mulch

Category: Content & Multimedia | Uploader: RuneweaverStudios | Downloads: 0 | Version: v1.0（Latest）

OpenRouter does not support image input endpoints',

Changelog: Source: GitHub https://github.com/RuneweaverStudios/mulch

Directory Structure

Current level: Root

📁 .learnings/
- 📄 ERRORS.md 75 B
- 📄 FEATURE_REQUESTS.md 84 B
- 📄 LEARNINGS.md 99 B
📁 assets/
- 📄 LEARNINGS.md 1.4 KB
- 📄 SKILL-TEMPLATE.md 3.3 KB
📁 bin/
- 📄 mulch 4.6 KB
📁 config/
- 📄 domains.json 3.0 KB
📁 docs/
- 📄 benchmark-comparison.html 15.5 KB
- 📄 benchmark-elevator-pitch.html 6.1 KB
📁 hooks/
- 📁 openclaw/
  - 📄 handler.js 5.3 KB
  - 📄 handler.ts 1.2 KB
  - 📄 HOOK.md 761 B
📁 references/
- 📄 benchmark.md 2.8 KB
- 📄 examples.md 8.1 KB
- 📄 hooks-setup.md 5.0 KB
- 📄 openclaw-integration.md 6.1 KB
📁 scripts/
- 📁 benchmark/
  - 📄 legacy-reminder.txt 633 B
  - 📄 mulch-reminder.txt 453 B
  - 📄 nanobot-baseline.sh 3.4 KB
  - 📄 nanobot-mulch.sh 2.3 KB
  - 📄 nanobot-style-baseline.sh 2.9 KB
  - 📄 nanobot-style-mulch.sh 3.7 KB
  - 📄 nanobot-troubleshooting.sh 2.9 KB
  - 📄 run-benchmark.sh 6.8 KB
- 📄 activator.sh 771 B
- 📄 docker-test-hook.js 4.3 KB
- 📄 docker-test.sh 5.8 KB
- 📄 error-detector.sh 1.2 KB
- 📄 extract-skill.sh 5.0 KB
📁 visual-explainer/
- 📁 prompts/
  - 📄 diff-review.md 8.2 KB
  - 📄 fact-check.md 4.6 KB
  - 📄 generate-web-diagram.md 913 B
  - 📄 plan-review.md 9.0 KB
  - 📄 project-recap.md 7.0 KB
- 📁 references/
  - 📄 css-patterns.md 27.0 KB
  - 📄 libraries.md 15.1 KB
  - 📄 responsive-nav.md 5.7 KB
- 📁 templates/
  - 📄 architecture.html 17.1 KB
  - 📄 data-table.html 15.9 KB
  - 📄 mermaid-flowchart.html 13.0 KB
- 📄 .gitignore 10 B
- 📄 banner.png 1.1 MB
- 📄 CHANGELOG.md 2.7 KB
- 📄 LICENSE 1.0 KB
- 📄 README.md 6.1 KB
- 📄 SKILL.md 18.3 KB
📄 .dockerignore 67 B
📄 .gitignore 194 B
📄 _meta.json 1.8 KB
📄 BENCHMARK.md 10.1 KB
📄 config.json 184 B
📄 Dockerfile 684 B
📄 mulch.js 7.7 KB
📄 package.json 244 B
📄 QUALIFICATION.md 4.7 KB
📄 README.md 4.6 KB
📄 SKILL.md 2.8 KB

SKILL.md

# Mulch Self Improver — Code Mode Edition 🌱

> Token-efficient memory system using Code Mode architecture. **99.9% token savings** — only relevant memories loaded, not entire history.

## Concept

Mulch records learnings in `.mulch/` (JSONL files). Code Mode lets you **search** and **execute** without loading all memories into context.

## Tools (Only 2!)

### search(query)
Filters memories using named safe filters (no arbitrary code execution). Pass a filter name as a string, or an object with `{filter, ...params}`.

```javascript
// Example: Find all failures related to OpenRouter
const learnings = await search({ filter: 'byType', type: 'failure' });

// Search by keyword across all domains
const results = await search('openrouter');

// Filter by domain
const apiLearnings = await search({ filter: 'byDomain', domain: 'api' });
```

### execute(action)
Records or updates learnings. Actions: `record`, `query`, `promote`.

```javascript
// Record a new learning
await execute({
  action: 'record',
  domain: 'api',
  type: 'failure',
  description: 'OpenRouter does not support image input endpoints',
  resolution: 'Use Ollama + Molmo/llava for local vision instead'
});

// Query domain
await execute({ action: 'query', domain: 'openclaw' });

// Promote to project memory
await execute({ action: 'promote', id: 'api:123', to: 'SOUL.md' });
```

## Setup

```bash
# Install
npm install -g mulch-cli

# Initialize
mulch init

# Add domains
mulch add api database testing openclaw marketing
```

## Workflow

1. **Session start:** `search()` relevant memories (not all!)
2. **During work:** When you learn something → `execute({action: 'record', ...})`
3. **Before finishing:** Review → promote high-value learnings

## Auto-Detection (Hook)

The hook detects:
- Errors/failures → prompts to record
- Corrections ("no", "actually", "wrong") → prompts to record
- Retries → prompts to record

## Domains (24 preset)

```
api, database, testing, frontend, backend, infra, docs, config,
security, performance, deployment, auth, errors, debugging,
workflow, customer, system, marketing, sales, content,
competitors, crypto, automation, openclaw
```

## Record Types

| Type | Use Case |
|------|----------|
| `failure` | What went wrong + fix |
| `convention` | Best practice ("use pnpm not npm") |
| `pattern` | Named reusable pattern |
| `decision` | Architecture/tech choice |
| `guide` | Step-by-step procedure |
| `reference` | Key file/endpoint |

## Token Savings

| Approach | Tokens |
|----------|--------|
| Old (all in context) | ~50,000+ |
| Code Mode (filtered) | ~500 |

**That's 99% savings!** 🎯

## Integration

Add to AGENTS.md/CLAUDE.md:
```
Run search() to query relevant memories at session start.
Use execute({action: 'record', ...}) to capture learnings.
Promote proven patterns to SOUL.md/AGENTS.md.
```

Comments 0

Latest | Hot

Please login before commenting.

No comments yet. Be the first one!

mulch

Directory Structure

SKILL.md

Report

Notice