dino-spec

Automatic context engineering for Claude Code

       __
      (_ \
  .-^^^  /
 /       /
<__|-|__|

dino-spec brings structured, context-aware development to Claude Code. It automatically maintains session state, provides 54 MCP tools, and enables intelligent context retrieval with real-time statusline display.

v10.0.0 Breaking Change: The dino mcp-serve command has been removed. Use dino mcp serve instead.

Features

• MCP Integration - 54 tools + 29 resources auto-configured during init
• Ralph Loop (v11.2.0) - Fresh-context iterative execution for large-scale development
  • Story-Based PRD - Define multi-story features with acceptance criteria
  • Fresh Context - Each iteration gets clean context (no history accumulation)
  • Git Checkpoints - Automatic checkpoints for rollback capability
  • Learning Propagation - Insights from failed iterations inform future attempts
  • Max 10 iterations - Hard limit with max 2 attempts per story
  • MCP tools: dino_ralph_init, dino_ralph_iterate, dino_ralph_status, dino_ralph_learn
• MCP Tool Aggressive Filtering (v10.3.0) - Phase-based tool context reduction (~50-70% savings)
  • Phase Filtering - Only load tools relevant to current workflow phase
  • Description Modes - Minimal/standard/full verbosity control
  • Lazy Details - Full descriptions loaded on-demand via dino://tools/details
  • Auto-Enabled - Filtering enabled by default, opt-out with DINO_AGGRESSIVE_FILTERING=false
• Lazy Context Loading (v10.1.0) - Defer file loading to save tokens (~60-80% initial context reduction)
  • Deferred Resources - Replace file content with lightweight placeholders (~5-10 tokens)
  • On-Demand Expansion - Load content only when needed with dino_expand
  • Token Budget - Control expansion with budget constraints
  • HUD Integration - Real-time display of deferred/loaded resources
• Self-Validate Hooks (v9.16.0) - Proactive validation system that catches edge cases before execution
  • Ambiguity Detection - Identifies vague terms, incomplete requirements, missing specifics
  • Edge Case Analysis - Null checks, empty collections, error handling coverage
  • Confidence Scoring - Weighted validation with thresholds (block < 0.4, ask < 0.6, proceed ≥ 0.6)
  • Impact Assessment - Breaking changes, security risks, cross-cutting concerns
  • Question Generation - Auto-converts issues to user questions with recommended options
  • Pre-validate hook blocks low-confidence operations before execution
• Phase 4 Advanced Optimizations (v9.14.0)
  • MCP Tool Filtering - Phase-based tool tiering for 60-70% context overhead reduction
  • Memory Recitation - SM-2 spaced repetition for long-term memory retention
  • Entropy Enhancement - Language-specific profiles and cross-section deduplication
  • External Vector DB - Qdrant/Pinecone adapter support for scalable embeddings
• Enhanced Observability - Execution tree, memory access log, context transitions with HTML/JSON export (v9.10.0)
• Session State - Automatic tracking of focus, test/build/lint status, blockers
• Persistent Memory - Decisions, patterns, learnings survive across sessions
• Interactive Memory Browser - Full-featured TUI for memory management with search, graph view, and actions (v10.0.0)
• Real-Time HUD - Statusline integration showing budget, phase, tests, focus, git status (v9.1.0)
• Session Handoff - Context continuity system for seamless session transitions (v9.1.0)
• Workflow Skills - Phases from discovery to completion (/dino.*)
• Task Delegation - Specialized agents with context handoffs
• Recursive Agents - Self-spawning agents with complexity-based depth control (v8.0.0)
• Recursive Retrieval (v10.4.0) - Enhanced RLM-style query decomposition with:
  • Semantic Decomposition - Intent-based query splitting by distinct concepts
  • Dependency-Aware Decomposition - Sequential sub-queries with prerequisite analysis
  • Composition Decomposition - Parallel execution for "A and B" style queries
  • Quality Confidence Bounds - Uncertainty tracking with Wilson score intervals
  • Task-Specific Weight Calibration - Automatic quality weight adjustment per task type
  • Provenance Tracking - Full source chain attribution for every result
  • Conflict Detection - Automatic identification of score variance and data mismatches
  • Decision Rationale - Explainability with alternatives considered
• Dynamic Partitioning - Intelligent codebase segmentation (directory, feature, dependency, time)
• Context REPL - Declarative DSL for model-driven exploration
• Spec-Kit: Intelligent Prompt Analysis (v10.5.0 - v10.8.0) - Auto-specification system that analyzes prompts before implementation
  • Complexity Scoring - Ambiguity, scope, risk, novelty analysis (0-100 scale)
  • Confidence Gating - 80% confidence threshold prevents premature implementation
  • Auto-Trigger - Automatic /dino.spec suggestion for complex prompts (3+ signals)
  • Multi-Part Decomposition - Splits "add X and Y" into separate tasks
  • Context-Aware Questions - References existing patterns and focus area
  • Learning System - Refines scoring weights based on outcome feedback
  • Confidence Visualization (v10.8.0) - Progress bar, factor breakdown, improvement suggestions
  • MCP tools: dino_spec_analyze, dino_spec_feedback, dino_spec_stats, dino_spec_weights
• Context Engineering - Research-backed token optimization (v9.6.0, v9.7.0, v9.14.0)
  • Tool result summarization (~50-70% better retention)
  • Positional context optimization (~15-25% recall improvement)
  • Just-in-time context loading (~60-80% initial context reduction)
  • Entropy-guided budgeting (~20-30% token reduction)
  • Virtual memory management (~40-50% memory overhead reduction)
  • Multi-agent context isolation (~70-80% token reduction for parallel agents)
  • Semantic memory validation (~20-30% better relevance)
  • Language-specific entropy profiles (v9.14.0)
  • Cross-section deduplication (v9.14.0)
  • Reference docs and skills (/research, /context-optimize)

Installation

npm install -g dino-spec

Quick Start

cd your-project
dino init          # Analyze codebase, install hooks, configure MCP

# MCP server is auto-configured for Claude Code
# Use /mcp in Claude Code to verify

CLI Commands

Workflow Skills

Use in Claude Code with /dino.<phase>:

Utilities: /dino.status, /dino.focus, /dino.next, /dino.handoff (v9.1.0), /dino.release (v9.4.0)

Spec-Kit: /dino.spec (v10.5.0) - Intelligent prompt analysis before implementation

Context Engineering: /research (v9.7.0), /context-optimize (v9.7.0)

MCP Resources

Access via @dino:// in Claude Code:

MCP Tools

Key tools available via Claude Code:

Specialized Agents

Memory System

Persistent memories in .dino/memories/:

dino memory list              # View all memories
dino memory search "auth"     # Search memories
dino memory decay             # Clean stale memories

Project Structure

your-project/
├── CLAUDE.md                 # Project context
├── .claude/
│   ├── skills/dino/          # Workflow skills
│   ├── rules/dino/           # Workflow rules
│   └── agents/               # Specialized agents
└── .dino/
    ├── session.md            # Session state
    ├── structure.json        # Codebase analysis
    ├── tasks/                # Task registry
    ├── checkpoints/          # Agent checkpoints
    └── memories/             # Persistent memory

Requirements

• Node.js >= 20.0.0
• Claude Code CLI

Troubleshooting

MCP not showing in /mcp

# Windows
claude mcp add-json --scope user dino '{"command":"cmd","args":["/c","npx","dino-spec","mcp","serve"]}'

# macOS/Linux
claude mcp add --scope user dino -- npx dino-spec mcp serve

Then restart Claude Code.

License

MIT

Built for developers who believe AI coding should be structured, reproducible, and context-aware.