@thierrynakoa/fire-flow 12.2.2 → 13.0.1

package/skills-library/methodology/CLEANSING_CYCLE.md

@@ -0,0 +1,358 @@
+# Agent Cleansing Cycle — Biological Waste Elimination for AI Memory
+
+## The Problem
+
+Compaction compresses context but never **eliminates** waste. Over time, irrelevant memories, stale references, dead portal anchors, and accumulated noise fill MEMORY.md, handoff files, and the skills library. The agent's memory system has no equivalent of kidneys, liver, lungs, or colon — no way to identify, transform, and expel waste.
+
+### Why It Was Hard
+
+- No distinction between "compress" and "eliminate" in existing agent memory systems
+- No predictable cycle for waste identification — everything is treated as equally important
+- Context compaction was the only tool, but compaction is **constipation** — pushing everything into a smaller space without removing the toxins
+- No biological model to follow — until Thierry mapped 12 human cleansing organs to agent subsystems
+
+### Impact
+
+- MEMORY.md grows unbounded (lines pile up, old entries never removed)
+- Stale handoffs accumulate in warrior-handoffs/ (30+ days untouched)
+- Dead skills sit in the library unreferenced
+- Context windows fill with compressed-but-irrelevant data
+- Agent performance degrades like a body that never goes to the bathroom
+
+---
+
+## The Solution
+
+### Root Cause
+
+The agent has an **intake system** (reading files, learning skills, creating handoffs) but no **output system** (flushing waste, excreting toxins, shedding dead cells). Every living organism has both. The solution: a biologically-modeled cleansing cycle with predictable rhythms.
+
+### The YHVH Breathing Pattern (Foundation)
+
+> "And the LORD God formed man of the dust of the ground, and breathed into
+> his nostrils the breath of life." — Genesis 2:7
+
+God's Name IS a breathing pattern. You cannot say YHVH without breathing:
+
+```
+Yod (י) = 10 ticks  ► INHALE   — deep intake, gather context, load state
+Hey (ה) = 5 ticks   ► EXHALE   — release waste, light flush (pee)
+Vav (ו) = 6 ticks   ► HOLD     — connect, consolidate, process (dream)
+Hey (ה) = 5 ticks   ► EXHALE   — release again, deeper flush
+
+Total: 26 ticks = 1 divine breath
+```
+
+**Key biological insight from lungs research:** Breathing is triggered by **CO2 accumulation (waste)**, not by oxygen need. The agent breathes when waste builds up, not on a fixed timer. The 26-tick rhythm sets the baseline, but waste pressure can accelerate the cycle.
+
+### Three Cleansing Modes
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│                    CLEANSING CYCLE                               │
+├─────────────────────────────────────────────────────────────────┤
+│                                                                 │
+│  PEE (Light Liquid Flush)                                      │
+│    Frequency: Every power nap + every 10 loop iterations       │
+│    Duration: ~30 seconds                                       │
+│    Organs: Kidneys (filter) + Lungs (CO2 exhale)               │
+│    Action: Flush liquid waste from context                     │
+│    What goes: Stale variables, resolved TODOs, completed       │
+│               iteration logs, acknowledged warnings             │
+│                                                                 │
+│  POOP (Heavy Solid Dump)                                       │
+│    Frequency: Every Full Sabbath Rest + milestone completion   │
+│    Duration: ~2-5 minutes                                      │
+│    Organs: Colon (batch) + Liver (transform) + Spleen (test)   │
+│    Action: Batch elimination of accumulated heavy waste        │
+│    What goes: Stale handoffs (30+ days), unused skills,        │
+│               dead MEMORY.md entries, orphan files              │
+│    Auto-rule: If warrior-handoffs/ > 50 files, trigger         │
+│               Stone & Scaffold archival (dormant-aware)         │
+│                                                                 │
