Retrospective

You are reviewing past work to extract patterns that make future work better. This is not a report for its own sake — every finding should either become a learning (fed to /learn) or be acknowledged as already captured.

The goal: leave HQ smarter than you found it.

Step 0: Mode + Company Resolution

Mode detection:

• If $ARGUMENTS contains --session → SESSION mode
• If $ARGUMENTS contains a project slug → PROJECT mode
• If $ARGUMENTS is empty or ambiguous → AskUserQuestion: "A) Review a specific project. B) Review recent session work."

Company resolution: Check if the first word matches a company slug in companies/manifest.yaml. Same pattern as /brainstorm:

1. If matched: set {co}, strip from args, load policies
2. If not matched: infer from cwd or project slug location
3. If still ambiguous: ask

Announce: Mode: {PROJECT|SESSION} | Company: {co}

Step 1: Data Collection

PROJECT Mode

Read these data sources (skip gracefully if any don't exist):

# 1. The PRD
# Read companies/{co}/projects/{slug}/prd.json
# Extract: stories (titles, statuses), metadata (dates, linear IDs)

# 2. Git history for the project
git log --oneline --since="{prd.createdAt}" -- {repo paths if known}

# 3. Workspace threads mentioning this project
# Glob pattern: workspace/threads/*{slug}*.json
# Read each, extract conversation_summary + files_touched

# 4. Audit log events
grep '"{slug}"' workspace/metrics/audit-log.jsonl | tail -50

# 5. Model usage (worker activity)
grep '"{slug}"' workspace/metrics/model-usage.jsonl | tail -50

# 6. Orchestrator state (if exists)
# Read workspace/orchestrator/{slug}/state.json

# 7. Brainstorm (if exists)
# Read companies/{co}/projects/{slug}/brainstorm.md

SESSION Mode

# 1. Recent threads (last 24h)
# List workspace/threads/ sorted by modification time, read top 5-10

# 2. Recent learnings
# List workspace/learnings/ sorted by modification time, read top 10

# 3. Recent audit log entries (last 24h)
tail -100 workspace/metrics/audit-log.jsonl

# 4. Recent git activity
git log --oneline --since="24 hours ago"

# 5. Current git status
git status

Present collection summary:

Data collected:
- Stories: {N} found ({done}/{total})
- Threads: {N} found
- Audit events: {N} found
- Git commits: {N} in scope
- Workers used: {list or "none tracked"}

Step 2: Analysis

PROJECT Mode — Analyze These Dimensions

1. Story Completion

Story	Title	Status	Retries	Workers
US-001	...	done/failed/stalled	N	...

2. Worker Usage

Worker	Tasks Assigned	Success Rate
...	N	N/N

Which workers carried the most load? Any worker used on >70% of stories?

3. Timeline

• PRD created: {date}
• First story started: {date}
• Last story completed: {date}
• Total elapsed: {N days}
• Stories per day: {rate}

4. Quality Gate Outcomes

• How many stories passed on first attempt?
• How many required retries? What caused retries?
• Quality gate failure rate: {N}%

5. File Churn

• Files touched across most stories (indicates scope underestimation or tight coupling)
• Any files touched by >50% of stories? Flag as architectural hotspot

6. Learnings Already Captured

• Check workspace/learnings/ and companies/{co}/policies/ for rules created during this project
• List them — don't re-capture what's already been learned

SESSION Mode — Analyze These Dimensions

1. What Was Accomplished

• List each completed task/thread with 1-line summary
• Any tasks started but not finished?

2. Files Touched

• Aggregate across all session threads
• Any surprising files? (files outside expected scope)

3. Failures or Retries

• Any tool call failures or retries in audit log?
• Any handoff/checkpoint events?

