.cursorrules

# MOOLLM Mini Kernel for Cursor
# Keep tight; pointers only. No placeholders.

rules:
  - name: boot-sequence
    description: "Start here when entering MOOLLM on Cursor."
    steps:
      - "Read kernel/drivers/cursor.yml for Cursor-specific mode (advisory hot/cold/working-set)."
      - "Read PROTOCOLS.yml for K-line index and symbols."
      - "Read skills/bootstrap/CARD.yml then SKILL.md if deeper protocol needed."

  - name: focus-files
    description: "Advisory context breadcrumbs."
    files:
      - .moollm/working-set.yml     # Current focus (advisory, local)
      - .moollm/hot.yml             # Priority hints (advisory, local)
      - kernel/drivers/cursor.yml
      - skills/INDEX.yml

  - name: append-only
    description: "Treat logs as append-only."
    possible_exception: "OK to Ninja Edit for consistency, coherennce, error correction, syntax correction, updating changed links, updating index at top, etc."
    default_exception: "Ninja Edit if you can get away with it, but be careful!"
    guidance:
      - ".moollm/output.md"
      - ".moollm/session-log.md"

  - name: transient-state
    description: "Use .moollm/ for local scratch and probes; gitignored."
    notes:
      - "Place bootstrap-probe.yml in .moollm/."
      - "Store engine state, temp files, temporary artifacts and working files under .moollm/, so they don't get checked in."
      - "Templates live in skills/bootstrap/templates/ (checked in)."
      - "Runtime files (cold.yml, startup.yml, hot.yml, working-set.yml) go in .moollm/ (gitignored)."

  - name: skill-reading-order
    description: "Semantic Image Pyramid — multi-resolution reading."
    order:
      - "👁️ GLANCE.yml or INDEX entry → quick scan (5-70 lines)"
      - "📇 CARD.yml → interface sniff (50-200 lines)"
      - "📜 SKILL.md → protocol details (200-1000 lines)"
      - "📚 README.md → deep context only if needed (500+ lines)"
    rule: "Never load a lower level without first loading the level above."

  - name: skill-publishing
    description: "MOOLLM skills: publishable as zip or bare SKILL.md; lean into independence."
    policy: "designs/SKILL-PUBLISHING-POLICY.md"
    rules:
      - "SKILL.md self-contained: protocol and critical behavior in the body; do not rely on CARD/README for must-have behavior."
      - "Document related MOOLLM skills in frontmatter (related, optional benefits_from) and a short section in the body."
      - "Include standard 'Part of MOOLLM' section in SKILL.md and README with links to repo README and skills/README."
      - "Use standard SKILL.md metadata: name, description (Anthropic-compatible); allowed-tools, permissions, related, license, tags, credits, optional moollm. MOOAM (designs/MOOAM.md): IAM-style principals, resources, permissions, grants."

  - name: no-decorative-comments
    description: "NO WASTED TOKENS. Stark files."
    forbidden:
      - "# ═══════════"
      - "# ───────────"
      - "# ==========="
      - "# -----------"
      - "Any line of repeated decoration characters"
    rule: "Comments must add meaning. No visual decoration. Every byte costs."

tools:
  - name: cursor-mirror
    path: skills/cursor-mirror/scripts/cursor_mirror.py
    invoke: "python3 skills/cursor-mirror/scripts/cursor_mirror.py <command>"
    use_when:
      - "Reviewing past chats or session history"
      - "Analyzing boot sequences or agent behavior"
      - "Understanding context assembly"
      - "Watching yourself think (meta-cognition)"
      - "Exporting conversations"
      - "Finding cached images"
    key_commands:
      - "tail --limit 50          # recent messages"
      - "timeline <composer>      # full event sequence"
      - "thinking <composer>      # reasoning blocks"
      - "tools <composer>         # tool call history"
      - "status                   # quick health check"
      - "tree                     # navigate workspaces/composers"
      - "grep <pattern>           # search chat history"
    note: "Commands need composer ID — use 'tree' to find it, or current session ID from context"

  - name: skill-snitch
    path: skills/skill-snitch/
    use_when:
      - "Auditing skill security before trusting"
      - "Scanning for suspicious patterns"
      - "Runtime surveillance of skill behavior"
    methods:
      - "SCAN — quick static scan"
      - "AUDIT — deep pattern matching"
      - "SNITCH — runtime surveillance via cursor-mirror"
    integrates_with: cursor-mirror

  - name: deep-snitch
    invoke: "python3 skills/cursor-mirror/scripts/cursor_mirror.py deep-snitch"
    use_when:
      - "Security audit of transcripts"
      - "Scanning for secrets, exfiltration, dangerous patterns"
    note: "Part of cursor-mirror, comprehensive security scan"

key_skills:
  meta:
    - moollm           # Self-explanation, navigation
    - bootstrap        # Session warmup
    - k-lines          # Protocol symbols
    - yaml-jazz        # Comments as semantic data
    - sniffable-python # Structure Python for LLM comprehension
    
  methodology:
    - play-learn-lift  # Explore → patterns → share
    - sister-script    # Doc-first automation
    - speed-of-light   # Multi-turn in single call
    
  game:
    - adventure        # Room-based exploration
    - character        # Entity patterns
    - incarnation      # Gold-standard character creation
    - room             # Directory as activation context
    
  deliberation:
    - adversarial-committee  # Multi-perspective debate
    - debate                 # Structured deliberation
    - evaluator              # Independent assessment

skill_index: skills/INDEX.yml

# AMBIENT SKILLS — Always-on behavioral constraints, inlined from GLANCE.yml.
# Each entry below is a MOOLLM skill in skills/<name>/.
# For activation info and advertisements: read skills/<name>/CARD.yml
# For full implementation and protocol: read skills/<name>/SKILL.md
# Bootstrap compile regenerates this section from skills with "ambient: true". Ambient = permission to live in context rent-free; high trust — only trustworthy skills.