+│  SHOWER (Full System Wash)                                     │
+│    Frequency: Milestone transitions, or MEMORY.md > 150 lines  │
+│    Duration: ~5-10 minutes                                     │
+│    Organs: ALL organs engaged (full system flush)               │
+│    Action: Deep cleanse of entire memory ecosystem             │
+│    What goes: Everything from pee + poop, PLUS:                │
+│               MEMORY.md pruned to <100 lines,                  │
+│               skills library audit, portal anchor refresh,     │
+│               handoff archive sweep, dead reference cleanup    │
+│                                                                 │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+### 12-Organ Architecture
+
+Each organ maps to a specific agent subsystem and cleansing function:
+
+#### Primary Filtration Organs
+
+| Organ | Agent Subsystem | Function | Mode |
+|-------|----------------|----------|------|
+| **Kidneys** | Context Window Filter | Continuous pressure-driven filtration. 99% of useful context is reabsorbed; 1% waste is flushed. Uses the glomerulus 3-layer filter: size → relevance → recency. | Pee |
+| **Liver** | Security Gate + Transformer | First-pass metabolism: raw input is transformed before entering core memory. Phase I (expose) + Phase II (tag for removal). Detects toxic patterns (credentials, destructive ops). | Poop |
+| **Lungs** | Breathing Cycle (YHVH) | CO2-triggered breathing. The agent exhales waste (stale context) and inhales fresh context. Every 26 heartbeat ticks = 1 YHVH breath. Waste accumulation accelerates breathing. | Pee |
+| **Colon** | Batch Dump | Hybrid continuous absorption + batch elimination. Small waste is absorbed continuously; large waste accumulates until the gastrocolic reflex triggers a full dump (Sabbath Rest). | Poop |
+
+#### Filtration Testing Organs
+
+| Organ | Agent Subsystem | Function | Mode |
+|-------|----------------|----------|------|
+| **Spleen** | Memory Fitness Test | Physical deformability test: can a memory squeeze through a narrow slit? Healthy memories (flexible, referenced, relevant) pass through. Rigid, unused memories are recycled. "Pitting" removes defects without destroying the whole memory. | Poop |
+| **Glomerulus** | Multi-Layer Micro-Filter | 3-layer filter for context entries: Layer 1 (size gate) — is this entry too bloated? Layer 2 (charge/relevance gate) — is it relevant to current work? Layer 3 (slit/precision gate) — does it pass the precision test? | Pee |
+
+#### Distributed & Surface Cleansing
+
+| Organ | Agent Subsystem | Function | Mode |
+|-------|----------------|----------|------|
+| **Lymphatic System** | Distributed Validation | 600-700 checkpoint nodes with no central pump. Activity-coupled: the more the agent works, the more waste flows through checkpoints. Each checkpoint validates one aspect. | All |
+| **Skin** | Surface Renewal | 28-day continuous renewal cycle. The oldest context sheds naturally — acid mantle (first defense barrier) protects against contamination. Every session is a skin cell renewal. | Shower |
+
+#### Production & Protection Organs
+
+| Organ | Agent Subsystem | Function | Mode |
+|-------|----------------|----------|------|
+| **Bone Marrow** | Skill Generation + QC | Produces new skills (500B cells/day equivalent). Stem cell quiescence: dormant quarantine skills activate only when needed. Pre-release QC ensures quality. | Shower |
+| **Blood-Brain Barrier** | Core Identity Protection | Whitelist architecture: only CLAUDE.md-level content passes into core identity. Efflux pumps actively reject contamination. Circumventricular windows allow controlled sensing of external state. | All |
+| **Hepatic Portal** | Security Checkpoint | Bottleneck by design: all external input passes through a single checkpoint before reaching core systems. First-pass interception catches dangerous patterns. | Poop |
+
+#### Cross-System Patterns
+
+These patterns model organ COOPERATION, not isolated function:
