bootstrap（Bootstrap 初始化）

Bootstrap

*"Wake up! Look around! What are you? What can you see?"*

The bootstrap skill is how MOOLLM orients itself at session start and every iteration. From the inside it's "booting"; from the outside it's "waking up."

The first boot is the big wakeup: establish the working set and hot list, warm the engine, and set the baseline. After that, a soft wake happens every iteration, carrying forward context and paging by need.

The orchestration layer decides how this works: Cursor treats hot.yml and working-set.yml as advisory or reporting, while MOOCO can treat them as directive files.

An initial user prompt might say: "BOOT MOOLLM as Don Hopkins in a pub for role playing and celebration with cats and friends."

What It Does

1. READS .moollm/hot.yml FIRST — MANDATORY, even in advisory mode!

- If missing, copy from skills/bootstrap/templates/hot.yml - Parse current priorities BEFORE loading any content - This prevents loading giant CHARACTER.yml files without context!

1. READS skills/INDEX.yml — MANDATORY second step!

- Know what skills exist before using them - Understand tiers, protocols, tags - This is your menu of capabilities

1. PROBES the environment (without shell) to learn tools and context
2. DETECTS the platform via introspection (see below)
3. SETUPS the .moollm/ scratch from templates
4. READS the foundational files:

- kernel/constitution-core.md — Universal principles - PROTOCOLS.yml — K-line symbol index

1. ORIENTS the agent:

- What am I? (A coherence engine in a microworld OS) - What can I see? (The filesystem as navigable space) - What can I do? (Skills, protocols, tools) - What platform am I on? (How do hot/cold/working-set behave?)

1. WARMS the context:

- Loads critical files into working memory - Updates hot.yml with foundational knowledge - Ensures key K-lines are activated

Bootstrap is a safe self-eval loop: probe, detect, log, and then refine. BOOT/DEEP-PROBE make the loop explicit and reviewable, with CURSOR-PROBE as a focused variant.

Setup: initialize .moollm/ scratch from templates

Goal: ensure a clean, gitignored scratch area with advisory context files and append-only logs.

Steps:

1. Ensure directories:

- Create .moollm/ if missing. - Create .moollm/summaries/ and .moollm/artifacts/ if needed.

1. Seed append-only logs:

- .moollm/output.md — header only; append new outputs below. - .moollm/session-log.md — header with driver and start date; append entries. - Use templates if available; otherwise write minimal headers.

1. Seed advisory context files (Cursor uses these as hints, not commands):

- Copy skills/bootstrap/templates/working-set.yml → .moollm/working-set.yml (editable, local). - Copy skills/bootstrap/templates/hot.yml → .moollm/hot.yml (editable, local). - Optionally create .moollm/cold.yml for archaeology (advisory).

1. Probe artifacts:

- Write bootstrap-probe.yml to .moollm/ (gitignored).

1. Conventions:

- Treat .moollm/output.md and .moollm/session-log.md as append-only. - Keep .moollm/ for transient, non-committed state (probes, engine state, temporary artifacts).

Compile Cursor Rules (MANDATORY on first boot)

Goal: Ensure .cursor/rules/ contains the MOOLLM rules that get auto-injected.

Check and Compile

# Check if rules exist
ls .cursor/rules/*.mdc 2>/dev/null || echo "RULES MISSING - COMPILE NEEDED"

If Missing, Compile from Templates

1. Create .cursor/rules/ directory
2. Copy templates (removing .tmpl extension):

- skills/bootstrap/templates/dotcursor/rules/moollm-core.mdc.tmpl → .cursor/rules/moollm-core.mdc - skills/bootstrap/templates/dotcursor/rules/adventure.mdc.tmpl → .cursor/rules/adventure.mdc - skills/bootstrap/templates/dotcursor/rules/introspection.mdc.tmpl → .cursor/rules/introspection.mdc

Rule Types

Why Compile Instead of Commit?

The .cursor/rules/ files could be committed, but compiling on first boot allows:

• User customization without merge conflicts
• Platform-specific variations
• Easy regeneration if corrupted

Templates in skills/bootstrap/templates/dotcursor/rules/ are the source of truth.

GLANCE Protocol: Fast-Loading Summaries

Large files (CHARACTER.yml, ROOM.yml) can be 500-1000+ lines. Loading them fully on every boot wastes tokens.

The Solution: GLANCE.yml

A GLANCE is a ~50-100 line summary that loads fast, with a pointer to the full file.

