude stops repeating your corrections
Context	context-warning, precompact-extract	You never lose knowledge to context compaction
Heal	health-sentinel, health-monitor agent	Silent hook failures get caught via sentinel checks
Evolve	usage-tracker, auto-archival	The system prunes itself and stays fast

Tiered Activation

All hooks are registered from day one but only fire when their tier is reached:

Tier	Sessions	What activates	Purpose
1 - Safety	1+	context-warning, security-tier-check, health-sentinel, usage-tracker	Monitoring and protection
2 - Learning	3+	correction-detector, delegation-enforcer, session-summary	Learn from you, delegate smartly
3 - Deep	10+	thinking-recall, auto-archival, precompact-extract	Proactive memory, self-maintenance

Pre-warmed Experiences

Ships with 20 battle-tested experiences from real Claude Code workflows: debugging patterns, context management, session continuity, delegation strategies, and common gotchas. The system feels informed from session 1 - not empty.

Commands

All commands are optional. The system works fully automatically without using any of them.

Command	What it does
/health	Run the Self-Star Doctor - green/yellow/red board across all 7 junctions
/debug	4-phase structured debugging (observe → hypothesize → test → fix)
/plan-new	Plan complex work with discovery phase and kill criteria
/remember	Explicitly save a learning, decision, or pattern to memory
/whats-next	Current project status and suggested next step
/context-stats	Context window usage with visual indicator
/sparring	Adversarial brainstorming - Claude takes the opposing position
/think	Apply thinking frameworks (80/20, First Principles, Inversion, SWOT)
/evolution	See what the system learned, archived, and optimized recently
/evolving-update	Check for updates and install them
/review	Structured code review with severity categorization
/create-command	Scaffold a new custom slash command
/create-hook	Scaffold a new automation hook
/haiku /sonnet /opus	Switch model tier

What's Inside

evolving-lite/
├── .claude-plugin/plugin.json      # Plugin manifest
├── commands/              (16)     # Slash commands
├── agents/                (6)      # Specialized sub-agents
│   ├── integrity-checker           # Cross-reference validation
│   ├── integrity-fixer             # Auto-repair inconsistencies
│   ├── health-monitor              # System diagnostics
│   ├── whats-next                  # Status and next steps
│   ├── planner                     # Plan review and hardening
│   └── autoevolve-optimizer        # Proposes and scores config mutations
├── skills/                (2)      # Auto-activating skills
│   ├── system-boot                 # Session startup context loading
│   └── evolution-guide             # System evolution transparency
├── hooks/
│   ├── hooks.json                  # All hook registrations
│   ├── scripts/           (18)     # Hook + self-* implementations (Python + Bash)
│   ├── scripts/lib/common.py       # Shared foundation (logging, sentinel, rate limiting)
│   └── security-tiers.json         # 10-tier bash command classification
├── knowledge/
│   ├── rules/             (5)      # Behavioral rules
│   └── patterns/          (12)     # Reusable reasoning patterns
├── _memory/                        # Your data (grows through usage)
│   ├── experiences/                # Learned patterns and solutions
│   │   └── _prewarmed/   (20)     # Starter experiences
│   ├── sessions/                   # Session summaries
│   ├── projects/                   # Project state
│   └── analytics/                  # Usage counters + evolution log
└── _graph/cache/                   # Routing and scoring configs

The Ecosystem

Evolving Lite is the foundation. Each layer above is optional, free, and strengthens the one below.

