reflect

name	reflect
description	Auto-reflective self-improvement skill — extracts learnings from corrections and success patterns, permanently encodes them into memory and skills. Philosophy: Correct once, never again.

Reflect — Agent Self-Improvement (Hermes)

Extrahe aprendizagens de correções e padrões de sucesso do conversation history. Transforma correções únicas em memória permanente. Philosophy: Correct once, never again.

Quick Commands

Command	Action
reflect	Analisa conversation para aprendizagens
reflect on	Ativa auto-reflection
reflect off	Desativa auto-reflection
reflect status	Mostra estado e métricas
reflect review	Review pending learnings

Quando Usar

• Depois de tarefas complexas
• Quando Álvaro corrige comportamento explicitamente
• Em session boundaries
• Quando padrões de sucesso merecem ser preservados

Workflow — 5 Passos

Step 1: Scan Conversation for Signals

Analisa conversation para correction signals e learning opportunities.

Signal Confidence Levels:

Confidence	Triggers	Examples
HIGH	Explicit corrections	"never", "always", "wrong", "stop", "the rule is"
MEDIUM	Approved approaches	"perfect", "exactly", "that's right", output accepted
LOW	Observations	Padrões que funcionaram mas não validado explicitamente

See: signal_patterns.md — full detection rules

Step 2: Classify & Match to Target Files

Map cada signal ao target apropriado:

Category	Target Files
Code Style	~/.hermes/skills/*/SKILL.md (relevant skill)
Architecture	bianinho-os + relevant skills
Process	USER.md, memory
Domain	USER.md ( Álvaro & Maryanne section)
Tools	memory (environment facts)
New Skill	Create new skill in ~/.hermes/skills/

See: agent_mappings.md — mapping rules

Step 3: Check for Skill-Worthy Signals

Skill-Worthy Criteria:

• Debugging non-trivial (>10 min investigation)
• Erro enganoso (root cause diferente da mensagem)
• Workaround descoberto por experimentação
• Config insight (difere do documentado)
• Padrão reutilizável

Quality Gates (must pass all):

• Reusable: Vai ajudar em tarefas futuras
• Non-trivial: Requer descoberta, não só docs
• Specific: Pode descrever condições exatas de trigger
• Verified: Solução realmente funcionou
• No duplication: Não existe já

Step 4: Generate Proposals

Apresenta findings no formato:

# Reflection Analysis

## Session Context
- **Date**: [timestamp]
- **Messages Analyzed**: [count]

## Signals Detected

| # | Signal | Confidence | Source Quote | Category |
|---|--------|------------|--------------|----------|
| 1 | [learning] | HIGH | "[exact words]" | Code Style |

## Proposed Changes

### Change 1: Update [target]
**Target**: `[file path]`
**Section**: [section name]
**Confidence**: HIGH

```diff
+ New rule from learning

Review Prompt

Apply these changes? (Y/N/modify/1,2,3)

### Step 5: Apply with User Approval

**`Y` (approve):**
1. Apply change using patch tool
2. Update metrics
3. Log to `~/.hermes/reflect/learnings.log`

**`N` (reject):**
1. Discard proposed changes
2. Log rejection

**`modify`:**
1. Present each change individually
2. Allow editing before applying

**Selective (e.g., `1,3`):**
1. Apply only specified changes
2. Log partial updates

## State Management

State stored in `~/.hermes/reflect/`:

~/.hermes/reflect/ ├── state.yaml # auto_reflect, last_reflection, pending_reviews ├── metrics.yaml # total_signals, acceptance_rate, etc. └── learnings.log # History of all learnings applied

## Hermes Integration

**State file:** `~/.hermes/reflect/state.yaml`
```yaml
auto_reflect: false
last_reflection: "2026-04-20T16:00:00Z"
pending_reviews: []

Metrics file: ~/.hermes/reflect/metrics.yaml

total_sessions_analyzed: 0
total_signals_detected: 0
total_changes_accepted: 0
acceptance_rate: 0
confidence_breakdown:
  high: 0
  medium: 0
  low: 0
skills_created: 0

Safety Guardrails

Human-in-the-Loop

• NEVER apply changes without explicit approval
• Always show full diff before applying
• Allow selective application

Incremental Updates

• ONLY add to existing sections
• NEVER delete or rewrite existing rules
• Preserve original structure

Conflict Detection

• Check if proposed rule contradicts existing memory
• Warn if conflict detected
• Suggest resolution strategy

Output Locations

• ~/.hermes/reflect/learnings.log — Log de aprendizagens aplicadas
• ~/.hermes/reflect/metrics.yaml — Métricas agregadas
• New skills → ~/.hermes/skills/<name>/SKILL.md
• Updates → existing skill SKILL.md or memory

Components

• signal_patterns.md — Regex patterns for detecting corrections by confidence
• agent_mappings.md — Category → target file mappings (adapted for Hermes)

Autores

• Original: stevengonsalvez (MIT)
• Hermes: Bianinho adapt. 20/04/2026