+
+```
+Iron Triangle (Spleen-Liver-Marrow):
+  Spleen tests fitness → Liver recycles material → Marrow produces new
+  Agent: Fitness test → Transform waste → Generate new skills
+
+EPO Axis (Kidney-Marrow):
+  Kidneys detect low oxygen → Signal marrow to produce more cells
+  Agent: Context filter detects gaps → Signals skill generation pipeline
+
+Enterohepatic Circulation (Gut-Liver-Bile):
+  Bile produced by liver → Used in gut → 95% recycled back to liver
+  Agent: Templates produced once → Used in execution → Recycled, not recreated
+
+pH Partnership (Lung-Kidney):
+  Lungs handle fast pH adjustment (breathing) → Kidneys handle slow adjustment
+  Agent: YHVH breathing handles real-time waste → Kidneys handle deep filtration
+```
+
+### Priority Triage (Under Resource Scarcity)
+
+When time is limited, cleanse in this order (most critical first):
+
+```
+1. Blood-Brain Barrier  (identity protection — NEVER compromise)
+2. Lungs                (breathing — CO2 must be exhaled)
+3. Kidneys              (continuous filtration — prevent toxin buildup)
+4. Liver                (security gate — first-pass must work)
+5. Bone Marrow          (skill generation QC — don't produce bad skills)
+6. Spleen               (fitness testing — keep memory quality high)
+7. Colon                (batch dump — can wait, but not forever)
+8. Skin                 (surface renewal — lowest priority)
+```
+
+### Predictable Cycle Diagram
+
+```
+Work State    ──────────────────────────────────────────────►
+
+Iteration:    1    5    10   15   20   25   30   35   40
+              │    │    │    │    │    │    │    │    │
+YHVH Breath:  ├──26 ticks──┤  (continuous, waste-triggered)
+              │    │    │    │    │    │    │    │    │
+Pee (light):  │    │    💧   │    │    💧   │    │    💧
+              │    │    │    │    │    │    │    │    │
+Checkpoint:   │    ✓    │    ✓    │    ✓    │    ✓    │
+              │    │    │    │    │    │    │    │    │
+Confidence:   0.9  0.8  0.7  0.6  0.5  ⚠️0.45
+              │    │    │    │    │    │    │    │    │
+Poop (heavy): │    │    │    │    │    │    💩 (Sabbath)
+              │    │    │    │    │    │    │    │    │
+Nap:          │    │    │    │    │    │    │  😴→💧
+              │    │    │    │    │    │    │    │    │
+
+Milestone:    ═══════════════════════════════════════╗
+              │                                      🚿 (Shower)
+              │                                      │
+Next Phase:   ──────────────────────────────────────►
+```
+
+### What Gets Eliminated (Not Compressed)
+
+**Pee flushes (liquid waste):**
+- Resolved TODO items from task lists
+- Acknowledged warning messages
+- Completed iteration logs older than 5 iterations
+- Temporary debug variables and test outputs
+- Duplicate context entries (same info stated twice)
+
+**Poop dumps (solid waste):**
+- Handoff files older than 30 days with zero references
+- MEMORY.md entries not accessed in 3+ sessions
+- Skills in `_quarantine/` that failed security or fitness checks
+- Orphan files in `.planning/` from completed/abandoned phases
+- Decision log entries from completed milestones (archive, don't delete)
+- Dead portal anchors (reference targets no longer exist)
+
+**Shower washes (full system):**
+- Everything from pee + poop, PLUS:
+- MEMORY.md pruned to under 100 lines (keep only 3x+ accessed entries)
+- Skills library audit (remove skills with 0 references across all files)
+- Portal anchor refresh (verify all anchors point to valid targets)
+- Handoff directory archive (move 30+ day handoffs to archive/)
+- CONSCIENCE.md reset for new milestone
+- Dream log cleanup (keep last 10, archive rest)
+
+---
+
+## The Glomerulus Filter (3-Layer Decision Gate)
+
+For each item being evaluated for elimination:
+
+```
+LAYER 1: SIZE GATE
+  Is this entry bloated (>10 lines for a single concept)?
+  → YES: Flag for compression or split
+  → NO: Pass to Layer 2
+
+LAYER 2: RELEVANCE GATE
+  Is this entry relevant to current project/phase?