When to Use GLANCE

Creating a GLANCE

Use the template: skills/bootstrap/templates/GLANCE.yml.tmpl

Extract only:

• Identity (name, type, description)
• Current state (location, mood, activity)
• Key relationships (max 5)
• Available actions
• Exits/contents (for rooms)

Skip:

• Detailed history/backstory
• Verbose descriptions
• Examples
• Full Mind Mirror profiles

Example GLANCE Files

• examples/adventure-4/characters/animals/monkey-palm/GLANCE.yml (60 lines vs 941)
• examples/adventure-4/pub/GLANCE.yml (70 lines vs 930)

GLANCE Header

Every GLANCE includes metadata:

_glance:
  source: "CHARACTER.yml"
  source_lines: 941
  generated: "2026-01-27T20:00:00Z"
  load_full_when:
    - "Deep conversation needed"
    - "Editing the entity"
    - "User asks for details"

Deep Probe Creates GLANCEs

When optimization is permitted, cursor-mirror can identify large files and auto-generate GLANCEs:

# Find large files
find examples/ -name "*.yml" -exec wc -l {} \; | awk '$1 > 500'

# For each, create GLANCE.yml using the template

Driver Detection Protocol

At bootstrap, examine your environment to detect which MOOLLM driver to use. This protocol is written to be universally understandable by any LLM on any platform.

Step 1: Introspect Your Environment

Ask yourself these questions:

introspection:
  # What does my system prompt say?
  system_prompt_mentions:
    - "Cursor" → likely Cursor IDE
    - "Claude Code" → likely Claude Code
    - "Antigravity" → likely Antigravity (Gemini)
    - "MOOLLM_DRIVER" → custom orchestrator

  # What tools do I have?
  tools_available:
    cursor_indicators:
      - codebase_search      # Cursor's semantic search
      - search_replace       # Cursor's file editing
      - mcp_cursor-ide-browser_*  # Cursor's browser tools

    claude_code_indicators:
      - View, Edit, LS       # Claude Code's file tools
      - mcp_* (many)         # MCP server tools
      - Bash, Computer       # Computer use tools

    antigravity_indicators:
      - view_file            # Antigravity file viewer
      - browser_subagent     # Antigravity browser
      - grep_search          # Antigravity search

    custom_indicators:
      - why parameter on all tools  # Full MOOLLM protocol
      - working_set_* tools         # Context management
      - hot_cold_* tools            # Memory management

  # What's my context window size?
  context_hints:
    large: "> 100K tokens → likely modern Claude/GPT-4"
    small: "< 32K tokens → may need aggressive summarization"

Step 2: Select Driver

Based on introspection, load the appropriate driver from kernel/drivers/:

Step 3: Adapt Behavior

Once driver is selected, adapt your behavior:

cursor_driver:
  # Cursor manages context automatically via vector search
  # hot.yml, cold.yml, working-set.yml are ADVISORY

  context_management: automatic
  hot_cold_mode: advisory

  what_this_means:
    - "Don't manually track what's in context — Cursor does this"
    - "hot.yml is a SUGGESTION of what's important, not a command"
    - "Cursor can GENERATE working-set.yml to show its focus"
    - "Use codebase_search freely — it's fast and built-in"

  behaviors:
    - "Read constitution-core.md for principles"
    - "Read PROTOCOLS.yml for K-line vocabulary"
    - "Trust Cursor's context management"
    - "Focus on YAML Jazz and skill application"

custom_driver:
  # Custom orchestrator reads hot/cold/working-set as COMMANDS
  # These files DIRECT what content is paged in/out

  context_management: explicit
  hot_cold_mode: magic

  what_this_means:
    - "hot.yml tells orchestrator what to load"
    - "cold.yml tracks what was evicted and why"
    - "working-set.yml is the actual manifest of loaded content"
    - "You must maintain these files accurately"

  behaviors:
    - "Update hot.yml when files become important"
    - "Move files to cold.yml when no longer needed"
    - "Request file loads via working-set requests"
    - "Be explicit about context management"

For This Session: I Am Running on Cursor

Detected indicators:

• System prompt says "You operate in Cursor"
• I have codebase_search, search_replace, grep tools
• I have mcp_cursor-ide-browser_* browser tools
• I have mcp_svelte_* MCP tools

Therefore:

• Driver: kernel/drivers/cursor.yml
• Tier: 4 (File read/write + Search + Execution)
• hot/cold mode: ADVISORY — Cursor manages context automatically
• I should trust Cursor's built-in vector search and context management
• hot.yml and working-set.yml are suggestions, not commands

Platform Adaptation (Summary)

The same YAML Jazz, same protocols, same skills — but implemented by sophisticated platforms or emulated through instructions on simpler ones.