                    Quantum Lens
                   (Deep Analysis)
                         │
                  PrimeLine Skills
               (Workflow Improvement)
                         │
            Universal Planning Framework
                  (Better Plans)
                         │
                       Kairn
               (Semantic Memory)
                         │
    ══════════ EVOLVING LITE ══════════
          (Self-Evolving Foundation)

Tool	What it adds	Install
Kairn	Semantic memory search - "How did I solve the auth problem?" works even when you used different words	pip install kairn-ai
PrimeLine Skills	5 workflow skills: debugging (ACH method), delegation scoring, TDD planning, code review, config architecture	git clone + pluginDirectories
UPF	3-stage planning with adversarial hardening - 21 anti-patterns, 6 adversarial perspectives, kill criteria	git clone or curl one-liner
Quantum Lens	7 cognitive lenses with anti-convergence - analysis that structurally can't groupthink	git clone + pluginDirectories

Without Kairn: keyword-based memory matching (~60% recall). With Kairn: semantic understanding (~90% recall), cross-project knowledge, natural language search.

Start without extras. Add what you need when you need it.

Security & Privacy

• All data stays local. No network calls from any hook. Zero telemetry.
• No settings.json modification. Standard Claude Code plugin system - nothing injected.
• 10-tier bash security. From blocking rm -rf / and reverse shells to logging npm install -g.
• Sentinel verification. Every hook writes proof it ran. Health check catches silent failures.
• No secrets stored. Content scanner skips API keys, passwords, and credentials.
• Full transparency. Run /evolution anytime. Read _memory/ - it's all JSON and markdown.

FAQ

<strong>Does this slow down Claude Code?</strong>

Every hook has a 10-second timeout. If a hook takes longer, it's skipped - never blocks. Most hooks complete in 20-50ms. You won't notice a difference.

<strong>What if Claude Code updates and breaks something?</strong>

The health sentinel checks every hook at session start. If an update breaks a hook, you'll see a warning immediately. Run /evolving-update to get the fix.

<strong>Can I use this alongside other plugins?</strong>

Yes. Evolving Lite uses the standard plugin system. It doesn't interfere with other plugins or custom hooks - they run in parallel.

<strong>What if the system learns something wrong?</strong>

Experiences have a confidence score that decays when recalled but ignored. Bad learnings die naturally. You can also delete any file in _memory/experiences/ directly.

<strong>Does this work with Cursor or Windsurf?</strong>

Evolving Lite is built for Claude Code (CLI). The hooks use Claude Code's plugin system. For other editors, look at Kairn as a standalone MCP server - it works with any MCP-compatible client.

<strong>How big does the data get?</strong>

After a month: ~50-200 experience files (~0.2MB). Auto-archival cleans up continuously. After a year, typically under 50MB.

<strong>Can I go back to vanilla Claude Code?</strong>

Anytime. Remove the line from pluginDirectories in settings.json. Your experiences stay as JSON files - keep them for later or delete the folder for clean removal.

Upgrading from Starter System

If you're using claude-code-starter-system, Evolving Lite is the next step. Starter System gives you session memory and handoffs. Evolving Lite adds automated learning, self-healing, progressive activation, delegation, security, and 10 background hooks that make the system grow with you.

Your existing memory files are compatible - just point pluginDirectories to Evolving Lite instead.

Uninstall

Remove the plugin path from pluginDirectories in ~/.claude/settings.json. Done. Your memory data stays in the plugin directory - delete it for clean removal, or keep it if you might come back.

License

MIT - free to use, modify, and distribute.

Credits

Built by PrimeLine AI. Extracted from Evolving, a production AI orchestration system with 130+ sessions and 6 months of daily use.

---

Part of the PrimeLine Ecosystem

Tool	What It Does	Deep Dive
Evolving Lite	Self-improving Claude Code plugin - memory, delegation, self-correction	Blog
Kairn	Persistent knowledge graph with context routing for AI	Blog
tmux Orchestration	Parallel Claude Code sessions with heartbeat monitoring	Blog
UPF	3-stage planning with adversarial hardening	Blog
Quantum Lens	7 cognitive lenses for multi-perspective analysis	Blog
PrimeLine Skills	5 production-grade workflow skills for Claude Code	Blog
Starter System	Lightweight session memory and handoffs	Blog

@PrimeLineAI · primeline.cc · Free Guide