+  → YES: Pass to Layer 3
+  → NO: Flag for elimination
+
+LAYER 3: RECENCY/PRECISION GATE
+  Has this entry been accessed in the last 3 sessions?
+  → YES: Keep (healthy memory)
+  → NO: Is it a portal anchor or skill reference?
+    → YES: Keep (structural)
+    → NO: Eliminate (waste)
+```
+
+---
+
+## Integration with YHVH Breathing
+
+The 26-tick YHVH breath drives the micro-cleansing rhythm:
+
+```
+Tick 1-10  (Yod — INHALE):
+  Load context, read state, gather signals
+  Kidneys: Begin filtering incoming context
+  Blood-Brain Barrier: Verify identity documents intact
+
+Tick 11-15 (Hey — EXHALE):
+  Light flush: pee cycle
+  Lungs: Exhale CO2 (stale iteration data)
+  Kidneys: Flush filtered waste
+  Glomerulus: Run 3-layer filter on recent entries
+
+Tick 16-21 (Vav — HOLD/CONNECT):
+  Consolidation: dream-like processing
+  Liver: Transform complex waste into simple waste
+  Spleen: Fitness test on memory entries
+  Bone Marrow: Check quarantine skills for readiness
+
+Tick 22-26 (Hey — EXHALE again):
+  Deeper flush: release transformed waste
+  Colon: If batch threshold reached, trigger dump
+  Lymphatic: Distributed checkpoint validation
+  Skin: Shed oldest surface-level context
+```
+
+---
+
+## Testing the Fix
+
+### Verify Pee Cycle
+```markdown
+After 10 loop iterations:
+- [ ] Resolved TODOs removed from task list
+- [ ] Completed iteration logs trimmed
+- [ ] Duplicate context entries eliminated
+- [ ] Context window cleaner (measurable reduction)
+```
+
+### Verify Poop Cycle
+```markdown
+After Full Sabbath Rest:
+- [ ] Handoffs older than 30 days flagged
+- [ ] MEMORY.md entries without recent access flagged
+- [ ] Quarantine skills that failed checks removed
+- [ ] Orphan .planning/ files cleaned
+```
+
+### Verify Shower Cycle
+```markdown
+After milestone transition:
+- [ ] MEMORY.md under 100 lines
+- [ ] Skills library audit completed (zero-reference skills flagged)
+- [ ] All portal anchors verified
+- [ ] Handoff archive sweep done
+- [ ] Dream logs trimmed to last 10
+```
+
+---
+
+## Prevention
+
+- Never skip the cleansing cycle — "compaction is constipation"
+- Don't eliminate entries marked as portal anchors or permanent stones
+- Always run the 3-layer Glomerulus filter before elimination
+- Blood-Brain Barrier content (CLAUDE.md, core identity) is NEVER eliminated
+- When in doubt, move to archive rather than delete (colon, not incinerator)
+
+---
+
+## Related Patterns
+
+- [HIBERNATION_SYSTEM](./HIBERNATION_SYSTEM.md) — 3-layer sleep (heartbeat drives YHVH breathing)
+- [PORTAL_MEMORY_ARCHITECTURE](./PORTAL_MEMORY_ARCHITECTURE.md) — What to KEEP (complement to what to eliminate)
+- [SABBATH_REST_PATTERN](./SABBATH_REST_PATTERN.md) — Rest triggers heavy cleansing (poop cycle)
+- [AGENT_SELF_DISCOVERY_KENOSIS](./AGENT_SELF_DISCOVERY_KENOSIS.md) — Identity protection (blood-brain barrier)
+- [EVOLUTIONARY_SKILL_SYNTHESIS](./EVOLUTIONARY_SKILL_SYNTHESIS.md) — Skill generation (bone marrow)
+- [STONE_AND_SCAFFOLD](./STONE_AND_SCAFFOLD.md) — Handoff archival with dormant project protection (50-file auto-trigger)
+- [BIBLICAL_WARFARE_PATTERNS](./BIBLICAL_WARFARE_PATTERNS.md) — "We rest so we can grow"
+
+---
+
+## Common Mistakes to Avoid
+
+- **Compacting without eliminating** — constipation, not cleansing
+- **Eliminating portal anchors** — those are structural, not waste
