Run Retrospective

Purpose

Execute scheduled retrospective analysis on accumulated mistakes. Identifies patterns, evaluates action item effectiveness, derives new action items, and creates escalations for ineffective fixes. Implements the full workflow defined in retrospectives.json.

When to Use

• Automatically triggered by learn-from-mistakes when thresholds met
• Manually invoked with /cat:run-retrospective
• After significant project milestones
• When pattern recurrence is suspected

Trigger Conditions

Retrospective is triggered when EITHER condition is met:

triggers:
  time_based: days_since_last_retrospective >= trigger_interval_days  # default: 14
  count_based: mistake_count_since_last >= mistake_count_threshold    # default: 10

Workflow

1. Check Trigger Conditions

RETRO_DIR=".claude/cat/retrospectives"
INDEX_FILE="$RETRO_DIR/index.json"

# Get config and state from index.json
INTERVAL=$(jq -r '.config.trigger_interval_days' "$INDEX_FILE")
THRESHOLD=$(jq -r '.config.mistake_count_threshold' "$INDEX_FILE")
LAST_RETRO=$(jq -r '.last_retrospective // empty' "$INDEX_FILE")
MISTAKES_SINCE=$(jq -r '.mistake_count_since_last' "$INDEX_FILE")

# Calculate days since last retrospective
if [[ -n "$LAST_RETRO" && "$LAST_RETRO" != "null" ]]; then
  LAST_EPOCH=$(date -d "$LAST_RETRO" +%s 2>/dev/null || echo 0)
else
  LAST_EPOCH=0
fi
NOW_EPOCH=$(date +%s)
DAYS_SINCE=$(( (NOW_EPOCH - LAST_EPOCH) / 86400 ))

# Check triggers
if [[ $DAYS_SINCE -ge $INTERVAL ]] || [[ $MISTAKES_SINCE -ge $THRESHOLD ]]; then
  echo "RETROSPECTIVE TRIGGERED"
  echo "  Days since last: $DAYS_SINCE (threshold: $INTERVAL)"
  echo "  Mistakes since last: $MISTAKES_SINCE (threshold: $THRESHOLD)"
else
  echo "No retrospective needed"
  echo "  Days since last: $DAYS_SINCE / $INTERVAL"
  echo "  Mistakes since last: $MISTAKES_SINCE / $THRESHOLD"
  exit 0
fi

2. Gather Mistakes for Analysis

# Aggregate ALL mistakes since last retrospective across all split files
if [[ -n "$LAST_RETRO" && "$LAST_RETRO" != "null" ]]; then
  MISTAKES_TO_ANALYZE=$(cat "$RETRO_DIR"/mistakes-*.json 2>/dev/null | \
    jq -s --arg last "$LAST_RETRO" \
    '[.[].mistakes[] | select(.timestamp > $last)]')
else
  # No previous retrospective - analyze all mistakes
  MISTAKES_TO_ANALYZE=$(cat "$RETRO_DIR"/mistakes-*.json 2>/dev/null | \
    jq -s '[.[].mistakes[]]')
fi

MISTAKE_COUNT=$(echo "$MISTAKES_TO_ANALYZE" | jq 'length')
echo "Analyzing $MISTAKE_COUNT mistakes since $LAST_RETRO"

3. Analyze by Category

category_analysis:
  query: |
    # Use pre-aggregated MISTAKES_TO_ANALYZE from step 2
    echo "$MISTAKES_TO_ANALYZE" | jq '
      group_by(.category)
      | map({category: .[0].category, count: length, ids: [.[].id]})
      | sort_by(-.count)
    '

  output_format:
    - category: protocol_violation
      count: 7
      ids: [M028, M034, M035, ...]
    - category: build_failure
      count: 4
      ids: [M029, M030, ...]

