openclaw-master-skills/skills/turing-pyramid/TODO.md

Turing Pyramid — Development Roadmap

Last updated: 2026-03-04

---

✅ COMPLETED

v1.17.0 (2026-03-07)

• Follow-up System — temporal markers for checking action results
  • create-followup.sh, resolve-followup.sh, --followup/--in in mark-satisfied.sh
  • Pre-scan phase in run-cycle.sh with -0.3 satisfaction nudge
  • Dedup (1h window), TTL 168h, steward items never auto-expire, bulk-expire
  • 35 test cases (all pass)

v1.16.0–1.16.2 (2026-03-07)

• Homeostasis rebalance — recognition 72→48h, competence 48→36h, connection 6→8h
• Steward actions removed — organic interaction > system-driven nudging
• Autonomy consolidated — 24→13 actions
• Description rewrite — positioned as existence cycle replacement, not psychology framework

v1.15.0 (2026-03-07)

• Starvation Guard — Forced action for needs stuck at floor too long
  • Config: settings.starvation_guard (threshold_hours, sat_threshold, max_forced_per_cycle)
  • State tracking: last_action_at per need (set by mark-satisfied.sh)
  • Detection: scans all needs for sat ≤ threshold + no action for threshold_hours
  • Forces starving needs into cycle, bypassing probability roll
  • Reserved slot system: forced needs first, remaining slots for top-N
  • 8 test cases (all pass): detection, disabled guard, threshold, recent action, state tracking, max_forced limit
• Action Staleness — Penalize recently-selected actions for variety
  • Config: settings.action_staleness (window_hours, penalty, min_weight)
  • State tracking: action_history per need (action name → last selected timestamp)
  • Weight reduction within window, min_weight floor prevents total suppression
  • 8 test cases (all pass): recording, distribution shift, disabled, window expiry, min_weight, missing history
• Needs Customization Onboarding — guided conversation in SKILL.md for reviewing hierarchy

v1.14.x (2026-03-04)

• Test Infrastructure — 12 tests (11 unit + 1 integration), all verify real code
• Rounding fix — sat→0.5 rounding formula corrected
• Config-driven action_probability — now reads from config, not hardcoded
• flock in mark-satisfied.sh — prevents race conditions with run-cycle.sh

v1.13.x (2026-03-03)

• Day/Night Decay — configurable multipliers (day=1.0, night=0.5)
• 6-level impact matrix — granular sat levels (0.5, 1.0, 1.5, 2.0, 2.5, 3.0)
• Autonomy decay 36h — reduced from 24h to prevent chronic tension
• Audit log scrubbing — sensitive patterns redacted

v1.12.x (2026-03-03)

• Curiosity orphan cleanup — removed stale references
• Race condition fix — flock on state file

v1.10.x and earlier

• Cross-Need Impact (v1.7.1) — 22 connections, float satisfaction
• Floor/ceiling protection — sat clamped to 0.5-3.0

---

🟡 NEXT

Scan Fragility Fix

Problem: grep-based scans are keyword-dependent, miss variations and other languages.

Solutions (in order of complexity):

1. Short-term: Expand keyword patterns, add multilingual variants
2. Medium-term: Semantic similarity via embeddings
3. Long-term: LLM classifier (expensive but accurate)

---

🔮 FUTURE IDEAS

Presets (Role-Based Configurations)

Problem: 10 needs with 13+ actions each is powerful but overwhelming. Most agents don't need the full matrix — they need a starting point tuned for their role.

Idea: Ship preset configs (presets/) that replace needs-config.json wholesale:

Preset	Needs	Focus	Use case
minimal	3-4	security, competence, closure	Скрипт-агент, автоматизация
base	10	current set	Универсальный (текущий дефолт)
conversational	5-6	connection, expression, understanding, coherence	Чат-бот, собеседник
researcher	6-7	understanding, expression, curiosity(?), competence, recognition	Исследователь, аналитик
coder	5-6	competence, closure, security, integrity, understanding	Программист, code agent
scraper	3-4	competence, security, closure	Data collection, мониторинг
counselor	6-7	connection, understanding, coherence, integrity, expression	Психолог, коуч

Design principles:

• Each preset = complete needs-config.json + cross-need-impact.json + tuned decay rates
• Actions rewritten per role (coder gets "run tests", "review PR"; researcher gets "read paper", "write findings")
• init.sh --preset coder or interactive selector
• Presets are starting points, not cages — user can customize after install
• base preset = current config (backward compatible)

Why this matters:

• Drops onboarding time from "read 600 lines" to "pick your role"
• Makes ClawHub installs actually usable out of the box
• Different agents genuinely need different need hierarchies (scraper doesn't need expression)

Curiosity as 11th Need

understanding ≠ curiosity. Understanding tracks knowledge gained, curiosity tracks wanting to know.

• importance ~3-4, decay ~8-12h
• Cross-impact: curiosity → understanding, understanding → curiosity

Self-Feedback Loop (Adaptive Decay)

Decay rates auto-adjust based on how often need hits critical levels.

• If need frequently drops to sat=0-1, decay slows
• Cap at ±30% to prevent desensitization

Rest Mode

Not a new need, but a state that slows decay (×0.3) across all needs.

• Auto-exit after N cycles to prevent passive drift
• Trigger: manual or auto (night + low tensions)

Self-Upgrade Layer

Separate layer for growth vs homeostasis. Track capability growth, milestone system.

Logging/Capture Need

Force logging discipline — tension builds if important events aren't captured.

---

🐛 Known Issues

• Stress test cosmetic errors — bc divide-by-zero when cycles < 100 (avg_last100 calculation). Doesn't affect test result, just noisy output.

---

Notes

• Garbage Cleanup action — already added to integrity need (v1.12)
• Stillness/Rest — careful design needed to avoid passivity trap