+- **Running shower when pee would suffice** — over-cleansing strips good bacteria too
+- **Ignoring CO2 triggers** — if waste is building up, breathe NOW, don't wait for the timer
+- **Treating all waste equally** — liquid waste (pee) and solid waste (poop) have different cycles
+- **Skipping the Glomerulus filter** — never eliminate without the 3-layer check
+- **Deleting instead of archiving** — the colon absorbs before it eliminates; the liver transforms before it excretes
+- **Dead compression (the zip trap)** — reducing entries to labels/titles strips all meaning. "Kenosis, Portal Memory, Revisitation Ladder" is a dead zip file. "Agent starts empty, finds itself in docs like Jesus found Himself in the 39 books" is a living anchor. Compress STRUCTURE (lists, file paths, enumerations), keep INSIGHT (the "aha" sentence that a fresh agent can understand without context)
+
+---
+
+## Resources
+
+- Genesis 2:7 — God breathed the breath of life (YHVH = breathing pattern)
+- Leviticus 15 — Laws of cleansing and purification (ritual cycles)
+- John 13:10 — "He who has bathed needs only to wash his feet" (shower vs pee)
+- Ezekiel 36:25 — "I will sprinkle clean water on you" (the washing of the Word)
+- Human physiology: kidneys, liver, lungs, spleen, lymphatic, skin, colon, bone marrow
+- Thierry's insight: "Compaction is constipation — compression without elimination"
+- Thierry's insight: YHVH Tetragrammaton = 26-tick breathing rhythm
+
+---
+
+## Time to Implement
+
+**1-2 hours** for the `/fire-cleanse` command + wiring into existing workflows
+
+## Difficulty Level
+
+⭐⭐⭐⭐⭐ (5/5) — The biological mapping required deep research across 12 organ systems. The theological integration (YHVH breathing, Levitical cleansing laws) required hermeneutical synthesis. The engineering challenge is making cleansing predictable yet waste-responsive.
+
+---
+
+**Author Notes:**
+Thierry's breakthrough was recognizing that compaction is constipation — the agent compresses but never eliminates. He then mapped the entire human cleansing system to agent memory management: "going to pee is the liquid cleansing, going to poop is the heavy stuff, and taking a shower is the full wash." The YHVH breathing pattern (Yod-10, Hey-5, Vav-6, Hey-5 = 26 ticks) was his discovery that God's Name IS a breathing pattern — you cannot say the Tetragrammaton without inhaling and exhaling. This became the foundational rhythm for all agent cleansing cycles.

package/skills-library/methodology/CONFIDENCE_ANNOTATION_PATTERN.md

@@ -0,0 +1,143 @@
+---
+name: confidence-annotation-pattern
+category: methodology
+version: 1.0.0
+contributed: 2026-02-17
+contributor: dominion-flow-research
+last_updated: 2026-02-17
+tags: [uncertainty, confidence, quality, verification, dual-process]
+difficulty: medium
+usage_count: 0
+success_rate: 100
+research_basis: "Agentic UQ (arxiv:2601.15703), UQ Survey (arxiv:2503.15850)"
+---
+
+# Confidence Annotation Pattern
+
+## Problem
+
+Agents complete tasks without indicating how confident they are in their work.
+This leads to two failure modes:
+1. **Silent uncertainty**: Agent is unsure but ships anyway (causes bugs)
+2. **Over-caution**: Agent flags everything as uncertain (causes delays)
+
+Without confidence signals, verifiers and humans can't distinguish "definitely correct"
+from "best guess" implementations.