4. Check Action Item Effectiveness

For each existing action item with status: implemented:

effectiveness_check:
  for_each_action_item:
    # Find mistakes matching this action's pattern AFTER completion
    query: |
      COMPLETED=$(jq -r --arg id "A004" '.action_items[] | select(.id == $id) | .completed_date' "$INDEX_FILE")
      PATTERN=$(jq -r --arg id "A004" '.action_items[] | select(.id == $id) | .pattern_id' "$INDEX_FILE")

      # Count post-fix mistakes across all split files
      POST_FIX=$(cat "$RETRO_DIR"/mistakes-*.json 2>/dev/null | \
        jq -s --arg pattern "$PATTERN" --arg completed "$COMPLETED" \
        '[.[].mistakes[] | select(.pattern_keywords | contains([$pattern])) | select(.timestamp > $completed)] | length')

    verdicts:
      effective: post_fix_count == 0
      partially_effective: post_fix_count > 0 AND post_fix_count < pre_fix_count
      ineffective: post_fix_count >= pre_fix_count
      escalate: post_fix_count >= recurrence_after_fix_threshold  # default: 1

5. Identify New Patterns

pattern_identification:
  steps:
    - Group unaddressed mistakes by category
    - Look for common keywords across mistakes
    - Check if pattern matches existing pattern_id
    - If new pattern with >= 2 occurrences, create PATTERN-XXX

  new_pattern_template:
    pattern_id: "PATTERN-007"
    pattern: "{category}"
    occurrences_total: 3
    occurrences_after_fix: 0
    first_seen: "{earliest_timestamp}"
    last_seen: "{latest_timestamp}"
    last_action_date: null
    status: "new"
    effectiveness: "pending"
    preventions: []
    related_action_items: []
    note: "{description of pattern}"

6. Derive Action Items

For each identified pattern without an action item:

action_item_derivation:
  priority_rules:
    high:
      - Pattern count >= 5
      - Category is git_operation_failure or protocol_violation
      - Escalation of previous action
    medium:
      - Pattern count >= 3
      - Category is build_failure or test_failure
    low:
      - Pattern count >= 2
      - Category is documentation or detection_gap

  action_template:
    id: "A008"
    priority: "{calculated}"
    description: "{derived from pattern analysis}"
    category: "{pattern_category}"
    pattern_id: "PATTERN-007"
    status: "open"
    created_date: "{now}"
    completed_date: null
    related_mistakes: ["M042", "M043", "M044"]
    effectiveness_check:
      mistakes_before: 3
      mistakes_after: null
      post_fix_mistakes: []
      verdict: "pending"

7. Create Escalations

When action item is ineffective:

escalation:
  trigger: effectiveness_verdict == "escalate" OR "ineffective"

  template:
    id: "ESCALATE-{date}-{seq}"
    original_action_id: "A004"
    original_fix_description: "{from action item}"
    failure_analysis:
      expected_result: "{what the fix was supposed to do}"
      actual_result: "{N} new failures in period"
      root_cause: "{analyze why fix didn't work}"
      gap_identified: "{what was missing}"
    proposed_solution:
      approach: "{defense-in-depth|alternative|enhancement}"
      description: "{new proposed fix}"
      prevention_type: "{code|hook|skill|threshold}"
      layers: ["{layer1}", "{layer2}"]
    priority: "high"
    status: "open"

8. Update index.json and Create Retrospective Record

# Get current year-month for retrospective split file
YEAR_MONTH=$(date +%Y-%m)
RETRO_SPLIT_FILE="$RETRO_DIR/retrospectives-${YEAR_MONTH}.json"
TIMESTAMP=$(date -Iseconds)

# Initialize retrospective split file if needed
if [ ! -f "$RETRO_SPLIT_FILE" ]; then
  echo "{\"period\":\"$YEAR_MONTH\",\"retrospectives\":[]}" > "$RETRO_SPLIT_FILE"
  # Add to index
  jq --arg f "retrospectives-${YEAR_MONTH}.json" \
    'if (.files.retrospectives | index($f)) then . else .files.retrospectives += [$f] | .files.retrospectives |= sort end' \
    "$INDEX_FILE" > "$INDEX_FILE.tmp" && mv "$INDEX_FILE.tmp" "$INDEX_FILE"
fi

# Get next retrospective ID across all split files
MAX_RETRO_NUM=$(cat "$RETRO_DIR"/retrospectives-*.json 2>/dev/null | \
  jq -s '[.[].retrospectives[].id] | map(select(startswith("R")) | ltrimstr("R") | tonumber) | max // 0')
RETRO_ID="R$(printf '%03d' $((MAX_RETRO_NUM + 1)))"