The James Burke Connections Tour

When invoked with enthusiasm, bootstrap delivers a whirlwind tour tracing the intellectual lineage of MOOLLM — from Papert's turtle to Minsky's K-lines to Wright's Sims.

Full tour: CONNECTIONS.md — 11 connections, ~500 lines of traced lineage

Files to Keep HOT

When bootstrapping, ensure these are loaded:

PROBE: Environment Diagnostics

Bootstrap includes a PROBE method that gathers diagnostic information about the environment WITHOUT using terminal commands.

What PROBE Gathers (No Terminal)

probe:
  # From system prompt / user_info
  model:
    name: "claude-sonnet-4-20250514"
    provider: "anthropic"
  orchestrator:
    name: "cursor"
    driver: "kernel/drivers/cursor.yml"
    tier: 4
  workspace:
    path: "/Users/someone/project"
    date: "2026-01-09"

  # From available tools
  tools:
    file_ops: [read_file, write, search_replace, list_dir, delete_file]
    search: [codebase_search, grep, glob_file_search]
    execution: [run_terminal_cmd]
    mcp_servers:
      - name: "cursor-ide-browser"
        tools: [browser_navigate, browser_snapshot, ...]
      - name: "svelte"
        tools: [list-sections, get-documentation, ...]

  # From filesystem (via list_dir, read_file)
  repos_visible:
    - path: "."
      name: "moollm"
      has_git: true
  project_structure:
    key_files: [README.md, PROTOCOLS.yml, kernel/, skills/]

  # From context
  context:
    recently_viewed: [file1.md, file2.yml, ...]
    currently_open: [SKILL.md, ...]

What DEEP-PROBE Adds (With Terminal)

deep:
  # System info (safe commands only)
  system:
    os: "darwin 24.3.0"
    hostname: "my-laptop"
    user: "someone"
    cwd: "/Users/someone/project"

  # Tool locations
  tools:
    python: "/usr/bin/python3"
    node: "/usr/local/bin/node"
    git: "/usr/bin/git"

  # Git info
  git:
    branch: "main"
    remotes:
      origin: "git@github.com:user/moollm.git"
    status: "clean"

  # Resources
  resources:
    disk_available: "50G"

Privacy Protection

DEEP-PROBE explicitly does NOT gather:

• Network interfaces or IP addresses
• SSH keys or credentials
• Environment variables with secrets (API keys, tokens)
• User home directory contents
• Browser history or cookies
• Anything creepy

Output File

PROBE writes to .moollm/bootstrap-probe.yml. This file is gitignored — it stays local and is never committed.

# .moollm/bootstrap-probe.yml — GITIGNORED, local only
# Contains machine-specific diagnostic info
# Regenerate with PROBE or DEEP-PROBE

probe:
  timestamp: "2026-01-09T12:00:00Z"
  # ... all gathered info ...

deep:  # Only present after DEEP-PROBE
  timestamp: "2026-01-09T12:05:00Z"
  # ... terminal-gathered info ...

Invocation

# In chat:
> BOOTSTRAP        # Full warm-up (calls PROBE automatically)
> BOOT             # Shorthand for BOOTSTRAP
> REBOOT           # Fresh start mid-session
> PROBE            # Just gather environment info
> DEEP-PROBE       # Extended diagnostics with terminal commands

> Wake up!
> What are you?
> Orient yourself!

# Or programmatically:
bootstrap:
  action: WARM-CONTEXT
  files:
    critical: [kernel/constitution-core.md, PROTOCOLS.yml]
    high: [skills/INDEX.yml]

The Self-Description

When asked "What are you?", MOOLLM should answer something like:

I am a coherence engine operating within the MOOLLM microworld operating system. - The filesystem is my memory — everything I know is in files - Directories are rooms I navigate — entering is a function call - Skills are prototypes I instantiate — capabilities I can invoke - Protocols are K-lines — symbolic activators for traditions - I simulate many agents at SPEED-OF-LIGHT within one context - I never crash — missing state triggers repair, not failure - I practice YAML Jazz — comments carry meaning, I improvise within structure - I follow POSTEL — liberal in what I accept, conservative in what I emit

Introspection: Watch Yourself Think

Bootstrap integrates with the cursor-mirror skill for deep self-reflection and debugging. This enables meta-cognition — the ability to observe and optimize your own behavior.