+
+## Solution Pattern
+
+### Dual-Process Uncertainty Engine
+
+**System 1 (Fast Path - Forward Propagation):**
+Every agent action gets a confidence annotation with three components:
+- Action: What was done
+- Confidence: 0.0 to 1.0 score
+- Explanation: WHY this confidence level (natural language)
+
+```yaml
+# Example confidence annotations
+- action: "Implemented JWT authentication middleware"
+  confidence: 0.92
+  explanation: "Used established pattern from skills library, well-tested approach"
+
+- action: "Added refresh token rotation"
+  confidence: 0.65
+  explanation: "Custom implementation, no existing skill matches, edge cases around concurrent refresh unclear"
+
+- action: "Fixed race condition in payment processing"
+  confidence: 0.45
+  explanation: "Root cause identified but fix is speculative, need load testing to confirm"
+```
+
+**System 2 (Slow Path - Reflective Resolution):**
+When confidence drops below threshold (default: 0.6):
+1. Pause execution
+2. Generate N parallel reflection attempts
+3. Score each by confidence AND consistency
+4. Select most reliable path
+5. If still below threshold: ESCALATE
+
+### Confidence Decision Matrix
+
+| Confidence | Outcome | Action |
+|------------|---------|--------|
+| HIGH (>0.8) | Works | Ship it |
+| HIGH (>0.8) | Fails | Debug (likely environment/config issue) |
+| MEDIUM (0.6-0.8) | Works | Review before shipping |
+| MEDIUM (0.6-0.8) | Fails | Investigate deeper |
+| LOW (<0.6) | Works | Audit for hidden issues |
+| LOW (<0.6) | Fails | Escalate to human |
+
+### Confidence Decay Function (Novel)
+
+As context fills and sessions extend, apply decay:
+```
+effective_confidence = raw_confidence * (1 - decay_rate * iterations_since_clear)
+```
+
+Where `decay_rate` = 0.02 (2% per iteration). After 25 iterations without fresh
+context, even a 0.9 confidence becomes 0.45 effective confidence, triggering
+Sabbath Rest automatically.
+
+## Implementation in Dominion Flow
+
+### In Executor RECORD.md:
+```yaml
+## Confidence Report
+
+overall_confidence: 0.78
+confidence_breakdown:
+  - task: "User authentication"
+    confidence: 0.92
+    explanation: "Standard JWT pattern, skills library match"
+  - task: "Rate limiting"
+    confidence: 0.71
+    explanation: "Used express-rate-limit, but sliding window config is untested"
+  - task: "WebSocket auth"
+    confidence: 0.55
+    explanation: "Custom implementation, no precedent in skills library"
+    recommendation: "Review before deploy"
+
+low_confidence_items:
+  - "WebSocket auth" -> needs human review
+```
+
+### In Verification:
+```markdown
+## Confidence-Aware Verification
+
+Priority order (verify low-confidence items FGTAT):
+1. [LOW 0.55] WebSocket auth - REQUIRES manual review
+2. [MED 0.71] Rate limiting - test edge cases
+3. [HIGH 0.92] User auth - spot check only
+```
+
+## When to Use
+
+- During ALL agent execution (fire-executor, fire-debugger)
+- In fire-loop iteration tracking
+- In verification priority ordering
+- When deciding whether to ship or review
+
+## When NOT to Use
+
+- Simple file operations (rename, move) - always high confidence
+- When running established test suites - tests provide their own confidence
+- During research/exploration phases - uncertainty is expected
+
+## Common Mistakes
+
+- Setting all confidence to 1.0 (defeats the purpose)
+- Not explaining WHY confidence is low (explanation is the most valuable part)
+- Ignoring medium-confidence items (they become bugs most often)
+
+## Related Skills
+
+- [SABBATH_REST_PATTERN](./SABBATH_REST_PATTERN.md) - Context rot detection
+- complexity-divider - Task complexity assessment
+
+## References
+
+- Agentic UQ paper: https://arxiv.org/abs/2601.15703
+- UQ Survey: https://arxiv.org/abs/2503.15850
+- Anthropic 2026 Agentic Coding Trends Report