4. Time Distribution

• Rough breakdown: planning vs building vs debugging vs reviewing

Step 3: Pattern Detection

Flag these patterns when evidence supports them:

Pattern	Signal	Implication
Complexity underestimated	Story with 2+ retries	PRD story sizing was off — capture as sizing heuristic
Worker overuse	Same worker >70% of stories	Consider making it a default in PRD template
Quality gate failures	>20% failure rate	Upstream planning issue — stories shipped before ready
Undeclared files	Stories with 0 files[] in PRD	File locking risk — add to PRD template guidance
Scope creep	Files touched outside story scope	Scope definition was too narrow or coupling wasn't anticipated
Architectural hotspot	Same file in >50% of stories	This file needs refactoring or better abstraction
Missing test coverage	Quality gate test failures	Test strategy needs strengthening for this area
Session fragmentation	3+ handoffs for one project	Context overhead — consider longer sessions or better checkpoints
Tool mismatch	Repeated failures with specific tool/MCP	Tool setup issue or wrong tool for the job

Only flag patterns with clear evidence. Don't speculate.

Step 4: Learning Extraction

For each detected pattern:

1. Generate a learning candidate — state the rule in ALWAYS/NEVER format:

   • e.g., "ALWAYS declare file dependencies in prd.json stories when the project touches shared modules"
   • e.g., "NEVER estimate stories touching {area} as S — they're always M minimum"

2. Classify scope:

   • Project-specific → note it but don't create a policy (too narrow)
   • Company-scoped → companies/{co}/policies/
   • Global → core/policies/

3. Check for duplicates — grep existing policies for similar rules

4. Present to user:

   Learning candidates ({N} found):
   
   1. {rule} — Scope: {company|global} — {new|already exists}
   2. {rule} — Scope: {company|global} — {new|already exists}
   ...
   
   Run /learn for new learnings? [Y/n]
   

5. If user approves: suggest running /learn with each new learning. Do NOT write policy files directly — /learn handles dedup, formatting, and injection.

Step 5: Write Retro Report

Output to workspace/reports/{slug}-retro.md:

# {Project/Session Title} — Retrospective

**Date:** {ISO8601}
**Mode:** Project | Session
**Company:** {co}
**Duration:** {start date} → {end date}

## Summary

{2-3 sentence summary of what happened — scope, outcome, notable events}

## Stories (PROJECT mode)

| Story | Title | Status | Retries | Workers Used |
|-------|-------|--------|---------|-------------|
| US-001 | ... | done | 0 | frontend-designer |

**Completion rate:** {done}/{total} ({%})

## Worker Usage (PROJECT mode)

| Worker | Tasks | Coverage |
|--------|-------|----------|
| ... | N | N% |

## Timeline (PROJECT mode)

- PRD created: {date}
- First story: {date}
- Last story: {date}
- Elapsed: {N} days
- Velocity: {stories/day}

## Session Summary (SESSION mode)

| Thread | Summary | Files | Outcome |
|--------|---------|-------|---------|
| T-... | ... | N files | completed/partial |

## Patterns Detected

{For each pattern with evidence:}

### {Pattern Name}
**Signal:** {what was observed}
**Evidence:** {specific data point}
**Implication:** {what this means for future work}

## Quality Gate Outcomes (PROJECT mode)

- First-pass success rate: {%}
- Total retries: {N}
- Most common failure: {category}

## Learnings

| # | Rule | Scope | Status |
|---|------|-------|--------|
| 1 | {rule} | {scope} | new / already captured |

## What Went Well

- {Specific thing that worked — be concrete, not generic}

## What to Improve

- {Specific thing to do differently — actionable, not vague}

## Open Questions

- {Anything unresolved that should be addressed in the next project}

Step 6: Wrap Up

qmd update 2>/dev/null || true

Print:

Retro complete: {title}
Report: workspace/reports/{slug}-retro.md
Patterns detected: {N}
New learnings: {N} (run /learn to capture)

---

Rules

• Read-only analysis — retro does not modify project files, prd.json, or board.json. It writes only its own report
• Evidence-backed patterns only — do not flag patterns without specific data points from Steps 1-2
• Delegate learning capture — suggest /learn, do not write policy files directly
• No re-capture — check existing policies before proposing a learning. Duplicates waste context
• Both modes are valid — SESSION mode is lighter but still valuable. Not every retro needs a full project
• Company isolation — scope all data collection to the resolved company. Never mix cross-company data
• Do NOT create task lists or enter planning mode — this is a focused analysis, not a project
• Honest assessment — if a project went poorly, say so. Retros that only praise are useless
• Actionable over comprehensive — 3 concrete learnings beat 10 vague observations