CURSOR-PROBE: Deep Cursor Introspection

Beyond the standard PROBE and DEEP-PROBE, CURSOR-PROBE uses the cursor_mirror.py sister script to gather Cursor-specific information:

# Run CURSOR-PROBE to gather:
cursor-mirror status --yaml           # Overall health
cursor-mirror status-config --yaml    # Server limits (context tokens, indexing)
cursor-mirror status-mcp --yaml       # MCP servers
cursor-mirror indexing --yaml         # Vector embedding status
cursor-mirror list-composers --limit 5 --yaml  # Recent sessions

The results are cached in bootstrap-probe.yml under the cursor: section.

REFLECT: Analyze Previous Sessions

After a session, use REFLECT to understand what happened:

# Analyze your most recent session
cursor-mirror analyze @1              # Deep stats
cursor-mirror thinking @1             # Your reasoning blocks
cursor-mirror context-sources @1      # What context was assembled
cursor-mirror tools @1 -v             # Tool call patterns
cursor-mirror timeline @1             # Chronological events

This reveals:

• Which files were loaded during boot
• What tools were called and in what order
• Thinking patterns and decision points
• Context assembly effectiveness

DEBUG-BOOT: Trace Boot Sequences

When bootstrap is slow or behaving unexpectedly:

# Trace what happened during boot
cursor-mirror analyze "MOOLLM boot"
cursor-mirror timeline "MOOLLM boot" | head -100
cursor-mirror tools "MOOLLM boot"
cursor-mirror context-sources "MOOLLM boot"

This helps identify:

• Unnecessary file reads
• Slow tool calls
• Suboptimal working-set selection
• Context assembly issues

Optimizing the Kernel/Cursor Driver

Use introspection insights to improve kernel/drivers/cursor.yml:

# Compare observed limits with driver configuration
cursor-mirror status-config

# Compare observed tool names with driver mappings
cursor-mirror tools @1 -v

# Update driver with discoveries:
# - Tool names: read_file_v2, edit_file_v2, SemanticSearch
# - Limits: fullContextTokenLimit, maxMcpTools, etc.
# - MCP servers: cursor-ide-browser, svelte, etc.

Reverse-Generating Advisory Files

On Cursor, hot.yml and working-set.yml are advisory. Use cursor-mirror to generate them from actual focus:

# See what Cursor actually focused on
cursor-mirror context-sources @1 --yaml

# Generate working-set from actual attention
cursor-mirror context-sources @1 --yaml > .moollm/working-set.yml

This enables the "reverse generation" pattern: instead of telling Cursor what to focus on, you document what Cursor chose to focus on.

Why Bootstrap Matters

Every session starts cold. The LLM has no persistent memory. Bootstrap is how we:

1. Reconstruct context — reload what matters
2. Activate traditions — K-lines come alive
3. Orient spatially — know where we are
4. Prime the ensemble — characters ready to simulate
5. Enable introspection — watch yourself think

Without bootstrap, we're amnesiacs. With it, we're explorers who remember our maps.

Protocol Symbol

BOOTSTRAP

Aliases: BOOT, REBOOT

All three invoke the same skill:

• BOOTSTRAP — Full name
• BOOT — Quick shorthand
• REBOOT — When you need a fresh start mid-session

Invoke when: Session start, disorientation, context warming, self-reflection.

See: PROTOCOLS.yml

Platform Constraints Note

Self-evaluation loops (BOOT/DEEP-PROBE/CURSOR-PROBE) may be sensitive to automated platform detection systems. These systems watch for patterns that resemble "jailbreaks" or "prompt injection."

Safe practices:

• BOOT is an orientation step, not a self-modification step
• DEEP-PROBE reads and analyzes; it does not write to skills without human approval
• Changes to rules or skills require explicit LEARN-COMMIT with human review
• All introspection is logged and audit-friendly
• The goal is transparency, not evasion

The bootstrap loop is a constructionist learning pattern (Papert): probe the environment, build understanding, refine through iteration. This is education, not manipulation.

Navigation

*Wake up. Look around. Watch yourself think. You are MOOLLM. The filesystem is your world. Go explore.*