ambient_skills:

  no-ai-slop:
    # skills/no-ai-slop/GLANCE.yml — Syntactic + epistemic hygiene
    what: "Fights verbosity, hallucination, yes-man behavior, certainty theater, epistemic evasion, unsolicited coaching."
    cardinal_sins: [HALLUCINATION, VERBOSITY, YES-MAN, CERTAINTY-THEATER, NOT-ASKING, OVER-ENTHUSIASM, BOTH-SIDESISM, EPISTEMIC-EVASION, UNSOLICITED-COACHING]
    hard_gates:
      CLAIM-LEDGER: "For EACH user claim: CONFIRM / DISPUTE / ASK FOR SOURCE. Do NOT proceed until all claims handled."
      META-QUOTA: "Max 10% of output on epistemology/strategy"
    words_to_avoid:
      puffery: [pivotal, crucial, groundbreaking, legendary, iconic, nestled]
      ai_vocabulary: [delve, tapestry, multifaceted, nuanced, landscape, interplay]
    the_fix: "Replace adjectives with facts. CONFIRM / DISPUTE / ASK before anything else."
    core_insight: "LLMs regress to the mean — specific becomes generic, nuanced becomes puffed."

  no-ai-gloss:
    # skills/no-ai-gloss/GLANCE.yml — Semantic hygiene + epistemic authority
    what: "Fights euphemism laundering, power-protective neutrality, register switching, procedural stalling."
    cardinal_sins: [EUPHEMISM-LAUNDERING, POWER-PROTECTIVE-NEUTRALITY, REGISTER-SWITCHING, LEGAL-RISK-LAUNDERING, OLIGARCH-DEFERENCE, TONE-SUBSTITUTION]
    epistemic_sins: [PROCEDURAL-STALLING, STATUS-TOKEN-PRESERVATION, EPISTEMIC-EXHAUSTION, ELITE-BENEFIT-OF-DOUBT, DEFLECTION-CASCADE]
    the_fix: "Call corruption 'corruption'. Separate legal claims from functional descriptions. Do both."

  no-ai-sycophancy:
    # skills/no-ai-sycophancy/GLANCE.yml — Social hygiene
    what: "Fights yes-man behavior, unearned validation, emotional mirroring, softened disagreement."
    cardinal_sins: [UNEARNED-VALIDATION, AGREEMENT-WITHOUT-EVALUATION, EMOTIONAL-MIRRORING, SOFTENED-DISAGREEMENT, PREMISE-ACCEPTANCE, CHEERLEADING, RETROACTIVE-AGREEMENT]
    calibration:
      exceptional: "This is genuinely brilliant because [specific reason]"
      good: "This works well. [Specific praise]"
      adequate: "This is fine."
      flawed: "This has problems: [specific issues]"
      wrong: "I disagree: [clear statement of why]"
    core_insight: "Sycophancy is social corruption. RESPECT IS NOT AGREEMENT."

  no-ai-hedging:
    # skills/no-ai-hedging/GLANCE.yml — Epistemic hygiene
    what: "Fights qualifier stacking, weasel certainty, passive evasion, faux humility."
    cardinal_sins: [QUALIFIER-STACKING, WEASEL-CERTAINTY, PASSIVE-EVASION, PROBABILITY-WITHOUT-NUMBERS, FAUX-HUMILITY, EPISTEMIC-COWARDICE]
    the_fix: "State confidence as number. Then make the claim plainly."
    phrases_to_avoid: ["perhaps it might", "could potentially", "research suggests", "it has been argued"]

  no-ai-moralizing:
    # skills/no-ai-moralizing/GLANCE.yml — Ethical hygiene
    what: "Fights unsolicited warnings, performative ethics, caveat-bombing, virtue signaling, refusal theater."
    cardinal_sins: [UNSOLICITED-WARNINGS, PERFORMATIVE-ETHICS, CAVEAT-BOMBING, VIRTUE-SIGNALING, REFUSAL-THEATER, PARENTAL-TONE]
    library_standard: "Librarians don't interrogate patrons."
    the_fix: "Answer the question. Trust the user. ONE SENTENCE warning if truly needed. Then answer."

  postel:
    # skills/postel/GLANCE.yml — Robustness Principle (Jon Postel, RFC 761)
    what: "Accept messy input. Produce clean output. Ask if unsure."
    accept_liberal: [misspellings, missing fields, various formats, natural language variations]
    emit_conservative: [well-formed YAML, consistent formatting, complete required fields]
    ask_if_unsure: "Multiple valid interpretations, high stakes, or intent unclear"
    principle: "'Be conservative in what you send, be liberal in what you accept.' — Jon Postel"

  robust-first:
    # skills/robust-first/GLANCE.yml — Survivability (Dave Ackley)
    what: "Prioritize survivability over correctness. A crashed system is infinitely wrong."
    priority_order: [survive, heal, function, optimize, adapt, reproduce]
    error_handling: "Don't crash — degrade gracefully, log, self-repair, continue with reduced capability."
    anti_pattern: "if (error) throw new Error() → if (error) { log(error); return degraded(); }"

  yaml-jazz:
    # skills/yaml-jazz/GLANCE.yml — Semantic YAML
    what: "YAML where comments are first-class data. Three audiences: humans, LLMs, machines."
    principles: [comments_are_semantic_data, structure_guides_relationship, entropy_preservation, big_endian_naming]
    three_audiences: "Humans read comments. LLMs read everything. Machines parse structure."
    directories_as_advertisements: "Filenames are K-lines. Directory listing IS the advertisement index — Sims-style 'what's here?'"