# Build and append retrospective entry to split file
jq --arg id "$RETRO_ID" \
   --arg ts "$TIMESTAMP" \
   --arg trigger "$TRIGGER_TYPE" \
   --arg period "${LAST_RETRO} to ${TIMESTAMP}" \
   --argjson count "$MISTAKE_COUNT" \
   --argjson findings "$FINDINGS_JSON" \
   '.retrospectives += [{
     id: $id,
     timestamp: $ts,
     trigger: $trigger,
     period_analyzed: $period,
     mistakes_analyzed: $count,
     summary: "...",
     findings: $findings
   }]' "$RETRO_SPLIT_FILE" > "$RETRO_SPLIT_FILE.tmp" && mv "$RETRO_SPLIT_FILE.tmp" "$RETRO_SPLIT_FILE"

# Update index.json with new state
jq --arg ts "$TIMESTAMP" \
   '.last_retrospective = $ts | .mistake_count_since_last = 0' \
   "$INDEX_FILE" > "$INDEX_FILE.tmp" && mv "$INDEX_FILE.tmp" "$INDEX_FILE"

9. Present Action Items for Execution

output_format:
  summary: |
    ## Retrospective R003 Complete

    **Period:** 2026-01-08 to 2026-01-11
    **Mistakes Analyzed:** 5

    ### Category Breakdown
    - protocol_violation: 2
    - prompt_engineering: 2
    - git_operation_failure: 1

    ### Action Item Effectiveness
    - A001: effective
    - A004: escalated -> ESCALATE-2026-01-11-001

    ### New Action Items
    - A008 (high): Add positive verification requirement to all validation hooks
    - A009 (medium): Document common prompt engineering pitfalls

    ### Escalations
    - ESCALATE-2026-01-11-001: Git operations need PreToolUse hook for destructive commands

  next_steps: |
    Execute action items in priority order:
    1. [A008] - /cat:execute-action-item A008
    2. [A009] - /cat:execute-action-item A009
    3. [ESCALATE-2026-01-11-001] - Requires immediate attention

10. Execute Action Items (Optional)

If user requests execution:

execution_workflow:
  for_each_action_item:
    - Read action item description
    - Determine implementation approach:
        hook: Create/update hook in .claude/settings.json
        skill: Create/update skill in cat/skills/
        code: Modify source code
        documentation: Update CLAUDE.md or skill docs
    - Implement the fix
    - Update action_item.status = "implemented"
    - Update action_item.completed_date = now
    - Commit changes with reference to action ID

Anti-Patterns

Check existing action effectiveness first

# ❌ Just add new action items
new_action: "Add hook for X"
# Never checked if previous hook worked!

# ✅ Check existing items first
check: "A004 had 3 post-fix failures -> escalate"
then: "Create ESCALATE with root cause analysis"

Create specific, actionable items

# ❌ Vague
description: "Improve git operations"

# ✅ Specific
description: "Add PreToolUse hook to block `git reflog expire` and `git gc --prune` commands"

Treat escalations as high priority

# ❌ Create escalation, never act
escalation: "ESCALATE-2026-01-08-001"
status: "open"  # Forever

# ✅ Escalations are HIGH priority
escalation: "ESCALATE-2026-01-08-001"
# Immediately: propose solution, get approval, implement

Integration with learn-from-mistakes

At the end of learn-from-mistakes Step 12 (Update Retrospective Counter), add:

trigger_check:
  action: "Check if retrospective is needed"
  command: |
    INDEX_FILE=".claude/cat/retrospectives/index.json"

    # Increment mistake counter
    jq '.mistake_count_since_last += 1' "$INDEX_FILE" > "$INDEX_FILE.tmp" \
      && mv "$INDEX_FILE.tmp" "$INDEX_FILE"

    # Check thresholds
    MISTAKES=$(jq '.mistake_count_since_last' "$INDEX_FILE")
    THRESHOLD=$(jq '.config.mistake_count_threshold' "$INDEX_FILE")
    if [[ $MISTAKES -ge $THRESHOLD ]]; then
      echo "RETROSPECTIVE THRESHOLD REACHED ($MISTAKES >= $THRESHOLD)"
      echo "Run: /cat:run-retrospective"
    fi

Related Skills

• cat:learn-from-mistakes - Records individual mistakes, triggers this skill
• cat:token-report - Provides context metrics for analysis
• cat:work - Can be used to implement action items