Feedback Learning System v2#

A complete, zero-LLM pipeline for agents to learn from user feedback. Track what works, catch what doesn't, promote durable rules.

Architecture#

text

User feedback / exec error
        ↓
detect-feedback.py   ←── error-catcher.sh (PostToolUse hook)
        ↓
  log-event.sh  ──────────────────────────────────────────→ events.jsonl
                                                                   ↓
                                            analyze-patterns.py (nightly)
                                                                   ↓
                                                          patterns.json
                                              (positive + negative patterns)
                                                                   ↓  (≥3 hits, delta test)
                                                             genes.json
                                                    (structured rules: condition→action)
                                                                   ↓
                                            weekly-report.py (Sundays)
                                                                   ↓
                                                          reports/WEEKLY_*.md

Quick Reference#

SituationActionUser gives positive feedbacklog-event.sh <agent> positive user_nlp "<ctx>" "<signal>"User corrects/complainslog-event.sh <agent> correction user_nlp "<ctx>" "<signal>" "<lesson>"Exec command failedlog-event.sh <agent> error exec_fail "<ctx>" "<stderr>" "<lesson>"Detect feedback from textpython3 detect-feedback.py "переделай это"Run pattern analysis nowpython3 analyze-patterns.pyGenerate report nowpython3 weekly-report.pyCheck active rules (genes)python3 check-genes.pyMark gene as resolvedpython3 check-genes.py --resolve <gene_id>

Setup#

1. Install files#

bash

DIR="${FEEDBACK_LEARNING_DIR:-$HOME/.openclaw/shared/learning}"
mkdir -p "$DIR/reports"
cp scripts/* "$DIR/"
chmod +x "$DIR/log-event.sh" "$DIR/error-catcher.sh"
touch "$DIR/events.jsonl"

2. Initialize data files#

bash

DIR="${FEEDBACK_LEARNING_DIR:-$HOME/.openclaw/shared/learning}"

[ -f "$DIR/patterns.json" ] || cat > "$DIR/patterns.json" << 'EOF'
{"version": "2.1", "updated": "", "patterns": {"negative": [], "positive": []}}
EOF

[ -f "$DIR/genes.json" ] || cat > "$DIR/genes.json" << 'EOF'
{"version": "2.1", "rules": []}
EOF

[ -f "$DIR/capsules.json" ] || cat > "$DIR/capsules.json" << 'EOF'
{"version": "2.1", "capsules": []}
EOF

3. Add to AGENTS.md boot sequence#

markdown

## Feedback Learning
Before tasks: check `$FEEDBACK_LEARNING_DIR/genes.json` for applicable rules.

Auto-detect and log signals:
- Positive words/emoji → `bash $DIR/log-event.sh <agent> positive user_nlp "<ctx>" "<signal>"`
- Negative/correction → `bash $DIR/log-event.sh <agent> correction user_nlp "<ctx>" "<signal>" "<lesson>"`
- Exec fail (exit≠0) → `bash $DIR/log-event.sh <agent> error exec_fail "<ctx>" "<stderr[:200]>" "<lesson>"`

4. Set up crons#

text

# Pattern analysis (nightly 3:30 AM)
schedule: cron 30 3 * * * @ Europe/Moscow
payload: python3 ~/.openclaw/shared/learning/analyze-patterns.py

# Weekly report (Sundays 4:00 AM)
schedule: cron 0 4 * * 0 @ Europe/Moscow
payload: python3 ~/.openclaw/shared/learning/weekly-report.py

5. (Optional) Hook integration for auto-error capture#

For Claude Code / Codex hooks:

json

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Bash",
      "hooks": [{"type": "command", "command": "bash ~/.openclaw/shared/learning/error-catcher.sh"}]
    }]
  }
}

Usage#

Log events manually#

bash

DIR="${FEEDBACK_LEARNING_DIR:-$HOME/.openclaw/shared/learning}"

