Weekly Retro

Structured weekly retrospective. Reflects on system health, output quality, and operational patterns. Runs after graph-hygiene on Sundays.

Tone & Register

This is operational telemetry, not external content. Match the operator's casual register: short fragments, no corporate hype, no inflated symbolism, no em-dashes. Internal log voice. grep-able, scannable on phone. If the week was quiet, say "quiet week". don't invent ceremony. Frontmatter type: summary flows through the operational-telemetry exemption in /review so the slop gate is advisory not blocking.

Source Material (background data)

The retro reads from these source-of-truth files. Pair this skill with its inputs:

Cross-reference with skills/self-audit/SKILL.md (nightly, finer-grained). weekly-retro is a weekly aggregate, not a duplicate of self-audit.

When to Run

• Scheduled: weekly (Sunday 09:00 local, after graph-hygiene at 08:00)
• Manual: /retro or "run retro"

Procedure

1. Gather Data (automated)

Pull from the past 7 days:

• git log on your repo (what changed in skills, knowledge, config)
• logs/session.jsonl (commands used, personas activated, skills invoked)
• logs/errors.jsonl (unresolved errors, failure patterns)
• outputs/raw/ and outputs/reviewed/ (what was produced, what's still pending review)
• Scheduler execution log via schedule_execution_log (task success/failure rates)
• Graph hygiene report from earlier today (if available)

2. Analyze (5 categories)

SHIPPED: What was built, created, or delivered this week?

• New skills, knowledge files, config changes
• Outputs produced (by persona, by type)
• Items progressed, outreach sent, content published

BROKE: What failed or caused friction?

• Error patterns from logs
• Failed scheduled tasks
• Tool/MCP issues
• Anything that needed manual intervention

IMPROVED: What got better vs. last week?

• Error count trend (up/down)
• Pipeline throughput (raw → reviewed → brain)
• Response quality observations
• System changes that reduced friction

PATTERNS: What recurring behaviors are worth noting?

• Most-used personas and skills
• Time-of-day patterns (when is the operator most active?)
• Common intent categories
• Requests that had no matching skill

NEXT WEEK: What should change?

• Skills to create or improve
• Knowledge gaps to fill
• System issues to fix
• Scheduled tasks to add/modify/remove

3. Output

Save to outputs/raw/agent/YYYY-MM-DD-weekly-retro.md with frontmatter:

---
title: Weekly Retro. YYYY-MM-DD
persona: agent
type: summary
status: raw
created: YYYY-MM-DDTHH:MM:SS+00:00
week: YYYY-WNN
---

4. Chat Summary

Send a concise summary to the chat channel:

• 📦 X items shipped
• 🔴 X errors (Y unresolved)
• 📊 top persona: {name}, top skill: {name}
• 💡 recommendation: {one actionable suggestion}

Keep it under 500 chars. Link to the full report.

Rules & Constraints (detailed task rules)

1. Data-grounded: every metric must trace to a source file. If logs/session.jsonl doesn't have the data, the report says "data missing". never fabricate counts.
2. Adversarial stance: assume the week had problems until evidence proves it was clean. Don't paper over by counting only ships and ignoring breaks.
3. One improvement per week: the NEXT WEEK section caps at 1-2 actionable items. Compounding > overhaul.
4. No external slop: this is INTERNAL telemetry. Skip the corporate-therapist tone. "shipped 3, broke 1" beats "successfully delivered 3 important capabilities while encountering 1 minor incident".
5. Time math is in one timezone: pick the week boundary (e.g. Sunday 00:00 → Saturday 23:59) in your local timezone. Don't mix UTC + local in the same report.
6. Operational-telemetry exemption: output goes through /review with type: summary, the slop scan is advisory only. Em-dashes won't block. But don't add ceremony. internal grep-able prose is the bar.

Example output (positive)

---
title: Weekly Retro. 2026-05-04 → 2026-05-10
persona: agent
type: summary
status: raw
week: 2026-W19
---

## SHIPPED
- 12 substantive items this week (highest: a Saturday build session, 12 ships in 6h+)
- Top categories: skills/, scripts/, knowledge/
- Key ships: temporal-rerank on the recall index + a chat-reply failover + a batch of repo audits + the /prd skill + a 9-element auditor

## BROKE
- 14 MCP disconnect events (a plugin orphan pattern. fix shipped this week)
- 5x a hook-dispatch timeout (fix applied: 10s→25s)
- 2x a stop-hook reply miss (pre-existing fix awaiting operator review)

## IMPROVED
- Error count: 23 (week prior) → 9 (this week). 61% drop. driver: the reply failover + the timeout fix.
- /review queue depth: 8 raw → 0 raw (Sunday flush)

## PATTERNS
- The pattern-lift discipline fired twice (one SHIP, one PARK). gate working
- Verification-over-recitation fired 3x in 10d. now structural, an early-phase lift shipped

## NEXT WEEK
- 1 item: close out the outstanding follow-up flagged as a repeat failure

Example output (anti-pattern, do NOT do this)

## SHIPPED
This week proved to be a transformative period in which the stack
successfully delivered numerous important capabilities. leveraging cutting-edge
patterns to ensure operational excellence across all dimensions...

If you find yourself writing "transformative" or "cutting-edge" or "leveraging". stop. Start over.

What NOT To Do

• Don't make up metrics. If data isn't available, say so.
• Don't sugarcoat. If the week was quiet, say "quiet week."
• Don't propose massive overhauls. One small improvement per week compounds.