Retrospective Command

--help Handler

If the user passes --help as an argument, display this help and stop:

/coconut-rules:retrospective - Session Retrospective Analysis

DESCRIPTION
  Analyze the current session to extract learnings, mistakes, patterns,
  and decisions. Proposes updates to rules and knowledge graph.

USAGE
  /coconut-rules:retrospective [--depth LEVEL] [--dry-run]
  /coconut-rules:retrospective --help

OPTIONS
  --depth LEVEL    Analysis depth: quick, deep, ultrathink (default: ultrathink)
  --dry-run        Show proposed changes without applying them

ANALYSIS CATEGORIES
  M - Mistakes: Errors and their root causes
  G - Good Decisions: What worked and why
  P - Patterns: Recurring code/workflow patterns
  D - Debugging: Investigation techniques that worked
  A - Architecture: Design decisions and rationale

OUTPUT
  Grouped by action type:
  - Rules to add (with placement: CLAUDE.md or .claude/rules/)
  - Rules to update
  - Knowledge graph entries to add
  - Items skipped (with reason)

EXAMPLES
  # Full retrospective with ultrathink
  /coconut-rules:retrospective

  # Quick analysis
  /coconut-rules:retrospective --depth quick

  # Preview changes without applying
  /coconut-rules:retrospective --dry-run

Run a retrospective analysis on the current session to extract learnings and update rules/knowledge graph.

Parse Arguments

Extract options from $ARGUMENTS:

• --depth quick|deep|ultrathink (default: ultrathink)
• --dry-run flag (default: false)

Workflow

Execute the retrospective in phases:

Phase 1: Initialize

1. Create todo list for tracking progress
2. Ensure .claude/retrospectives/ directory exists:

   mkdir -p .claude/retrospectives
   

3. Generate retrospective filename: YYYY-MM-DD-HHmm.md

Phase 2: Parallel Analysis

Launch three agents in parallel using Task tool:

[session-analyzer] - Analyze conversation for learnings
[rule-reviewer] - Review current rules against learnings
[knowledge-updater] - Propose knowledge graph updates

All agents receive the same depth setting.

Phase 3: Collect Results

Wait for all agents to complete. Collect outputs:

• Learnings extracted (from session-analyzer)
• Rule changes proposed (from rule-reviewer)
• KG entries proposed (from knowledge-updater)

Phase 4: Present Batch

Format results for user review:

## Retrospective Results

### Session Summary
- Events analyzed: [count]
- Learnings extracted: [count by category M/G/P/D/A]

### Rules to Add
| Rule | Placement | Rationale |
|------|-----------|-----------|
| [text <50 words] | [CLAUDE.md/.claude/rules/X.md] | [why] |

### Rules to Update
| Current | Proposed | File |
|---------|----------|------|
| [old text] | [new text] | [path] |

### Knowledge Graph Entries
| Type | Content | Tags |
|------|---------|------|
| [decision/pattern/convention] | [text] | [relevant tags] |

### Skipped Items
- [Item]: [Reason - too specific/already covered/low impact]

Phase 5: User Confirmation

Use AskUserQuestion to get user confirmation:

Review the proposed changes above.

Options:
1. Apply all changes
2. Apply rules only
3. Apply KG entries only
4. Review individually
5. Cancel

Phase 6: Apply Changes

If user confirms (and not --dry-run):

For rules:

• Use /coconut-rules:add-rule or Edit tool
• Follow concise-rule-writing skill for formatting
• Place critical rules in CLAUDE.md, domain rules in .claude/rules/

For knowledge graph:

• Use /coconut-knowledge-graph:add command
• Include type and tags

Phase 7: Save History

Save retrospective summary to .claude/retrospectives/YYYY-MM-DD-HHmm.md:

# Retrospective: [Date Time]

## Session Focus
[Brief description of what was worked on]

## Learnings by Category
### Mistakes (M)
- [learning]

### Good Decisions (G)
- [learning]

### Patterns (P)
- [learning]

### Debugging (D)
- [learning]

### Architecture (A)
- [learning]

## Changes Applied
### Rules
- Added: [list]
- Updated: [list]

### Knowledge Graph
- Added: [list]

## Notes
[Any additional context]

Also sync to knowledge graph if available:

/coconut-knowledge-graph:add "retrospective-YYYY-MM-DD" --type decision --content "[summary]"

Agent Descriptions

For Task tool invocations:

session-analyzer: "Analyze this conversation session to extract learnings. Categories: Mistakes (root cause), Good Decisions (why worked), Patterns (recurring), Debugging (techniques), Architecture (decisions). Use [depth] analysis level. Output structured learnings."

rule-reviewer: "Review .claude/rules/ and CLAUDE.md against the session learnings. Identify: rules to add (with placement), rules to update, rules to remove. Use concise-rule-writing skill. Each rule <50 words."

knowledge-updater: "Propose knowledge graph updates based on session learnings. Use /coconut-knowledge-graph:search to check existing entries. Avoid duplicates. Output: entries to add with type (decision/pattern/convention) and tags."

Tips

• Run retrospective at natural session breakpoints
• Use --depth quick for short sessions with obvious learnings
• Use --dry-run to preview before applying
• Review skipped items - they may become relevant later
• Retrospectives are most valuable after debugging sessions