# Error
bash "$DIR/log-event.sh" anton error exec_fail \
  "updating openclaw.json" "SyntaxError: trailing comma" \
  "Always validate JSON with python3 -c before writing"

# Positive
bash "$DIR/log-event.sh" anton positive user_nlp \
  "generated weekly report" "🔥 огонь!"

# Correction
bash "$DIR/log-event.sh" anton correction user_nlp \
  "sent message in wrong format" "не так, в маркдауне давай" \
  "Confirm output format before sending to Telegram"

Detect feedback from text (no LLM)#

bash

echo "круто, зашло!" | python3 detect-feedback.py
# → {"type": "positive", "source": "user_nlp", "signal": "круто", "confidence": 0.8}

python3 detect-feedback.py "переделай это, не тот формат"
# → {"type": "correction", "source": "user_nlp", "signal": "переделай", "confidence": 0.8}

# Pipe mode for hook usage
echo "$TOOL_OUTPUT" | python3 detect-feedback.py --pipe | bash log-event.sh auto

Check active rules before a task#

bash

python3 check-genes.py
# Lists active rules, signals stale ones

python3 check-genes.py --filter exec_fail
# Filter by type

python3 check-genes.py --resolve gene_20260310_120000_0
# Mark a resolved rule as inactive

Data Files#

FilePurposeevents.jsonlAppend-only event log (all feedback), deduped by content hashpatterns.jsonGrouped patterns: BOTH positive and negative, with countsgenes.jsonPromoted structured rules (condition → action → context)capsules.jsonSuccessful reasoning paths to avoid re-computationreports/Weekly synthesis reports

Event Schema#

json

{
  "ts": "2026-03-20T12:00:00Z",
  "id": "sha256_first8",
  "agent": "anton",
  "type": "error|correction|positive|requery",
  "source": "exec_fail|user_nlp|user_emoji|requery|auto",
  "context": "what agent was doing",
  "signal": "the trigger text or emoji",
  "hint": "suggested fix or rule",
  "heat": 1
}

Gene (Promoted Rule) Schema v2#

json

{
  "id": "gene_20260310_120000_0",
  "status": "active|stale|resolved|wont-fix",
  "origin": "original signal/pattern text",
  "type": "error|correction|positive",
  "condition": "When doing X",
  "action": "Do Y instead of Z",
  "context": "Additional context",
  "agents": ["anton"],
  "occurrences": 3,
  "last_seen": "2026-03-20T...",
  "promoted_at": "2026-03-20T...",
  "expires": null,
  "active": true
}

Promotion Flow (v2)#

1. Events accumulate in events.jsonl (deduped by hash)
2. analyze-patterns.py groups similar events (both positive AND negative)
3. Pattern hits ≥3 in 30 days → Behavioral Delta Test: would this rule change a future decision? If yes → promote.
4. Promoted gene has structured fields: condition, action, context
5. Stagnation check: if gene exists but same pattern keeps recurring → mark gene as stale and escalate
6. Genes auto-expire after 90 days of inactivity (no new events matching)
7. weekly-report.py includes gene health: active / stale / resolved counts

Supported Languages#

• Russian: 20+ negative, 19+ positive triggers, correction patterns
• English: 10 negative, 8 positive triggers
• Emoji: Universal positive/negative reactions

What's New in v2#

Featurev1v2Positive pattern tracking❌ skipped✅ tracked separatelyGene structure"AVOID: key_text"``condition → action → contextGene lifecycleactive onlyactive / stale / resolved / wont-fixBehavioral Delta Test❌✅ promotes only if rule changes future behaviorStagnation detection❌✅ re-occurring genes flagged as stalePath configurationhardcoded$FEEDBACK_LEARNING_DIR env varEvent deduplication❌✅ content hashHook integration❌✅ error-catcher.sh for PostToolUseGene check utility❌✅ check-genes.pyGene expiry❌✅ 90-day inactivity auto-expire